i i

OUP CORRECTED PROOF – FINAL, //, SPi

i i

Deontic Modality

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

Deontic Modality

edited by
Nate Charlow and
Matthew Chrisman

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

3
Great Clarendon Street, Oxford, OX DP,
United Kingdom
Oxford University Press is a department of the University of Oxford.
It furthers the University’s objective of excellence in research, scholarship,
and education by publishing worldwide. Oxford is a registered trade mark of
Oxford University Press in the UK and in certain other countries
© the several contributors 
The moral rights of the authors have been asserted
First Edition published in 
Impression: 
All rights reserved. No part of this publication may be reproduced, stored in
a retrieval system, or transmitted, in any form or by any means, without the
prior permission in writing of Oxford University Press, or as expressly permitted
by law, by licence or under terms agreed with the appropriate reprographics
rights organization. Enquiries concerning reproduction outside the scope of the
above should be sent to the Rights Department, Oxford University Press, at the
address above
You must not circulate this work in any other form
and you must impose this same condition on any acquirer
Published in the United States of America by Oxford University Press
 Madison Avenue, New York, NY , United States of America
British Library Cataloguing in Publication Data
Data available
Library of Congress Control Number: 
ISBN ––––
Printed in Great Britain by
Clays Ltd, St Ives plc
Links to third party websites are provided by Oxford in good faith and
for information only. Oxford disclaims any responsibility for the materials
contained in any third party website referenced in this work.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

Contents

List of Contributors vii

Introduction 
Nate Charlow and Matthew Chrisman
. Deontic Modals and Probabilities: One Theory to Rule Them All? 
Fabrizio Cariani
. Decision Theory: Yes! Truth Conditions: No! 
Nate Charlow
. Linguistic and Philosophical Considerations on Bayesian Semantics 
Daniel Lassiter
. Contextualism about Deontic Conditionals 
Aaron Bronfman and J. L. Dowell
. Objective and Subjective ‘Ought’ 
Ralph Wedgwood
. ‘Ought’: Out of Order 
Stephen Finlay
. On a Shared Property of Deontic and Epistemic Modals 
Jessica Rett
. Modalities of Normality 
Seth Yalcin
. Extreme and Non-extreme Deontic Modals 
Paul Portner and Aynat Rubinstein
. Rationalization and the Ross Paradox 
Benj Hellie
. Dynamic Foundations for Deontic Logic 
Malte Willer
. Dynamic Expressivism about Deontic Modality 
William B. Starr
. Metanormative Theory and the Meaning of Deontic Modals 
Matthew Chrisman

Index 

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

List of Contributors

Aaron Bronfman, University of Nebraska–Lincoln

Fabrizio Cariani, Northwestern University

Nate Charlow, University of Toronto

Matthew Chrisman, Edinburgh University

J. L. Dowell, Syracuse University

Stephen Finlay, University of Southern California

Benj Hellie, University of Toronto
Daniel L assiter, Stanford University

Paul Portner, Georgetown University

Jessica Rett, University of California Los Angeles

Aynat Rubinstein, Hebrew University of Jerusalem

William B. Starr, Cornell University
Ralph Wedgwood, University of Southern California

Malte Willer, University of Chicago

Seth Yalcin, University of California Berkeley

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

Introduction
Nate Charlow and Matthew Chrisman

Philosophers, linguists, and logicians have traditionally conceived of deontic modality

as having to do with what is morally or legally obligatory and permissible, where
these are treated as species of necessity and possibility. For example, what’s morally
obligatory is treated as what’s necessary to conform to the rules of morality, or what’s
legally permissible is treated as what’s possible to do without breaking the (relevant)
law. In this respect, we might distinguish deontic modality from metaphysical and
epistemic modalities. The former concerns what is necessarily and possibly true
given various metaphysical laws, such as the laws of mereology, logic, mathematics,
or nature. The latter concerns what is necessarily and possibly (and perhaps also
probably/improbably) true given various bodies of knowledge or information.
This volume is about deontic modality, but in a somewhat extended sense. We
are concerned not only with what is necessary and possible in light of morality and
law but also with the sorts of necessities and possibilities that attach to all kinds of
practical reasoning. For example, if one were to reason about how to get to London
from Edinburgh and conclude that leaving Scotland is necessary for achieving that
end, here we count this kind of necessity as a species of deontic modality although it’s
not moral or legal. Similarly, on the plausible assumption that morality and prudence
can pull apart, we might want to allow that some of what one may do prudentially is
nonetheless morally forbidden; nevertheless, here we count prudential possibility as a
further species of deontic modality, even though it’s not moral or legal.
We use this extended sense of “deontic modality” here because we believe it helps
to unify and illuminate connected phenomena all having to do with what’s necessary
and possible in light of various rules or norms that potentially govern our actions. This
phenomenon is an interesting instance of the way we can “displace” from actuality to
consider what could, should, or must in various senses be the case.
The chapters in this volume are based on the supposition that the phenomenon
of deontic modality can be investigated as a linguistic phenomenon, and probed
using linguistic methods. In English, words such as “must”, “may”, “can”, “should”,
“ought” form a class of modal auxiliary verbs capable of taking deontic modal meaning

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 nate charlow and matthew chrisman

(but usually also capable of taking epistemic or metaphysical modal meaning). But
deontic modal notions can also be expressed with adjectives such as obligatory and
permissible, adverbs such as “necessarily” and “possibly”, and various constructions
such as “Have a drink!” and “Unless the proposed payout is very high, one is to
reject the offer.” One of the central topics of this volume is how to understand the
meaning of the bits of language used to express deontic modal notions. We focus
by default on English, but the lessons drawn are usually intended to extend to other
languages as well to expedite the search for an overall theory of deontic modal language
(syntax, semantics, pragmatics, metasemantics) capable of integrating this language
with overlapping bits of language used to express other modal notions while explaining
notable differences.
The study of deontic modal language is interesting and important in its own right
as a central linguistic phenomenon in need of explanation. However, this volume also
concerns more philosophical topics about the operation of deontic modal concepts and
the logic of deontic reasoning. For instance, regardless of how one models the meaning
of a word such as “ought”, we face questions about what application of the concept(s) it
expresses commits one to. In what sense, if any, does ought imply can? Does a belief
about what you yourself ought to do commit you to having an intention to do this
thing? Deontic necessities and possibilities do not seem to follow all of the same logical
patterns as metaphysical necessities and possibilities (e.g., must(p) doesn’t seem to
entail must(p or q), when the must is deontic). But which logical patterns do deontic
modalities follow?
It is an interesting psychological fact that humans engage in modal thought. We
might wonder what, in general, that ability does for us, but we might also wonder how
the more specific ability to displace from what is actually the case and think about what
may, could, should, or must be the case in light of various rules or practical laws does for
us. Notably, unlike metaphysical and epistemic necessities, a deontic necessity doesn’t
guarantee actuality; as everyday life repeatedly teaches, the fact that some action is
morally, legally, prudentially, and so on, obligatory doesn’t guarantee that it will be
performed. Also, deontic modal reasoning seems to be tied up crucially in our ability
to reason practically. A good theory of deontic modality would explain all of this.
A background assumption behind the organization of this volume is that this ability
is importantly connected to the way humans are subject to norms. So developing
an adequate theory of deontic modality will also—we think—be a crucial part of
understanding the elusive but all-important topic of normativity. Moreover, given that
humans do think with deontic modal concepts, philosophers and logicians face an
important challenge: explain the logic of this thought. What are the principles by
which we should reason when deploying deontic modal concepts? How are these
related to the principles by which we should reason when determining what to do?
Are the formal tools of modal logic adequate for representing these principles of
reasoning?

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

introduction 

We take the study of deontically modal language to be a window into the phe-
nomenon of deontic modality proper (and vice versa). Deontic modality, we said,
is a phenomenon that encompasses normative concepts and thought (e.g. practi-
cal reasoning), the normative rules that govern deontically modal thought, and an
account of the role deontic modality plays in our cognitive and social lives. For
philosophers, this general line of thinking is familiar from the work of C. L. Stevenson,
R. M. Hare, H.-N. Castañeda, Wilfrid Sellars, G. H. von Wright, Christine Korsgaard,
Peter Railton, Allan Gibbard, Simon Blackburn, Ruth Millikan, and Sharon Street.
For linguists (and philosophers steeped in the methods of modern linguistics) it
is probably more unfamiliar. Such questions have perhaps not traditionally been at
the forefront of the minds of those who work on the meaning of deontically modal
language. Indeed, it might seem that those interested in the phenomena surrounding
deontic modality proper are interested in something altogether different from the
things that occupy the attention of linguistic work on deontic modality. Is that right?
In part, yes. Modern linguistic semantics tries to derive correct truth conditions for
a whole sentence as a function of the meanings and arrangement of that sentence’s
syntactic parts. The task of the linguist, faced with a sentence such as (), is straight-
forward.
() You must wash your hands.
We identify ()’s syntactic parts (and their mode of arrangement), and assign ‘must’
an interpretation that allows it to combine with the other parts of (), so that () is
assigned a truth condition of the appropriate sort.
In this vein, Angelika Kratzer (see, e.g., Kratzer , , ) has developed and
defended what we might call the “Standard” or “Orthodox” or “Classical” semantics
of deontically modal language. Kratzer’s theory—endorsed by acclaim in certain areas
of linguistic semantics—is subtle and multi-faceted (see Portner , chapter ). The
driving ideas are that, first, deontic modality is of the same species as, for example,
metaphysical and epistemic modality and, secondly, that any differences in meaning
between these species of modality can be explained by appeal to different ways of
determining a domain of quantification.1 The idea is that all types of modality involve
some type of quantification over a domain of possibilities that is determined contextu-
ally, by appeal to a contextually fixed range of relevant possibilities and a contextually
fixed way of ordering or organizing those relevant possibilities according to relevant

1 Neither this idea—that modal expressions exhibit something like polysemy—nor the treatment of the
modalities as quantifiers whose specific interpretation is resolved by context is due originally to Kratzer.
These ideas tend, however, to be most closely associated with Kratzer, and with good reason. Kratzer was
the first to develop a detailed and predictive account of the nature of this polysemy, the nature and properties
of the relevant context-sensitive parameters (her “modal base” and “ordering source”), and the relationship
between these parameters and the domain of possibilities over which modals are thought to quantify. The
fact that the account integrates seamlessly with Kratzer’s “Restrictor” account of indicative conditionals
(which is also effectively standard in certain circles in linguistic semantics) renders this achievement nothing
short of monumental.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 nate charlow and matthew chrisman

criteria. (Potentially relevant criteria can include, on the non-normative side of things,
comparative normality or likelihood; and on the normative side of things, how well,
comparatively speaking, a possibility does at realizing designated goals, manifesting
various values, satisfying particular preferences, or upholding moral ideals, etc.).
Hence, on Kratzer’s view, deontic modality is distinguished primarily by the fact that
its necessities and possibilities order the relevant possibilities according to normative
criteria: criteria that bear somehow on the preferences, motivations, plans, or intentions
of some relevant agent. So, according to Kratzer, a deontic necessity modal such as
() says, roughly, that the relevant possibilities that are best (according to the relevant
criteria) are all possibilities in which you wash your hands.
Kratzer’s is an elegant, explanatory, and predictive theory of deontically modal
language (and modal language more generally). By itself, however, it does not seem
to shed much light on the distinctive features of deontic modality proper. (Notice, for
example, that the “structural” characteristics of deontic modality are the same as those
associated with any other kind of modality.)
Several papers in this volume represent broadly critical reactions to this feature
of the Orthodox account of deontic modality. Fabrizio Cariani, Nate Charlow, and
Daniel Lassiter each propose to replace Kratzer’s theory with a theory that tries
to illuminate the connection between deontic modality and substantive theories of
practical rationality (such as, for example, normative decision theories). This would
seem a very natural thing to do with deontic modally language, insofar as a normal
utterance of a sentence such as () can be plausibly thought to express something
like a verdict of a theory of practical rationality (applied to a certain kind of decision
problem).
But, these authors argue, this is a fact that is not appropriately reflected in the
Orthodox account. Cariani argues that the truth condition of a deontic modal should
reflect the fact that whether a deontic modal such as () is true can depend on the
availability of certain probabilistic information, e.g., whether or not washing your
hands is likely to leave them cleaner than before. Charlow, on the other hand, argues
that the truth condition of a deontic modal should reflect the fact that whether a
deontically modal claim such as () is acceptable can depend on the sort of theory
of normative practical rationality that it is appropriate to deploy in evaluating that
sentence. (Ultimately he sees this fact as pushing us to an Expressivistic interpretation
of the “truth condition” that is computed by semantics.) Lassiter, extending the work
of his important  dissertation (as well as Portner ), argues that the truth
condition of a deontic modal should model the gradability of deontic value—as
reflected in, e.g., the fact some action can be more required than another, or very
required—and suggests that the right kind of scalar structure is provided by expected
(deontic) value, roughly of the sort familiar from Bayesian decision theory.
Such theories raise important questions about whether (and to what degree) a
semantics for deontic modals should presuppose particular commitments in nor-
mative decision theory. For it’s debatable how far competent use of deontic modals

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

introduction 

requires one to have specific views about which decision theory is right (to say nothing
of requiring endorsement of some specific decision theory or other). Cariani and
Charlow are each careful to avoid presupposing a certain substantive account of the
relationship between the information that is relevant for evaluating a deontic modal
and the truth of that deontic modal—that is to say, a substantive account of practical
rationality. Each provides his own reasons for this—Cariani’s of an empirical variety,
Charlow’s of a more theoretical variety. Lassiter is mostly unmoved by these: he argues
that there is no genuine obstacle, theoretical or empirical, to writing down a correct
theory of practical rationality in our semantics for deontic modals. Whether it is
appropriate to write down theories of practical rationality in one’s semantics (and, if
so, how to do so) represents a fundamental point of disagreement between Cariani
and Charlow, on the one hand, and Lassiter, on the other.
Though this inevitably cuts things too coarsely, we find it helpful to view Cariani,
Charlow, and Lassiter as trying to motivate different, more or less radical, alternatives
to Orthodoxy. On the other hand, it is helpful to read Aaron Bronfman/Janice Dowell
and Ralph Wedgwood as offering (qualified) defenses of certain important pieces of
Orthodoxy, while highlighting the degree to which accounts in the “Kratzerian” vein
can be adapted to challenges of the sort highlighted by Cariani, Charlow, and Lassiter
(among others).
Bronfman and Dowell set out to defend Kratzer’s own theory of the meaning
of deontic conditionals: conditionals that typically express an obligation conditional
on the obtaining of some relevant fact or other. Many recent articles take Kratzer’s
semantics to task for yielding apparently incorrect truth conditions for deontic
conditionals, in, for example, the Miner Puzzle of Kolodny and MacFarlane ().
Bronfman and Dowell continue the project2 of demonstrating how a subtler under-
standing of (i) the logical forms of deontic conditionals and (ii) the manner in
which the domain of quantification for a deontic modal is resolved in context can
yield a theoretically and empirically satisfying treatment of the meaning of such
constructions.
Both Wedgwood and Finlay, like Cariani and Lassiter, make use of broadly proba-
bilistic notions such as expectation to motivate alternative semantic analyses of deontic
modals; these authors agree, roughly, that deontic modals have a semantics that is, at
least in part, “probabilistic” in nature. Unlike Cariani and Lassiter, Wedgwood suggests
that it is possible to accommodate the probabilistic character of deontic modals in a
manner consistent with the “classical” semantics (by ordering the relevant possibilities
roughly according to their expected value). Though, Wedgwood suggests, it is essential
to semantically represent the (probabilistic) “epistemic perspective” of an agent—
to account, inter alia, for the meaning of the oft-discussed “subjective ‘ought”’—we

2 Recently pursued by Kratzer herself (see her ). See also von Fintel (), Dowell (), and
references therein.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 nate charlow and matthew chrisman

can nevertheless accomplish this while endorsing an account that retains the overall
structure of the standard semantics.3
Finlay represents a kind of return to first principles. Virtually all work on the
semantics of deontic modals has assumed, with Kratzer, that (i) the distinction
between relevant information and relevant priorities must be somehow represented in
the truth condition of a deontic modal, (ii) the truth condition for a deontic modal
must be doubly-relative—sensitive not only to relevant possibilities, but also to ways of
organizing or ordering these possibilities according to some body of relevant priorities.
Finlay rejects this particular piece of orthodoxy—en route to developing an account of
the contrast between so-called “strong” deontic modals such as “must” and so-called
“weak” deontic modals such as “ought”. For Finlay, a strong deontic necessity modal of
the form “mustE (p)” (read: in order that E, it must be that p) quantifies universally over
the relevant E-possibilities (i.e. asserts that these are all p-possibilities). A weak deontic
necessity modal of the form “oughtE (p)” (read: in order that E, it ought to be that p)
asserts that, given E, p is likelier than any alternative—roughly, that achieving end E
makes it probable that p. Finlay, like Bronfman and Dowell, aims to show how much
power a formally unornamented “classical” account can achieve, with a sufficiently
subtle understanding of the relevant empirical phenomena and the flexibility of the
semantic apparatus at hand.
These approaches are, we think, in accord regarding the assertion that an under-
standing of deontic modality proper should inform a semantics for expressions used
to express deontically modal opinions. Consider the the introduction of a notion
such as expectation into one’s semantics for deontic modals. The introduction of
such a notion cannot obviously be motivated on ordinary “compositional” grounds.
It is rather motivated by the consideration that what one ought to do depends on
appropriate expectation. Where these approaches disagree, primarily, is over how well
the Orthodox semantics performs in representing those factors that are relevant in
determining what one ought to (or must) do.
The techniques and methods of compositional semantics also enrich our under-
standing of deontic modality—a theme evident in the papers of Jessica Rett, Seth
Yalcin, and Paul Portner and Aynat Rubinstein.
Rett argues on linguistic grounds—having to do with the interaction of deontic
modals with exclamatives (e.g. “wow!”)—that deontic modals share with epistemic
modals something like an “evidentiality” requirement—the requirement, namely,
that their complements be inferred as the conclusion of an episode of reasoning.
(The reasoning is, roughly, practical in the case of deontic modals, theoretical in
the case of epistemic modals.) If right, this should bear on debates concerning the
epistemic status of deontically modal judgments. Like epistemically modal “states of
affairs”, representations of deontically modal “states of affairs” are not formed, e.g.,

3 Cariani’s account also strives to maintain this structure, but the theoretical departures seem to us much
more significant than those in Wedgwood’s proposal.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

introduction 

perceptually. They must, rather, be formed on the basis of normative commitments

one already happens to endorse.
Portner and Rubinstein argue that the relationship between weak deontic necessity
and strong deontic necessity is an instance of the general relationship between non-
extreme and extreme gradable expressions (e.g. the relationship between “big” and
“huge”, or between “smart” and “brilliant”). Theirs is destined to become an important
analysis of the strong/weak distinction for deontic modals. Portner and Rubinstein
also build usefully on the earlier-mentioned work of Portner and Lassiter concerning
the gradability of deontic modality. They argue that an appropriate account of the
scalar structure that underlies deontic modality is, in fact, compatible with Kratzer’s
earlier-described doubly relative account of deontic modality. The existence and char-
acteristics of this underlying scalar structure are themes that permeate this volume.
Future work on deontic modality must pay closer attention to it.
Yalcin, finally, argues that, while there is perhaps a clear distinction between weak
and strong deontic necessity, there is no parallel distinction to be drawn between weak
and strong epistemic necessity. The reading that is normally classified as weak epistemic
necessity is ordinarily associated with discourses of the following type:
() A: Where is Sam?
B: She ought to be in her room.
Instead, this reading is, Yalcin suggests, better classified as expressing what the speaker
would reasonably expect to hold—in this case, the modal expresses where B would
reasonably expect Sam to be. Yalcin calls this a “normality” reading of the modal, and
suggests an illuminating analysis of this reading, on which it expresses something like
a default (compare Horty ): a defeasible expectation of what follows from some
pattern of priors, and which potentially admits of both normative or nonnormative
varieties. The points of contact with Rett, and the possibilities for developing this
analysis as an account of the strong/weak distinction, will be evident to the reader.
A final important theme in this volume concerns the way the distinctive linguistic
phenomena surrounding deontic modals bear on more fundamental theorizing about
that in virtue of which modal sentences have the meanings they have, and more
generally about the way deontic modals function as part of normative language. One’s
views about these issues typically guide or at least influence the methodology of one’s
linguistic theorizing about the semantics and pragmatics of deontic modal language.
Accordingly, several of the papers in this volume address semantic and pragmatic
questions about deontic modality with metasemantic and methodological questions
firmly in view.
In philosophy, there is a familiar distinction from metaethics between views that
treat normative language as mainly means for expressing practical attitudes influ-
encing motivation and views that treat normative language as mainly means for
describing what is the case normatively. The former are often called “noncognitivist”,
“prescriptivist”, or “expressivist”, while the latter are often called “cognitivist”, “realist”,

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 nate charlow and matthew chrisman

or “representationalist”. Deontic modal claims with normative import (e.g. claims

about what someone morally/prudentially must, ought, or may do) strain this tra-
ditional theoretical division. For deontic modals are typically treated as some kind of
intensional operator, which means that sentences with a modal taking widest scope
are a means for displacing from more ordinary descriptions of what is actually the
case. So deontic modal language is not straightforwardly descriptive. But that doesn’t
obviously settle the metaethical issue in favor of the view that deontic modal claims
express practical attitudes, since most semantics for deontic modals still attempt to
identify the contribution such words make to the truth conditions of the declarative
sentences in which they figure.
In this volume there are two kinds of reactions to this strain to traditional metaeth-
ical categories. On the one hand, Benj Hellie, Malte Willer, and William Starr
argue that consideration of deontic modals motivates abandoning truth-conditional
semantics in favor of an alternative based in the logic of (or rational relations between)
preferences or intentions. On the other hand, Nate Charlow and Matthew Chrisman
argue that we can hold onto the tools and explanations of truth-conditional semantics,
as long as we can develop an expressivistic or inferentialist construal of the “truth
conditions” of deontic modal sentences to rival the more familiar representationalist
construals. (Charlow is, however, more reticent than Chrisman about calling the con-
dition delivered by the compositional semantics a “truth condition”.)
Thematizing Ross’s paradox, Hellie argues that the logic of intentions and the way
this logic interacts with the pragmatics of commands provide a rich and underexplored
structure that can serve as the basis for a novel kind of semantics for deontic modals.
As Ross famously noted, although the proposition that you post the letter entails the
proposition that you post the letter or you burn it, the command to post the letter
doesn’t entail (or “support”) the command to post the letter or to burn it; and similar
observations have been made about the deontic modal claims “You must post the
letter” which similarly doesn’t seem to entail “You must post the letter or you must
burn it.” So, Hellie argues, rather than focusing on truth conditions of propositions
and the entailment relations they are meant to explain, we should instead focus
on the way intentions can support various procedures—conceived of as structured,
nonpropositional representations of a plan for action. On the back of the structure
of this kind of practical/rational support, Hellie develops a semantics for deontic
necessity claims according to which they are something like a modalized imperative.
They say what imperatival content (which on Hellie’s view is a kind of collectively
intended procedure) is supported across all of a range of mental states, individual
or collective. As such, the resulting semantics has new resources for avoiding Ross’s
paradox and it is intended as a “support-semantic” alternative to traditional truth-
conditionalist semantics—one motivated in large part by Hellie’s metasemantic views
about the function of deontic modal language.
Willer’s concern is less with natural language deontic modals and more with the
logic of obligations, which appears to be interestingly nonmonotonic. This means that,

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

introduction 

roughly, given a set of premises that entails a deontic modal conclusion, adding a
further premise won’t necessarily preserve the entailment. If Jones promised to meet
Mary for lunch, then he ought to do so. However, if he promised and breaking the
promise is necessary to save a drowning child, then it’s not the case that he ought to
keep the promise. Willer argues that the nonmonotinicity of deontic logic extends
beyond the nonmonotinicity we are familiar with from reasoning with defeasible
generalizations. It also arises—he argues—owing to the violability of obligations and
the sensitivity of obligations to epistemic uncertainty. In response, he develops a
dynamic account of deontic logic capable of handling these additional sources of
nonmonotinicity. This generates some support for the methodology of dynamic logic
over the methodology of static logic.
In a similar vein (though with more focus on natural language), Starr’s paper
explores a dynamic semantics for a fragment of deontic modal language. The general
framework is one wherein one attempts to model the meaning of a sentence in terms
of a function specifying how accepting the sentence would change an agent’s overall
state of mind. Starr argues that accepting a deontic modal sentence primarily functions
to change one’s preferences rather than one’s beliefs about the world. In this sense,
he intends this semantics as an expressivist semantics for deontic modal language,
whereby the discourse function of this language is to coordinate on preference
orderings rather than to coordinate on beliefs about the world.
Charlow and Chrisman are motivated by some of the same intuitions as Hellie
and Starr about the preference/plan-coordinating, rather than world-representing
function of deontic modals. Accordingly, they agree with Hellie’s and Starr’s anti-
representationalist outlook regarding deontic modals. However, they develop this in
ways (different from one another) meant to be more friendly to truth-conditionalist
approaches to compositional semantics (largely leaving dynamic aspects of the mean-
ing of deontic modals at the pragmatics/semantics interface).
As mentioned above, Charlow proposes to enrich standard possible worlds seman-
tics with elements of decision theory in order to reflect the way deontic modals seem
inextricably tied up with the practical language of planning and action. Then he puts
this to service of a metasemantic view about why deontic modal language has the
semantics he predicts it to have. More specifically, he argues for an expressivistic gloss
on his semantics for deontic modals: the purpose of the semantics for deontic modals
is to represent a relation of competent acceptance between agent and deontic modal.
Chrisman, by contrast, outlines what he views as three metasemantic orientations
one could take towards truth-conditional semantics: the representationist view that
truth conditions articulate ways reality might be, the ideationalist view that truth con-
ditions articulate what one asserting a sentence ought to think, and the inferentialist
view that truth conditions articulate a place in a logical space of implications. He
argues that the fact that deontic modals are intensional operators encourages the view
that they function primarily not as devices for representing bits of reality but rather as
“metaconceptual” devices for performing various logical or quasi-logical operations

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 nate charlow and matthew chrisman

on more basic pieces of content. In the inferentialism of Sellars and Brandom he

finds a natural home for this idea, while allowing that there are also congenial ways
expressivism might be developed to make sense of it.
An extraordinary amount of recent work by philosophers of language and logic,
metaethicists, and linguistic semanticists has focused on the meaning and function
of deontic modals. This volume represents an attempt to gather together many of the
most active and distinctive voices in this discussion. The papers are on diverse topics
about natural language semantics, the psychology and logic of deontic and normative
thought, and deontic modality as a window into the foundations of meaningfulness.
We believe bringing that diversity into one place is generating significant prospects for
interesting new lines of research and a better understanding of deontic modality as a
distinctive phenomenon.

References
Dowell, J. L. () Contextualist solutions to three puzzles about practical conditionals. In
Shafer-Landau, R. (ed.) Oxford Studies in Metaethics. Volume . Oxford: Oxford University
Press.
Fintel, Kai von () The best we can (expect to) get? Challenges to the classic semantics
for deontic modals. [online] Available from: http: //mit.edu/fintel/fintel--apa-ought.pdf.
Manuscript. MIT. [Accessed Nov  .]
Horty, John F. () Reasons as Defaults. Oxford: Oxford University Press.
Kolodny, Niko and John MacFarlane () Ifs and oughts. Journal of Philosophy. . pp. –.
Kratzer, Angelika () The notional category of modality. In Eikmeyer, H. and Rieser, H. (eds.)
Words, Worlds, and Contexts. pp. –. Berlin: De Gruyter.
Kratzer, Angelika () Modality. In von Stechow, A. and Wunderlich, D. (eds.) Semantics: An
International Handbook of Contemporary Research. pp. –. Berlin: De Gruyter.
Kratzer, Angelika () Modals and Conditionals. Oxford: Oxford University Press.
Portner, Paul () Modality. Oxford: Oxford University Press.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i


Deontic Modals and Probabilities:
One Theory to Rule Them All?
Fabrizio Cariani

Introduction
An important chapter of the semantics and pragmatics of modals concerns deontic
sentences such as:
() You should move that pawn.
() You may not move a pawn that way, you must move it this way.
() It is better to move a pawn than to eat it.
A theory of the meaning of these sentences must specify when they are acceptable,
the kind of content they express, how they contribute to more complex expressions
that embed them, and what systematic effects they may have on the evolution of a
conversation.
In this paper, I compare some ways of developing such a theory on the basis of
empirical considerations as well as philosophical and methodological ones. Philo-
sophically, the main theme of my discussion is the connection between the conven-
tional meanings of deontic vocabulary and substantive theories about what one ought
to do. Accounts of the meaning of deontic vocabulary are not alone in making pre-
dictions about deontic sentences. Such sentences may also be implied by substantive,
nonlinguistic theories. For example, () may follow, depending on the circumstances,
from (i) the rules of chess (ii) from a theory of tactically correct play or (iii) from
a more general theory of rational decision making. Say that a practical theory is any
theory that issues verdicts of the form,
() Given α’s circumstances, goals and information, α ought to ϕ
Given α’s circumstances, goals and information, α is allowed to ϕ.
What is the relationship between practical theories and the semantics of deontic
modals? Should a practical theory be built into the conventional meaning of deontic

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 fabrizio cariani

modals? The hypothesis that has informed much research in formal semantics1 is
that the semantic theory should be, within certain limits, neutral between different
practical theories. This is the hypothesis I aim to clarify, articulate and defend here.
It is important to distinguish at the outset two different degrees of neutrality.
One may note that practical theories can conflict with one another about particular
sentences, and maintain that it is not up to the semantics to settle these conflicts.
Consider again: “You should move that pawn”. If my desire is to lose the game and go
home, a decision theory yields different verdicts from a theory of tactically correct play.
To be neutral in the first sense, a semantic theory should be above these differences.
A precise way of putting the requirement is that for any agents α, and predicate ϕ, such
that it is contingent whether ϕ applies to α, α should ϕ is consistent.
This degree of neutrality is usually achieved by adopting a flexible semantics
for modals. For example in the classical, contextualist ordering semantics, deontic
modals—and, in fact, modals in general—are analyzed relative to contextual param-
eters. Among these contextual parameters is an ordering of possibilities. In context,
practical theories can help fix this ordering, but that is the entire extent of their
contribution. However, there are no context-independent constraints on what can be
the source of this ordering. On this picture, there is a thin lexical meaning for deontic
modals that is largely independent of practical theories, but practical theories can
affect the sharpenings and connotations that individual deontic modals acquire in
context. Given some orderings, “You should move that pawn” is true; given others,
it is not.2
I will argue that we need a slightly stronger degree of neutrality. There are reasonable
practical theories that make simultaneous predictions about sets of sentences. If the
semantic theory is to be compatible with these theories, it seems, it should make those
sets of sentences consistent. Accordingly, the stronger degree of neutrality is this: for
every coherent practical theory P, if there are circumstances in which P entails a set
of deontic sentences S, then the semantic theory ought to imply that S is consistent.3
So, for example, if there is a coherent practical theory that entails “You should move
that pawn unless your king is threatened” and “If your king is threatened, you should
move the queen”, then these sentences ought to come out as jointly consistent. It
is important to note at this stage that even the stronger degree of neutrality does

1 See Kratzer (, , b, ); Lewis (, ).

2 I do not assume that a semantic theory for deontic modals must appeal to truth-conditions. The
ascription of truth here is in the contextualist’s voice.
3 Other authors have also suggested a similarly strong degree of neutrality (Carr, , ; Charlow,
). However, I want to emphasize a potential difference. Carr criticizes the classical ordering semantics
for modals for implementing a particular “decision rule”. That is: because the classical semantics operates
by selecting the “best” worlds out of the ordering, Carr thinks it implements the decision rule maximax.
Similarly, she argues that the semantic proposal of Cariani et al. () implements the decision rule
maximin. I will not tackle these arguments directly (I want the reader to focus on a different set of details).
I do want to make a methodological point: these arguments cannot be made simply on the basis of structural
features of a lexical entry. A semantic theory might select the best worlds out of an ordering, but it does
not mean that it implements maximax. To sustain an argument that a theory is not neutral, there have to
be specific examples of coherent practical theories that are ruled out by the semantics in specific contexts.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

deontic modals and probabilities: one theory to rule them all? 

not amount to the suggestion that the verdicts of every practical theory should be
consistent. Some practical theories are straightforwardly incoherent. No one claims
that their predictions ought to be made consistent.
My main negative thesis is that analyses that attempt to explain the meaning of
deontic modals in terms of expected value are in violation of this stronger, but not
universal, requirement of neutrality.4 The argument against expected value accounts
will not assume neutrality as an abstract principle. What I hope to show, instead, is
that violating it has specific problematic empirical consequences (§).
On the other hand, expected value accounts have some attractive features. They
systematically capture the idea that our dispositions to accept and reject deontic
statements are tightly linked with probabilistic judgments. Indeed, before criticizing
them, I will describe some facts concerning the interaction of probability operators
and deontic modals that seem to motivate expected value accounts (§).
The main positive thesis of the chapter is that it is possible to preserve the virtues of
expected value accounts while avoiding their problems. I show how to characterize a
probabilistic premise semantics that shares the features that motivate expected value
accounts without their problematic consequences (I develop an initial version of my
proposal in § and a refinement in §).5
In addition to these programmatic theses, the paper also has a more modest
ambition. It is to show, as the semantics of §§– does, that it is possible to merge the
insights of a classical premise semantics for deontic modals with the current explosion
of work on probabilistic semantic theories for epistemic vocabulary. In other words,
to show that it is a mistake to view the space of possible theoretical options as neatly
partitioned in two—with classical non-probabilistic theories on the one hand and
expected value analyses on the other.

 Deontic and probabilistic talk

Many (though not all) theories of deontic modals relativize the interpretation of
ought and should to a probabilistic state. To get a sense of the variety of proposals,
I will list some examples of probabilistic theories. According to one version of the
expected value analysis (see § for more details): should A is true if and only if the
expected value of A (as calculated on the basis of contextually given probability and
value function) is greater than the expected value of the relevant alternatives.6 This

4 Expected value accounts are developed, among other places in Goble (); Cariani (); Lassiter
(); Wedgwood (). This kind of view is also explored without endorsement in Yalcin (a). The
theoretical proposals in these references differ on several substantial points.
5 The proposal generalizes the theory of Cariani et al. (). A preliminary sketch has appeared in my
Cariani (b).
6 In this paper, metalinguistic variables A, B, C range over sentences, while A, B, C range over sets of
worlds. In informal discussion and within the same stretch of discourse the value of A is the set of worlds
at which A is true. In more formal discussion, I use the more standard notation [[A]] to denote A.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 fabrizio cariani

theory counts as probabilistic because expected values are defined partly in terms
of probability functions. Another probabilistic theory is the end-relational analysis
(Finlay, , ). According to this view, should A is true in a context c just in
case the (contextually salient) probability of A given the contextually salient end E
is greater than the probability of the relevant alternatives given E. According to yet
another analysis (see Carr ) should A is true iff A is entailed by all the “best”
alternatives where “best” is calculated on the basis of contextually supplied decision
rule d, probability Pr, and value function v. To these theories, I will add a fourth
in §.
More precisely, these analyses appear to commit to:
Probabilize Deontic Semantics (in short: Probabilize):
(a) In defining an intepretation function [[·]]· for a fragment of natural language
including deontic modals, we should include a probabilistic parameter (e.g.
a probability function) among the parameters relative to which we interpret
sentences.
(b) Deontic sentences depend non-trivially on this probabilistic parameter (to
be precise this means that for some deontic sentence A there are parameter
assignments π and π ∗ that differ only in their Pr coordinate, such that
∗
[[A]]π = [[A]]π ).
What motivates this idea? In this section, I propose two arguments in its favor.
I emphasize that neither argument is intended as an unconditional proof of Prob-
abilize. In other words, neither argument aims to motivate every kind of theorist
regardless of their prior commitments. Instead, both arguments depend on substantive
assumptions that I do not intend to defend here.

. Probabilistic Deontic Conditionals

The first argument turns on examples in which deontic modals appear embedded in
conditionals that use both probabilistic and deontic operators (the core observation
of this argument is from Yalcin a). Imagine an experienced soccer coach advising
a less experienced one about an upcoming game. Suppose that the more experienced
coach says:
() If it is likely that your opponents will attack you on the right flank, you should
concentrate your defense on the right side.
() If it is not likely that your opponents will attack you on the right flank, you
should not concentrate your defense on the right side.
We have no difficulty interpreting these conditionals. We clearly judge them to be
consistent and expect them to figure in certain inferential patterns. For example, the
conclusion that you should concentrate your defense on the right follows from ()
together with:

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

deontic modals and probabilities: one theory to rule them all? 

() It is twice as likely that your opponents will attack you on the right flank
as not.
These are the kinds of facts we would like a joint theory of modals, conditionals, and
probability operators to explain.7
Accounting for the meanings of embeddings such as ()–(), I claim, motivates
Probabilize. The argument I advance here is that, given a natural view of the meanings
of their probabilistic antecedents, we must handle their consequents as sensitive to an
underlying probabilistic structure.
Let me first discuss the probabilistic antecedents. Much recent literature has devel-
oped reasons in favor of a probabilistic semantics for such sentences as:8
() It is likely that they will attack you on the right.
For example, Yalcin () shows that a probabilistic semantics is a natural fit for a
core of inferences involving probability operators.9 Swanson () notes that it seems
difficult, if not impossible, to reduce the quantitative aspect of probabilistic discourse
to a purely qualitative basis. On a simple probabilistic semantics, the semantic value
of the probability operator likely (short for ‘It is likely that. . .’) may be characterized
as in ().
() [[likely A]]c,Pr,w = T iff Pr(A) > .
If a probabilistic semantics such as () is on the right track, we should ask what
sort of local context is created by evaluating conditional antecedents of the form
If it is likely that A, . . . .
One way to get a grip on this problem is to think about what it is to suppose
that It is likely that A. One option is factualism—the view that in making such
suppositions one entertains a standard possible world proposition. Factualism is
compatible with the semantics in (), provided that one is a contextualist of a certain
kind. One can maintain that context determines a probability function Prc , and so

7 It is sometimes said that the goal of semantic theory is to predict which inferences are deemed
acceptable by competent speakers of a language (Chierchia and McConnell-Ginet, , p. ; Yalcin, c;
Holliday and Icard, , §). From this point of view, it may appear unclear exactly what needs to be
explained about ()–(). Surely, a minimal bar would be to explain why they are consistent. However, and
crucially, a semantic theory must deliver at least two more things: first, together with the metasemantics,
it must make predictions about speakers’ dispositions to accept sentences in context; secondly, it must
connect smoothly to an account of communication and conversation, so as to explain phenomena such as
disagreement, assertion, retraction and to form a basis for the calculation of implicatures. It is with respect
to the first of these two additional tasks that ()–() pose the biggest problems.
8 Yalcin (, , b); Swanson (); Lassiter (); Rothschild (); Moss (). In addition
to these authors, some think that probabilistic resources might be needed for the semantics of conditionals:
see Edgington (), Bennett () among others.
9 But see Holliday and Icard (, §), for an account that retrieves essentially the same core inferences
identified by Yalcin in a comparative probability setting. In reply, Lassiter (forthcoming) notes that there
might still be inferences involving sentences with ratio modifiers (A is twice as likely as B) that elude the
comparative probability approaches.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 fabrizio cariani

It is likely that A expresses in c the proposition that is true at a world w just in case
Prc (A) > .. As for how context might determine such a probability function, one
option is to claim that Prc is the subjective credence of the speaker of c. More complex
accounts are possible (e.g. the collective credence of some salient group in c).
Despite the familiarity of this proposal, there are well-known reasons to doubt
it (Yalcin, , ; Swanson, ; Rothschild, ). For instance, Yalcin (,
pp. –) and Rothschild (, p. ) note that factualism seems to make
attitudes towards modal and probabilistic contents be about the wrong sorts of things.
To make a supposition that A is likely is to (provisionally) enter into a particular state of
uncertainty about the world. According to the factualist, however, when Dana makes
such a supposition, she enters into a state of certainty about a credence, either her own
or the group’s. This seems to be the wrong content for the supposition in question, and
it motivates a different model of attitudes towards probabilistic contents.
I don’t wish to debate the factualist here about whether her position has been
thoroughly refuted by these objections. These arguments are meant to open up new
lines of inquiry, not close off the old ones. They motivate a non-factualist model
of probabilistic discourse. Ultimately, the score can be settled only globally—by
considering a large variety of phenomena and fully fleshed-out theories.
The baseline non-factualist idea for probabilistic sentences is this: to suppose it will
likely rain is provisionally to adopt a state that supports the probability of rain. Moving
back to the case of conditionals, the basic idea will be to conceive of probabilistic
antecedents as operating on an epistemic (and specifically probabilistic) state that
is given as a coordinate of semantic evaluation. Evaluating an antecedent of the
form likely A involves creating a local context according to which A is likely. The
consequent of the conditional is evaluated relative to such a local context. The details
of this operation vary from semantic theory to semantic theory. However, a simple
preliminary model of how this might work (not the one I will end up adopting in §)
is as follows. Assume that probability operators are evaluated relative to a probability
function, as in () above. To evaluate conditionals of the form (if likely A)(B), shift the
probability function Pr, to some other function Pr  that is suitably related to Pr and
makes A likely.
If this picture of the behavior of probabilistic antecedents is on the right track,
an account of the meaning of () and () requires a semantics for should on which
the deontic modal is sensitive to the probabilistic update introduced by likely A.
Otherwise, we would not be able to predict the consistency of ()–(). The same
holds for other modal expressions with deontic interpretations, such as may, must,
better, etc.
The first argument for Probabilize, then, is that it constitutes a natural extension
of a plausible picture concerning the meaning and conversational role of expressions
of probability. The extension is needed to make sense of what pairs such as ()–()
mean. It should be obvious that this argument for a probabilistic framework is only
as strong as the underlying picture of probability claims. I take this as a welcome
result: Probabilize does not seem to be an uncontroversial thesis, but it is supported

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

deontic modals and probabilities: one theory to rule them all? 

by independently interesting and well-developed views concerning the language of

subjective uncertainty and is a natural extension of these views.
. Information Sensitivity and Probabilistic Structure
My second argument for Probabilize involves embeddings of deontic modals in the
consequent of ordinary deontic conditionals. Kolodny and MacFarlane () argue
that deontic modals are information sensitive. Very roughly, their central hypothesis
is that a correct account of deontic conditionals requires a quantificational semantics
for ought and, I suppose, should, on which the domain of quantification is specified
as a function of an underlying information state. This proposal is motivated by an
analysis of the Miners puzzle (Regan, ; Parfit, unpublished, ; my variant is
loosely inspired by Jackson, ):
Rescue: in a faraway land, there is a natural disaster and you are the only rescuer available.
Before you go on your mission, you are offered a choice of taking one of two pills (you may also
refuse and take no pills). If you refuse, you will rescue nine people. The outcome of taking either
of the pills depends on your genetic makeup. The red pill will give you extra energy if you have
the red gene (you will rescue ten people), but it’ll make you faint if you have the blue gene (you
will rescue zero people). The blue pill will give extra energy if you have the blue gene and make
you faint if you have the red one. No one knows your genetic makeup and there is not enough
time to find out: as far as we all know, probabilities are /.

It is helpful to visualize the scenario as a decision matrix:

Red Gene (.) Blue Gene (.)

Take Red Pill Rescue  Rescue 
Take Blue Pill Rescue  Rescue 
Refuse Pills Rescue  Rescue 

We want to derive that the following judgments are jointly consistent:

() You should refuse the pills.
() If you have the red gene, you should take the red pill.
() If you have the blue gene, you should take the blue pill.
An information-sensitive semantics, on which the domain is specified only as a
function of an information state i, provides an intuitive account of these examples.10
This is Kolodny and MacFarlane’s information-sensitive semantics:

10 There is a complex debate to be had about what sort of state i is and how it gets fixed. Towards the end
of their (), Kolodny and MacFarlane hint towards relativism. According to the relativist, the knowledge
state of the assessor sets the initial value for i (other operators might shift it away from its initial value).
However, their compositional semantics is explicitly neutral on this issue: a contextualist and even a certain
kind of expressivist might deploy it. Likewise, the semantic theories presented here do not require a stand on
these issues. Relativism is more explicitly avowed in MacFarlane (, §.ff.), but again the compositional
semantics presented there does not require a relativist construal.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 fabrizio cariani

() [[should A]]c,d,i,w = T iff ∀w ∈ d(i), [[A]]c,d,i,w = T
Here d is a deontic selection function that inputs the information state i and out-
puts a domain. Since Kolodny and MacFarlane assume that the job of conditional
antecedents is to shrink the information state i, deontic conditionals may quantify
over a significantly different domain from bare deontic claims. This, in turn, allows an
explanation of the consistency of (), () and (). It turns out to be pretty difficult
to explain how to derive all three verdicts at once if, instead of being information-
sensitive, deontic modals quantify over a domain that is fixed by context in a more
direct way.11
The move to an information-sensitive semantics is, in my view, well-motivated.12
Since the present aim is to develop the information-sensitive approach, rather than to
justify it, I do not discuss ways of resisting arguments for information-sensitivity (see
the references in footnote ). My focus here is to argue that if we adopt an information-
sensitive approach, there is reason to take an extra step and adopt a probabilistic model
of information.
Kolodny and MacFarlane’s semantics makes deontic modals responsive to qualita-
tive information states. Consider, for example, this variant on Rescue:
Knowledge: Everything is exactly as it is in Rescue but now you (the agent) as well as all the
participants in the conversation acquire conclusive and irrefutable evidence that you have the
red gene.

In Knowledge, we judge () to be unacceptable: intuitively, you should take the

red pill. Kolodny and MacFarlane’s apparatus can deliver this verdict: acquiring the
information that the agent has the red gene may shift the domain. In the resulting
domain, every world may be one in which the agent takes the red pill.
In Knowledge, we add information that has a simple qualitative effect on i: we know
that you have the red gene. The situation of perfect ignorance in Rescue becomes one
of perfect knowledge. But there are cases whose effects seem best described in terms
of graded states:
Slanted: Everything is exactly as it is in Rescue but now you have strong evidence (known to
you and to all the participants to the conversation) that leads you to assign . probability to
your having the red gene.

As in Knowledge, in Slanted, it is plausible to judge () to be unacceptable.13 If there

is a . probability that you have the red gene, it is plausible that you should take the

11 In addition to Kolodny and MacFarlane (), see also Charlow (); Cariani et al. ();
Silk () for developments of the pro-information sensitivity arguments. For defenses of the classical
semantics, see von Fintel (unpublished); Dowell ().
12 In my preferred account, the parameter i does not have to be tied to information. But whatever it is
tied to, the job of prioritizing possibilities must be executed as a function of i.
13 I am not claiming that the semantic theory must deliver this verdict of unacceptability for (). To do
so would be to say that risk-averse decision theories are false as a matter of meaning. The point in the main
text is not to exclude these theories. It is, instead, that, in context, there are natural practical theories that
reject () in Slanted.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

deontic modals and probabilities: one theory to rule them all? 

red pill (of course, if . is not enough we could choose a larger probability value short
of ).
On the basis of these cases, I advance the following argument:
(P) If the difference between Knowledge and Rescue is traced to the value of an
informational parameter in the semantics, then so is the difference between
Slanted and Rescue.
(P) The difference between Knowledge and Rescue is traced to the value of an
informational parameter in the semantics.
(P) If the difference between Slanted and Rescue is traced to the value of an
informational parameter in the semantics, then Probabilize holds.
(C) Probabilize holds.
I think (P) is intuitively plausible: the cases are not sufficiently different to warrant
altogether different treatments. Those who accept Kolodny and MacFarlane’s proposal
that deontic modals are information sensitive also accept (P). But I suppose that one
may—and Kolodny and MacFarlane () do—disagree with (P). As they point out,
they adopt:
an epistemic and non-probabilistic model of information states; it takes information states to
be sets of known facts. We have chosen this model because we think that what one ought to
do (relative to an information state) supervenes on what is known: mere differences in beliefs
(or partial beliefs) or perceptual states, unaccompanied by differences in what is known, cannot
make a difference to what an agent ought to do (footnote , p. ).

Armed with this picture, one may diagnose Slanted as involving acquisition of a piece
of non-probabilistic evidence which affects the probabilistic facts only via the relevant
supervenience function. Imagine, for example, that the context of Slanted arises after
some reliable test for the red gene has come back positive.
I do not know whether the supervenience claim Kolodny and MacFarlane endorse
is true. Luckily, I do not think we need to engage with the debate over its truth. Even
granting the supervenience claim, it still does not follow that we should theorize at
the level of the supervenience basis (i.e., qualitative information states). It is a familiar
point that there are supervenience structures in which the “higher” levels are explana-
torily autonomous. For example, chemical facts might supervene on physical facts
without undermining the autonomy of chemical explanations. Closer to our concerns,
the explanatory constraints on semantics dictate relativizations that are independent
of whatever supervenience relations might hold. For instance, one may relativize the
truth-conditions of “Someone owns a guitar” to worlds and assignments but that
is perfectly compatible with the claim that existential facts about guitar ownership
supervene on worlds alone. The case of the relationship between probabilistic states
and deontic judgments may be one such case.
In sum, Kolodny and MacFarlane’s argument does not motivate resistance to my
premise (P), and without such motivation, my argument for Probabilize stands.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 fabrizio cariani

. Taking Stock

Consideration of embeddings of deontic modals judgments, e.g. in the consequents
of conditionals, provides two significant arguments for Probabilize. Since deon-
tic modals can occur (§.) in the consequents of conditionals with probabilistic
antecedents, a semantic theory must explain how these are connected and a theory
that satisfies Probabilize seems ideally placed to do so. Furthermore (§.), there is
independent reason to maintain that deontic judgments may have to be evaluated
relative to information states. Once that is granted, it is difficult to resist the fine-
graining of these states.

 Expected value accounts

In this section, I introduce an expected value framework for deontic modals in more
detail. I raise three criticisms that apply to every version of the theory I am aware of.14
. Spelling out the Expected Value Accounts
Suppose that context provides a probability function Pr and a value function v that
assigns to each possible world w a real number representing the value of that world
according to some contextually salient standard. With these tools (and assuming for
simplicity that there are finitely many worlds) we can calculate one kind of expected
value for a possible-world proposition A as:

EV(A) = v(w) · Pr(w|A)
w∈A

This quantity is the weighted average of the value of the worlds that belong to A,
weighted by their probability (conditional on A). Some versions of decision theory
evaluate the rationality of choices on the basis of what might appear to be a sharpening
of the above concept of expected value.15 The sharpening in question involves several
additional commitments. For example, v is identified with a quantity called utility—a
numerical representation of the subjective desirability of an outcome. It is also claimed
in these theories that the rationally permissible choices are exactly those choices that
maximize expected utility.
Neither of these commitments is essential to use expected values in semantics.
According to the best versions of this approach, the choice of value function is not
encoded by the semantics and is left up to context (hence the value function need

14 For different critical discussions of the approaches I discuss in this section, see Rubinstein (,
§..) and Yalcin (a, §IX).
15 When one looks closer, it turns out to be quite different. Before formulating a precise notion of
expected utility, a decision theorist will generally set up a sophisticated modeling discipline concerning
what counts as a decision problem (see, inter alia, Joyce , chapter ). This modeling discipline affects
how one formulates the relevant notion of expected value. For example, utility does not attach to individual
possible worlds, but to coarser objects called outcomes.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

deontic modals and probabilities: one theory to rule them all? 

not be identified with an agent’s utility function).16 Proponents emphasize that the
point of the expected value analysis is not to stipulate that deontic modals implement
a particular decision theory—after all, they accept the milder of our two degrees of
neutrality outlined in the introduction. It is, rather, that we must do semantics with
a finer set of resources and more structure. For similar reasons, the semantic analysis
is not wedded to the particular way of computing expected values I have identified
here. One may well develop a semantic theory based on a notion of expected value
that is closer to how causal decision theorists (e.g. Skyrms, ; Joyce, ) define
the notion.
Expected value analyses come in many different variants. On one variant, should-
sentences directly express comparisons of expected value: should A may be inter-
preted as meaning that A has higher expected value than the relevant alternatives;
or perhaps it may mean that A has an expected value higher than a contextually set
threshold.
Both of these approaches yield a non-classical deontic logic (for example, should is
not closed under logical consequence and does not agglomerate over conjunction).
For those who prefer more canonical logics,17 it is easy to sketch an expected value
analysis that generates a relatively classical deontic logic. One just needs to adopt two
ideas:
(i) should/ought are analyzed as universal quantifiers
and
(ii) the domain over which they quantify is the union of those salient alternatives
that maximize expected value (relative to the given contextual parameters).
All of this is to highlight an important point: favoring expected value analyses does not
require choosing a particular semantic account and it does not determine a particular
logic.
In the rest of this section, I discuss expected value accounts generally, regardless of
differences in the logics that they project. This makes my critical task more difficult,
because I restrict myself to raising objections that apply to all expected value theories
(and there are other objections that apply to some theories, but not to others). To
narrow down the field of opposing views, I limit myself to views that entail:
(FB) should A is true if (but not necessarily only if ) the set of salient A-worlds
has higher expected value than each of the salient alternatives.

16 This is why Wedgwood () helpfully suggests using the terms “value” and “expected value” rather
than “utility” and “expected utility”. I follow Wedgwood on this score.
17 Even though I have argued (Cariani, a) that closure under logical consequence is not a desidera-
tum in a logic for should and ought, I also proposed that agglomeration is desirable over conjunction—with
the possible exception of moral dilemmas. See Cariani (forthcoming).

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 fabrizio cariani

(FB) should A is false if (but not necessarily only if ) for some B, the set of salient
A-worlds and the set of salient B-worlds (a) are disjoint and (b) have the
same expected value.
In less formal terms, (FB) says that if o is better than every other option, then o is what
you should do. (FB) says that if two incompatible options have equal expected value,
then it cannot be that one of them is exactly what you should do. These constraints are
satisfied by all the expected value accounts I know of.
Expected value theories can easily account for Rescue and its variants (Lassiter, ,
§...). Consider again:
() You should refuse the pills.
() If you have the red gene, you should take the red pill.
() If you have the blue gene, you should take the blue pill.
Suppose that context determines the value function v as follows v(w) = the number of
people you rescue in w (it is an arbitrary choice, in the sense that we might have picked
a different v, but it will do as an illustration) and that Pr is a probability function that
satisfies the constraints stated in the description of Rescue (e.g., the probability of your
having the red gene is .). Then,
• () is predicted to be acceptable because refusing the pills has the highest
expected value among the alternatives.
• ()–() are also predicted to be acceptable. On one way of construing condi-
tionals, conditional antecents can update the salient probability function (see the
account of the conditional in the system in §, which is available to the expected
value theorist).18 In the local context created by the antecedent of (), the option
that maximizes expected value is taking the red pill. In the local context created
by the antecedent of (), the option that maximizes expected value is taking the
blue pill.
Expected value theories also predict the variants of Rescue I have considered. For
instance, the reversal on () in Slanted and Knowledge is predicted because the
expected value of taking the red pill increases as the probability of your having the red
gene increases. Once it increases far enough, taking the red pill might become the
option with the highest expected value.
Summing up: Rescue and its variants involve decision-theoretic verdicts. A theory
that mirrors the structure of Bayesian decision theory is well suited to explain them.

18 This assumes that one of the effects of evaluating a conditional is to conditionalize the underlying
probability function. I think it is plausible that this is one possible construal of deontic conditionals. But
I also think that there is another construal on which conditional antecedents restrict a covert necessity
modal, even when should is present. See my discussion in fn. .

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

deontic modals and probabilities: one theory to rule them all? 

. Against Expected Value Accounts

Despite these advantages, there are important reasons to explore alternatives to these
accounts. I consider three kinds of implausible verdicts they force.

attitudes of non-bayesian agents

Suppose that John is an extremely risk-averse subject. According to John, in every
decision problem, one should choose one of the options that yield the least bad worst-
case scenario (John abides by the decision-theoretic rule maximin and thinks everyone
else ought to as well). According to John’s decision-making extremism, you should
refuse the pill even in Slanted (this was the case in which there is . probability
that you have the red gene). Suppose, now, that John deliberates about Slanted on
your behalf; because of his commitment to maximin, he comes to the conclusion that
refusing is the thing to do. As a consequence, the following seems acceptable:
() John thinks you should refuse the pill.
How should an expected value theorist handle ()?
The first task is to find an appropriate model for j thinks that A, when A is a
sentence including deontic vocabulary. Expected value theories relativize composi-
tional semantic values (at least) to quintuples of the form i, Pr, Alt, v, w, where i is a
qualitative state (a set of worlds), Pr a probability function, Alt a set of alternatives, v
a value function, and w a world. In this setting, a prominent option for the semantics
for thinks would be to extend a Hintikka-style operator (Hintikka, ) treatment to
these complicated points of evaluation. A first stab might be:19

[[thinks(j, A)]]i,Pr,Alt,v,w = T iff for all worlds z in ij(z) ,

j(z) ,Pr j(z) ,Alt,v j(z) ,z
[[A]]i = T.
Notation: ij(w) is j’s qualitative state in w, Pr j(w) is j’s credence in w, vj(w) is j’s value
function in w.
Let us concede for the sake of argument that the above descriptions (e.g. ‘j’s value
function in w’) are all uniquely referring.20 The theory’s prediction is that () says
something roughly like:

19 The analysis that follows is built on a modification of Hintikka’s account based on ideas by Stephenson
() and Yalcin (). The innovation is to model modal belief by shifting not just the world of evaluation
to match the belief worlds, but also the entire belief state i.
20 A possible complication is relevant here: plausibly, thinks(j, should(A)) should not switch the value
function v to one based on j’s relevant priorities, but rather to the value function that j assigns to whoever is
the agent of the deontic modal (if there is one). To address this point, Yalcin (a) formulates the expected
value theory by replacing v, with functions h(A, w) (modeled after the hyperplans of Gibbard ) that take
each agent A and world w to a value function. We could then say that thinks shifts the salient h-function,
to hj , i.e., the hyperplan of the subject of the attitude ascription. This all strikes me as correct, but it isn’t
necessary to cover the cases I deal with here: I assume, instead, that j’s value function in w is based on the
agent’s priorities.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 fabrizio cariani

() For every probability function and value function compatible with John’s state,
refusing the pill has higher expected value than the alternatives.
At this point, the problem should be clear: () does not capture the meaning of ().
() is not an ascription to John of the content that refusing the pill maximizes expected
value. Rather, () is an ascription to John of a way of prioritizing alternatives relative
to which refusing the pill is the best option.
Expected value theorists might reply that the attitude ascription expresses a com-
parison among expected values relative to a value function that builds in John’s risk
aversion. After all, as I emphasized, they need not think that one’s value function
coincides with one’s utility in the decision theorist’s sense, and it may appear that my
argument above requires this.
Although I agree that this move is available, it does not solve this particular problem.
To see why, we just have to change some features of the case, by taking the option of
refusing the pill off the table. Consider this variant context:
Rescue− : Everything is exactly as it is in Rescue but now someone is going to force-feed you a
pill (either the red one or the blue one). You get to choose which pill you will be given.

In this context, it is plausible to accept:

() John does not think you should take the red pill and does not think you should
take the blue pill.
After all, John thinks that either pill would be equally good (or equally bad), so he
thinks you should be indifferent. Since no other options are salient, this means that
taking the red pill and the blue pill must each have the same expected value.
Now, suppose that some evidence becomes common ground that makes it . likely
that you have the red gene. That is, Rescue− gets updated to a context that resembles
Slanted, except that, again, you do not have the option of refusing. Call this context
Slanted− . Despite the fact that Pr(red gene) = . (in Slanted− ), maximin does not
prioritize taking the red pill over taking the blue pill. That is to say that in Slanted−
John remains indifferent in respect of choice between the pills, and hence () is still
acceptable. But the expected-value semantics does not predict this. The expected-value
theory must be bound by the following structural commitment:
If John thinks you should be indifferent in respect of choice between taking red or blue in
Rescue− , and if Slanted− results only by shifting the relevant probability function to make
it . likely that you have the red gene, the expected value of taking the red pill in Slanted−
must be higher than the expected value of taking the blue pill.21

21 Technically, it is possible to define a value function that violates this. For example:
v(w)= if, in w, you are certain that you have the red (viz. blue) gene and you take the red (viz. blue)
pill.
v(w)= if, in w, you are uncertain about your genetic state and take neither pill.
v(w)=– if, in w, you either take a pill in a state of uncertainty, or take the wrong pill.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

deontic modals and probabilities: one theory to rule them all? 

Therefore, the expected-value theorist predicts, contrary to our judgments, that () is
unacceptable in Slanted− .
To avoid this result, the expected-value theorist might claim that, in updating from
Rescue− to Slanted− , we must also change the value function. I see two problems
with this reply. First, this is neither intuitive nor theoretically motivated: nothing in
the update to Slanted− suggested that we had to affect the value function. Secondly,
even in Rescue− , we can say things like:
() Even if it were more likely than not that you have the red gene, John would
still not think you should take the red pill.
It is implausible that the local context created by the antecedent of () involves a
change in value function.
Another reply that is available to the expected-value theorist might be to reject the
operator treatment of epistemic attitudes. In the literature on propositional attitudes,
the standard alternative to operator treatments is a relational account. Let C be a
context, however conceived. Then, the first stab at a relational semantics for thinks is
given by:
[[thinks(j, A)]]C,w = T iff in w, j stands in the thinks-relation to the structured
content of A in C.
There are two main reasons to adopt a relational analysis of thinks over an oper-
ator account. First, to slice contents more finely than is allowed by a possible-
worlds representation (e.g., if  and  are distinct tautologies, we may want
to accept John thinks that   but not John thinks that  ). Secondly, to block
problems of logical omniscience (e.g., if A is a contingent sentence, it need not follow
from John thinks that A that John thinks that  ). The problem I raised for the
expected-value theories does not appear to be of either kind: it does not involve the
claim that the expected-value approaches slice contents too coarsely, and it it does not
appear related to issues of logical omniscience.
This observation leads to a more general point: the appeal to the relational analysis
is merely evasive, unless it is complemented by an account of what content is expressed
by a deontic sentence in a given context. And it is difficult to imagine an account of the
content of should A that (i) is plausible in light of the meaning that the expected-
value semantics assigns to deontic sentences and (ii) makes the right predictions
about () and all the related cases in which an agent has what I have been calling a
“non-Bayesian” attitude. Let me emphasize that this is in no way a criticism of the

This kind of value function is counter to the spirit of information-sensitivity theories. All these theories
separate the contribution of the information state from the contribution of whatever device we use to
prioritize possibilities. This objection has an empirical component, in that the expected-value approach,
with v as value function and the natural assignments to the other parameters, wrongly predicts that
(i) John thinks that if you have the red gene, you should take the red pill.
should come out false.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 fabrizio cariani

relational analysis. My target is the claim that the relational analysis helps with the
problem. The point that supports it is that the kinds of facts that justify the relational
analysis seems orthogonal to the present argument.
My opponent may reply here that they are not orthogonal, after all—perhaps John
really does have inconsistent deontic beliefs. But there is no reason to accept this
claim (the points that follow are more extensively discussed in Cariani, , which
was written as an elaboration of this argument). As a preliminary point, the contents
of John’s beliefs are strictly speaking consistent even according to expected value
analyses: these analyses make them inconsistent only relative to background facts
about the context. More importantly, there is a clear intuitive difference between how
we judge John’s beliefs and how we judge the beliefs of someone who ranks alternatives
in a totally incoherent way (e.g., by minimizing the minimum value). There is no
pressure for the semantics to predict the consistency of the deontic beliefs of such
an agent. From this point of view, the “miniminer” is just like someone who has
incoherent views about, say, universal quantification. I am happy to rest my argument
on the judgment that not every deviation from the expectational paradigm is a form
of logical inconsistency.

disagreement about decision rules

The previous argument can be extended and simplified by considering judgments
about disagreement. Suppose that Will, like John, has a favorite decision rule. Unlike
John, Will thinks that everyone ought to maximize expected utility. Intuitively, Will
and John agree in Rescue− that:
() It is not the case that you should take the red pill and it is not the case that you
should take the blue pill.
Similarly, it is intuitive that they disagree on:
() If it is likely that you have the red gene, you should take the red pill.
I argue that if expected-value analyses model the agreement on () correctly, then
they cannot model the disagreement on ().
To set up this argument, I need a skeletal model of agreement and disagreement
between different parties.22 Given an evaluator α and a point of evaluation b, we
might associate with α a shifted point of evaluation bα —similar to what we obtained
in the operator semantics for thinks (but now without the aim of giving a semantics
for attitudes). Earlier, I considered points of evaluation of the form i, Pr, Alt, v, w.
So, for instance, bWill might be iWill , Pr Will , Altb , vWill , wb . We can then say:

22 As with the account of attitudes I entertained in the previous section, this is an extremely simplified
approach that is nonetheless sufficient to highlight a difficulty that remains problematic even when we
complicate our account of disagreement. For a much more nuanced story about normative disagreement,
see Plunkett and Sundell ().

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

deontic modals and probabilities: one theory to rule them all? 

α and β disagree on A (relative to b) iff [[A]]bα = [[A]]bβ .23

In this setting, ()–() pose the following problem: given that it is settled that you will
take one of the pills, the only way for John and Will to agree on () is if the expected
values of taking red and taking blue are identical (in their respective states). But, if that
is the case, John and Will should not disagree on ()—for reasons that parallel those
I discussed in the case of the attitudes. The conditional antecedent in () operates on
the probabilistic coordinates of bA , bB , creating local contexts in which having the red
gene is more likely than having the blue gene. So, α and β should both accept ().24
In sum, the expected-value analysis can account for some disagreements by postu-
lating value functions that differ from the agent’s utility function. If, however, we use
patterns of agreement to constrain the value function, as we did to get the agreement
on (), the analysis cannot account for disagreements for all sentences that involve a
shift in the background state.

zero-probability events and decision problems

In the philosophy of probability, several authors have stressed the distinction between
probability  and epistemic impossibility (for a recent discussion, see Easwaran,
). This observation has important implications in decision theory (Skyrms, ;
Easwaran, ; Hájek, unpublished; see Briggs, , §., for a nice summary of this
discussion). These extensions have a direct impact on deontic semantics, especially on
expected-value theories. Consider this case:
Darts. Jeff and Zara are playing a game. Zara throws infinitely fine darts at the [,] interval.
Zara is a perfect randomizer on such dart throws: her throw is fair. Jeff makes predictions about
the outcome of Zara’s throw. Let D∞ =“this dart thrown at the [,] interval will hit π/”. If
Jeff raises his right hand, he is predicting that D∞ will happen. If Jeff raises his left hand, he is
predicting that it will not. Suppose that Jeff gets prizes according to the following matrix.

D∞ ∼D∞

Raise right hand $ M nothing

Raise left hand nothing nothing

23 Note that this notion of disagreement concerns a sentence in context. In my view the most fundamen-
tal notion of disagreement concerns contents of speech acts and attitudes, but I intend to run the present
argument without adding to the expected-value framework controversial views about contents. Given this,
the notion in the text seems fairly intuitive (see Willer, , for a use of a similar notion in the context of
disagreement with epistemic modal sentences).
24 Dan Lassiter points out to me that this is not technically true on the account of Lassiter (): on his
view, should-sentences can be false if there are small differences in expected value among options. However,
I don’t think that this response avoids the problem, though it may patch some of its occurrences. Even
on Lassiter’s view, there are many contexts in which two options A and B have similar expected value
unconditionally, but have vastly different expected values if we shift or update the underlying probability
function.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 fabrizio cariani

In such a case, dominance considerations support:

() Jeff should raise his right hand.
No amount of fiddling with value functions yields this verdict for expected value
analyses. Since D∞ has probability , it contributes nothing to the expected value of
raising one’s right hand. I will not try to spell out possible technical solutions for the
expected-value theorist: none are as simple as the solution I offer in §..
But perhaps it is worth considering a conceptual objection suggested by a reviewer
(the following is not a direct quote).
Let us grant that in uncountable spaces some events must have probability . It does not follow
that the possible events that ordinary agents focus on are ever -probability events. After all, the
contextually determined probability function might assign positive probability to D∞ . Insofar
as we get the intuition that Jeff should raise his right hand, it’s because we treat D∞ as having
positive probability.

I concede everything except for the last sentence. Maybe we do often treat events such
as D∞ as if they had positive probability. However, the decision theorists are competent
speakers of English. Other people who are well acquainted with the relevant measure-
theoretic facts are also competent speakers of English. It seems doubtful that their
judgments must be based on surreptitiously assigning positive probability to D∞ . To
put the point another way: if the objection successfully blocked my use of () as a
constraint on deontic semantics, it would be just as forceful in blocking the original
decision-theoretic arguments. But it does not do the latter, so it does not do the former.
Examples such as Darts support a more general objection. Expected-value accounts
would gain plausibility if they could be supported by the claim that they approxi-
mate the correct normative theory of rationally permissible choice. But this is not
a concession that we should grant. The decision-theory literature supports the view
that an expectational analysis of rational decision-making makes correct predictions
in a large class of one-off choice problems, restricted to individual agents and given
probabilities that are (i) determinate and (ii) regular (no contingent proposition gets
probability ). Whether it generalizes beyond that to life long plans (e.g., being a
vegetarian), collective choices or non-regular probability models is a (series of) wide-
open question(s). Darts is one example in which it seems not to, but the general
point is that it is premature to think that a formal model that is intended to have a
very clearly delimited domain of application should be constitutive of the meaning of
deontic modals.

. Section Summary

I have outlined the basic moves in the space of expected-value accounts and their core
advantages. Expected-value analyses come in a variety of logical strengths and none
of them have any trouble handling the phenomena in §. From a critical point of view,
however, I have advanced three objections. To start, they do not combine well with

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

deontic modals and probabilities: one theory to rule them all? 

attitude reports for non-Bayesian agents. I have used subjects who accept maximin as
a decision rule, but the point applies to any subject who does not rank alternatives
according to their expected-value. For similar reasons, expected value analyses do
not interface well with models of disagreement. Finally, they predict that we compare
expected values even in decision problems involving zero-probability states (cases in
which even Bayesian decision theorists are reluctant to go with expectations).

 Probabilistic premise semantics: a first pass

Let us try to collect all our goals in one place. The classical accounts of deontic vocabu-
lary satisfy the most basic degree of neutrality, but they are not information-sensitive
and they are not probabilistic (and so violate Probabilize). Expected-value analyses
provide probabilistic structure, but at conceptual and empirical costs. I propose a
middle ground which, hopefully, has the benefits of both.25
In this section, I develop a system that is capable of elegantly modeling cases like
Rescue and Slanted and more generally the phenomenon of information sensitivity.
But the system also integrates well with an operator treatment of the attitudes and
can easily represent different ways of prioritizing alternatives (as is called for in
Darts). Because of its relative simplicity, the proposal in this section cannot handle
the probabilistic conditionals of §.. This is not a disadvantage compared to the
alternatives (all of which face a similar challenge). But it would be a serious problem if
the theory could not explain cases that were essential in motivating it. To explain the
meanings of those conditionals and the inferential relations between probabilistic and
deontic language, the next section (§) explores a natural extension of the system.
. Formal Preliminaries
Assume that our target English sentences are translated into a sentential multi-modal
language L generated by atomic sentences, Boolean operators, epistemic modals may
and must, a deontic modal should,26 a probability operator likely, and a conditional
(if ·)(·).27 L has the usual formation rules, except that conditional antecedents are

25 Carr (, ) also advances a probabilistic account that is motivated by considerations of
neutrality. For Carr, the semantic theory should have a “decision rule” parameter. I prefer my proposal on a
number of specific counts, but I will not make the comparisons explicit. There also appear to be similarities
between my framework and the explanatory approach of researchers in the program of inquisitive semantics
(Ciardelli et al., ). I defer the exploration of these similarities to separate works.
26 Schroeder () argues that ought is ambiguous, and that, at least in one sense, ought should not
be understood as denoting a function from propositions to truth-values, but rather as denoting a relation
between a subject and an action. The systems I describe here can be easily reformulated to fit Schroeder’s
view of the syntactic and semantic category of ought. For the opposing side of the argument, see Chrisman
(); Finlay and Snedegar ().
27 I adopt this “translational” view of logical form for its familiarity to the philosophical audience. The
alternative, which is prevalent among linguists and I think preferable, is the view that logical form is a level
of syntactic representation of the English sentences themselves, without mediation from an intermediate
formal language.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 fabrizio cariani

momentarily restricted to non-modal and non-probabilistic sentences. If the transla-

tion of an English sentence s does not have a wide-scope modal, we prefix it with a
must when translating conditionals with s as their consequent.28
The function [[·]]· assigns truth-values to sentences of L relative to points of
evaluation constituted by a context, an index and a world of evaluation. To streamline
notation, I omit the context, because all of the parameters we will care about are
bundled in the index.29 In the background of [[·]] is the set of possible worlds W,
which I assume, for simplicity, to be finite (except where noted). Possible worlds are
objects that assign a truth-value (T or F) to every atomic sentence of L. Throughout,
“proposition” abbreviates “set of possible worlds” (i.e., “subset of W”). Substantial
theses about linguistic content are not intended to follow from this modeling choice.
. A Classical Setup
The framework I develop in this section follows the spirit, although not quite the letter,
of Kratzer’s premise semantics for modals. In Kratzer’s framework, the compositional
analysis of modals is given in terms of a pair of parameters, the modal base and the
ordering source whose job it is to determine the modal’s domain. Formally, the modal
base and the ordering source are the same type of object—they are sets of propositions
(modal bases and ordering sources are more properly treated as functions from worlds
of evaluation to sets of propositions. I simplify here because the added complexity is
not necessary to model the core cases of this paper). However, the two parameters
differ significantly in interpretation and in the role they play in the theory. The modal
base delimits a range of salient possibilities (where salience can be characterized in
different ways according to the context). Different choices of modal base and ordering
source create different flavors for the modals. In the modal flavors of interest here,
the ordering source is a set of contextually salient binary preferences among worlds
(following Portner , I call the propositions in the ordering source priorities). Its
job is to determine an ordering of worlds for deontic modalities to quantify over.
. Three Unorthodox Ideas
My framework deviates from the standard Kratzer framework in three major respects.
I introduce them here without giving a specific justification for each deviation: they
are justified as a package by the explanations they allow (§.).
My first deviation is to replace modal bases with fine-grained states.

28 This is an attempt to replicate Kratzer’s idea that conditional antecedents always restrict modals
in their consequents, without commitment to Kratzer’s (a) syntactic hypothesis that there are no
conditional operators at the level of logical form.
29 Two remarks. First, to define certain concepts of logical consequence, it would be useful to also have a
notion of proper point of evaluation (that is a triple whose index and world coordinate are directly assigned
by context), but we will not need this here. Secondly, implementing some deontic semantic theories might
require double indexing in the sense of Kamp (). An example of this is the actualism of Jackson ();
Jackson and Pargetter ().

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

deontic modals and probabilities: one theory to rule them all? 

Definition  (Fine-grained states). A fine-grained state is a pair i, Pr consisting of

a set of worlds i, and a probability function Pr defined on the algebra of subsets A
determined by some subset A of P(W).
Like modal bases, fine-grained states need not be determined by salient information.
In some contexts, i might be determined by relevant facts, or by some other objective
factor (ditto for Pr). However, when the modal is interpreted deliberatively (e.g.,
as expressing the conclusions of practical deliberation),30 i and Pr are plausibly
understood to be respectively the agent’s qualitative epistemic state and that agent’s
credence.
It is reasonable to worry that replacing the modal base with a single probability space
picks up a heavy metasemantic debt. The worry is that it is implausible to suppose that,
in every context, the factors that assign values to contextual parameters determine a
single probability space. For instance, this would clearly be a problem if one believed
that modal bases are determined by the attitudes of conversational participants. In
reply, I note that it is very easy to weaken my model so that fine-grained states are
treated as sets of probability spaces (I do this in §).31
For my second deviation, I propose that we order alternatives, and not individual
worlds. I model alternatives as propositions. When the deontic modal is interpreted
deliberatively, these alternatives represent the courses of action available to an agent
in context.32
Definition  (Alternatives). An alternative set Alt is a set of mutually exclusive
propositions.
The idea of ranking alternatives is inspired by Horty (). Unlike Horty, however,
I do not restrict myself to dominance orderings.33
For it to be possible to order alternatives, alternative sets must be available in the
process of semantic evaluation. To satisfy this requirement, I will introduce in the
indices (see Definition  below) a parameter that records the relevant deliberative
alternatives. Notice that ranking alternatives does not imply that all uses of modals
exhibit alternative-sensitivity—much as adding ordering sources in modal semantics
does not imply that all uses of modals exhibit ordering-source-sensitivity.
My last deviation from the classical framework flows naturally from the second
(though it is not forced by it). In the classical semantic framework, we order individual
worlds: to this end, we use ordering sources consisting of properties that determine
binary preferences among worlds. In my framework, as I have indicated, we order

30 See Schroeder () for some criteria that govern deliberative interpretations.
31 Further weakenings are also possible: in fact, my semantics does not require any probabilistic structure
at all. The design principle is to make it compatible with probabilistic structure, not to require it.
32 See Cariani (a, §) for my preferred way of understanding deliberative alternatives in deontic
semantics.
33 In Cariani et al. (), we have also urged this kind of analysis to model the information sensitivity
of modals. The theory I advance in this paper is a probabilistic generalization of the view in that article.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 fabrizio cariani

alternatives: to keep as much of the structure of the classical semantics, I induce the
ordering by means of properties that determine binary preferences among alternatives.
Since alternatives are modeled as propositions, the simplest extension would be to
model priorities as sets of propositions (I call these elevated priorities).
Definition  (Elevated Priorities and Ordering Sources).
(i) An elevated priority is a set of propositions whose extension is fixed relative to
a probability function Pr.34
(ii) An elevated ordering source o is a set of elevated priorities.
These moves open up the possibility of ranking alternatives on the basis of genuinely
probabilistic considerations. For example, the following is a property that an alterna-
tive A might have (relative to Pr):
(*) {A | given A, it is likely (according to Pr) that Joe will be less hungry}
In many ordinary contexts, the alternative Joe eats a sandwich has this property, while
the alternative Joe runs a marathon does not. Because Pr is a parameter, changes in Pr
may affect the composition of sets such as (*).
By moving to elevated priorities we can distinguish between (*) and:
(**) {A | given A, it is guaranteed that Joe will be less hungry}
(***) {A | A is compatible with Joe’s being less hungry}
As we will see, these distinctions play a crucial role in my account.
. The Semantics
Informed by the discussion in §., we can define our indices.
Definition  (Indices). An index r is a quadruple r = i, Pr, Alt, o consisting of a set
of worlds i, a probability function Pr, a set of alternatives and an elevated ordering
source.
Notation: Given r, ir , Prr , Altr , or denote the respective parameters in r.
The non-modal part of the characterization of [[·]]· is straightforward: relative to r and
w, atomic sentences get T or F according to how things are in w. Boolean operators
are given their standard clauses. It follows that non-modal sentences only depend on
the world w in a point of evaluation r, w. Here are the entries for the modal part of the
vocabulary:

34 In other words, elevated priorities are functions from probability functions to sets of propositions.
I thank Malte Willer and a reviewer for this volume for making me notice that I need this complication.
The reviewer also wonders whether the fact that elevated priorities are such functions makes my view a
notational variant of the expected-value analysis. It does not: my resulting semantics is more permissive
than (i.e., lacks some of the validities of) all the expected-value analyses discussed above, as I verify towards
the end of §..

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

deontic modals and probabilities: one theory to rule them all? 


[[must(A)]]r,w = T iff ∀w ∈ ir , [[A]]r,w = T

[[may(A)]]r,w = T iff ∃w ∈ ir , [[A]]r,w = T
[[likely(A)]]r,w = T iff Prr ([[A]]|ir ) > 

[[should(A)]]r,w = T iff ∀w ∈ Selected(r), [[A]]r,w = T
[[(if A)(B)]]r,w = T iff [[B]]r+A,w = T

First Pass: Lexical Entries

Completing the entry for the conditional. The entry for if appeals to an operation
r + A we have not yet defined. To define it, intersect ir with the set of A-worlds and
conditionalize Pr on the resulting qualitative state (while leaving Alt and o unchanged).
Definition  (Intersective Update). r + A = ir ∩ [[A]], Prr (· | ir ∩ [[A]]), Altr , or .
It might appear at first sight that conditionals with non-modal consequents are settled
by what happens at the world of evaluation w, but recall that we assumed that for such
conditionals we always add in must when translating.
Note also that there is independent evidence that we need another construal for
indicative conditionals (in addition to the one we just characterized). In particular, we
might need a construal on which the antecedent does not have this “conditionaliza-
tion” effect.35 The conditionalization reading, however, is the one that is salient here.

Completing the entry for should. The entry for should appeals to a domain selection
function Selected(·): this function inputs an index and outputs a domain for should (if
we want to change the system to make deontic sentences contingent, Selected might
have to be relativized to a world as well).
Following the Kratzerian playbook, Selected(·) is determined by the ordering source
(but recall that we use elevated ordering sources).
Definition  (Elevated Preorder).
A r B iff {π ∈ or | (A ∩ ir ) ∈ π } ⊇ {π ∈ or | (B ∩ ir ) ∈ π }
Informally, A is at least as good as B (relative to r) just in case A ∩ ir satisfies all of the
elevated priorities (in r) that are satisfied by B ∩ ir . If, in addition, A ∩ ir satisfies some

35 Within Kratzer’s framework, Frank (); Geurts (unpublished); von Fintel () all maintain that
there is an alternative reading for conditionals on which a covert must is always added to the logical form of
the salient conditionals, whether the consequent has an overt modal or not. If we allow each modal to have its
own modal base, this will have a significant truth-conditional difference: (if A)(should B) would have the

following truth condition: for every epistemically possible world w in which A is true, [[should B]]r,w = T.
This is different from the truth-condition in the text, partly because it does not update r to r + A, and hence
does not conditionalize Pr. There is some evidence for this reading, but it does not affect any of the cases I
discuss here. See the discussion of the interplay of conditionals and deontic modals in Cariani et al. (),
§...

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 fabrizio cariani

priorities that are not satisfied by B ∩ ir , we can say that A is ranked strictly above B
in the preorder (relative to r).
Definition  implements another central idea of my probabilistic premise semantics:
the relative ranking of an alternative A in the preorder depends in part on ir . This is
because when we evaluate whether a priority π applies to A, we do not check whether
it applies to the whole set of A-worlds. Rather, we check whether it applies to the
distinguished subset A ∩ ir . To make this intuitive, consider two deontic alternatives,
say flying to France and flying to Japan. The claim is that, to determine which alternative
is better (relative to some contextual priorities, say, making it probable that you see
cherry blossoms), we only consider the worlds in which you fly to France that belong
to ir and the worlds in which you fly to Japan that belong to ir . If it is early spring,
the elevated priority might favor flying to Japan; if it is late fall, it does not favor either
alternative. This idea is a crucial element of the account of Rescue, Knowledge and
Slanted to be given below (§.).
The preorder in Definition  determines the domain for should:
Definition  (Deontic Selection).
Selected(r) = {v ∈ W | ∃B ∈ Altr [∼ ∃A ∈ Altr (A r B & v ∈ B)]}36

The informal gloss on Definition  is that the domain for should consists of the worlds
that belong to the maximally ranked alternatives. This completes my discussion of the
lexical entries.
The last thing we need to check are the predictions of the semantics relative to given
indices is a notion of acceptance at an index, which is defined as follows:

Definition . A is accepted in r iff for all w ∈ ir , [[A]]r,w .
Two quick remarks: although I described must and may as having epistemic inter-
pretations, the semantics can be modified to add entries for their deontic readings. As
in Kratzer’s semantics, there is no need to postulate lexical ambiguities (at least not on
the basis of anything I have said here). The so-called “strong” necessity modal must and
the possibility modal may get their own elevated ordering source o∗ which is a subset
of o (as proposed by von Fintel and Iatridou ) and selection function Selected∗ .
The entries are altered as follows:

[[must(A)]]r,w = T iff ∀w ∈ Selected∗ (r), [[A]]r,w = T

[[may(A)]]r,w = T iff ∃w ∈ Selected∗ (r), [[A]]r,w = T

36 Two remarks on Definition :

(i) the correctness of this definition depends on the limit assumption for , which follows from the
(stipulated) finiteness of W—hence Alt is finite. When we model Darts below, we will have to
consider a case in which W is infinite—in fact uncountable. However, even in that case the set of
alternatives is still finite. Since it is the alternatives that are ranked by , the limit assumption still
applies.
(ii) ≺k means that the inclusion in the definition of k is proper.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

deontic modals and probabilities: one theory to rule them all? 

In epistemic flavors, we could stipulate that o∗ is empty, and hence Selected∗ (r) = ir . In
deontic flavors, o∗ is a designated subset of the elevated priorities in o. This completes
my presentation of a probabilistic premise semantics in the context of a relatively
standard deontic logic.
A second, more radical modification is possible. The First Pass semantics validates
the logical schema known as Inheritance.37
Definition  (Inheritance). If A | B, then should(A) | should(B).
In Cariani (a), I challenged this principle and gave a theory that invalidates
Inheritance while validating Agglomeration:
Definition  (Agglomeration). should(A), should(B) | should(A & B).
It is easy to expand the present framework to implement that non-classical theory.
Since the data I discuss here are orthogonal to that debate, I stick to the Inheritance-
satisfying variant.
. Applying the Semantics
Let us check how the account handles information sensitivity. My first goal is to show
that there are natural assignments of values to the contextual parameters that deliver
the expected verdicts in Rescue, Slanted and the whole family of cases from §. In each
case, I identify indices that plausibly represent the salient conversational contexts and
show that they make the expected predictions. Recall that the target sentences are:
() You should refuse the pills.
() If you have the red gene, you should take the red pill.
() If you have the blue gene, you should take the blue pill.
For Rescue, consider an index r with:
i = the set containing any world that is compatible with the agent’s information
in the context specified in Rescue
Pr = any probability function that respects the constraints stated in the
description of the context of Rescue
In particular: Pr(red gene) = Pr(blue gene) = ..
Alt = {take the blue pill, taking the red pill, refuse the pills}
o ={ {A | given A, it is at least . likely that nine people are saved},
{A | given A, it is at least . likely that ten people are saved}}

37 Although there are debates on the appropriate notion of validity for modal systems such as mine
(Veltman, ; Yalcin, ; Kolodny and MacFarlane, ; Willer, ; Dowell, ), every notion
I know of entails that an argument with premise A and conclusion B is valid if there is no pair r, w such
that [[A]]r,w = T and [[B]]r,w = F. This sufficient condition is enough to derive Inheritance, given the
semantics I spelled out in this section.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 fabrizio cariani

I have chosen the priorities somewhat arbitrarily here, but it is easy to see that there
are many other assignments that would work
Fact . In Rescue, r accepts (), (), and ().
I.e., for any w in ir , [[()]]r ,w = [[()]]r ,w = [[()]]r ,w = T.
In the initial state, taking the red pill fails to satisfy either of the elevated priorities in
r (conditional on taking the red pill you are only . likely to save a person; similarly
for taking the blue pill). By contrast, refusing the pills satisfies the first priority: it
guarantees saving nine people. So [[()]]r = T. When we evaluate the conditional
(), however, the deontic modal appears in an embedding. We must consider the
result of updating r —and in particular ir  and Prr  —with the proposition that you
have the red gene (as in Definition ). Relative to this updated index, refusing the
pill still satisfies the first elevated priority (it guarantees the rescue of nine people)
and nothing else (you cannot save ten if you refuse). However, taking the red pill
guarantees (and hence makes probable) the rescue of ten people, so satisfies both
elevated priorities. For this reason, [[()]]r ,w = T—and by parallel reasoning the
same holds of ().
It is important to appreciate the significance of this reasoning. In my notation, the
technical notion of serious information-dependence in Kolodny and MacFarlane ()
amounts to the following property:
Definition  (Serious Information Dependence). Selected(·) is information-
dependent iff for some A and r, Selected(r) ∩ A  Selected(r + A).
Fact . Selected is seriously information-dependent.
The reasoning in support of Fact  establishes Fact . The witnesses for the existential
quantifier are (i) the proposition that you have the red gene (for A) and r (for r).
To model Slanted, let r be the result of changing Pr in r to:

Pr = some probability function that respects the constraints stated in the
description of the context of Slanted.
In particular, s.t. Pr(red gene) = .

This change suffices to select taking the red pill as the highest-ranked alternative,
without changing anything else in the index. To see this, note that the option of taking
the red pill satisfies both elevated priorities, while refusing can only make it probable
that nine people are saved.
Fact . In Slanted, r accepts () and (), the negation of () and in fact, it also
accepts “You should take the red pill”.
I.e., for any w in ir , [[()]]r ,w = F and [[()]]r  ,w = [[()]]r  ,w = T. Moreover,
[[should (take red)]]r ,w = T

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

deontic modals and probabilities: one theory to rule them all? 

The verification is routine and is omitted here. The broader conclusion is that there is a
natural assignment of values to contextual parameters such that, both unconditionally
and conditionally, you should take the red pill. Moreover, this natural assignment
results from the index in Rescue by modifying only the probabilistic coordinate of
context.
Fact . There is an index r that differs from r only in its elevated ordering source,
and that accepts (), () and ().
Suppose that r is obtained by replacing o in r with:

o ={{A | given A, it is guaranteed that nine people are saved},

{A | given A, it is guaranteed that ten people are saved}}

According to r  , you ought to take neither pill, even in Slanted. Effectively, r involves
priorities that are distinctive of a more risk-averse evaluation of the deliberative
situation. Even though in Slanted there is . probability that you have the red gene,
that is not enough to trigger either priority.
This showcases a precise sense in which my framework is neutral. Different com-
binations of elevated priorities can match the verdicts of a large variety of decision
rules.
As a consequence, on this theory, we can correctly model:
() John thinks you should refuse the pill.
The effect of “John thinks” is to shift the index to one that is compatible with John’s state
(we write h(x, r, w) for the index containing x’s information state in w, x’s probability
function in w, Altr , and x’s priorities in w; we do not shift the alternative set).

[[thinks(x, A)]]r,w = T iff ∀w ∈ ih(x,r,w) , [[A]]h(x,r,w),w = T

Even if the initial context is Slanted, the embedded deontic claim in () gets inter-
preted relative to something like r . In particular, it gets interpreted relative to a state
whose elevated priorities reflect John’s risk aversion.
Finally, the probabilistic premise semantics allows a simple account for Darts. First,
include models in which W is uncountably infinite. Since in Darts we still have only
finitely many alternatives, no other changes to the semantics are needed. Now, consider
an index whose only elevated priority:
() {A | A is compatible with Jeff ’s winning $ million}.
This elevated priority applies, in context, to Jeff ’s raising his right hand, but not to Jeff ’s
raising his left hand, so that we can accurately derive the judgment on ().
This example also illustrates a particularly intuitive feature of my probabilistic
premise semantics: the theory allows us to model priorities that involve assigning low
probability to a certain goal. Consider this dialogue:

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 fabrizio cariani

() A: I would like to sign up for a softball league.

B: The registration for softball leagues is now closed. There is, however, a small
chance that we will be able to get you into a league if a spot opens up. So, you
should sign up for the waitlist.
It is natural to interpret B as conveying that signing up for the waitlist is the agent’s
only way to confer a small probability to her salient goal (playing in a league). We
could model this idea by supposing that the salient priority was:
() {A | the probability of entering a league, given A, is greater than r.}
[with r being some low probability value, e.g. .]
As a result, relative to the given context, signing up for the waitlist, which satisfies (),
ranks higher than, say, throwing a tantrum, which does not satisfy ().

. First Pass Wrap-up

I conclude that the first pass of the probabilistic premise semantics can handle the
cases that motivate information dependence as well as their probabilistic variants. As
anticipated, the theory cannot model the probabilistic conditionals of §.. The most
immediate reason for this is that so far I have allowed only non-modal and non-
probabilistic antecedents.
The deeper reason is that, even if we abandon this syntactic restriction, our cur-
rent theory cannot possibly work well with those sentences. On the current theory,
sentences of the form likely A do not vary with w (they are either T at every w
or F at every w). While this may not be problematic per se, it becomes problematic
when it interacts with the semantics of §.. If conditional antecedents restrict ir , the
restriction effected by a probabilistic conditional antecedent is either trivial (no worlds
are ruled out) or total (every world is ruled out). It follows that conditionals of the form
(if likely A)(should B), such as ()–() (repeated here), are either equivalent to their
consequents or else vacuously true.
() If it is likely that your opponents will attack you on the right flank, you should
concentrate your defense on the right side.
() If it is not likely that your opponents will attack you on the right flank, you
should not concentrate your defense on the right side.
However, this result is undesirable: intuitively, ()–() is a jointly consistent pair, but
neither member is equivalent to its consequent, and neither is vacuously true.

 Probabilistic premise semantics: second pass

We just saw that the First Pass semantics does not derive plausible meanings for
()–(). The problem, I am going to argue, does not lie in the theory’s account of

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

deontic modals and probabilities: one theory to rule them all? 

deontic modals, but rather in the theory’s modeling of probabilistic antecedents. In this
section, I briefly sketch an approach to probabilistic antecedents that, when conjoined
with my account of deontics, can handle ()–()—I do not claim it to be the only
possible approach.
The approach I have chosen for this illustration is Yalcin’s account of conditionals
with probabilistic antecedents, which he formulates as an update semantics (in the
style of Veltman , but with several important and subtle modifications). I reformu-
late Yalcin’s system using the indices from § as states and using p as a metalinguistic
variable ranging over the atomic sentences of L. In this system, we restrict [[·]] to just
the atomic sentences and give a recursive characterization of an update operation [·]
on states.

r[p] = ir ∩ [[p]], Prr (· | ir ∩ [[p]]), Altr , or 

r[∼A] = ir − ir[A] , Prr (· | ir − ir[A] ), Altr , or 
r[A & B] = r[A][B]
r[may(A)] = r, if ir[A] = ∅, else ∅, Prr (· | ∅), Altr , or .
r[likely(A)] = r, if Prr (ir[A] ) > ., else ∅, Prr (· | ∅), Altr , or .
r[(if A)(B)] = r, if r[A][B] = r[A], else ∅, Prr (· | ∅), Altr , or .

Update System with Sharp Probabilities

Note that Prr (· | ∅) is not and cannot be a probability function. To remain consistent
with the treatment of ir , we can suppose that Prr (A | ∅) =  for all A. This means
that we have to qualify the claim that the second coordinate of our indices is always a
probability function: this is only true if understood as restricted to non-degenerate
states. In degenerate states such as ∅, Prr (· | ∅), Altr , or  we allow the function
Prr (· | ∅).
This system is designed for epistemic modalities, but we can add a clause for deontic
should by assuming that, like the other modals, it performs a test on the state r (in the
sense of Veltman ).

r[should(A)] = r, if (ir ∩ Selected(r)) ⊆ ir[A] , else ∅, Prr (· | ∅), Altr , or .

Informally, if one thinks of ir ∩ Selected(r) as a quantificational domain, what is tested

is that this domain is a subset of ir[A] . Note that when A is non-modal and non-
probabilistic, ir[A] = ir ∩ [[A]] (this is a consequence of Fact  below). It follows that
what is tested is whether ir ∩Selected(r) is a subset of [[A]], which is just what a defender
of a quantificational semantics for should would want to test for.
We can complete the account with some standard definitions.
Definition  (Acceptance).
(i) A is accepted in r iff r[A] = r.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 fabrizio cariani

(ii) the sentences in set  are jointly accepted in r iff for all A ∈ , A is accepted
in r.
The technical machinery is best illustrated by exploring some of its consequences.
Fact . When A is non-modal and non-probabilistic,

r[A] = ir ∩ [[A]], Prr (· | ir ∩ [[A]]), Altr , or 

Fact  states that if we restrict attention to the non-modal and non-probabilistic

fragment of our language, the update operation coincides with the update operation
we have ascribed to atomic sentences. It updates the information state ir and condi-
tionalizes the probability function, leaving the rest unchanged. Fact  is proven by
an easy induction on the complexity of the sentences in the non-modal and non-
probabilistic fragment of the language.
Yalcin’s system involves a subtle treatment of negation. The clause for ∼A yields
different behaviors according to the kind of sentence that is negated. If A is from the
Boolean fragment of the language, update on ∼A is covered by the generalization in
Fact . If A is probabilistic or modal, ∼A works as a test (as summarized by Facts 
and ).
Fact . r[∼likely(A)] = r if Pr(A | ir ) ≤ ., else ∅, Prr (· | ∅), Altr , or .
Fact . r[∼may(A)] = r if ir[A] = ∅, else ∅, Prr (· | ∅), Altr , or .
The proofs of these and the remaining facts are trivial and generally omitted in the
interest of space.
The motivation for considering this style of update semantics was the hope for an
account of the joint acceptability of ()–() that did not require them to be vacuous or
to collapse on their consequents. We are not there yet: updating on likely A relative
to index r must yield either r itself or the degenerate point ∅, Prr (· | ∅), Altr , or . In
the first case, (if likely A)(B) is accepted in r if and only if B is accepted in r. In the
second case, (if likely A)(B) is accepted in r for every B (it is vacuously true).
Fortunately, this system is not the last word. Drawing on ideas from Willer (),
Yalcin extends his system for epistemic vocabulary to one involving sets of fine-grained
states. His motivation is to represent uncertainty without buying into the idea that
context supplies a perfectly sharp probability function Pr.
Suppose, for instance, that we want to deploy a probabilistic analogue of the
common ground to determine fine-grained states. A probabilistic common ground
may include all the constraints (qualitative as well as probabilistic) that are mutually
presupposed by conversational participants. According to this picture, it is almost
certain that our mutual presuppositions are not satisfied by a single probability
function and are compatible with a large class of probability functions. You and I may
mutually presuppose that rain in Paris is more likely than snow, and not much else. It
would appear natural, then, to represent the relevant state of uncertainty with a set of

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

deontic modals and probabilities: one theory to rule them all? 

probability functions—those functions that assign greater probability to rain in Paris

than they assign to snow in Paris.
The formal system built to support these ideas can offer the desired account of
()–(). This extension works by first considering sets of states r.
Definition  (Blunt States). A blunt index R is a set of sharp states such that any two
r, r ∈ R differ at most in their i, Pr or o coordinates.
Luckily, we do not need to characterize the update operation again. Building on an
idea by Willer (), we can update each member of R and then collect only those
members that accept A (in the sense of Definition ).
Definition  (Blunt Update). R[A] = {r[A] ∈ R | r[A] = r}.
The following facts reveal some key bits of the behavior of blunt updates.
Fact . R[likely A] = {r ∈ R | Prr ([[A]]) > .}
Fact . R[∼likely A] = {r ∈ R | Prr ([[A]]) ≤ .}
Fact . R[(if A)(B)] = {r ∈ R | r[A][B] = r[A]}
In general, when a sentence performs a test on r, Blunt Update retains exactly those
states that successfully pass the test.
Acceptance and joint acceptance can be redefined for blunt updates by lifting
Definition .
Definition  (Blunt Acceptance).
(i) A is accepted in R iff R[A] = R.
(ii) the sentences in a set  are jointly accepted in R iff for all A ∈ , A is accepted
in R.
(iii) the sentences in  are jointly acceptable iff there is a blunt state R such that the
sentences in set  are jointly accepted in R.
The definition of blunt acceptance has implications for what it takes for a conditional
with a probabilistic antecedent to be accepted:
Fact . R accepts (if likely A)(B) iff there is no r ∈ R with: r[likely A] = r but
r[B] = r.
With this fact in hand, our pair ()–() can be shown to be jointly acceptable without
treating either conditional as vacuous or as collapsing on its consequent. In this setting,
(if likely A)(B) is vacuously accepted iff there is no r ∈ R such that r[likely A] = r;
it collapses on its consequent iff for every r ∈ R, r[likely A] = r.
Fact . There is an R such that ()–() are jointly acceptable in R . Additionally,
R can be chosen so that ()–() are not vacuously accepted and so that they do not
collapse on their consequents.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 fabrizio cariani

For reasons of space, I do not take up the detailed construction of R . However, the
idea should be obvious in light of the earlier facts. Let R contain some sharp states r
according to which the opponents are likely to attack on the right, and some according
to which the opponents are not likely to attack on the right. Then () is accepted in
R provided that all of the sharp states according to which the opponents are likely
to attack on the right pass the test for “you should concentrate the defense on the
right”. Moreover, () is accepted in R provided that all of the sharp states according to
which the opponents are not likely to attack on the right fail the test for “You should
concentrate the defense on the right”. It is easy to construct an elevated ordering source
that accomplishes this.
The upshot of this section is that the challenge posed by ()–() can be factored into
two separate problems. The first problem concerns update on probabilistic informa-
tion. The other concerns the interface between deontic judgments and probabilistic
states. The semantics of § solves the latter problem. This is all that it should be
reasonably asked to do. If we borrow an answer to the first problem, we can inject it
in the probabilistic premise semantics and explain the meanings of these troublesome
conditionals.

 Conclusion
So, is there one theory to rule them all? Probably not. For one thing, the motivational
arguments that justify my proposal depend on significant assumptions about semantic
theories for the language of subjective uncertainty, about information-sensitivity for
deontic modals, about how to integrate deontic semantics with off-the-shelf treat-
ments of attitudes and so on. For another, even granting those assumptions, a few
different systems are compatible with the ideas I have developed.
Nonetheless, there is much to recommend the core ideas of the semantic account
I developed in §. The neutrality-based desiderata are plausible and can be reflected
in specific empirical predictions. My theory does as well as one can with respect to
these desiderata. On the one hand, it is comfortably able to handle the dependence
of deontics on probabilistic states. On the other, it can model judgments that track
orderings based on expected utility, maximin, dominance and what have you, without
assuming that these practical theories are built into the semantics. It strikes a solid
balance between the desire to develop a probabilistic semantics for modals and the
desire to keep the conventional meanings of deontic modals as thin and flexible as
possible.

Acknowledgments
Special thanks to Nate Charlow, Matthew Chrisman, Daniel Lassiter, Paolo Santorio, Ralph
Wedgwood, and Malte Willer for detailed feedback on previous versions of this paper, as

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

deontic modals and probabilities: one theory to rule them all? 

well as to Thony Gillies and Shyam Nair for public commentaries and to Magdalena and
Stefan Kaufmann for our collaboration, which helped me sharpen the ideas in this paper. I am
also indebted to conversations and e-mail exchanges with: Mike Caie, Janice Dowell, Kai von
Fintel, Michael Glanzberg, Jeff Horty, Graham Katz, Angelika Kratzer, Steve Kuhn, Manuel
Križ, Hanti Lin, Peter Ludlow, Dilip Ninan, Paul Portner, Aynat Rubinstein, Alex Silk, Seth
Yalcin, audiences at Georgetown Linguistics Colloquium, Maryland Philosophy Colloquium,
the  Rutgers Semantics Workshop, Kai von Fintel and Sabine Iatridou’s  MIT Graduate
Seminar on Deontic Modals and Imperatives, the  Pacific APA, Northwestern Deontic
Modality Workshop and USC Deontic Modality Conference. Finally, thanks (again!) to both
Nate Charlow and Matthew Chrisman for editing this volume.

References
Bennett, Jonathan () A Philosophical Guide to Conditionals. Oxford: Oxford University
Press.
Briggs, Rachael () ‘Normative Theories of Rational Choice: Expected Utility’. In
E.N. Zalta (ed.) The Stanford Encyclopedia of Philosophy, Fall. [online] Available from:
http://plato.stanford.edu/archives/fall/entries/rationality-normative-utility/. [Accessed
Oct  .]
Cariani, Fabrizio () The Semantics of ‘ought’ and the Unity of Modal Discourse. Ph.D. thesis.
UC Berkeley.
Cariani, Fabrizio (a) ‘Ought and Resolution Semantics’. Noûs. . . pp. –.
Cariani, Fabrizio (b) ‘Epistemic and Deontic Should’. Thought. . . pp. –.
Cariani, Fabrizio () ‘Attitudes, Deontics and Semantic Neutrality’. Pacific Philosophical
Quarterly. . . pp. –.
Cariani () ‘Consequence and Contrast in Deontic Semantics’. Journal of Philosophy.
Cariani, Fabrizio, Kaufmann, Magdalena, and Kaufmann, Stefan () ‘Deliberative Modality
under Epistemic Uncertainty’. Linguistics and Philosophy. . pp. –.
Carr, Jennifer () ‘Deontic Modals without Decision Theory’. Proceedings of Sinn und Bedeu-
tung deta: Chemla, E., Homer, V. and G. Winterstein, (eds.); . École normale supérieure,
Paris, – September. Semantics Archive. . pp. –.
Carr, Jennifer () ‘Subjective Ought’. Ergo. . . pp. –.
Charlow, Nate () ‘What We Know and What To Do’. Synthese. . pp. –.
Charlow, Nate () ‘Decision Theory: Yes! Truth-Conditions: No!’ In Charlow, N. and Chris-
man, M. (eds.). Deontic Modality. Oxford: Oxford University Press.
Chierchia, Gennaro and McConnell-Ginet, Sally () Meaning and Grammar: an Introduction
to Semantics. MA: MIT Press.
Chrisman, Matthew () “‘Ought” and Control’. Australasian Journal of Philosophy. . .
pp. –.
Ciardelli, Ivano, Groenendijk, Jeroem, and Roelofsen, Floris () ‘Inquisitive Semantics: a
New Notion of Meaning’. Language and Linguistics Compass. . . pp. –.
Dowell, Janice () ‘Contextualist Solutions to Three Puzzles about Practical Conditionals’.
In Shafer-Landau, Russ (ed.) Oxford Studies in Metaethics. volume . pp. –. Oxford:
Oxford University Press.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 fabrizio cariani

Easwaran, Kenny () The Foundations of Conditional Probability. Ph.D. thesis. UC Berkeley.
Easwaran, Kenny () ‘Regularity and Hyperreal Credences’. Philosophical Review. .
pp. –.
Edgington, Dorothy () ‘On Conditionals’. Mind. . pp. –.
Finlay, Stephen () ‘Oughts and Ends’. Philosophical Studies. . . pp. –.
Finlay, Stephen () A Confusion of Tongues. Oxford: Oxford University Press.
Finlay, Stephen and Snedegar, Justin () ‘One Ought too Many’. Philosophy and Phenomeno-
logical Research. . . pp. –.
Frank, Annette () Context Dependence in Modal Constructions. Ph.D. thesis. University of
Stuttgart.
Geurts, Bart (unpublished) ‘On an Ambiguity in Quantified Conditionals’. Manuscript.
University of Njemegen.
Gibbard, Allan () Thinking How To Live. Cambridge, MA: Harvard University Press.
Goble, Lou () ‘Utilitarian Deontic Logic’. Philosophical Studies. pp. –.
Hájek, Alan (unpublished) ‘Staying Regular’. Australian National University.
Hintikka, Jaako () Knowledge and Belief: an Introduction to the Logic of the Two Notions.
Ithaca, NY: Cornell University Press.
Holliday, Wesley and Icard, Thomas () ‘Measure semantics and qualitative semantics for
epistemic modals’. Proceedings of SALT . pp. –.
Horty, John F. () Agency and Deontic Logic. Oxford: Oxford University Press.
Jackson, Frank () ‘On the Semantics and Logic of Obligation’. Mind. . . pp. –.
Jackson, Frank () ‘Decision Theoretic Consequentialism and the Nearest Dearest Objection’.
Ethics. . . pp. –.
Jackson, Frank and Pargetter, Robert () ‘Oughts, Options and Actualism’. The Philosophical
Review. . . pp. –.
Joyce, James M. () The Foundations of Causal Decision Theory. Cambridge: Cambridge
University Press.
Kamp, Hans () ‘Formal Properties of “Now”’. Theoria. . pp. –.
Kolodny, Niko and MacFarlane, John () ‘Ifs and oughts’. Journal of Philosophy. . .
pp. –.
Kratzer, Angelika () ‘What “Must” and “Can” Must and Can Mean’. Linguistics and Philoso-
phy. . . pp. –.
Kratzer, Angelika () ‘The Notional Category of Modality’. In Partee, B. and Portner, P. (eds.)
Formal Semantics: the Essential Readings. London: Blackwell.
Kratzer, Angelika (a) ‘Conditionals’. In von Stechow, A. and Wunderlich, D. (eds.), Seman-
tics: An International Handbook of Contemporary Research. Berlin: DeGruyter Mouton.
Kratzer, Angelika (b) ‘Modality’. In von Stechow, A. and Wunderlich, D. (eds.), Semantics:
An International Handbook of Contemporary Research. Berlin: De Gruyter Mouton.
Kratzer, Angelika () Modals and Conditionals. Oxford: Oxford University Press.
Lassiter, Daniel () Measurement and Modality: The Scalar Basis of Modal Semantics. Ph.D.
thesis. NYU.
Lassiter, Daniel (forthcoming) ‘Epistemic Comparisons, Models of Uncertainty, and the Dis-
junction Puzzle’. Journal of Semantics.
Lewis, David K. () ‘Reply to McMichael’. Analysis. . . pp. –.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

deontic modals and probabilities: one theory to rule them all? 

Lewis, David K. () ‘Ordering Semantics and Premise Semantics for Counterfactuals’. Journal
of Philosophical Logic. . pp. –.
MacFarlane, John () Assessment Sensitivity. Oxford: Oxford University Press.
Moss, Sarah () ‘On the Semantics and Pragmatics of Epistemic Vocabulary’. Semantics &
Pragmatics. . . pp. –.
Parfit, Derek () On What Matters. Oxford: Oxford University Press.
Parfit, Derek (unpublished) ‘What We Together Do’. Manuscript. Oxford University.
Plunkett, David and Sundell, Tim () ‘Disagreement and the Semantics of Normative and
Evaluative Terms’. Philosophers Imprint. . pp. –.
Portner, Paul () Modality. Oxford: Oxford University Press.
Regan, Donald () Utilitarianism and Cooperation. Oxford: Oxford University Press.
Rothschild, Daniel () ‘Expressing Credences’. Proceedings of the Aristotelian Society. .
pp. –.
Rubinstein, Aynat () Roots of Modality. Ph.D. thesis. UMass, Amherst.
Schroeder, Mark () ‘Oughts, Agents and Actions’. The Philosophical Review. . . pp. –.
Silk, Alex () ‘Evidence Sensitivity in Deontic Modals’. Journal of Philosophical Logic. .
pp. –.
Skyrms, Brian () Causal Necessity. New Haren, CI: Yale Univeristy.
Skyrms, Brian () ‘Causal Decision Theory’. The Journal of Philosophy. . . pp. –.
Stephenson, Tamina () ‘Judge Dependence, Epistemic Modals, and Predicates of Personal
Taste’. Linguistics and Philosophy. . . pp. –.
Swanson, Eric () ‘How Not to Theorize About the Language of Subjective Uncertainty’. In
Egan, Andy and Weatherson, Brian, (eds.) Epistemic Modality. pp. –. Oxford: Oxford
University Press.
Veltman, Frank () ‘Defaults in Update Semantics’. Journal of Philosophical Logic. .
pp. –.
von Fintel, Kai () ‘Conditionals’. In von Heusinger, Klaus, Maienborn, Claudia and
Portner, Paul (eds.) Semantics: An international handbook of meaning. Berlin: DeGruyter
Mouton.
von Fintel, Kai (unpublished) ‘The best we can (expect to) get? Challenges to the classic
semantics for deontic modals’. Presented at the  Central APA, Chicago, IL. Available from:
http://mit.edu/fintel/fintel--apa-ought.pdf. [Accessed October .]
von Fintel, Kai and Iatridou, Sabine () ‘How to Say Ought in Foreign: the Composition
of Weak Necessity Modals’. In Guéron, Jacqueline and Lecarme, Jacqueline (eds.) Time
and Modality. (Studies in Natural Language and Linguistic Theory ). pp. –. Dordrecht:
Springer.
Wedgwood, Ralph () ‘Subjective and Objective Ought’. In Charlow, N. and Chrisman, M.
(eds.) Deontic Modality. Oxford: Oxford University Press.
Willer, Malte () ‘A Note on Iffy Oughts’. Journal of Philosophy. . pp. –.
Willer, Malte () ‘Dynamics of Epistemic Modality’. The Philosophical Review. . pp. –.
Yalcin, Seth () ‘Epistemic Modals’. Mind. . . pp. –.
Yalcin, Seth () ‘Probability Operators’. Philosophy Compass. pp. –.
Yalcin, Seth () ‘Nonfactualism About Epistemic Modality’. In Egan, A. and Weatherson, B.
(eds.) Epistemic Modality. Oxford: Oxford University Press.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 fabrizio cariani

Yalcin, Seth (a) ‘Bayesian Expressivism’. Proceedings of the Aristotelian Society. .
pp. –.
Yalcin, Seth (b) ‘Context Probabilism’. In Aloni, M. et al. (ed.) Logic, Language, and Meaning:
Proceedings of the th Amsterdam Colloquium. University of Amsterdam, pp. –. Berlin:
Springer.
Yalcin, Seth (c) ‘A Counterexample to Modus Tollens’. Journal of Philosophical Logic. .
pp. –.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i


Decision Theory: Yes! Truth
Conditions: No!
Nate Charlow

 Introduction
This essay attempts to leverage a special theoretical pressure involved in accounting
for the meaning and function of language with loosely-speaking normative or action-
guiding meaning. A core case—and my focus here—is that of deontic modals such
as should and ought.1 Many authors claim that a “standard” semantics for deontic
modals—on which they express varieties of quantification over contextually deter-
mined domains of deontically admissible possibilities—is unviable. In particular, the
standard semantics is said to ignore—in virtue of lacking the structure to represent—
factors relevant to determining the truth of a deontic modal in a given situation. To
yield plausible truth-verdicts for claims expressing what should or may be done in
various kinds of situation, a semantics for such claims should enrich itself with the
fruits of the normative literature on how to determine and reason about what should
or may be done in such situations. By and large, existing accounts suggest the way to
fix this is by adding structure—much implicitly borrowed from normative theory—to
the standard semantics.
The literature has grown thick with proposals in this vein (and with theoretical and
empirical justifications for complications that would otherwise appear unparsimo-
nious). But it is comparatively thin on methodological and philosophical reflection,

1 Linguists used to standard ways of carving up modals in natural language will think “deontic modal” a
poor name for the focus of this chapter. Better would be “prioritizing modal,” a label used to cover modals
with a broadly action-guiding or advisory function (namely: proper deontic, bouletic, and teleological)
(Portner ). Philosophers tend to use “deontic modal” to cover roughly the same ground as the linguist’s
“prioritizing modal.” Although I tend to prefer “practical modal” here (see Charlow a), I follow the
philosophers here. (Most of the chapters of this volume do so too.) Allow me to also note that my primary
focus in this chapter is on so-called weak deontic necessity modals (such as should and ought) whose content
is directly paraphrasable in terms of comparative betterness (on which more below; cf. especially Kratzer
() on “weak necessity”). Strong deontic necessity modals (for example, must) are a more difficult case.
(Like most who have written on the distinction, I am inclined to treat strong deontic necessity as a special
case of weak deontic necessity.)

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 nate charlow

whether about the endeavor itself, where the significant theoretical divisions and
methodological defaults lie,2 or the principles—besides the constraint that requires a
correct semantics to match considered judgments of correctness and consequence—
that might govern this sort of theorizing. Should we augment our theorizing about the
semantics of deontic modals with the fruits of normative theory? If so, how should we
do it? Have past efforts gone wrong in important ways?
This chapter tries to address such questions. It first attempts to build the strongest
possible case for packing the fruits of the normative literature into the semantics
of deontic modals. Unfortunately, the case crumbles on examination. There are
metasemantic reasons to demand that semantics remain neutral on certain normative
matters—namely, all of them. We get closer to the truth about the semantics of deontic
modals by simply excising normative assumptions from our semantic theory. This,
however, seems to mean giving up on the project of matching the verdict of truth
delivered by a semantics to the one delivered by intuition (or so I will argue).
This is a pessimistic conclusion, but this chapter has an optimistic response in mind:
a view that might be clumsily called Nonpropositionalist, Minimalist Expressivism. I will
suggest that deontic modals support a reorientation of semantic theorizing, away from
the aim of delivering an account of a sentence’s truth-condition, toward the aim of
modeling the state of competently accepting a sentence (“Nonpropositionalism”). More
precisely: the correct semantics for deontic modals is one that lays bare the manifold
ways in which a normative view can be combined with linguistic competence with
respect to deontic modals to generate a concrete linguistic judgment of the form
A should do X. The core compositional semantics for deontic modals is exhausted
by the idea that deontic modals quantify over a domain determined by a sequence
of parameters (of basically any length) whose values are fixed by the context of use
(namely, by the specific normative view of a subject) (“Minimalism”). The project
of attempting to understand the “content” or “truth-condition” of a deontic modal
on an occasion of use by investigating its compositional semantics should, in large
part, be abandoned. (This, I argue, appears less of a cost when married to the right
pragmatics—an “Expressivist” pragmatics.)
What is the upshot, if I am right? Corresponding issues clearly arise for language
with broadly normative or action-guiding or advisory meaning—moral language,
imperatives, and so forth. If I am right, much of the work in these areas needs to be
reconsidered, on the grounds that its authors are engaged in normative theorizing,
rather than semantics. How far this kind of critique can be extended to recent,
“cognitively laden” work on, e.g., modals, probability operators, and conditionals is
a question I will not address here, but which will merit future attention.

2 A vignette: the type of amendment to Kratzer’s view suggested in, for example, Charlow (b);
Cariani () is variously portrayed as a “challenge” and a “conservative amendment” to the “classical
semantics” (see respectively von Fintel ; Carr ). This is a theoretically confounding state of affairs
(about which I will have more to say below). (Let me note, however, that Carr’s work is an increasingly
influential exception to the trend I am referencing here.)

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

decision theory: yes! truth conditions: no! 

 The Standard Semantics

What does a sentence like (), which I will refer to as a deontic necessity modal (DNM),
say?
() You should wash your hands.
In one sense, we know pretty well: () says roughly that washing your hands is better
(given the relevant information and body of standards) than not (see especially Kratzer
, ). To see why these claims are approximately equivalent, here is an informal
proof. First, suppose you should wash your hands, and suppose for reductio it is not
the case that washing your hands is better than not. Then not washing your hands is at
least as good as washing your hands (or they are incomparable). In either case, it hardly
seems in any sense obligatory that you wash your hands; so, contrary to supposition,
it is not the case that you should wash them. In the other direction, suppose washing
your hands is better than not and suppose for reductio that it is not the case that you
should wash your hands. Then it would seem permissible for you not to wash your
hands. This cannot be the case if washing your hands is better than not: something is
optional only if not outranked by the alternatives.
While this apparently innocent gloss on the meaning of DNMs hardly seems likely
to tell us anything of interest about their semantics, it in fact leads more or less directly
to a substantive semantics for them. It leads, namely, to the familiar story on which
DNMs express universal quantification (or something close enough) over a domain
of best possibilities. I will present the case for this formally, but, for those inclined to
skip ahead, the basic idea is this: given some relatively innocent assumptions about
the relevant better than relation, p is better than ¬p iff seeing to it that ¬p yields a sub-
optimal outcome, iff all of the best (i.e. good-enough) possibilities are possibilities in
which p. The remainder of this section firms up this gloss while introducing a bit of
apparatus that will be used at various points in the chapter; it can be skipped without
much harm.
To begin, notice that our rough gloss seems to involve just two core notions. First,
better than, indicating ranking according to some kind of preference or standard.
Second, than not, indicating that the ranking applies to some range of relevant
options or alternatives. The oldest (and still standard) way of modeling these notions
in linguistic semantics understands them modally (Kratzer , , ; Lewis
, b). Begin with a domain of relevant alternative possibilities R (than not),3
ordinarily typed as a set of possible worlds,4 commonly called a Modal Base.

3 I will remain silent on what it takes for a possibility to be a relevant alternative when interpreting a
DNM.
4 We will understand possible worlds as functions from atomic sentences to truth-values—i.e., as
something that settles whether grass is green, sheep scream, etc. It is open to us to understand it as a set
of more complicated entities—worlds centered on a time coordinate (or agent or whatever), relations on
worlds (i.e., actions), and so on. For a sense of the different options here, see Cariani (); Cariani et al.
(); Charlow (, §).

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 nate charlow

Definition . A modal base R is a set of relevant possibilities.

Next, we must rank the stuff in R according to some criteria (better). Usually possibili-
ties are ranked relative to a weak preorder  on R, where ‘w  v’ means approximately
that w is as least as good as v. The weak preorder is usually assumed to obey certain
axioms (analogous to the axioms of order governing preference orders in decision
theory): at minimum, it is Reflexive (w  w) and Transitive (w  v and v  u imply
w  u).5 Such a ranking is commonly called an Ordering.
Definition . An ordering  on R is a weak preorder on R that is Reflexive and
Transitive.
From this weak preorder, a strict order ≺ may be defined:
Definition . The strict ordering ≺ is the relation w bears to v iff w  v and v  w.
Note: since ≺ is by definition Asymmetric (w ≺ v implies v ≺ w) and Transitive, it
is Acyclic: for no w , . . . , wn ∈ R such that w ≺ w , . . . , wn− ≺ wn is it the case
that wn ≺ w . It is easily lifted to propositions (sets of possibilities) as follows (see
especially Kratzer ):
Definition . The lifted ordering  is the relation p (p ⊆ W) bears to q (q ⊆ W) iff,
for all v ∈ q there is some w ∈ p such that w  v.

Definition . The lifted strict ordering ≺ is the relation p bears to q iff p  q and
q  p.
The definition most important for our purposes here is Definition . According to
our innocent gloss, DNMs are interpreted relative to a Modal Base and Ordering. What
conditions does a DNM express on a Modal Base and Ordering? In other words, under
what conditions is a DNM p true at R, ? According to our gloss, p says that p is
better than ¬p, relative to R, . What does this mean? It would seem to mean that
there’s no ¬p possibility (∈ R) that cannot be improved on by some p-possibility (also
∈ R); p is better than ¬p relative to R,  just if, if ¬p, it is always possible to better
(i.e., if p). Formally (letting [[φ]] give the set of possibilities φ maps to true):
(Q) ∀x ∈ [[¬p]] ∩ R : ∃y ∈ [[p]] ∩ R : y ≺ x

Informally: seeing to it that ¬p will always leave things sub-optimal. It is easy to see p
is better than ¬p, by the lights of Q, iff p ≺ ¬p, by the lights of Definition . Recall
that p ≺ ¬p iff p  ¬p and ¬p  p. Hence, iff each ¬p world is equaled or bettered by
some p world, but not vice versa (so that at least one p-world betters any ¬p-world).
In other words, it seems that a condition of the form X ≺ Y is a reasonable way of
encoding the content of a claim of the form X is better than Y.

5 Possibly, according to Kratzer, it is partial: for some w and v: w   v and v   w. Partiality introduces
complications that I do not wish to discuss in the main text—and its original motivation is frankly quite
weak. I will implicitly assume that the ordering is non-partial whenever I need it to be total (see, e.g., the
sort of case mentioned in Footnote ).

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

decision theory: yes! truth conditions: no! 

Now suppose we make the Limit Assumption (Lewis ), so that there is always
some non-empty set of -best worlds in R:6

(Limit) ∃x ∈ R : ¬∃y ∈ R : y ≺ x

. . .and we let min (R) designate the set of -best worlds in R.

Definition . w ∈ min (R) iff ¬∃v ∈ R : v ≺ w
On this assumption, a condition of the form p ≺ ¬p is equivalent to the follow-
ing truth-condition, which says that the domain of sufficiently good possibilities is
p-entailing:7

(Q) min (R) ⊆ [[p]]

Here, finally, is an informal proof of the main “result” of this discussion:

p ≺R ¬p ⇔ min (R) ⊆ [[p]]
⇒ Suppose p ≺R ¬p. Suppose for reductio that for some v ∈ min (R) : v ∈ ¬p.
Then there is no world in R that betters v. This contradicts the first supposition.
⇐ Suppose min (R) ⊆ p. Suppose for reductio that p ≺R ¬p. Then either p R ¬p
or ¬p R p. Suppose p R ¬p. Then some ¬p-worlds are not equaled/bettered
by any p-world. Contradiction. So suppose ¬p R p. Then each p-world is
equaled/bettered by some ¬p-world. Contradiction. 

Signpost. The point of that was this: the Standard Semantics—on which deontic
necessity involves universal quantification (and deontic possibility existential quan-
tification) over min (R)—seems to follow from a very minimal understanding of

6 The Limit Assumption is reasonable insofar as the domain of relevant alternative possibilities should
be presumed to be finite (for its negation implies that the domain is infinite). I think the Limit Assumption
is reasonable, at least for the deontic case, and I will make it here. Nothing of ultimate importance will turn
on this, however.
7 I do not mean to suggest that this should be regarded as a revelation. However, standard reasons
for thinking so are arguably mistaken. von Fintel and Iatridou (), for example, write: “Kratzer ()
distinguishes between necessity and weak necessity as well. Her informal characterization is similar to ours:
p is a weak necessity iff p is a better possibility than not p. The technical implementation is different from ours
and crucially involves not accepting that there is always a set of most favored worlds (what is known as the
Limit Assumption in the trade). It appears to us that if one makes the Limit Assumption, Kratzer’s definitions
collapse, leaving no distinction between simple necessity and weak necessity” (p.  n. , emphasis added).
Contra the view expressed in this passage, the Limit Assumption is not sufficient for the relevant definitions
to collapse. Consider a finite model (a fortiori satisfying the Limit Assumption) with R = {x, a, b} in which
p = {x, a} and in which a and b are equally good (a  b and b  a), but x is incomparable to both a
and b. Obviously, since a  b, p  ¬p. But since b   x, ¬p   p. Hence p ≺ ¬p. Obviously, however,
min (R) = R—since no world in R can be strictly improved on—hence min (R)  p. The easiest way
around this is to assume away the troublesome forms of partiality in , and I have implicitly done so in the
main text. Thanks here to Fabrizio Cariani, Alex Silk, and Eric Swanson.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 nate charlow

what deontic modals are in the business of saying (together with some relatively
innocent assumptions about how to structure the formal semantic apparatus). This
fact could, moreover, very reasonably be taken as a kind of roundabout vindication for
the semantic value that, e.g., Kratzer (, ) assigns to necessity modals generally.
On Kratzer’s classic treatment, necessity modals exhibit something like polysemy. The
necessity modal is assigned a context-independent, doubly-parametrized semantic
value, that, in context, maps values for these two parameters—the Modal Base (analo-
gous to our relevant alternative space) and Ordering Source (analogous to our Order-
ing) into a universally quantificational truth-condition. DNMs are distinguished from
E(pistemic)NMs, not by quantificational force, rather by being interpreted in contexts
in which deontic-specific assignments of the relevant pair-values are appropriate. The
fact that an independently motivated Standard Semantics for DNMs can be derived
from the Kratzer semantics for NMs is no small point in favor of the latter.

 Refinement
It would be natural to draw the following moral from the foregoing: accounting for
facts for which the Standard Semantics fails to account will mean refining the basic
apparatus of the Standard Semantics, often times in ways that derive insight and
plausibility from phenomena falling within the domain of practical rationality. Let us
consider some examples.

. Strong and weak DNMs

Notice that, given a fixed R and , the Standard Semantics has difficulty distinguishing
the truth-condition of a “strong” DNM (you must wash your hands) and a “weak”
DNM (you should wash your hands). (The standard story is that these must be
truth-conditionally distinct, since strong DNMs apparently entail corresponding weak
DNMs but not vice versa.) Thus von Fintel and Iatridou () suggest weak DNMs
are sensitive to (at least) two kinds of Ordering—one of which (Primary) ranks
alternatives according to their realization of a “designated goal”, the other of which
(Secondary) further refines the domain generated by the Primary Ordering by ranking
possibilities in it according to how well they meet various subsidiary considerations
(e.g., cost or ease). On this view, the Standard Story is right for what we might call
the basic deontic modality—the one whose interpretation is as sketched above—but
must be refined before it can be extended to what we might call secondary or derived
deontic modalities, such as weak DNMs, which have roughly the following as their
truth-condition (with n the nth -ary Ordering):

(Q) min (min (R)) ⊆ [[p]]

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

decision theory: yes! truth conditions: no! 

This refinement seems to derive plausibility from reflection on the parallel character
of practical rationality—more precisely, on an aspect of practical rationality that the
unrefined semantics seems to ignore. Treating some goal as designated in practical rea-
soning, as evoked in the following sort of instruction, means treating it as something
that (in Dworkinian, Rawlsian, and Nozickian idioms, respectively) trumps, is lexically
prior to, or functions as a side-constraint on the pursuit of secondary considerations.
To illustrate, consider:
() Suppose my goal is to get to Harlem from Brooklyn. What should I do?
It is beside the point to note that you will get some enormous prize if you stay in
Brooklyn; the practical question what should I do? queries which ways of getting to
Harlem are best, while taking no note of rankings amongst the various ways of not
getting to Harlem. Similarly, notice that, no matter how highly an alternative w ∈ R
ranks according to  , if w does comparatively poorly by the metric of  , so that
w∈ / min (R), then necessarily w ∈
/ min (min (R)).
. Epistemic and deontic orderings
Here is another illustration.8 Consider the following impeccable practical inference.
() If you can wash your hands or can wipe them on your pants, you should wash
them. (if X ∨ Y)(A)
If you can wash your hands or can use hand sanitizer, you should wash them.
(if X ∨ Z)(A)
So, if you can wash your hands or can wipe them on your pants or can use
hand sanitizer, you should wash them. (if X ∨ (Y ∨ Z))(A)
() is apparently underwritten by the following truism of practical rationality: if p is
preferable to both q and r, given a choice between p and (q ∨ r), p rationally must
remain preferable. Relatedly, and in fact more generally, if p is preferable given all
the relevant alternatives individually, p rationally must remain preferable given all
the relevant alternatives collectively, as illustrated by, e.g., a case in which the relevant
alternatives are given by the circumstances of the decision problem:
() If your fellow prisoner defects, you should defect. (if X)(A)
If your fellow prisoner cooperates, you should defect. (if Y)(A)
So, whatever your fellow prisoner does, you should defect. (if X ∨ Y)(A)

8 Inspired by discussion in Lassiter () of Kratzer’s treatment of the better possibility than relation. A
difference worth noting is that Lassiter is not worried about inferences involving modals and conditionals in
the object language. He is rather worried about the fact that, regardless of whether better has a preferential
or probabilistic meaning, Kratzer’s treatment of the metalinguistic better possibility than relation validates
the inference from (i) A is a better possibility than B, (ii) A is a better possibility than C, to (iii) A is a better
possibility than B or C.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 nate charlow

The reasoning of which this example is apparently an instance is apparently related to

what is generally known as the Statewise Dominance constraint on choice functions
in the literature on rational choice (which is related to the Sure Thing Principle of
Savage ).9
Now consider a minimal variation of (), which is rather less than impeccable.
Note: should in () is meant to express a form of epistemic necessity concerning one’s
expectations about Al.
() If Al can wash his hands or can wipe them on his pants, he should wash them.
(if X ∨ Y)(ep A)
If Al can wash his hands or can use hand sanitizer, he should wash them.
(if X ∨ Z)(ep A)
#So, if Al can wash his hands or can wipe them on his pants or can use hand
sanitizer, he should wash them. (if X ∨ (Y ∨ Z))(ep A)
In (), a formally similar inference fails, owing to something like a truism of epistemic
rationality: even if p is considerably likelier than both q and r—sufficiently so to license
the claim that if either p or q may be true, it should be p, and the claim that if either
p or r may be true, it should be p—it doesn’t follow that p is considerably likelier (to
the same degree) than (q ∨ r).
Suppose with Kratzer (, ) (and most proponents of the Standard Semantics)
that epistemic and deontic should express contents that are resolutions-in-context of a
single parametrized semantic value—i.e., resolutions obtained by fixing, relative to a
context of interpretation, values for parameters in the semantic value for should. This
pair would seem strongly to suggest that the status of such inferences should follow
not merely from facts about the parametrized semantic value of should. Instead, they
should follow from these facts together with, on the one hand, stipulations about the
character of the domain of quantification for modals expressing deontic necessity,
as well as, on the other hand, stipulations about the character of the domain of
quantification for modals expressing epistemic necessity.10 Such stipulations, whatever
shape they take, will apparently be underwritten by the above truisms of practical and
epistemic rationality.

Signpost. The natural spin on the cases we considered in this section is that they
motivate certain emendations to the Standard Semantics. More specifically, they seem

9 The relationship between Statewise Dominance and the Sure Thing Principle is in fact complicated
(as is the matter of their relationships to the examples mentioned here). Since none of the points I make will
turn on the fact that it is the Sure Thing Principle, as opposed to Statewise Dominance (or indeed another
dominance principle entirely), that is encoded within a semantics for DNMs, I will tend to speak loosely
about these relationships in the main text.
10 Lassiter () draws the lesson that epistemic and deontic should probably do not share a common
semantic core (and that these modals do not ultimately have a quantificational semantics). While I am
sympathetic to the second part of his conclusion, it is several steps removed from the bare data we are
considering here.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

decision theory: yes! truth conditions: no! 

to highlight that the Standard Semantics is too coarse-grained to represent phenomena

of normative significance which seem to bear clearly on the truth-conditions of DNMs.
The way out is to refine the Standard Semantics to allow it to represent these sorts of
normative phenomena, preliminarily along the lines suggested in this section. A large
literature is devoted to pursuing a research program broadly in this vein.

 Uncertainty
A key opportunity for refinement explored in the literature on DNMs concerns the
semantic role of uncertainty. Specifically, the Standard Semantics has a hard time
accommodating the manifest importance of uncertainty in determining what we
should do. While some—Dowell (forthcoming); von Fintel (), not to mention
Kratzer ()—claim the Standard Semantics can accommodate uncertainty as-is,
it is commonly accepted that some sort of refinement is needed (as I’ll explain).
A number of people have attempted to provide one (see Carr ; Cariani et al. ;
Charlow b; Silk ). Here I will review the major issues and build a case for, in
Kratzer’s phrase, “packing in” some of the fruits of decision theory into the semantics
of DNMs.
. MaxiMax
Here’s a simple schematic case. Possibilities where g is realized are preferred to
possibilities in which g is realized, which are in turn preferred to possibilities in which
g is realized:
g ≺ g ≺ g

Your available actions: A and B. The relevant contingency—the contingency that

determines the respective outcomes of A-ing and B-ing—is whether p. B-ing brings
a sure outcome in which g is realized, while A-ing brings a good outcome (g ) if p,
but a relatively bad outcome (g ) if ¬p.

p ¬p
A g g
B g g

In short, you prefer A-ing if p to B-ing, which you in turn prefer to A-ing if ¬p. Now:
what should you do in this sort of decision problem? We cannot, it seems, answer this
question without further information regarding, at least, the likelihood of p (as well
as information about how preferable g is to g and how preferable g is to g ). Absent
such information, what you should do is indeterminate.
The Standard Semantics—on a perhaps naïve reading (see §.)—seems, wrongly,
to say that the answer is determinate. Possibilities in which g is realized are strictly
preferred to possibilities where it is not. The best possibilities (according to the relevant

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 nate charlow

ordering ≺) are thus all possibilities where g is realized. But all the possibilities
where g is realized are possibilities where you A. And so all the best possibilities
are possibilities where you A: A ≺ B. As Carr puts it, the Standard Semantics seems
to encode the decision rule MaxiMax. “This is a straightforward consequence of the
more basic commitment...that we should always simply bring about the best possible
outcome in the modal background” (Carr , p. ). If we had to pick any decision
rule for our semantics for DNMs to encode, MaxiMax would be far down the list.
. Decision-theoretic quietism
I bring up this sort of case to introduce a possible response on behalf of the Standard
Semantics before approaching a more well-known case. The response is familiar from
the decision-theoretic dialectic surrounding, e.g., the Ellsberg Paradox: the decision
table—more precisely, its preferential dimension—is said to be under-described. In the
Ellsberg Paradox, the charge is that the payoffs explicitly represented in the decision
problem ignore the role of risk-aversion in informing the preferences of apparently
rational agents. Analogously, in this case, the charge might be that the relevant payoffs
should be enriched. What is (dis-)valuable outstrips what can be represented with
payoffs g –g . Also valuable is choosing an action that maximizes expected value (or a
suitable stand-in for expected value). More generally, agents manifestly value choices
that meet procedural conditions on how such choices are made (call such preferences
rational preferences). Rational preferences manifestly bear on the desirability of actions
that implement them (or fail to).11
Let’s take this a bit further. In the schematic case in §., it is natural to ascribe first-
order preferences—preferences that supervene on your preferences among outcomes—
of broad shape:
(A ∧ p) ≺FO B ≺FO (A ∧ ¬p)
Nevertheless, it is still clearly possible (for, e.g., an expected-utility maximizer) to have
considered preferences among actions of shape:
B ≺ATC A
Given Definition  (and, indeed, any reasonable understanding of the properties of
≺) necessarily ≺FO = ≺ATC . So long, then, as we allow ≺ATC , rather than ≺FO , to
determine the domain of quantification for the DNM, we can avoid the problem posed
above.
This reply is decision-theoretically quietist, in the following sense. Contra what
was suggested in §., MaxiMax is in fact rejected as a “theorem” of the semantics.12
However, no replacement decision theory is encoded in its stead. Indeed, it would seem

11 Such an account is presented approvingly in an unpublished reply by Kratzer to Cariani et al. (),
and echoed in von Fintel (). For criticisms similar to the one developed here, see Carr (); Silk ().
12 Compare Cariani (): “A semantic theory might select the best worlds out of an ordering, but [this]
does not mean that it implements MaxiMax.”

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

decision theory: yes! truth conditions: no! 

that effectively any recipe for transforming ≺FO into ≺ATC must be invisible to the
compositional semantics of DNMs.
Kratzer’s unpublished argument for this sort of invisibility is a question whose
answer is intended as obvious: “Why pack information about rational decision making
into the meaning of modals?” A perhaps more perspicacious version of this question
would be as follows: why pack any more information about rational decision making
into the meaning of modals than we have to in order to make our semantics mini-
mally predictive? After all, Kratzer’s stipulations about the ordering  encode certain
“axioms” of rational decision-making, including:
• Kratzer Dominance (p ≺ q ∧ p ≺ r) ⇒ p ≺ (q ∨ r)
• Acyclicity of ≺ p  ≺ . . . ≺ p n ⇒ p n  ≺ p
• Independence of Irrelevant Alternatives (see Charlow b):
(w ∈ min (R) ∧ w ∈ R ∧ R ⊆ R) ⇒ w ∈ min (R )
Any and all object-language expressions of these features of the preference ordering
are thus treated as theorems within Kratzer’s semantics. These, however, are intuitively
“good” things to treat as theorems within the relevant formal system (whereas it would
be intuitively bad to render, e.g., a particular decision rule such as Expected Utility
Maximization or MaxiMin a theorem).
Why pack any more into the compositional semantics of DNMs than we have to
in order to generate the sorts of theorems we regard as desirable? I will refer to the
proponents of the “obvious” answer to this rhetorical question as Decision-Theoretic
Quietists. In the next section I will apply Quietism to a well-known puzzle case in the
semantics of DNMs.

. Applying decision-theoretic quietism: the miners

Kolodny and MacFarlane () famously describe the following case. Ten miners are
trapped in a shaft—A or B, we don’t know which—and threatened by rising waters. We
have time to block exactly one shaft or neither. If we block the correct shaft, everyone
lives. If we block the wrong shaft, everyone dies. If we do nothing, exactly one miner
dies. Thus:

in_A in_B
block_A All live All die
block_B All die All live
¬(block_A ∨ block_B) Nine live Nine live

This is actually an elaboration of the sort of case described in the previous section.
Your first-order preferences among outcomes look like this:

(in_x ∧ block_x) ≺FO ¬(block_A ∨ block_B) ≺FO (in_x ∧ block_y)

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 nate charlow

Now, what should we do? Absent information about the probability, e.g., that the
miners are in A, what we should do is indeterminate (or, perhaps, it is determinate
that we should block neither). But evidently the best possibilities are all possibilities
where all miners live. So the best possibilities are all possibilities where we block A or
block B. Apparently, then, according to the Standard Account, the following is true:
() We should either block A or block B. (block_A ∨ block_B)
Not only does this sound false (or at least not determinately true); it seems that, on
certain assignments of probabilities to the claim that the miners are in A (e.g., %),
the following is instead true:
() We should do nothing. ¬(block_A ∨ block_B)
As we discussed above, a fan of the Standard Account might reply: there are things,
other than the lives of the miners, but which we manifestly do care about (e.g., dealing
with uncertainty, managing risk), that the decision table fails to represent. In this vein,
von Fintel (), addressing cases specifically in which we judge () true, writes: “It
is not obvious that [a world where all ten miners are saved] is a better world than one
where nine miners survive because we choose not to run the risk.” Thus our considered
preferences are, in fact, of the following shape:

¬(block_A ∨ block_B) ≺ATC (block_A ∨ block_B)

This, the claim goes, underwrites our judgment of the truth of () and the falsity of ().
Impressively, this sort of thought can be extended to explain why we tend
(“instantly,” in the phrase of von Fintel ()) to judge the following conditionals
true:
() If the miners are in A, we should block A. (if in_A)(block_A)
() If the miners are in B, we should block B. (if in_B)(block_B)
All we need do is combine it with the further thought—for which there is an ever-
expanding body of evidence (see, e.g., Yalcin ; Gillies ; Charlow a, b)—
that the if-clause of an indicative conditional functions, in part, to shift the information
relevant for evaluating information-sensitive expressions (such as DNMs). This allows
us to exploit this claim of Kolodny and MacFarlane ():
A world may be more ideal than another relative to one information state and less ideal than it
relative to another. For example, a world in which both shafts are left open may be more ideal
than one in which shaft A is closed relative to a less informed state, but less ideal relative to a
more informed state. (p. )

Let ≺ATC[φ] give our considered preferences given a hypothetical update with infor-
mation φ—i.e., the sort of update induced by the if-clause of an indicative conditional

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

decision theory: yes! truth conditions: no! 

of the form (if φ)(ψ). If we take on Kolodny and MacFarlane’s quoted claim, we have
a rationale for endorsing the following:
block_x ≺ATC[in_x] ¬(block_A ∨ block_B) ≺ATC[in_x] block_y
Bracketing details, it isn’t hard to use this sort of claim about the structure of the sup-
positional or hypothetical preference order to underwrite a semantics for conditional
DNMs on which () and () come out true, even while () comes out false (relevant
formal tools can be found in Charlow b; Cariani et al. ; Silk ).13
. Decision theory and truth-conditions
Here we have a Quietist account of the Miners Puzzle that seems, at a first pass,
adequate. Kratzer’s question presents itself with even greater force: “Why pack infor-
mation about rational decision making into the meaning of modals?” Here is a
response (also pursued in Charlow b).
A preliminary answer: in fact, it seems that a key motive for introducing orderings
into the semantics for DNMs is to recapitulate, if partially, the structure of practical
reasoning. Practical reasoning takes place against preferences over outcomes as well
as information about outcome-relevant circumstances. This explains, in part, why the
semantics of DNMs must (and does) distinguish between something analogous to
a preference-state and something analogous to an information-state. Note, further,
the fundamental theoretical purpose of a normative decision theory to bridge the
gap between, on the one hand, descriptions of the agent’s preferences over outcomes
(and the agent’s information about outcome-relevant circumstances) and, on the other
hand, an account of what that agent should decide, intend, do. (Important note
to the reader: by “decision theory” I will mean nothing more than something that
determines a choice function. A choice function is an object that, given a specification
of preferences over outcomes and information about outcome-relevant circumstances
as input, generates a set of permissible actions as its output.)
If the recapitulation claim is right, there is reason for the semantics of DNMs—
claims which serve to express verdicts about what agents should do—to represent
the normative role of a decision theory. Note that, in a non-semantic context, it
would be something like a category error to build the decision theorist’s conception
of the preferred way to bridge the gap between preference/information and action
into the description of what outcomes are preferable for the agent. Further, it is clear
that verdicts of what an agent should do, in light of a description of what outcomes

13 Kolodny and MacFarlane () argue that this means giving up modus ponens, since (in_A ∨ in_B)
together with () and () apparently entail (). But this view—note I include my concurrence in Charlow
(b)—is mistaken. As Willer () notes, if our notion of entailment is dynamic in nature—roughly,
φ , . . . , φn |Dyn ψ iff the result of updating an information-state on φ , . . . , φn yields an information-
state that accepts ψ—this does not follow. Given one very natural treatment of the conditional, on which a
state σ accepts (if φ)(ψ) iff the result of updating σ with φ accepts ψ—modus ponens is dynamically valid—
(if φ)(ψ), φ |Dyn ψ. But constructive dilemma is clearly not—(if φ)(χ ), (if ψ)(ρ), (φ ∨ψ)  |Dyn (χ ∨ρ).
For some discussion, see Yalcin (b, n. ); Charlow (a, §.).

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 nate charlow

are preferable for the agent (and the information to which she has access), must be
generated by appeal to some decision theory or other. What goes for decision theory
goes, then, for the theory of what the agent should do: take care to distinguish between
first-order and rational preferences. There is, I would think, strong pressure to say that
what goes for the theory of what the agent should do goes, in turn, for a semantic
theory for sentences that conventionally serve to express views about the verdicts of
the theory of what the agent should do. Regardless, then, of whether we are doing
decision theory or theorizing about the meaning of modals that express the considered
verdicts of a decision theory, it seems we should take care to avoid building rational
preferences into our description of first-order preferences. But it seems like this is
exactly the conflation Kratzer is encouraging us to make.
To put it differently, consider a decision problem  for an agent with preferences
P , information I (which also specifies a range of available actions for the agent,
A = X , ..., Xn ); we will represent the decision problem as an ordered triple of the
form  = P , I , A . Let D be the selected decision theory,14 and let CHD be the
choice function determined by D from decision problems into sets of D-admissible
actions in a decision problem. Now consider:
The Master Argument for “Packing In” Decision Theory
i. X (given ) iff X ∈ CH D (P , I , A )
ii. “X” is true at  iff X (given )
iii. X (given ) iff ¬¬X (given )
iv. So “X” is true at  iff ¬X ∈/ CHD (P , I , A )
Because which actions are permissible within a decision problem is in part a function
of a choice of decision theory (i), by an application of the T-Schema15 (ii) and the
metalinguistic duality of deontic necessity and possibility (iii), the truth-condition of
a sentence describing what an agent should do in a decision problem is also such a
function (iv).16

14 D can be any decision theory whatever—any recipe from transforming preferences over outcomes
and information into a choice function, including recipes that, e.g., treat information as irrelevant. D need
not be an expected utility calculus, or indeed something borrowed from the decision-theoretic literature
at all.
15 Instances of (ii) are not, strictly, instances of the T-Schema, which, as Fabrizio Cariani has pointed
out to me, concerns the application of a monadic, not relational, truth-predicate. Still: schema (ii) is correct
and clearly related to the T-Schema (and seems to me to be endorsed by Tarski himself when he discusses
monadic truth as a special case of relational satisfaction). Consider an index-sensitive truth-condition for
the sentence grass is green. This sentence is true at w iff [λv.grass is green at v](w) =  iff grass is green at w.
Generalizing, φ is true at some sequence of indices S iff [λS .φ(S )](S) =  iff φ(S).
16 I concede, of course, that work needs to be done to fill out this basic picture. Here I will mention two
outstanding issues. First, if it turns out that which things are chairs depends on the distribution of a certain
kind of fundamental physical particle P, this argument could thereby be used to show the need to represent P
in the semantics of a sentence such as that’s a chair. That would be a reductio of this argument. (I regret that I
cannot recall the name of the person who raised this worry to me.) Avoiding this sort of overgeneration is not

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

decision theory: yes! truth conditions: no! 

Two related lessons can be drawn from this. First, if a kind of parameter—more
precisely, a kind of distinction between parameters—is constitutively relevant to the
determination of what the agent should do, it is thereby semantically relevant to the
truth-condition of a sentence describing what she should do. Second, a semantics that
refuses to distinguish between first-order and rational preferences cannot represent
the distinct contribution of first-order and rational preferences to the truth-condition
of a DNM—a state of affairs that is at once both theoretically embarrassing and liable
to lead to empirical difficulties. (I will outline an empirical difficulty below.)
This is the case for “pack[ing] information about rational decision making” into the
meanings of modals—at least into the meanings of modals that express verdicts about
rational decision making. Quoting myself, cases like the Miners (and the schematic
case on which it is based) . . .
. . . recommend the introduction of some sort of independent theoretical apparatus taking some
specification of (i) substantive, first-order ends and (ii) an information-state as its input, and
generating a modified ordering on possibilities as output. We could, of course, conflate what
is “best” or “ideal” with what is minimal with respect to the modified ordering [i.e. ≺ATC ].
Semantically, this would amount to restyling the deontic selection [min] as selecting, not worlds
that are best sans phrase [i.e. relative to ≺FO ], but with respect to some mixture of [first-order
and rational preferences] . . . [But we] should take care to avoid running together theoretical
notions (and pieces of theoretical apparatus) that, like actual and expected utility, a theory ought
not to run together. (Charlow b, pp. ff.)

To be clear, I am not arguing against allowing ≺ATC , rather than ≺FO , to determine
the DNM’s domain of quantification. I am arguing against a decision-theoretically
quietest semantics—one that refuses to represent anything more than ≺ATC in the
truth-condition for the DNM. Even if we allow ≺ATC to determine the domain of
quantification for the DNM, we must also represent the distinct contributions of (at
least) first-order preferences and rational preferences in determining ≺ATC . A story
that fails to represent the distinctive roles of first-order preferences and preferences
concerning ways of reasoning about what to do in the truth-conditions of claims
invoking DNMs cannot be the whole story about their meanings.

so hard, if we understand the premises of the argument as involving special equivalencies—e.g., conceptual
or analytic—that do not hold between chairs (and claims about chairs) and claims about Ps.
Second, CHD (P , I , A ) must be closed under conditions that are appropriate given the logic of DMs.
For instance, if X entails Y, then for any CHD and  such that X ∈ CHD (P , I , A ), it must be that
Y ∈ CHD (P , I , A ). Giving a precise description of the operations under which CHD (P , I , A ) is
closed requires settling quite a lot of disputed issues in the logic of DMs—e.g., whether permission itself is
closed under arbitrary disjunction (i.e., the problem of Free Choice Permission)—and I will avoid it here.
I will note that imposing closure conditions on CHD (P , I , A ) will make it less natural to say that the
elements of CHD (P , I , A ) are actions—rather, they will be constructs out of the range of permissible
actions delivered directly by the relevant decision theory, often, but not always, corresponding themselves
to actions. In no way does this vitiate the argument of this section: a DM expresses a judgment about what
should be done, which in turn depends on which actions are permissible in light of the relevant decision
theory.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 nate charlow

. Rational fetishists

To see this better, I will describe an empirical difficulty. Note first that von Fintel’s claim
that it’s “not obvious that [a world where all ten miners are saved] is a better world than
one where nine miners survive because we choose not to run the risk” (and Kolodny
and MacFarlane’s parallel claim that “a world in which both shafts are left open may be
more ideal than one in which shaft A is closed relative to a less informed state”) must
be wrong. In the most immediate senses of better and ideal, the best, or most ideal,
outcome is the outcome in which all ten miners live; the central difficulty of the case
is that we don’t know how to bring the best outcome about.
Not, I think, that anyone would deny this. But it is easy to imagine someone—call
him the Rational Fetishist—who did. The Fetishist values ordinary things (for instance
human life), as well as the demands of practical rationality, in exactly the same way.
For the Fetishist, ordinary things do not function as constraints on the pursuit of the
demands of practical rationality; these kinds of goals are comparable to (and, so, on
occasion, compete with) one another. (For the ordinary rational agent, the demands
of practical rationality do not compete with non-decision-theoretic goals; the former
rather structures the agent’s rational pursuit of the latter.) For the Fetishist, the value
of not risking the deaths of all ten miners in fact renders the outcome in which we
block neither shaft actually better, literally more desirable, than the outcome in which
we block the correct shaft.
Note that, so long as the same decision theory (hence the same decision-theoretic
verdicts) applies to the ordinary rational agent and the Fetishist, the considered
preferences of the Fetishist coincide—indeed, if it helps, we could imagine that they
necessarily coincide—with the considered preferences of the ordinary agent. But I
submit the following as a theoretical datum (although I will also argue for it): the
truth-condition of a DNM “in the mouth of ” the Fetishist is distinct from the truth-
condition of a DNM “in the mouth of ” the ordinary agent.17 A theory which refuses to
assign distinctive roles to first-order preferences and rational preferences in the truth-
conditions of claims invoking DNMs will find this hard to explain (since in such a
theory only ≺ATC is visible, and the properties of this ordering coincide for the kinds
of agents under consideration).
In support of this datum, note that, for the Rational Fetishist, there is, by stipula-
tion, a single preference-structure in play, whose implementation in, e.g., the Miners
scenario, requires blocking neither shaft. It is a necessary condition on implementing
the relevant preference-structure that neither shaft be blocked. According to at least

17 To be clear: I do not really mean to assume that the preference-structure relevant for interpreting a
DNM is supplied by the agent who utters the DNM (hence the scare quotes around “in the mouth of ”).
Here I intend something much weaker: that when we (qua semantic theorists) evaluate DNMs relative to
orderings that are supplied by Fetishistic and ordinary preference-structures, we should be able to identify
a semantic distinction between them.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

decision theory: yes! truth conditions: no! 

one influential story about the semantics of strong DNMs (e.g., must) (see von Fintel
and Iatridou ),18 a claim of roughly the form in (). . .
() To implement the basic preferences, we must block neither shaft.
. . .is therefore true relative to a Rationally Fetishistic preference-structure. For the
ordinary agent, however, there are two preference-structures in play: first-order and
decision-theoretic (the former valuing things like miners’ lives, the latter perhaps
valuing things like risk-sensitivity in pursuit of the former). It is not a necessary
condition on implementing the “primary” preference-structure that neither shaft be
blocked (in fact, as we’ve seen, it is a necessary condition on implementing the
“primary” preference-structure that at least one shaft be blocked). Blocking neither
shaft is, however, the best way of implementing the primary preference-structure,
given risk-averse rational preferences. So, applying the von Fintel and Iatridou ()
account of the semantics of weak DNMs (e.g. should), a claim of roughly the form in
() is therefore true when evaluated relative to an ordinary preference-structure.
() To implement the basic preferences, we should block neither shaft.
However, a claim of the form in () is judged (appropriately, in my view) false.19
There is, then, a fairly clear semantic difference between DNMs “in the mouths
of ” Rational Fetishists and those “in the months of ” ordinary agents. To explain
this, it seems we should augment the Standard Semantics with some of the fruits
of normative decision theory. For instance, in Charlow (b), to account for our
ordinary judgments about the Miners scenario, I suggested making use of a truth-
condition inspired by the truth-condition in Q, repeated here:

(Q) min (min (R)) ⊆ [[p]]

More precisely, I suggested semantically representing pairs of orderings—and, poten-

tially, sequences of orderings of arbitrary length—and suggested a recipe for construct-
ing a mixed ordering—corresponding here to the considered preference ordering—
from the sequence together with some specification of the relevant information. (In
this treatment, the mixed ordering was information-sensitive in roughly the way
described earlier in this section.) My aim was to state a truth-condition for the DNM
that tracked the verdict of a specific kind of decision theory—the knowledge-based
decision-theory of Hawthorne and Stanley ()—in the case. But other views,

18 Note: none of what I say here requires taking on the treatment of the strong/weak distinction in
von Fintel and Iatridou (), to which I will object below. My view is that on any adequate treatment
of the strong/weak distinction, () should come out true relative to a Rationally Fetishistic preference-
structure, false relative to an ordinary preference-structure. I will not attempt to defend any view of the
strong/weak distinction here.
19 For more on the importance of the strong/weak distinction in the Miners scenario, see Charlow
(b); Silk ().

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 nate charlow

making use of different decision theories, could easily be plugged into this basic
apparatus.

Signpost. I have argued that a semantics for DNMs should have, in the first instance,
the following sort of shape: X is true (given  according to D) just if ¬X is
impermissible in  by lights of D:
/ CHD (P , I , A )
X is true (given ) iff ¬X ∈
The truth-condition of a DNM will reference a choice function CHD , supplied by
the relevant decision theory D, and a decision problem  = P , I , A , given
presumably at a context.
Before moving on, I wish to emphasize, briefly, that this in no sense involves a
departure from the suggested lesson of §. Rather, for all we have seen, it reinforces it.
The Standard Semantics is too coarse-grained to represent phenomena of normative
significance which seem to bear clearly on the truth-conditions of DNMs. The way
to fix this is to refine it—in particular, by augmenting it with the fruits of decision
theory.20
Nor does it necessarily involve any departure from the idea that DNMs have quan-
tificational truth-conditions (cf. §). We are free to pursue an equivalence between a
quantificational truth-condition and a decision-theoretic truth-condition of the form
given here (analogous to the equivalence demonstrated in §).21 (While I am doubtful
we can retain anything resembling Kratzer’s treatment of comparative betterness, as
encoded in Definitions  and , the assumptions that endow these definitions with
predictive power are—as I explain in §.; see also Lassiter ()—independently
problematic. It is, in my view, a virtue of the account I am suggesting here that we
can move beyond these definitions. The account I advocate achieves predictiveness
without taking on problematic stipulations about the properties of the relevant order-
ing.)
Here we have what I take to be an intuitive, well-supported, and altogether fairly
conservative picture of how theorizing about the semantics of DNMs should proceed.
There are, however, deep difficulties in implementing it. I describe these in the next
section.

20 As I emphasized in Charlow (b), the representation of decision theories in the semantics of weak
DNMs is motivated by the same sorts of considerations that motivate the von Fintel and Iatridou analysis
of weak DNMs (§.). For this reason it is puzzling that von Fintel () portrays the kind of analysis
advocated in this section as a “challenge” to the classical semantics for DNMs (since he does not portray
his own understanding of weak DNMs similarly). The view I have defended does rely on a challenge to the
classical (Kratzer) semantics for indicative conditionals, but these are very different things.
21 How would that work? A choice function CH , when applied to a decision problem , selects a
D 
set of alternatives that are permissible according to D. The “disjunction” of these alternatives CHD ()
represents something like the “sphere of permissibility” (compare Lewis b) for the agent in , which
is formally akin to an ordinary domain of quantification, and which can therefore be linked to claims of
comparative betterness using the sort of strategy pursued in §.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

decision theory: yes! truth conditions: no! 

 Packing It In
The difficulties begin with the observation that this schematic sort of truth-condition
cannot be, strictly speaking, correct. It ignores a range of data suggesting that DNMs
have context- and index-sensitive truth-conditions. Here I will connect this rather
prosaic fact about DNMs to Carr’s (, ) (in my view successful) objection that
all semantics for DNMs in the mold I have described objectionably write normative
assumptions into the semantics.22 The scope of Carr’s objection is, I shall argue, rather
greater than you might expect. I shall argue for this by attempting to describe a
proposal broadly along the lines she suggests (on which DNMs are parametrized, not
only to decision problems, but to decision theories). It is fairly clear that taking this sort
of proposal seriously will mean excising all “substantive” normative assumptions—
basically, I will argue, all the normative assumptions—from the semantics. This is a
delicate matter, since it is, I’ll argue, prima facie incompatible with the sort of research
program sketched in the prior sections.
. Disagreement
It is a truism that agents with different preferences and priorities will tend toward
disagreement about what is to be done. It is also a truism that such agents are generally
linguistically competent with respect to DNMs and the broadly normative concepts
they function to express (note: I will henceforth refer to this mixture of linguistic and
conceptual competence simply as competence with respect to DNMs). While it would
be too strong to say that it is the job of the semantic theorist to account for this sort of
disagreement—here I disagree with, e.g., MacFarlane () for reasons broadly like
those given by Plunkett and Sundell ()—it is at least incumbent on the semantic
theorist not to state a theory which renders the disagreement unintelligible (compare
Gibbard , chapter ).
Consider the Newcomb Problem. For the uninitiated, a brief description. There are
two boxes, A and B. You get to keep whatever is inside any box you open. There is a
powerful Predictor, which makes extremely reliable predictions about your actions,
which at the time of your decision has already acted thus:
i. It has put $,  in box A.
ii. If it has predicted you will open just B, it has in addition put $,, in B.
iii. If it has predicted you will open both boxes, it has put nothing in box B.

22 I take myself here to be fleshing out an objection belonging to Carr (, ); this is also part of
the project of Cariani (), as I note at many points below. Carr’s main worry about writing a normative
assumption N into the semantics of DNMs is that it predicts a kind of linguistic incompetence in those who
use DNMs to express normative views distinct from N. (Intriguingly, she connects this to G. E. Moore’s
Open Question Argument against naturalistic analyses of “good”.) I quite agree with her. I only claim that
there is a lot worth saying about the issue beyond what Carr has said about it. I will try to say a small portion
of it here. I’ve no reason for thinking she would disagree with any of it.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 nate charlow

At the time of your decision, you know the Predictor has already acted thus. You are
given a choice between two actions: opening both boxes (two-boxing), or opening just
box B (one-boxing). The decision problem is given by roughly the following table:

Predicts one-boxing Predicts two-boxing

One-box $,, $
Two-box $,, $,

Following Gibbard and Harper (), Evidential Decision-Theorists (“EDT-ers”) cal-

culate the expected utility of an action X using, for each contingency C, the probability
of C conditional on the agent doing X, and they tend to assert (). Causal Decision-
Theorists (“CDT-ers”) calculate the expected utility of an action X using, for each
contingency C, the probability of C were the agent to do X—which, when the agent’s
action does not causally affect the probability of C, as in the Newcomb Problem, is
simply equal to the probability of C. CDT-ers tend to assert ():
() You should one-box.
() You should two-box.
Truism : each speaks correctly (in the minimal sense that each appropriately deploys
her competence to express her view about what to do). Truism : they disagree—
about what to do. Final truism: the disagreement stems from different views about
which decision theory—EDT or CDT—is correct. Assuming the truth-condition of
the DNM is of the shape for which I have argued, the disagreement originates, not
from a disagreement about what value to plug in for P —all parties agree about the
shape of the decision problem—but rather from disagreement about what value to
plug in for CH D .
Similar phenomena will, I conjecture, arise for any normative view about which
there can exist disagreement among agents that are competent with respect to DNMs.
(Note that whether the disagreement is rooted in one party’s irrationality is neither
here nor there, so long as the disagreement exists in the presence of the relevant seman-
tic and conceptual competence.) This plausibly describes any contentful normative
view whatever.
. Binding
In what will turn out to be a similar vein, next consider (cf. Carr , p. ):
() If EDT is correct, you should one-box.
() If CDT is correct, you should two-box.
These claims are intelligible and seem to have an obvious meaning. (In fact, they are
apparently correct to say.) To them we might add:
() If the correct decision-theory is risk-averse, we should block neither shaft.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

decision theory: yes! truth conditions: no! 

() If the correct decision-theory is risk-seeking, we should block a shaft.

Similar conditionals can be formulated and used to intelligibly (and correctly) express
claims about the verdicts of any contentful normative view whatever for any decision
problem whatever.
Such conditionals would, on the face of it, seem likely to be compositionally similar
to more prosaic examples such as:
() If the law forbids jaywalking, you should wait for the light.
() If you want to go to Harlem, you should take the A-Train.
These examples are apparently united by this feature: a hypothetically introduced
decision theory—in cases () and (), a hypothetically introduced preference—
apparently “binds” (in a very loose sense) a corresponding parameter in the matrix
DNM.23
. Parametrization
What morals should a semanticist draw from such cases? While Carr (, )
perhaps does not have in mind this exact question, her admonition against writing
substantive decision-theoretic assumptions into the semantics (and suggestion to
parametrize the truth-conditions of DNMs to decision theories in addition to decision
problems) gives clear guidance. We cannot represent the disagreement in () and ()
as intelligible, nor can we understand the apparent ability of hypothetically introduced
preferences to bind corresponding parameters in the matrix DNM, unless we allow—
contra a large body of orthodoxy—that the semantic variability in a DNM—a vague
notion, but one we will precisify shortly—outruns the specification of the relevant
decision problem.
There are various ways to see this. Let  be the Newcomb Problem, and let CHEDT
and CHCDT be the choice functions of EDT and CDT respectively. Suppose further
that the correct decision theory is CDT, and suppose we take this to motivate a truth-
condition for DNMs of this shape:
/ CHCDT (P , I , A )
X is true (given ) iff ¬X ∈
If CDT is the correct decision theory, this truth-condition is, in a very natural sense,
the correct truth-condition for the DNM. (I’ll elaborate below.) But it would seem that
we must resist this very natural thought. It seems to prevent us from understanding
the EDT-er’s utterance in (). The truth-condition the EDT-er intends is apparently

23 For differing views of how this “binding” happens, see von Fintel and Iatridou () and Charlow
(a). In the latter, I argue that the sort of doubly modalized account of these sorts of conditionals sketched
in von Fintel and Iatridou () cannot give a fully general account of the semantics of these conditionals,
roughly because it does not take the notion that this is a form of binding seriously enough. In a similar vein,
Yalcin (a); Cariani () cite the ability of probabilistic claims in antecedents to affect the interpretation
of modals in their consequents to motivate a treatment of probabilistic semantic parameters akin to the
treatment of the decision-theoretic semantic parameter for which I argue here.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 nate charlow

semantically ineligible—ruled out as a potential meaning of the DNM, by semantic

fiat. Insofar as one party to the disagreement cannot be interpreted as meaning
something intelligible, this apparently commits us to regarding the disagreement
between () and () as unintelligible. This is a disaster. People can, of course, express
mistaken normative views and engage in substantive normative disagreements by
using (and thereby exploiting their competence with respect to) DNMs.24
How can we accommodate decision-theoretic variability in the semantics? We have
options. Here, first, is a non-starter. Suppose we intensionalize CH, so that it expresses
a function from an ordinary circumstance of evaluation i (e.g., a world or world-time
pair) into the the choice function characterized by the correct decision theory at i.
/ CH(i)(P , I , A )
X is true at i (given ) iff ¬X ∈
This gets certain things right. It allows, for example, that the EDT-er manages
to express a proposition. More impressively, if we simply assume a vanilla strict-
conditional semantics for indicative conditionals—so that (if p)(q) is true iff, roughly,
for all i such that p is true at i, q is also true at i—intensionalization renders the truth-
condition of a conditional such as () roughly as follows: for all i such that EDT is
the correct decision theory at i, the DNM you should one-box is true at i. This is not
bad; indeed, it is quite similar to the truth-condition suggested by, e.g., von Fintel and
Iatridou () for preference-shifting, “anankastic” conditionals such as ().
But it gets certain things badly wrong. The proposition expressed by the EDT-er who
asserts () is apparently not one she would intend to express: it is a proposition true at i
just if CH(i) (which is by assumption the choice function of CDT for arbitrary i) selects
one-boxing. It is a proposition that is, moreover, false with respect to any circumstance
of evaluation (since, as everyone knows, CDT tells you unequivocally to two-box).
Why, then, would an EDT-er care to assert it? (Similar remarks apply mutatis mutandis
to a theory which treats CH as expressing a function from a context of utterance c into
the correct decision theory at c.) It is, in any case, not the right story about the sort of
binding into the matrix DNM involved in these sorts of constructions more generally
(for further details, see Charlow a).
A more promising strategy—and the one Carr herself adopts—simply introduces a
decision theory as a separate and sui generis semantic parameter.25 (I’ll refer to this

24 Similarly, Cariani () argues against a “Fully Bayesian” attempt to write a decision theory into
the semantics on DNMs (on which, roughly, X says the expected value of X is sufficiently high), by
noting that a Fully Bayesian semantics for DNMs cannot intelligibly represent the disagreement between
someone whose choices are characterized by a risk-averse choice function and someone whose choices are
characterized by a Fully Bayesian choice function. (Cariani additionally offers a convincing response to any
attempt to represent the view of the risk-averse agent as Fully Bayesian.)
25 The positive proposal of Cariani () is an interesting direct challenge to this claim, since it does
not involve the use of any decision-theory parameter, but still claims to achieve the right kind of decision-
theoretic neutrality. Apart from the sort of neutrality Cariani is interested in, however—which is in his essay
limited to being able to represent attitudes of and disagreements between agents with different attitudes
toward risk—there is still the issue of accounting for the meaning of ()–(). On this matter: I think there is
likely a tension between Cariani’s use of a binding-type argument to motivate parametrization to probability

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

decision theory: yes! truth conditions: no! 

strategy as Parametrization or Relativization.) There are different ways this sort of

strategy could be pursued. According to Dowell-style Contextualists (Dowell ),
we would parametrize the content of a DNM, not just to decision problems, but also
to decision theories. MacFarlane-style Relativists (e.g. Stephenson ; MacFarlane
) will also want to parametrize DNMs to decision theories, but will construe
this further parametrization as an elaboration of the context in which DNMs are
assessed for truth (rather than the context in which they are assigned interpretations).
Various other positions on the nature and proper theoretical interpretation of this
parametrization (and the mechanism by which conditionally introduced preferences
bind parameters in matrix DNMs) are possible.
. Competing truth-conditions
Parametrization to decision theories seems to me the way to proceed here. But there
is a rather fundamental difficulty with it. Parametrization to decision theories would
seem to entail allowing the possibility of a competing truth-condition associated with a
DNM “in the mouth of ” the EDT-er, of roughly the following shape:
X is true (given ) iff ¬X ∈
/ CHEDT (P , I , A )
Prima facie, it thus seems to mean allowing that the EDT-er speaks truly with (). Of
course that is something we should feel uncomfortable saying, if CDT is the correct
decision theory. If CDT is the correct decision theory, ()—even “in the mouth of ”
the EDT-er—is simply false.26 Indeed, this is precisely what the conditional () seems
to be (correctly) expressing. A similar sort of problem will arise in any case in which
conditionals can be formulated and used to express claims about the verdicts of a
contentful normative view.
Here I want to work through some options for dealing with this general sort
of problem (though I will maintain my focus on the issues arising from the
EDT/CDT disagreement). I will mostly be re-canvassing territory already covered in
the Contextualism-Relativism literature.

functions (see fn. ) and his decision not to parametrize to choice functions (since similar arguments can
be mustered for both).
26 One may object to the apparent assumption that the decision theory “of ” the speaker determines the
decision-theoretic parameter against which the speaker’s assertion of () is evaluated for truth. But there
is still the problem that evaluating what the EDT-er asserts, relative to CHEDT , yields a verdict of truth.
To avoid the conclusion that the EDT-er speaks truly, at, we need to prohibit evaluation of what the EDT-
er asserts relative to CHEDT . It is unclear on what grounds such a prohibition might rest. Phrased this
way, I suspect this is simply a vivid instance of a more general difficulty with model-theoretic semantics—
its difficulty deriving truth-conditions for false sentences on which they can be seen to come out false. As
Lepore (, p. ) puts it, “[F]rom a relativized truth-theory we cannot derive an absolute truth-theory.”
I do not want to be construed as endorsing this critique generally, but I think it is powerful in the case of
DNMs.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 nate charlow

No correct view? Perhaps there is no fact of the matter about what the correct deci-
sion theory is (perhaps owing to a meta-normative Error Theory, à la Mackie ).27
If this is right, then the supposition that CDT is correct is in some sense an ineligible
hypothetical.
This would come as a surprise to partisans of the CDT/EDT debate. It seems to
have the further consequence that commitments to and assertions of DNMs that are,
on the face of it, premised on an assumption of the correctness of either CDT or EDT—
() and ()—cannot be rationally maintained. I take this provisionally as a reductio
of the view under consideration. Whatever the meta-normative facts turn out to be,
the possibility of rational commitment to the correctness of a normative view (and
the ability to rationally assert this commitment) should not be relinquished.28 At the
least, the notion that recherché metasemantic considerations such as those in play here
would prompt us to relinquish it should not, absent much further argument, be taken
seriously.

Ecumenism? Perhaps, despite appearances, both EDT-ers and CDT-ers speak cor-
rectly (by which I mean something more than that they appropriately deploy their
competence with respect to DNMs). There are two main ways of making good on this
thought: Contextualism and Relativism.29
According to Contextualism, the EDT-er who asserts () makes a claim whose
content is glossed roughly as the choice function of EDT selects one-boxing, while the
CDT-er who asserts () makes a claim whose content is glossed roughly as the choice
function of CDT selects two-boxing. But, while I acknowledge that there are theorists
who would disagree, I will suppose that this does not correctly report the content of
the CDT-er’s claim. This would render the content of the CDT-er’s assertion of ()
trivial (likewise for the EDT-er’s assertion). It would thus render the purpose of any
article in which a CDT-er undertook to defend the correctness of such an assertion

27 This sort of response works better in cases involving disagreements with epistemically modal claims.
There is no fact of the matter about what the correct domain of quantification for epistemic modals is: there
is nothing over and above different bodies of evidence from whose vantage different epistemically modal
claims can be made. (A relatively more informative body of evidence is in some sense superior to a relatively
less informative body of evidence, but taking this idea seriously in a metasemantics for epistemic modals will
ultimately require us to shrink domains of quantification to maximally informative domains, i.e., singleton
sets of possibilities.) Similar remarks could be made for other sources of variability in epistemic claims. For
instance, assuming that knowledge-attributions are sensitive to stakes or to standards, most epistemologists
would agree that it is nonsensical to ask which stakes or standards are correct. The most we can say as theorists
is to describe which certain stakes or standards are operative in a context, by way of explaining the judgments
of truth and falsity that speakers who are competent with the relevant epistemic language tend to make.
(The skeptic, of course, disagrees with this general assessment, but this has earned the skeptic a fair share
of opprobrium in the contemporary epistemological literature on skepticism.) For reasons that will emerge
shortly, this is also why I think Contextualism about epistemic language remains viable in the face of this
sort of worry, while Contextualism about deontic modals does not.
28 For a defense of such commitments (within a broader program of anti-Realism about the normative)
that I take to be basically sound, see Blackburn (, ); Gibbard (, ).
29 There are, of course, many more than two ways, but I must confine my attention here to the main
contenders.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

decision theory: yes! truth conditions: no! 

mysterious. This is not exactly a palatable way to regard the work of your philosophical
colleagues.30
According to Relativism, however, the EDT-er who asserts () makes a claim
whose propositional content is glossed roughly as the correct choice function selects
one-boxing.31 This seems like exactly the right thing to say. For the Relativist, the
truth of this content is, however, said to be perspective-dependent. In particular, it is
dependent on a choice of choice function. The claim the correct choice function selects
one-boxing is true relative to CHEDT , false relative to CHCDT . Thus both the EDT-er
and the CDT-er speak correctly, since both assert contents that are true relative to the
choice functions supplied by their respective perspectives. So Relativism is a form of
Ecumenism.
Relativist Ecumenism seems to offer the best hope of a propositional semantics for
DNMs that is responsive to the worry developed in this section. In the next section,
I will explain why it is not a very good hope and go on to explain the implications of
this for theorizing about DNMs.

 Metasemantics
In this last section, I will first explain my dissatisfaction with propositional semantics
for DNMs. I will suggest a semantics—really, a metasemantics—that, rather than
attempting to deliver “truth-conditions” for DNMs, simply attempts to model linguistic
competence with respect to DNMs: to represent (very roughly) what state of mind
an agent is in when she judges a DNM to be the case. Because such competence
can co-exist with effectively any normative view, the core semantics for DNMs must
be extremely minimal: the compositional semantics of a DNM of the form φ
is nothing more than a function from sequences of parameters P , . . . , Pn into a

30 One suspects Plunkett and Sundell () will deploy an alternative way of rationalizing the
assertion—e.g., as an attempt to achieve a mutually acceptable conversational scoreboard (compare Lewis
c) on which CHCDT is mutually agreed to partly determine the content of assertions of DNMs, or as an
attempt to persuade one’s interlocutors to modify their own decision-theoretic views. (Their preferred term
here is “negotiation.”) I am, as a rule, basically happy with these sorts of pragmatic stories, but, for reasons
I will describe below, I think it makes better sense to embed them in an Expressivist (Charlow b, ;
Yalcin a; Rothschild ) or Constraint-Theoretic (Swanson forthcoming) metasemantics.
I note that the most convincing defenders of Contextualism for epistemic modals—von Fintel and
Gillies—choose as their ur-semantics for epistemic modals an apparently non-truth-conditional dynamic-
semantics borrowed from Veltman () (see, e.g., von Fintel and Gillies ). Insofar as Contextualism
about epistemic modals is committed to the doctrine that epistemic modals conventionally express propo-
sitions, their view is simply not a Contextualist view. However, insofar as Contextualism about epistemic
modals is not committed to the doctrine that epistemic modals conventionally express propositions, with all
of the dubious pragmatic commitments that are normally thought to come along with such a commitment—
e.g., that the essential effect of accepting an epistemic modal involves adding the proposition it expresses to
one’s beliefs (against this, see esp. Yalcin )—I would not object to being called a Contextualist.
31 As this suggests, I am understanding Relativism to encompass both MacFarlane-style Assessor
Relativism and various forms of Nonindexical Contextualism. The differences between these views do not,
I think, matter for my purposes here.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 nate charlow

condition glossable, roughly, as the domain characterized by P , . . . , Pn is φ-entailing

(equivalently: ¬φ is not among the set of admissibles characterized by P , . . . , Pn ).
Fixing definite values for these parameters generates a contentful instance of this
schematic condition (to which I will refer as the DNM’s satisfaction-condition relative
to these values; the schematic condition I will simply refer to as the DNM’s satisfaction-
condition).
But it is, I’ll argue, fairly clear that the satisfaction-condition thus characterized
cannot really be thought of as the DNM’s truth-condition, at least in the relatively com-
mon sense of that notion according to which a sentence’s compositionally determined
truth-condition:
i. determines its propositional content in context32
ii. subsequently determines a canonical discourse role (namely, adding the proposi-
tion so-determined to the Common Ground or some other body of information)
and canonical perlocutionary function (namely, proposing that one’s addressees
adjust their beliefs so that they come to accept the proposition so-determined)
for that sentence33
Since this is liable to lead to confusion, let me emphasize: it is irrelevant for my
purposes that the satisfaction-condition characterized by a DNM can be thought of
as a truth-condition in some other sense of that notion. My claims here concern only
the notion of truth-conditionality I have explicitly invoked. Importantly, this is the
same notion that is generally assumed by proponents of Relativism, which is my main
target in this section. (To give an example of a sense of truth-conditions with which
I am comfortable: I do not blanch at calling the assertions of EDT-ers false and the
assertions of CDT-ers true! I explain my behavior here as you’d expect: () is unsatis-
fiable relative to parameters that represent my state of mind, while () is satisfied.)

32 It will in §. become clear that, by propositional content, I mean a representational content. A repre-
sentational content is an entity that encodes what I term a locational perspective: a property that agents can
self-ascribe by way of self-locating in a space of centered possible worlds (cf. Lewis a). The state-type
of self-location is functionally distinguished from motivational state-types such as preference or desire, as
well as from broadly representational state-types that nevertheless cannot be propositionally individuated
(e.g., assigning p a credence of . conditional on q). There is, to be sure, a different sense of “proposition”—
namely, content that can “bear” truth-values, can serve as the denotation of a that-clause, and can (thereby?)
serve as the object of an attitude-type which we might neutrally term acceptance (as well as attitudes such as
doubt, uncertainty, etc.). Propositions in this latter sense potentially admit both representational and non-
representational varieties and might be understood as the semantic values of DNMs without running up
against anything I say here (see Schroeder ). Here I claim only that DNMs fail to encode perspectives
that help an agent to self-locate in a space of centered worlds.
33 Some readers may see a bit of slippage here. Note that I explicitly do not suggest that the right view of
the relation between compositional semantic value and communicated content is identity (compare Ninan
). I explicitly invoke determination in lieu of identity. More broadly, I take the view of the subject matter
of semantic theorizing I am articulating here to be basically consonant with Yalcin (); indeed, it draws
inspiration from Yalcin’s own work on Expressivism.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

decision theory: yes! truth conditions: no! 

If I am right that DNMs lack truth-conditions in the relevant sense, it is fairly clear
that the standard Stalnakerian pragmatics, on which assertions of sentences propose
their truth-conditions for addition to the Common Ground or some other body of
information, is not the right pragmatics for DNMs. In its place I briefly describe an
Expressivistic understanding of pragmatics, on which (i) asserting a DNM expresses
that DNM’s satisfaction-condition (by which I will not mean that it asserts that the
satisfaction-condition is met), (ii) the perlocutionary force of expressing a satisfaction-
condition is to propose that one’s addressees adjust their cognitive state so that the
satisfaction-condition of the DNM is met.
. Against propositions
I have no decisive objection to Relativism to air. But I do have an objection, which
I will describe briefly—the purpose being less to convince than to cajole.34 The claim
that there is a Relativistic propositional content in play for DNMs is ill-motivated. Such
contents are not really fit to play a good number of the theoretical roles we normally
require propositional contents to play.
• It is really not the right sort of thing to be asserted in cases where disagreement is
known to obtain (since it is known to exclude the perspective of one’s addressee,
and so simply cannot be consistently accepted by the addressee) (compare Egan
).
• Nor is it the right sort of thing to ground an account of why two agents who assert
incompatible perspective-dependent claims disagree (Dreier ).
• Nor, most significantly, is it the right sort of thing to be believed. At least the
standard account of what it is to believe a perspective-dependent proposition
(Lewis a) does not seem to apply. On the face of things, an agent who accepts
() is not locating herself in one region of perspectival space as opposed to
another—the region of perspectival space occupied by fans of EDT. An agent
who accepts () is just an agent who has this kind of perspective, not one
who self-ascribes the property of having this kind of perspective.35 Similarly, an
agent who accepts a conditionalized DNM such as () is not self-ascribing a

34 I develop the sort of view I describe here further in Charlow (a, b, ). See also Yalcin
(a); Rothschild (); Moss (); Swanson (forthcoming).
35 Compare Yalcin () on the attitude constitutively involved in acceptance of an epistemic modal.
Compare also Cariani (), who (in making a different point) notes that a sentence such as John thinks
you should refuse the pill “is not an ascription to John of the content that refusing the pill maximizes expected
value. Rather [it] is an ascription to John of a way of prioritizing alternatives relative to which refusing the
pill is the best option.” Egan () would likely disagree about whether Relativists can give a compelling
account of the state of mind of accepting a DNM along the lines suggested here. His case rests largely on
the claim that the de se-ified state of mind typically has the right functional profile—closely, but defeasibly,
connected to motivation. I cannot argue against the de se understanding of the relevant state of mind here—
although I do find it prima facie incredible. I can, however, do a bit of dividing and conquering. If you are (as
I am) inclined to a more robust form of motivational internalism—on which sincere and competent moral
judgments motivate, not defeasibly, but necessarily—you will not be happy with Egan’s account here.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 nate charlow

conditional proposition (whatever this would amount to); rather, her acceptance
consists in a fact about her preferences under the indicative supposition that the
conditional’s antecedent is true (Charlow a).
Summing up, accommodating the facts about competence and binding behavior
for DNMs seemed to require parametrizing DNMs to a choice of decision theory.
But this in turn makes it hard to see how to give an (in the phrase of Lepore )
“absolute truth-theory” for claims such as () and (). Not that I think there is
anything wrong with relativized truth per se. But none of the various ways of making
sense of a decision-theory-relative (and, more generally, preference-relative) notion
of truth for DNMs seems palatable. Relativism offers the best hope of a propositional
semantics for DNMs that is responsive to this sort of worry. I would, however, prefer
to look elsewhere.

. Modeling competent judgment

If the compositional semantic value for a DNM cannot be understood as represent-
ing its truth-condition, what can it be understood to represent? Stripped of truth-
conditional presuppositions, the satisfaction-condition delivered by the semantics has
it that X is satisfied at CHD ,  iff ¬X is not permitted in  by lights of D.
CHD ,  | X iff ¬X ∈
/ CHD (P , I , A )
But a satisfaction-condition of this form does not tell us much absent a metasemantic
account of what the relation | represents. Truth-conditional stories, of course, take
this to represent the truth of X relative to CHD , . I have argued that—given some
common assumptions about truth and propositions—this is not an appealing story
about what the relation | represents.
Instead, I will suggest that this condition is a representation of a psychological
characteristic in an agent α who bears a relation of the following sort to what I’ll call a
Representor. (For other iterations of this project, see see Charlow a, b, a, .)
Definition . CH D ,  is a Representor for α iff (i) α’s first-order preferences and
information can be represented with a decision problem , (ii) α’s rational preferences
can be represented with a choice function CH D .
When CHD ,  is a Representor for α and CHD ,  | X (and α is semantically
competent with respect to X), we will say α accepts X. The compositional seman-
tics for DNMs represents, not the conditions under which a DNM is true. Rather, it
represents the conditions under which the DNM is competently accepted by an agent.
I will not argue at any length that this is a viable metasemantics for DNMs (beyond
noting that it explains directly how competence with respect to DNMs can co-exist with
effectively any normative view). In view of the interactions that obtain between DNMs
and “ordinary” sentence-types, this would require arguing, inter alia, that this works
as a metasemantics, not only for DNMs, but all natural language sentence-types—
arguing, in other words, that the following sort of view is correct.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

decision theory: yes! truth conditions: no! 

Semantic Satisfaction Models Competent Judgment

A semantic theory T for L is a theory compositionally determining a satisfaction
relation |L,T between Representors R and sentences φ ∈ L, that meets the following
condition: R |L,T φ iff, for each α such that R is a Representor for α and α is
competent with respect to L, α accepts φ.
The task of developing, grounding, and extending this general type of theory is
enormous, but is being actively pursued by a range of theorists with empirical
ambitions.36 Central questions include: how and to what extent should we rewrite
compositional semantics for “ordinary” sentence-types?37 How should we introduce
a truth-predicate for the language?38 This represents a thriving area of research in
philosophical and linguistic semantics.
Obviously this is a promissory note rather than a worked-out defense. But it is a
promissory note for, at the very least, a relatively appealing view of DNMs (embedded
within an increasingly appealing view of the proper theoretical ambitions of semantic
theorizing).
. Expressivism
It is of course natural to ask for a story about what agents are doing when they accept
and assert DNMs. A natural first pass is this: an agent who asserts X is expressing a
competent judgment that X—she is expressing her acceptance of X.
Such acceptance is not individuable by appeal to a propositional content. Rather,
it is a non-propositionally-individuable state of mind. As there is, in general, no
proposition p such that an agent has a conditional credence of j in q given r iff she fully
accepts p (Lewis ); and as there is, in general, no proposition p such that q is an
epistemic possibility for an agent iff she accepts p (Veltman , Yalcin ); there is
(with even clearer reason) in general no proposition p such that an agent’s representor
R = CH D ,  is such that R | X iff the agent accepts p. The reason for the latter
is quite simple: having a state of mind representable with CHD and  such that ¬X ∈ /
CHD (P , I , A ) obviously cannot be regarded as a function of the propositions one
accepts—one’s full beliefs. It is, in large part, a matter of what one prefers.
So, agents who assert that X are expressing competent judgments holding, in
general, of an agent α iff α has a Representor R = CH D ,  such that ¬X ∈ /

36 For some initial attempts, see the citations above. Charlow (a) and Starr (forthcoming) argue
that the sort of reconceptualization of semantic theorizing being entertained here can be motivated by
independent considerations having to do with imperatives. I would also suggest, tentatively, that this is
the right sort of metasemantics for the Dynamic Semantic program (see, e.g., Kamp ; Groenendijk and
Stokhof ; Veltman ).
37 This turns out to be basically trivial for all the sentences ordinarily (and helpfully) thought to express
propositions, while less trivial, but extremely illuminating, for sentences—conditionals and the language
of subjective uncertainty are key cases—often thought to confound the ambitions of truth-conditional
semantics.
38 A deflationary truth-predicate is a natural option, but I am of the view that it should be resisted in
light of the “Hiyo” objection (Dreier ). Charlow (c) develops my own view.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 nate charlow

CHD (P , I , A ). To be frank, this does not really tell us anything very substantial
about the pragmatics of DNMs (indeed, many would regard it as a truism). But it
is a good start. For it allows us to formulate the following question: why do agents
express judgments of this type? More precisely, what, in general, is the illocutionary
and perlocutionary function of expressing a judgment of this type?
Different answers are possible here. Here is a sketch of the answer I favor (developed
in more detail in Charlow a, b,  see also Starr ). Speakers propose
concrete updates of the relevant state—understandable variously as the Conversational
Scoreboard (Lewis c), the cognitive states of their addressees, or perhaps some-
thing else. (I will assume here that states are just Representors.) There are two broad
kinds of updates that speakers propose: additive updates and tests. Additive updates, if
accepted, strengthen the Representor’s beliefs or preferences; imperatives and ordinary
declaratives are plausibly thought to involve such updates. Tests query the Representor
for the property of accepting a sentence; passing the test means that the sentence is
accepted, failing means that it is rejected. Modals of various stripes have long been
thought to be prime candidates for test-expressing sentences in natural language.
(Veltman , e.g., claims the only appropriate reaction to a modal is generally to
agree—if the modal is acceptable relative to a representation of your state of mind—or
to disagree—if not.)
Expressing a test might seem a trivial sort of speech act, since it does not propose
a change to the relevant state of mind. Not so. It is not hard to see how expressing
a test that yields disagreement, in particular, might set the stage for negotiation and,
ultimately, coordination about the relevant cognitive feature.39 There is much more to
say here, but this is a perfectly good start.
This is roughly what Expressivists (Gibbard) and Quasi-Realists (Blackburn) have
tended to say about these sorts of actions. I’ve chosen the Expressivist label because the
story here is directly inspired by the view of pragmatics articulated by Gibbard (espe-
cially in Gibbard ), who in turn seems to be drawing on Lewisian Conventionalism
(on which expressing a meaning is principally a means for achieving coordination in
attitude).
. Minimalism
No semantics for any sentence should encode presuppositions that rule out the
possibility of someone’s being competent with respect to a sentence, but substantively
mistaken in their acceptance or endorsement of it. For a sentence φ having any
normative, or action-guiding, “content” as part of its conventional meaning—roughly,
for a sentence that encodes any kind of normative view—the proper way to handle

39 This language is similar to that invoked by Plunkett and Sundell (). However, if this is the story we
tell about the pragmatics of DNMs, I see no reason to sustain the idea that the compositional semantic value
of a DNM is a proposition. The proposition qua proposition seems to be doing no explanatory work in a
theory of this sort; in fact, it is hard to make sense of a theory of DNMs in which the proposition supposedly
expressed by a DNM is given a central theoretical role qua proposition.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

decision theory: yes! truth conditions: no! 

this content in the semantics seems to be (i) to parametrize φ’s semantic value to
the relevant kind of normative view, (ii) to adopt a metasemantics on which the
fundamental satisfaction relation of the theory models competent acceptance of φ.
I take this to yield a kind of semantic minimalism about, not only DNMs, but any
kind of language functioning, even if only in part, to encode a normative view. For
sentences of such type, the job of the semantics is to lay bare the manifold ways in
which a normative viewpoint can competently ground acceptance of such sentences.
Whenever we try to say something substantive about the “content” of sentences of
such type, we will be stymied by the diversity of normative views compatible with
competent acceptance of such sentences.
To illustrate, return to the examples of §. These cases were designed to illustrate the
possibility of refining the basic semantics for DNMs to allow it to represent normative
phenomena of apparent significance for the content of DNMs. Neither of the sorts of
refinements I mentioned there seems compatible with the claim with which I began
this section.
The first case involved von Fintel and Iatridou’s () suggestion that weak DNMs
are sensitive to primary and secondary goals, the latter of which further refine the
ordering established by one’s primary goals—an insight borrowed from the theory
of practical rationality. Question: can there be competent use of weak DNMs by, for
instance, Rational Fetishists—agents who systematically fail to distinguish between
first-order and rational preferences? What about psychological value monists—agents
for whom the choice-worthiness of anything is a function of how much of a certain
quality (e.g., pleasure) it possesses, hence for whom it makes no psychological sense to
distinguish primary and secondary goals? On the face of it, (i) the answer to both ques-
tions ought to be yes, (ii) the von Fintel and Iatridou () framework seems to say no.
Simplifying somewhat, the second case involved the suggestion that the semantics
for DNMs be sensitive to differences between the is likelier than relation and the is
better than relation. In particular, the former does not, but the latter does, validate
something related to the Sure Thing Principle. Question: can there be competent use of
DNMs by agents for whom the Sure Thing Principle (or a related dominance principle)
fails—whose preferences cannot be accurately represented with a relation of this sort?
On the face of it, yes: in, for instance, the Allais and Ellsberg Paradoxes, the preferences
of actual agents seem to violate the Sure Thing Principle.40 Such agents are nevertheless
competent with respect to DNMs.
More generally, for just about any rational preference axiom you can state, it is pos-
sible to imagine an agent whose preferences cannot be represented with a preference

40 The Allais and Ellsberg Paradoxes are famously in tension with the Sure Thing Principle, which is why
I mention them here. Ultimately this is incidental to the point I am making: for any plausible dominance
principle, should we encode it in the semantics for DNMs, there will be many cases in which an agent who is
semantically competent with respect to DNMs competently, if irrationally, accepts claims whose negations
are nevertheless theorems from the point of view of the semantics. Gibbard (: Ch. ) was particularly
prescient on these points.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 nate charlow

relation satisfying that axiom.41 It is, on the face of it, simply wrong to try to account for
the impeccability of inferences () and () by claiming that these are semantic validities
in a compositional semantic system in which the Sure Thing Principle (or related
dominance principle) is a Theorem.42 How to understand the impeccability of these
inferences is another matter. (The obvious answer: these inferences are impeccable
because they are acceptance-preserving for a rational agent whose preferences respect
the relevant axioms.) But the way that they are most naturally understood—and
as Kratzer implicitly would understand them, given her apparent codification of a
dominance principle in her definition of the lifted ordering—cannot be sustained on
reflection. Again, I am not trying to suggest that we cannot criticize agents whose
preferences violate apparently obvious preference axioms on grounds of irrationality.
I am only insisting that, when we do so, we do not prevent ourselves from regarding
them as substantively mistaken about DNMs which they competently accept.
The overall effect here is to circumscribe the range of phenomena for which the
compositional semantic theory for DNMs can hope to account. This ripples out to
any kind of language functioning, even if only in part, to encode a normative view.
Should we parametrize the semantic value of some sentence φ to the probability
functions of Bayesian epistemology? We should be wary of this, unless we are prepared
to rule out the possibility of an agent who can access that semantic value via her
competence with respect to φ but whose partial beliefs are not representable with a
real-valued probability function. (How to account for the meaning of probabilistic
language while remaining sensitive to this point is a vexed question indeed.) Or
consider your favorite update semantic theory, on which semantic values are functions
from input information-states to output information-states. Such functions often
write a substantive theory of rational attitude-revision into the semantic value of any
sentence.43 For the same reasons, I think theories like this need to be reconsidered.
I do not say abandoned. But their subject matters need rethinking. Similarly,
theories which obscure Gilbert Harman’s bright line between logical validities and
rationally obligatory inferences (see, e.g., Harman )—e.g., theories that would

41 Probably even the minimal notion of a Representor (as defined in Definition ) assumes too much,
in light of the examples of the Rational Fetishist and value monist. I will not attempt to deal with this here.
42 This works as a criticism of Kratzer’s system, as pointed out in a different manner by Lassiter ().
The normative assumptions that are ironically simply written into Kratzer’s definition of the ordering ≺ have
received insufficient attention. Recall that the ordering is stipulated to be Transitive; to satisfy Independence
of Irrelevant Alternatives (see Charlow b); to forbid cardinal comparisons in value (see Carr );
and to necessarily yield incomparabilities-in-desirability between possibilities when the sets of propositions
satisfied respectively by those possibilities cannot be ordered by ⊆. Any object-language expression of these
features of the preference ordering is thus treated as a theorem within Kratzer’s semantics. This is bad: all of
these features of the preference ordering can be (and have been) disputed without impugning one’s semantic
competence; the latter two, in fact, are roundly rejected by most decision theorists.
43 I criticize Starr (forthcoming) for this in Charlow (a, §..). The theory of Veltman ()
actually is scrupulous about avoiding this. Veltman suggests that when genuine attitude revision is called
for—in cases where updating would lead to the absurd information-state—the semantics should be silent
about how best to proceed.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

decision theory: yes! truth conditions: no! 

appeal to something like the Sure Thing Principle to validate ()—should get a clearer
view on their subject matters. Semantic theories, when working correctly, ultimately
yield a theory of the broad cognitive directive associated with a sentence (as in
§.). Normative epistemology (along with other normative theories) furnishes a
substantive theory of synchronic and diachronic compliance for cognitive directives.
It is imperative to keep these separate, to the extent that you can.

References
Blackburn, Simon () Spreading the Word. Oxford: Clarendon Press.
Blackburn, Simon () Ruling Passions. Oxford: Oxford University Press.
Cariani, Fabrizio () ‘Ought’ and resolution semantics. Noûs. , pp. –.
Cariani, Fabrizio () Deontic modals and probabilities: One theory to rule them all? In
Charlow, N. and Chrisman, M. (eds.) Deontic Modals. Oxford: Oxford University Press.
Cariani, Fabrizio, Kaufmann, M. and Kaufmann, S. () Deliberative modality under epis-
temic uncertainty. Linguistics and Philosophy. , pp. –.
Carr, Jennifer () Deontic modals without decision theory. In Chemla, E. Homer, V. and
Winterstein, G. (eds.) Proceedings of Sinn und Bedeutung , pp. –. ENS Paris.
Carr, Jennifer () Subjective Ought. Ergo. , pp. –.
Charlow, Nate () Imperative statics and dynamics. [Online] Available from: http://
natecharlow.com/work/imperative_statics_and_dynamics.pdf. [Accessed  Nov. .]
Charlow, Nate (a) Conditional preferences and practical conditionals. Linguistics and
Philosophy. , pp. –.
Charlow, Nate (b) What we know and what to do. Synthese. , pp. –.
Charlow, Nate (a) Logic and semantics for imperatives. Journal of Philosophical Logic. ,
pp. –.
Charlow, Nate (b) The problem with the Frege-Geach problem. Philosophical Studies. ,
pp. –. doi:./s---.
Charlow, Nate () Prospects for an expressivist theory of meaning. Philosophers’ Imprint.
, pp. – [Online] Available from: http://hdl.handle.net//spo....
[Accessed  Nov. .]
Dowell, J. L. () A flexible contextualist account of epistemic modals. Philosopher’s Imprint.
, pp. –. [Online] Available from: http://hdl.handle.net//spo....
[Accessed  Nov. .]
Dowell, J. L. () Contextualist solutions to three puzzles about practical conditionals. In
Shafer-Landau, R. (ed.) Oxford Studies in Metaethics. Volume . Oxford: Oxford University
Press.
Dreier, James () Expressivist embeddings and minimalist truth. Philosophical Studies. ,
pp. –.
Dreier, James () Relativism (and expressivism) and the problem of disagreement. Philo-
sophical Perspectives. , pp. –.
Egan, Andy () Epistemic modals, relativism, and assertion. Philosophical Studies. ,
pp. –.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 nate charlow

Egan, Andy () Relativist dispositional theories of value. The Southern Journal of Philosophy.
, pp. –.
Fintel, Kai von () The best we can (expect to) get? Challenges to the classic semantics for
deontic modals. http://mit.edu/fintel/fintel--apa-ought.pdf. Ms. MIT.
Fintel, Kai von and Gillies A. S. () An opinionated guide to epistemic modality. In Gendler,
T. and Hawthorne, J. (eds.) Oxford Studies in Epistemology. Volume . Oxford: Oxford
University Press.
Fintel, Kai von and Iatridou, Sabine () What to do if you want to go to Harlem. Manuscript
MIT.
von Fintel, Kai and Iatridou, Sabine () How to say ought in foreign. In Guèron, J. and
Lecarme, J. (eds.) Time and Modality. Dordrecht: Springer.
Gibbard, Allan () Wise Choices, Apt Feelings. Cambridge, MA: Harvard University Press.
Gibbard, Allan () Thinking How to Live. Cambridge, MA: Harvard University Press.
Gibbard, Allan and Harper, W. L. () Counterfactuals and two kinds of expected utility. In
Harper, W., Stalnaker, R. and Pearce, G. (eds.) Conditionals, Belief, Decision, Chance and
Time. pp. –. Netherlands: Springer.
Gillies, Anthony S. () Iffiness. Semantics and Pragmatics. , pp. –.
Groenendijk, Jeroen and Stokhof, Martin () Dynamic predicate logic. Linguistics and
Philosophy. , pp. –.
Harman, Gilbert () Change in View: Principles of Reasoning. Cambridge, MA: MIT Press.
Hawthorne, John and Stanley, Jason () Knowledge and action. The Journal of Philosophy.
, pp. –.
Kamp, Hans () A theory of truth and semantic representation. In Groenendijk, J., Janssen, T.
and Stokhof, M. (eds.) Truth, Interpretation, and Information: Selected Papers from the Third
Amsterdam Colloquium. Amsterdam: Matematisch Centrum.
Kolodny, Niko and MacFarlane, John () Ifs and oughts. Journal of Philosophy. , pp. –.
Kratzer, Angelika. () What ‘must’ and ‘can’ must and can mean. Linguistics and Philosophy.
, pp. –.
Kratzer, Angelika () The notional category of modality. In Eikmeyer, H. and Rieser, H. (eds.)
Words, Worlds, and Contexts. pp. –. Berlin: De Gruyter.
Kratzer, Angelika () Modality. In von Stechow, A. and Wunderlich, D. (eds.) Semantics: An
International Handbook of Contemporary Research. pp. –. Berlin: De Gruyter.
Kratzer, Angelika () Modals and Conditionals. Oxford: Oxford University Press.
Lassiter, Daniel () Measurement and modality. Ph.D. dissertation. New York University.
Lepore, Ernest () What model theoretic semantics cannot do? Synthese. , pp. –.
Lewis, David () Counterfactuals. Malden: Blackwell.
Lewis, David () Semantic analyses for dyadic deontic logic. In Stenlund, S. (ed.) Logical
Theory and Semantic Analysis. Dordrecht: D. Reidel.
Lewis, David () Probabilities of conditionals and conditional probabilities. Philosophical
Review. , pp. –.
Lewis, David (a) Attitudes de dicto and de se. The Philosophical Review. , pp. –.
Lewis, David (b) A problem about permission. In Saarinen, E., Hilpinen, R., Niiniluoto, I.
and Hintikka, M. B. (eds.) Essays in Honour of Jaakko Hintikka. pp. –. Dordrecht: D.
Reidel.
Lewis, David (c) Scorekeeping in a language game. Journal of Philosophical Logic. ,
pp. –.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

decision theory: yes! truth conditions: no! 

MacFarlane, John () Epistemic modals are assessment-sensitive. In Weatherson, B. and

Egan, A. (eds.) Epistemic Modality. Oxford: Oxford Univesity Press.
Mackie, J. L. () Ethics: Inventing Right and Wrong. New York: Penguin.
Moss, Sarah () Epistemology formalized. Philosophical Review. , pp. –.
Ninan, Dilip () Semantics and the objects of assertion. Linguistics and Philosophy. ,
pp. –.
Plunkett, David and Sundell, Timothy () Disagreement and the semantics of normative and
evaluative terms. Philosopher’s Imprint. , pp. –.
Portner, Paul () Imperatives and modals. Natural Language Semantics. , pp. –.
Rothschild, Daniel () Expressing credences. Proceedings of the Aristotelian Society. CXII. ,
pp. –.
Savage, Leonard J. () The Foundations of Statistics. Dover.
Schroeder, Mark () Two roles for propositions: Cause for divorce? Noûs. , pp. –.
Silk, Alex () Evidence sensitivity in weak necessity deontic modals. Journal of Philosophical
Logic. , pp. –.
Starr, William B. () Dynamic expressivism about deontic modality. In Charlow, N. and
Chrisman, M. (eds.) Deontic Modality. Oxford: Oxford University Press.
Starr, William B. (forthcoming) A preference semantics for imperatives. Semantics and
Pragmatics.
Stephenson, Tamina () Judge dependence, epistemic modals, and predicates of personal
taste. Linguistics and Philosophy. , pp. –.
Swanson, Eric (forthcoming) The application of constraint semantics to the language of subjec-
tive uncertainty. Journal of Philosophical Logic.
Veltman, Frank () Defaults in update semantics. Journal of Philosophical Logic. ,
pp. –.
Willer, Malte () A remark on iffy oughts. Journal of Philosophy. , pp. –.
Yalcin, Seth () Epistemic modals. Mind. , pp. –.
Yalcin, Seth () Nonfactualism about epistemic modality. In Weatherson, B. and Egan, A.
(eds.) Epistemic Modality. pp. –. Oxford: Oxford University Press.
Yalcin, Seth (a) Bayesian expressivism. Proceedings of the Aristotelian Society. CXII. ,
pp. –.
Yalcin, Seth (b) A counterexample to modus tollens. Journal of Philosophical Logic. ,
pp. –.
Yalcin, Seth () Semantics and metasemantics in the context of generative grammar. In
Burgess, A. and Sherman, B. (eds.) Metasemantics: New Essays on the Foundations of Meaning.
pp. –. Oxford: Oxford University Press.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i


Linguistic and Philosophical
Considerations on Bayesian
Semantics
Daniel Lassiter

For the past several years I have found myself advocating what might be called a
“Bayesian” semantics for the deontic interpretations of modal expressions such as
ought, should, good, bad, may, must, and their modified forms (better, etc.). On the
version of this account that I have in mind, deontic modals are not quantifiers over
possible worlds; instead, they have a lexical semantics structured around degree scales,
just like gradable adjectives such as heavy/light, tall/short, and full/empty (Kennedy
and McNally, , etc.). Deontic scales are formally identical to the expected utility
scales used in Bayesian decision theory, but the function assigning values to states
is not necessarily interpreted as representing personal utility—it could be, as in the
case of teleological modals, but it could also represent other kinds of value, such
as moral value. I have argued that this approach accounts better than its main
rivals for a number of empirical phenomena including information-sensitivity, non-
monotonicity, and grammatical gradability (Lassiter, , c, ).1
In this chapter I will summarize several of the arguments for the Bayesian semantics
and discuss challenges that have been posed in four recent papers, two of which are
in this volume: Carr (); Charlow (); and Cariani (, ). The challenges
involve both empirical and high-level philosophical objections, and they have been

1 The Bayesian semantics made its first appearance (to my knowledge) in a brief paper by Jeffrey (a),
and was fleshed out and empirically motivated by Jackson (), formalized by Goble (), and integrated
into a compositional semantics for English, with special attention to the theory of gradability, by Lassiter
(, c, ). Related ideas can be found in Cariani (); Wedgwood ().
Note that the label “Bayesian” is not much used in these works, since the connection with Bayesian
decision theory is mostly formal in nature. But the term is used in subsequent reactions (e.g., Yalcin, a;
Cariani, ), and I find that it is not inappropriate as long as it is clear that this semantic theory does not
stand or fall with the empirical adequacy of a Bayesian theory of human decision-making. At most, there is a
philosophical connection in Jackson’s (), p.  observation that a moral value function can be thought
of as describing how people ought to be motivated in their subjective decision-making.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

bayesian semantics 

taken to motivate weakening the semantics so that it can mimic the predictions of
the Bayesian theory, or of other existing semantic theories, as desired. For obvious
reasons, it is not possible to show on empirical grounds that a strictly weaker theory is
incorrect. However, I will argue that the objections that Carr, Charlow, and Cariani
raise are not compelling. For the most part, the problems discussed are not even
specific to the Bayesian semantics, but rather very general issues that arise for anyone
who is trying to do lexical semantics within the framework of compositional model-
theoretic semantics, and we can reasonably expect that solutions to these general
problems will explain the special cases which involve deontic modals. Given this,
general methodological considerations favor adopting a more restrictive theory over
a more expressive one.
An important exception to this characterization is Charlow’s () objection
involving the deontic analogue of an evidential versus causal decision theory: e.g.,
if a behavior and a disease have a common genetic cause, should individuals refrain
from the behavior? Charlow argues that, to the extent that there is variation among
English speakers in their judgments on this point, no single decision theory can
be implicated in the semantics of ought. However, this argument assumes that the
distinction between causal and evidential decision theory must be located in the rule
used to compute expected values. As I will discuss in section , there is a prominent
version of decision theory according to which the causal/evidential divide is located
in the structure of a probabilistic model and the way that actions/interventions are
represented. On this construal, the Bayesian semantics can generate either kind of
prediction, depending on features of the probability distribution that is given as a
semantic parameter. If this general construal of causal and evidential decision theory
is successful, then, Charlow’s argument does not motivate adopting a more expressive
semantics.
A final category of objections involves real and apparent limitations in the Bayesian
theory’s ability to model certain conceivable kinds of reasoning about obligation. The
real limitations involve MaxiMax or MaxiMin choice rules, but I will argue that the
mere conceivability of someone employing such a choice rule in deontic reasoning is
not a sufficient reason to weaken the semantic theory dramatically. For a compelling
objection, empirical evidence that actual people do so would be needed, and this
evidence is conspicuously absent. The apparent limitation involves, for example, non-
consequentialist judgments in trolley problems. While there is abundant evidence that
many people have such judgments, the Bayesian theory can handle it straightforwardly
by locating variation in the value parameter: non-consequentialist intuitions are
associated with value functions which attach positive or negative value to the fact
that a particular individual takes a particular kind of action. In other words, the
Bayesian theory is not a semanticization of consequentialist ethics. It is a logical
and grammatical theory, and it stands or falls with the descriptive adequacy of its
validities, and with its ability to mesh with good theories of natural language syntax
and pragmatics.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 daniel lassiter

It may be useful for me to make clear from the beginning what my goals and
presuppositions are. I am a linguist and a cognitive scientist, and I am primarily
interested in understanding language, concepts, and reasoning as aspects of human
cognition. Metaphysical questions are relevant to these research interests, but only to
the extent that humans have metaphysical beliefs and assumptions which influence
their cognition and behavior. The question of the ultimate truth of one or another set of
metaphysical claims simply does not arise in this enterprise. Probably in part because
of this orientation, I am unsure to what extent I disagree philosophically with critics
of the Bayesian semantics. (I don’t have any particular objection to Charlow’s expres-
sivism, for example, but I also don’t know whether the distinction between truth-
conditional and expressivist frameworks does any work in a cognitively-oriented
theory of meaning and communication.) In any case, I hope that it will at least emerge
that the criticisms considered here are not special problems for the Bayesian semantics,
and that this theory does a pretty good job of regimenting the grammatical and
inferential features of deontic modals. Indeed, I think that it does this job much better
than any other theory that has been spelled out with a comparable level of specificity
and predictive power.

 Empirical and Theoretical Motivation for Bayesian

Semantics
As I am using the term, the Bayesian semantics relies on two basic claims: that deontic
modals have a semantics built around scales, and that the scales in question have a
particular logical structure. The motivation for building the semantics around degree
scales rather than (just) quantification over possible worlds goes back to the obser-
vations made by Lewis () and Kratzer (, ): many modal expressions have
graded meanings and combine grammatically with degree modifiers and form com-
paratives and equatives. This is true of epistemic, deontic, teleological, and bouletic
modals alike, but we will focus here on deontic expressions. Compare () to () and
() to (), illustrating parallel grammatical gradability among non-modal adjectives
and verbs.
() Murder is worse/better than jaywalking.
() Bill is sadder/happier than Mary.
() You ought very much to leave.
() Bill likes very much to play golf.
Lassiter (, §) gives much naturalistic evidence of grammatical gradability of deon-
tic modals and related expressions, and additional evidence (involving, for example,
neg-raising behavior and focus sensitivity) which supports the claim that many of
these expressions have a scalar semantics. There are some notable exceptions: while

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

bayesian semantics 

ought, should, good, bad, required, and supposed to are gradable, there is no clear
evidence for the gradability of may. It is still unclear what the situation is for must,
or have to: it’s usually thought that they are not gradable, but Portner and Rubinstein
(this volume, chapter ) argue that they are, and that their limited gradability is closely
related to the limitations on gradability of extreme adjectives like huge and gorgeous
(Morzycki, ). Regardless of how this empirical question comes out, gradability and
scalarity are not the same thing. Gradability is a grammatical property of expressions,
while scalarity is a property of the model-theoretic objects that expressions denote
(see Lassiter, c for discussion). It might turn out that deontic modals that are not
gradable nevertheless have a scalar component of their meanings, but the thresholds
that they invoke cannot be bound by other operators.
If this is correct, the next question to ask is what kinds of scales are relevant.
In the recent literature on degree expressions a great deal of attention has been
devoted to scale structure, i.e., the formal properties of degree scales (see surveys
in Lassiter, a; Morzycki, ). The bulk of attention in this literature has been
devoted to two issues. First, there is the question of how to model the difference
between one-dimensional adjectives like tall/short and higher-dimensional adjectives
like healthy/sick, where many factors go into the determination of the degree of the
property that an individual possesses (Kamp, ; Bierwisch, ; Sassoon, ).
Second, there is the question of whether a scale has or lacks lower and upper bounds,
which has been argued to influence the vagueness of adjectives and their potential
to combine with degree modifiers such as slightly and completely (Hay, Kennedy and
Levin, ; Rotstein and Winter, ; Kennedy and McNally, ; Kennedy, ).
In Lassiter (, d, ), I build on the insights of Measurement Theory
(Krantz, Luce, Suppes and Tversky, ) to identify a third parameter of scalar
variation: the logical part-whole structure of scales, i.e., systematic relationships
between the degree of a property that some object possesses and the degree to which
its parts possess the property. In particular, some scales are additive in the sense that
the degree to which an object has the property is the sum of the degrees to which the
object’s proper parts have the property. Clear examples of additive scales are weight
and size: the degree to which I am heavy is the sum of (i) the degree to which my right
arm is heavy, and (ii) the degree to which the rest of my body is heavy. In contrast,
predicates of temperature are not additive but intermediate. The degree to which an
object is hot is not the sum of the degrees of heat of the object’s parts, but rather
some value in between these degrees of heat. If my two arms, two legs, torso, and
head are all around  degrees Fahrenheit, then (fortunately) my body temperature
is not around  degrees Fahrenheit but around the same  degrees. It is easy to
imagine scales which behave differently with respect to part-whole relationships. One
important alternative possibility is maximality. If a scale is maximal, then whenever
an object z is fully covered by non-overlapping parts x and y, z’s degree on the scale
is either the same as x’s degree or as y’s, whichever is greater. Some scales do not
seem to encode any such systematic connection, e.g., happy/sad and beautiful/ugly.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 daniel lassiter

How and whether parts and wholes are systematically related constitutes an additional
parameter of variation in degree scales.
The issue of part-whole structure is vital for a scalar treatment of deontic modals
because many interesting questions in the study of deontic modals are about
monotonicity—when and whether embedding under a deontic modal preserves or
reverses entailment relations between propositions. In possible worlds semantics
entailment is ultimately about part/whole relations, because it is modeled as a subset
(part-of) relation between propositions construed as sets of worlds.
Given this, questions about the monotonicity of deontic modals can be refined into
questions about the part-whole structure of deontic scales. Both additive and maximal
scales have the property that the degree to which an object (proposition) has some
property is at least as great as the degree to which its constituent parts (subsets) have
the property. Suppose that ought means “has moral value greater than θ ” for some
threshold θ . (Under fairly light assumptions, this subsumes the common “better than
not” gloss as a special case.) Then additivity or maximality of deontic scales would
entail that ought is upward monotonic: if φ entails ψ then ought(φ) entails ought(ψ).
This is because the entailing sentence φ denotes a subset of the denotation of the
entailed sentence ψ, and maximal and additive scales require a set to have a degree
of the relevant property at least as great as any of its subsets.
If deontic scales are either additive or maximal, then, we expect inferences such as
the following to be valid on the threshold semantics for ought just mentioned.
() a. You ought to mail this letter.
b. So, you ought to mail this letter or burn it.
Ross’s puzzle (generalized from a related issue around imperatives, Ross, ) is
precisely that these inferences are intuitively invalid. Now perhaps this intuitive
invalidity can be explained away on Gricean lines as due to the presence of disjunction;
if so, an upward monotonic semantics for ought remains plausible (Hare, ;
Wedgwood, ; von Fintel, , but see Cariani,  for objections). However,
Jackson Pargetter’s () Professor Procrastinate scenario provides a structurally
identical failure of monotonicity which cannot be explained away by Gricean con-
siderations about disjunction. Here is a version which I think brings out the relevant
intuitions more clearly than the original (from Lassiter, d, inspired by a scenario in
Cariani, ).
Juliet is considering whether to feign death by taking the drug that Friar Laurence has offered
her. If she does, it will put her in a coma, and she will die unless Friar Laurence administers the
antidote exactly  hours later. If she takes it and the Friar does administer the antidote, she will
succeed in convincing her family of her death and she will be able to live happily ever after with
Romeo. If she does not take the drug, she will live a long life without Romeo and will be less
happy; this is much better than being dead, though. Unfortunately, the Friar is known for being
cruel and capricious, and it is extremely likely (though not totally certain) that he will “forget”
to administer the antidote if she takes the drug.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

bayesian semantics 

Many people find both of the judgments in () reasonable in this scenario.
() a. It ought to be that Juliet does not take the drug.
b. It ought to be that Juliet takes the drug and the Friar administers the
antidote.
But if ought is upward monotonic this pattern of judgments should be incoherent. The
sentence embedded under ought in (b) entails the negation of the embedded sentence
in (a), and so the truth of (b) should entail the truth of It ought to be that Juliet takes
the drug and the falsity of (a).2
This is one of several arguments that have been advanced in favor of the claim that
ought is not upward monotonic; see Lassiter (, chs –) for several more. As a
corollary, if we were correct to suppose that the meaning of ought is built around a
scale of moral/practical value, this scale cannot be either additive or maximal. (By the
way, the theories of Lewis () and Kratzer (, ) make use of qualitative scales
that are effectively maximal.)
A third option, if the value scale is neither additive nor maximal, is that it is
intermediate. If so, we predict that () is a coherent pattern of judgments: φ ∧ ψ can
have value greater than θ while φ does not, as long as the value of φ ∧ ¬ψ (Juliet takes
the drug and the Friar does not administer the antidote) is lower than θ . Then, the
value of φ—equivalently, of the disjunction (φ ∧ ψ) ∨ (φ ∧ ¬ψ)—will be intermediate
between those of φ ∧ ψ and of φ ∧ ¬ψ, and may well fall below the threshold. For
similar reasons, if deontic scales are intermediate we also do not predict the validity
of the Ross inference in ().
But it would be nice to go beyond simply rendering intuitively true judgments such
as () coherent, and rendering intuitively false sentences such as (b) not entailed
by obviously true ones. As Charlow () emphasizes, it would be much better if
we could also explain why the former are felt to be true, and the latter false, in
the relevant scenarios. To do this we need to propose a specific intermediate scale
for the relevant expressions whose structural properties are related to independently
motivated features of the context; we also need to make some assumptions about how
the lexical meanings of the various deontic expressions make use of this scale. We
can make progress on this front if we adopt the proposition that deontic scales have
the structure of expected value—including, as a special case for moral judgments,
expected moral value.3

2 That is, on the usual assumption that ought(φ) and ought(¬φ) cannot both be true. Note that this
version of the puzzle problematizes von Fintel’s () attempt to explain away Procrastinate cases within a
quantificational semantics by supposing that the domain of quantification shifts between the two sentences.
The fact that it is specifically mentioned that the Friar might administer the drug makes it look pretty ad
hoc to stipulate that worlds where he does so are not considered in forming this ought-judgment.
3 To simplify the math, I’m pretending that the set of possible worlds W is finite. I will also continue to be
sloppy about the distinction between sentences and the propositions that they denote relative to a context.
It should always be clear in what follows which is intended.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 daniel lassiter

First, we assign a value V(w) to each possible world w in the set of all worlds W
(or perhaps to coarser-grained objects such as cells in a partition of the set of possible
worlds; the difference will not be crucial here). The value of a world is a real number
representing how morally or practically desirable it would be if all of the facts of the
world were arranged as in w. The use to which we put value assignments will ensure
that they are unique up to positive affine transformation. That is, the numerical value
V(w) assigned to a world w is not meaningful per se, but it gains meaning by virtue
of the relative sizes of the gaps between V(w) and the values of other worlds V(w )
and V(w ).
Secondly, we assume that a probability measure P is provided (“by context”, however
this is spelled out). This function maps propositions A ⊆ W to real numbers in
the [, ] interval. This function obeys the usual constraints: it is additive for disjoint
propositions, P(W) = , and its domain is a (σ -)algebra, closed under (countable)
union and complement. This could be a probability measure representing the state
of knowledge of the speaker or of the holder of an obligation, but we probably need
to allow that it could belong to another relevant individual or be abstracted from the
information available to a group. It could even be the trivial probability measure of an
omniscient being, assigning probability  to the unit sets containing the actual world
and  to all others. (This conceit can plausibly be used to model the “objective” use of
ought: see Wedgwood, .)
Next we define the expected value of a proposition, i.e., the average value of the
worlds in the proposition, where each world’s value is weighted by the probability that
that world will be actual on the condition that the proposition is true.

EV (φ) = V(w) × P({w}|φ)
w∈φ

We can derive from this definition a formula for calculating the expected value of a
disjunction of disjoint propositions: it is a weighted sum of the expected values of the
individual disjuncts, where the weights are given by the probabilities of the disjuncts
conditional on the disjunction itself (Jeffrey, b, §).

EV (ψ ∨ χ ) = EV (ψ) × P(ψ|ψ ∨ χ ) + EV (χ ) × P(χ |ψ ∨ χ )

Expected value is an intermediate scale. Suppose that φ and ψ are mutually
exclusive, EV (φ) > EV (ψ), and that neither has probability zero conditional on the
disjunction φ ∨ ψ. Then it is easy to show, as a consequence of the previous equation,
that φ ∨ ψ has an expected value intermediate between those of φ and ψ.

EV (φ) > EV (φ ∨ ψ) > EV (ψ)

If we are seeking an intermediate scale, expected value is a candidate. Why would we

think that it is the right choice? One reason is that this semantics effortlessly encodes
the empirical phenomenon of information-sensitivity: our judgments about the truth
or appropriateness of deontic sentences frequently depend on what information is

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

bayesian semantics 

available to the deliberating agent or to us as evaluators of the sentences. (Note

however that the expected-value semantics is not “seriously information-sensitive”
in the sense of Kolodny and MacFarlane (): information can reverse preferences
over propositions or actions, but not preferences over worlds. As Charlow ()
discusses, this is a desirable limitation.) Information-sensitivity is an issue that has
exercised philosophers considerably in recent years, leading some to weaken their
semantic theories considerably or adopt other extraordinary measures. But it is an
automatic consequence of the semantics just sketched, for exactly the same reason that
information plays a crucial role in Bayesian decision theory. When deciding whether
some proposition ought to hold (including the proposition that some agent takes some
action), we must consider among other factors how good or desirable the proposition
is. In doing so, we should not only take into account the best outcomes that might
result if the proposition is true (as the Lewis/Kratzer semantics does) or the worst
(as the risk-averse theory of Cariani, Kaufmann, and Kaufmann  does). Rather,
we must consider all possible outcomes and weight their contribution to the overall
result by the probability that they will be the actual outcome if the proposition does
hold. This bit of common sense is faithfully encoded by Bayesian reasoning, and we
simply port its structure over from theories of subjective decision-making to theories
of moral and practical language.
The Juliet scenario given above can also be used to illustrate the information-
sensitivity of deontic scales. Crucially, the story was arranged so that the expected
value of Juliet takes the pill and the Friar administers the antidote is very high, and the
expected value of Juliet takes the pill and the Friar does not administer the antidote is
very low. Juliet takes the pill is logically equivalent to the disjunction of these sentences,
and so its expected value is a probability-weighted average of the expected value of
these sentences. In the account given, we supposed that it was highly likely that the
Friar would fail to administer the antidote if Juliet took the pill, and a reasonable
judgment was that Juliet ought not to take the pill. But if we change the story so that
the Friar almost certainly will administer the antidote, intuitions may shift toward the
judgment that Juliet ought to take the pill, especially if it is clear from the story that
being with Romeo is a great good.
More generally, as we manipulate the continuous range of probabilities for the
Friar’s actions systematically, intuitions about the relative goodness of Juliet’s possible
actions seem to vary continuously (compare Carr, ; Cariani, ; Lassiter, d).
In addition, there is a clear interaction with the relative goodness of the possible
outcomes: the importance of being with Romeo affects the probability needed for a
clear judgment that Juliet ought to take the pill.
This account immediately generalizes to the core of the much-discussed Miners’
Puzzle (Regan, ; Kolodny and MacFarlane, ). I will not pause to repeat the
prose version of this now-familiar scenario. For reference, Table  summarizes the
space of actions and outcomes specified in the story.
As Charlow (), Lassiter (), and Cariani et al. () discuss, the most basic
semantic puzzle in the Miners’ scenario is not the interaction with conditionals on

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 daniel lassiter

Table  Outcomes of possible actions, by world-

type, in the standard Miners’ Puzzle.
Block A Block B Block neither

Miners are in A  survive  survive  survive

Miners are in B  survive  survive  survive

which Kolodny and MacFarlane focus. A more fundamental problem is how the
judgment that We ought to block neither is true can be so robust, given that it is clear
that the best outcomes are all of the following form: either we block A (and the miners
are in fact in A) and we block B (and the miners are in fact in B). In either case, all
miners survive, while doing nothing ensures that someone will die.4 On the usual
assumption that “ought” means “true in all of the best accessible worlds”, this set-up
entails that We ought to do nothing should be clearly false, and We ought to either block
A or block B should be clearly true. In the Miners’ Puzzle, the thing that we ought to
do is guaranteed suboptimal: it is something that we do in none of the best worlds.
When we focus on this issue, the puzzle is a variant of Jackson’s () Medicine
scenario, involving a doctor who must choose whether to prescribe a promising
but very risky experimental treatment or a safe but mediocre treatment. Even if
the best possible result of the experimental treatment is clearly better than the best
possible result of the safe treatment, most people have the intuition that the doctor
ought to prescribe the safe treatment, given that there is a significant chance that the
experimental treatment will (say) cause the patient to die.
In the Medicine puzzle, as Jackson (, p. ) points out, “[t]he obvious answer
is to take a leaf out of decision theory’s book”. First, weigh the values of the possible
outcomes of the available actions against the probability that they will be actual if the
action is taken. Whichever action has the greatest expected value is the one which we
ought to take. This procedure is encoded in the Bayesian semantics we have discussed,
and it combines reasoning about values with reasoning about probabilistic infor-
mation in a common-sensical way. The information-sensitivity of ought is predicted
immediately, with no need to add additional mechanisms or assumptions to the theory.

4 This analysis is challenged by von Fintel (), who claims that the worlds where we do nothing
and some miner(s) die actually are better than the worlds where we pick right and they all live. This
judgment strikes me as incredible. But the discussion around it suggests that the judgment is based not
on consideration of possible worlds qua fully specified states of affairs with no remnant uncertainty, but on
sets of worlds where we perform certain actions, where uncertainty about which world is actual remains
a relevant consideration. (Kolodny and MacFarlane (, §.) seem to slip similarly between intuitions
about worlds and propositions: “a world in which both shafts are left open may be more ideal than one in
which shaft A is closed relative to a less informed state, but less ideal relative to a more informed state”.) This
is not the question, though: we already knew what ordering on worlds a classical semantics would have to
deliver in order to generate the right predictions, but the problem was that it is implausible that a ranking
of worlds in terms of (moral) value could give us that ordering. What we need is a rule that can derive the
relevant judgments about which action is best (taking into account uncertainty about consequences) from a
plausible ranking of fully specified worlds. The Bayesian semantics delivers exactly this. However, if we want
to complicate the account by taking into account whether individuals with certain kinds of information take
certain actions, this is no barrier: see section . for discussion.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

bayesian semantics 

The same treatment extends immediately to the Miners’ Puzzle, and to variants
which involve detailed manipulations of probabilities and outcomes which are prob-
lematic for non-Bayesian theories (Lassiter, d). A plausible analysis of Kolodny
and MacFarlane’s () ought-conditional interactions is also available in terms of
conditional expected value. See Lassiter  for details.

 “Controversial Normative Assumptions” and Binding

Carr () and Charlow () object to this project in various ways. Carr argues that
all of the usual theories of deontic modals “build controversial normative assumptions
into the semantics,” and offers a generalized theory that places most of the contentful
semantics of ought into the contextual parameters. Charlow takes Carr’s objection
a step further, arguing that we can convert Carr’s arguments together with some
problems around disagreement into a general argument against a truth-conditional
semantics for deontic modals and in favor of a deontic expressivism. Both seem to
agree that the Bayesian theory, along with its competitors, confuses the meaning of
deontic modals with the way that people reason about obligations.
This criticism is far-reaching if correct. The Bayesian semantics, obviously, builds
the formal structure of Bayesian decision theory into the semantics. The semantics
advocated by, e.g, Lewis (); Kratzer (); von Fintel (), since it focuses on
what happens in the best accessible worlds where an action is performed, emulates
a decision theory with a MaxiMax choice rule. Cariani et al.’s () alternative
effectively enforces a MaxiMin choice rule, paying attention only to what happens in
the worst accessible worlds where an action is performed.5
Carr’s empirical focus is information-sensitivity, which she elaborates and uses to
argue convincingly that neither the MaxiMax nor the MaxiMin semantics is able to
account for a fuller range of variants on the Miners’ Puzzle (varying information and
outcomes, along the lines discussed briefly above). An obvious response would be to
search for a better semantics, as we did in the last section. Carr acknowledges obliquely
the possibility of a more empirical successful semantics with a Bayesian character, but
does not offer any empirical objections to this project; instead, she moves immediately

5 Note—as a number of authors have recently pointed out, including Cariani () and Charlow
()—that these semantic theories do not technically force deontic judgments to emulate (respectively)
MaxiMax or MaxiMin decision procedures. But this is for a boring reason: until recently, everybody
assumed that deontic judgments were related to something that was independently motivated, such as
moral judgments about the relative values of certain fully specified outcomes, for which uncertainty is not
a relevant consideration. If we weaken our metasemantic theory so that the parameter controlling values is
not tied to anything that can be independently motivated and examined empirically, we grant ourselves the
freedom to reverse-engineer whatever value parameter (ordering source, etc.) is needed to support whatever
the observed judgments are in a given context. There does not seem to be any technical barrier to making
this move, but if it is the only way to save the formal structure of a best- or worst-worlds semantics, the
choice seems clear: a theory which delivers the right judgments on the basis of information about the values
of fully specified outcomes—without requiring extra degrees of freedom—is greatly preferable on general
methodological grounds of restrictiveness and predictive power.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 daniel lassiter

from the failure of MaxiMax and MaxiMin to the proposal that we should parametrize
the semantics by decision rules, allowing that any rational decision theory could be
supplied by “context” and used to calculate truth-conditions.6 In this way the account
effectively subsumes the Bayesian semantics as a special case, invoked whenever the
context supplies a decision rule based on expected value. Perhaps the context always
happens to be like this, in which case the data would appear to support the Bayesian
theory; but Carr’s position entails that this would be merely apparent.
It is striking that Carr moves without argument from the premise that neither
of two salient alternative semantic proposals explains information-sensitivity to the
conclusion that no proposal can solve the problem, unless we greatly increase the
expressiveness of the semantics—equivalently, unless we temper the predictive ambi-
tions of our theory. What is missing is a generic argument showing that no semantic
theory of ought that encodes substantive normative assumptions could be successful.
Without such an argument, we would be justified in insisting on the most empirically
restrictive semantics available that is compatible with the available data (which is, I
have argued, the Bayesian semantics).
A Moorean Open Question argument might do the trick here: if the meaning of
deontic modals encoded controversial normative assumptions, they would not be
controversial, since it would be apparent to all that these assumptions are incorrect.
Like many before me, I don’t find this kind of argument very convincing. It assumes
that we are able to reliably introspect meaning facts, and it relies on a sharp distinction
between meaning facts and facts about the world (compare §. below). However,
there is a hint of an empirical argument in Carr’s (, §) discussion of binding
in conditionals, which is elaborated by Charlow (, §.). Carr points out that it
makes sense in the Miners’ scenario to say such things as
() a. If MaxiMax is right, we ought to either block shaft A or block shaft B.
b. If MaxiMin is right, we ought to do nothing.
This can be explained if the meaning of ought is sensitive to a decision-rule parameter,
and the antecedent of this conditional binds this parameter temporarily for the
purpose of evaluating the consequent. What’s more, it could be taken to show that
the meaning of ought could not build in a specific decision rule: if it did, conditionals
of this form would all be trivial. That is, a conditional whose antecedent describes
the decision rule given by the correct semantics would be equivalent to (a), and any
antecedent which describes a decision rule incompatible with the correct semantics
would be equivalent to (b).

6 One could reasonably disagree with the implication that MaxiMax and MaxiMin are examples of
rational decision rules. The general point remains that there might be multiple rational decision rules which
make different recommendations in some situations. Certainly it would be interesting and relevant for
empirical semantics if we could find evidence that people vary in the procedures that they use to integrate
uncertainty and values in reasoning about obligation, regardless of whether such people would count as
“rational.”

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

bayesian semantics 

() a. If  = , we ought to perform action X.

b. If  = , we ought to perform action Y.
Since this is obviously not a correct prediction, the meaning of ought cannot build in
any specific decision rule, and the parametrization solution is preferable.
This argument is initially quite plausible, but it is an instance of a very general
problem involving conditionals and meaning variation, rather than a special problem
for certain accounts of deontic modals. For example, in Nix v. Hedden () the U.S.
Supreme Court considered the question of whether tomatoes are fruit. The Tariff Act
of  had placed an import duty on fruits, but not on vegetables, and a family of
importers named Nix had sued Hedden, the collector of the Port of New York, arguing
that they should not have been forced to pay the import duty on their tomatoes.
Botanical usage favored the classification of tomatoes as fruits, while ordinary usage
favored classification as a vegetable.
Crucially, both sides of this dispute could happily agree on the following two
conditionals: the dispute was about determining which antecedent was true.
() a. If tomatoes are fruit, the duty applies to imported tomatoes.
b. If tomatoes are not fruit, the duty does not apply to imported tomatoes.
The same semantic puzzle illustrated for () and () arises here. I take it that, if
tomatoes are fruit, then—for reasons familiar from Kripke () and Putnam ()—
they are fruit in all (metaphysically) possible worlds. So, we expect the conditionals in
() to be equivalent either to those in () or to those in (), depending on whether or
not tomatoes are in fact fruit.
() a. If  = , the duty applies to imported tomatoes.
b. If  = , the duty does not apply to imported tomatoes.
() a. If  = , the duty applies to imported tomatoes.
b. If  = , the duty does not apply to imported tomatoes.
Perhaps this is correct, but if so it is not much help in deciding the case. (In fact, the
Court ruled in favor of the Nix family, on the grounds that “the common language of
the people” was more relevant than botanical usage to the interpretation of the Tariff
Act. Arguments over whether tomatoes are “really” a fruit continue to this day.)
The fact that this disagreement involves a natural kind term is not essential.
For example, Ludlow (, p. ) discusses disagreement about whether the horse
Secretariat should have been included in Sports Illustrated’s list of “ greatest athletes
of the th century”. The issue was not whether Secretariat was fast or successful
enough, but about whether athlete can include non-humans in its extension. Here
again, both sides of the debate could agree on the conditionals:
() a. If horses can be athletes, then Secretariat belongs on the list of  greatest
athletes of the th century.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 daniel lassiter

b. If horses cannot be athletes, then Secretariat does not belong on the list of
 greatest athletes of the th century.
Plunkett and Sundell () come to a similar conclusion, and point out that a similar
argument could be run involving midwest, war, and many other ordinary terms.
Standard theories tell us that the conditionals in () and () must receive a
trivializing interpretation, similar to those in () and (). This is not a problem for
anyone’s theory of the meaning of ought, tomato, or athlete: it is a problem for standard
theories of conditionals. What is really going on here is that the antecedents of the
troublesome conditionals somehow bind the interpretation of certain terms for the
purpose of evaluating the consequent. This process can also produce side effects by
influencing the interpretation of other terms that are semantically related to the term
in question. () and () illustrate.
() If Bill ran to the store, Mary positively sprinted. (“We both saw how fast
Bill and Mary were moving; if we agree to call Bill’s style of movement
‘running’, then—given the semantic relationship between these terms—we’re
committed to describing Mary’s faster movement as ‘sprinting’. ”)
() If Sam is small, Al is tiny. (“We both know how big they are; if we decide to call
Sam ‘small’, then—given the semantic relationship between these terms, and
the fact that Al is substantially smaller than Sam—we’re committed to calling
Al ‘tiny’. ”)
Neither of these conditionals commits the speaker to endorsing the consequent; there
may even be an implicature that the speaker thinks the consequent is not a good
description of the relevant situation.
I predict that a solution to the general problem exemplified by (), (), and
especially () and () will account for the binding relationship in () automatically.
The supposition that the antecedent MaxiMax is right is true has an effect on the
interpretation of the consequent, via the supposition that a certain semantic issue has
been resolved in a particular way. This is parallel to () and (), which can be used
even if the speaker is strongly inclined to judge the relevant consequents false. For
similar reasons, someone who is implicitly committed to the falsity of the antecedents
of () and ()—because they describe an incorrect theory of the meaning of ought—is
still able to consider the truth-values of ought-sentences on the supposition that ought
is interpreted differently.
How to make sense of the metalinguistic use of conditionals within a compositional
model-theoretic semantics is a fascinating question about which I have little to say at
the moment. Whatever the best story turns out to be, though, the problem clearly
goes beyond the narrow issue of the lexical semantics of deontic modals; trying to
resolve the problem in () by parametrizing the semantics of ought (or athlete, tomato,
or run) is missing the generality of the issue. My response here is one which will recur
throughout the rest of this essay: critics have mistaken very general semantic puzzles
for special problems of particular theories.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

bayesian semantics 

 Causal and Evidential Expected Value

The version of the binding argument that Charlow () gives is interestingly differ-
ent, because it brings in a deontic analogue of the debate between evidential and causal
decision theorists. Modifying a classic example, suppose that drinking milk does not
cause obesity, but that there is an undetectable gene G which makes teenagers inclined
to drink milk and also, in adulthood, makes them obese. Figure  depicts this common-
cause structure.
If you are a teenager, you don’t know if you have G, but you have to decide whether
to drink milk or not. Milk is tasty and drinking it will give you great pleasure, whether
or not you later become obese; and of course you would strongly prefer not to become
obese. Now consider the following:
() a. You ought not to drink milk.
b. You ought to drink milk.
If you think that (a) is true here, I’ll say that you have a evidential-deontic
judgment. This judgment presumably relies on the fact that, if you make the choice
to drink milk, this gives evidence that you have G, and this is bad news since it means
you will become obese.
If you think that (b) is true, you have a causal-deontic judgment about this
scenario: since the choice to drink milk has no influence on whether you have G, you
should go ahead since it gives you pleasure and does not causally influence whether or
not you will become obese.
There are several formalizations of causal decision theory in the literature. The rule
for calculating expected utility given above was derived from Jeffrey (b), and is
usually assumed to generate evidential judgments exclusively. A prominent approach
to modeling causal judgments, originally proposed by Robert Stalnaker, is to replace
Jeffrey’s conditional probability-based weighting with a subformula referring to the
probability of certain counterfactuals (Gibbard and Harper, ). Another option,
proposed by Lewis (), is to weight utilities by certain well-chosen unconditional
probabilities. If these formal modifications to the expected-value calculation are the
only way to account for the difference between evidential and causal judgments, it
seems that at most one of the following conclusions must hold:

gene G

milk obese

Figure  A standard causal model for the milk/obseity example.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 daniel lassiter

a. The correct semantics for deontic modals refers to a scale of expected value
calculated using conditional probabilities.
b. The correct semantics for deontic modals refers to a scale of expected value
calculated using probabilities of counterfactuals.
c. The correct semantics for deontic modals refers to a scale of expected value
calculated using unconditional probabilities.
d. The correct semantics for deontic modals is noncommittal between these
options—either by being parametrized by a decision theory (Carr) or because
deontic modals can be used to express, in a non-truth-conditional fashion, the
recommendations of these decision theories and many more (Charlow).
Charlow argues that, since native speakers show a mixture of causal-deontic and
evidential-deontic judgments, none of (a)–(c) can be right, on pain of rendering these
speakers’ judgments incorrect as a matter of meaning. He concludes that (d) must be
the correct approach.
This argument misses the fact that there is another approach to causal decision
theory, relying on causal models, which does not require us to modify Jeffrey’s
formula for calculating expected values. This approach and its implications for the
causal/evidential debate is discussed clearly by Meek and Glymour (). It is closely
related to Pearl’s (, §) formulation of causal decision theory, which relies on
the addition to the probability calculus of a do operator representing an intervention
which modifies the probabilistic dependencies among variables in a causal graph. (See
Sloman, ; Sloman and Lagnado, , and the papers in Gopnik and Schultz,
, among many others, for empirical evidence of the relevance of this approach
to causal modeling to human cognition.)
In brief, Meek and Glymour () argue that the causal/evidential divide reduces
to the choice of how actions are modeled in a given probabilistic model: expected
values can be calculated with respect to a measure conditionalized on the observation
that an action has been performed (evidential) or the existence of an intervention
designed to produce that action (causal). Expected utilities are calculated in the same
way in both cases, but when an action is treated as an intervention, the action does
not influence the probability of nodes which are not causally dependent on it. In other
words, the expected utility of an action construed as the result of an intervention will
not be influenced by the “news value” of the action about its causes and their other
effects: interventions render actions independent of their other causes, and so do not
influence their probability or those of other effects which are not also effects of the
action in question.
More concretely, in the case of the milk/obesity example: if we model drinking milk
as an ordinary event, the choice to drink milk is evidence that you have G, and having G
means that you will become obese (Figure , left). This means that the Jeffrey expected
value of drinking milk will be lowered by the fact that it raises the probability of having
G, and so of becoming obese. This gives rise to the evidential-deontic prediction you
ought not to drink milk.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

bayesian semantics 

Imilk gene G P(G milk=T) T Imilk gene G P(G)

T milk obese P(milk Imilk=T ) milk obese

P(obese G, milk = T) P(obese G)

Figure  Effect of conditioning on an observation (left) vs. an intervention (right), following

Meek and Glymour ().

gene G P(G milk = T) gene G P(G)

T milk obese T milk obese

P(obese G, milk=T) P(obese G)

Figure  Effect of conditioning on an observation (left) vs. an intervention (right), following

Pearl ().

On the other hand, we could also model milk-drinking as the result of an exogenous
intervention, i.e., an uncaused variable Imilk with three possible values: T, F, and
None, where the latter represents no intervention. On Meek and Glymour’s ()
account, setting Imilk to T or F has two effects: it sets the probability of milk to  or
 (respectively), and it renders milk independent of G. As a result, the existence of
an intervention Imilk = T is not informative about the probability of the undesirable
outcome obese (Figure , right). (Note that this independence does not follow from
the structure of the causal model depicted, but rather from the definition of the full
joint distribution: see Meek and Glymour, , pp. –.) In this case, the expected
utility of drinking milk is no longer sensitive to the possibility of obesity. Since we
are assuming that drinking milk is intrinsically desirable, we wind up with the causal-
deontic prediction you ought to drink milk.
The difference between an observation and an intervention can be represented
in a visually more perspicuous—but equivalent—way by adopting Pearl’s () do
operator, where an intervention triggers a “surgery” which removes edges from the
causal graph. This is pictured in Figure .
Meek and Glymour’s () approach to modeling the difference between evidential
and causal judgments does not require varying the formula for calculating expected
values. Rather, the difference is that we are conditioning on different events: drinking

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 daniel lassiter

milk, or the existence of an intervention which sets P(milk) = . Since the internal
structure of the probability distribution is not something that is fixed by the Bayesian
theory—like other theories, we treat it as a contextual parameter—the Bayesian theory
makes room for both causal-deontic and evidential-deontic judgments. As a result, the
observation that there is (or at least may be) variation among speakers in intuitions
along these lines is not problematic for the Bayesian theory, and does not support
the conclusion that we should weaken the semantic theory in the ways that Carr and
Charlow suggest.
For the case of binding specifically, it is natural to analyze Charlow’s examples as
follows:
() a. If causal decision theory is right, you ought to drink milk.
b. If evidential decision theory is right, you ought not to drink milk.
While I am strongly inclined to endorse both the antecedent and the consequent
of (a), I also think that (b) is clearly true. Is this a problem? Not at all, if the
antecedent if evidential decision theory is right is interpreted as “If the right way to
model milk-drinking here is to treat it as a caused event, rather than an exogenous
intervention.” If we require the P(·) parameter to conform to that supposition, (b)
will come out true on the Bayesian semantics. For similar reasons, someone with
deontic-evidential judgments should be able to see (a) as true because he can
consider the effects of adopting a probabilistic model in which milk-drinking is treated
as an intervention.

 Variation in Judgments and (Conceivable)

Disagreement
Cariani () and Charlow () argue that the Bayesian semantics cannot make
sense of certain disagreements involving ought. (In Charlow’s case, the argument
is meant to apply more broadly.) How can we understand disagreement about an
ought-claim between two individuals who have the same information and values
(preferences over fully specified possible worlds)? To the extent that a disagreement
is possible, holding values and information fixed, at least one of the parties must not
be employing the EV (·) function defined in § to map probabilities and preferences
over worlds to preferences over propositions. Furthermore, it is implausible that such
disagreements can take place only when one of the two parties has misunderstood
the language; we must allow that a disagreement could occur even when both are
correctly employing their linguistic competence in forming the relevant judgment. So,
the choice of EV (·) as opposed to some other mapping cannot be part of the semantics.
The same argument can be run against any contentful semantics for ought (in a sense
of “contentful” to be made more precise below). So, no contentful semantics for ought
can be correct.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

bayesian semantics 

Charlow’s discussion focuses on causal versus evidential judgments, but I do not

think that these judgments are problematic for the Bayesian theory, for reasons
discussed above (§). In this section I will discuss two further potential empirical
problems: the existence of non-consequentialist deontic judgments and disagreements
involving them, and the conceivability of further patterns of judgment and disagree-
ment involving deontic modals. §. shows that major classes of empirically motivated
non-consequentialist judgments are readily modeled by the Bayesian theory. §.
holds a discussion of patterns of judgment that are ruled out by the Bayesian theory,
notably MaxiMax and MaxiMin: perhaps the conceivability of these systems is a
problem for the Bayesian theory. I will argue that, unless evidence is found that
English speakers actually have judgments of this type, the mere conceivability of such
judgments is not something that we should build a semantic theory around.
It may be useful to flag a distinction between two kinds of disagreements: reflective
theory-laden disagreements among semanticists and/or moral philosophers, and
disagreements about specific cases among non-theorists. Some philosophical work on
this topic seems to take for granted that both kinds of disagreement are semantically
relevant. For example, one of the most frequent criticisms of the Bayesian theory that I
have encountered is that it (supposedly) renders the meta-ethical positions of certain
philosophers analytically false. Along similar lines, Charlow (this volume, chapter )
argues at one point that a certain position is not “palatable” because it entails that “your
philosophical colleagues” have adopted a position that is trivially true (§.). This kind
of disagreement is, in my opinion, not very interesting for an empirically-oriented
theory of meaning. We are not trying to model the full range of conceivable theoretical
positions that someone could take on a topic: the object of interest is the knowledge
that ordinary speakers bring to bear in producing and interpreting language, including
grammatical and inferential features of the language as they understand it. I will
assume in what follows that the only kind of disagreements that are of interest are
those that can or could be motivated by empirical evidence involving non-theorists.
. Non-consequentialist Judgments
One kind of disagreement that has been given extensive empirical motivation involves
consequentialist and non-consequentialist judgments. A good survey is May (),
which includes a discussion of empirical work on five types of issues along these lines:
intentional vs accidental harm; action vs omission; means, force, and agents; means vs
side-effect; and personal vs impersonal harm. While evidence is mixed, non-theorists’
judgments frequently align with the predictions of non-consequentialist theories. Is
this a problem for the Bayesian semantics, as some have suggested?7

7 Note that the claim that it is a problem is not made explicitly by the authors that I am responding to
here. However, I have heard it from numerous philosophers in public and private communications on the
topic, and I suspect that some assumption along these lines is widespread: judging by his choice of title,
even Goble () seems to make it in “Utilitarian deontic logic,” one of the first statements of the Bayesian
semantics.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 daniel lassiter

The answer is “no,” and the reasoning is essentially the same as Lewis’s in his
response to McMichael (Lewis, ). McMichael had accused Lewis of encoding a
problematic utilitarianism into the deontic theory of Counterfactuals (Lewis, , §).
This objection would be valid if Lewis’s theory contained, in addition to a semantic
parameter representing deontic value as a world-ordering, some metasemantic restric-
tion on the kinds of information about a world that can go into determining its value.
But Lewis alleges no such restriction: in principle, any piece of information about
two worlds can play a role in determining their relative goodness. This could include
whether Bobby harmed Jimmy intentionally or accidentally, whether he harmed him
actively or simply allowed him to be harmed, and so on. These are facts about actions,
to be sure, but actions are part of the world, and a semantic theory which has access to
information about the whole world will of course have access to information about
properties of actions that take place in that world. As Lewis (, pp. –) puts
it: “The semantic analysis tells us what is true (at a world) under an ordering. It
modestly declines to choose the proper ordering. That is work for a moralist, not a
semanticist.”
Similarly, the Bayesian theory has no difficulty with the fact that people’s deontic
judgments are frequently sensitive to more fine-grained information about the world
than (for instance) the overall balance of good and evil. No metasemantic restriction
is in place to prevent the moral value function from ranking worlds in which the harm
that Bobby inflicted on Jimmy was intentional below worlds which are identical except
that this harm was accidental. The information that distinguishes these worlds involves
psychological facts about people in them, but nothing in the theory restricts the moral
value function from taking such information into account. If we wanted to encode
(say) a utilitarian ethical theory into the Bayesian semantics sketched above, we would
have to do it by adding a metasemantic restriction which constrains admissible value
functions in this way. Even radically nonconsequentalist theories—those in which the
consequences of actions are never taken into account—can be modeled within the
Bayesian semantics by placing constraints on allowable value functions; all we would
have to do is to supply a value function which ignores the consequences of actions
in choosing how to rank two worlds, and pays attention only to the universalizability
of agents’ actions, their conformance with God’s will, or whatever other feature we
choose to focus on.
There are important limitations to the Bayesian theory’s expressiveness, but these
limitations involve logical relations among sentences. The theory rules out, for exam-
ple, the possibility that φ ∨ ψ can have value greater than that of φ. If God decrees
that worlds are classified according to whether people aim to praise Him and to help
the poor, He still cannot make praying or giving to the poor better than each of
the following individually: praying, and giving to the poor. Such logical predictions
are the main selling point of the Bayesian semantics. Perhaps there is some strange
ethical theory which is not compatible with this prediction; if so, and if we could find
non-theorists whose moral judgments coincided with the predictions of this theory,

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

bayesian semantics 

we would be justified in rejecting the Bayesian semantics. As far as I know, no such

evidence exists.
. Conceivability Arguments and Other Disagreements
So far we have focused on empirical objections; but there is a well-known philo-
sophical argument which—if successful—would allow us to dispense with empirical
evidence and argue from conceivability facts alone. Clearly, a non-theorist could be
implicitly committed to Bayesian reasoning about obligation, in the sense that she
reasons about obligation and related concepts by considering the full range of possible
scenarios, weighing their value against the conditional probability of their occurrence
given various actions. (Indeed, I would hypothesize that this is how nearly everyone
reasons intuitively about obligation, perhaps excluding people with strong theoretical
commitments.) Such a person would be committed to the following judgments about
the standard Miners’ Puzzle and Jackson’s Medicine Puzzle, assuming that information
and values are fixed as described above.
() a. We ought to block neither shaft.
b. The doctor ought to prescribe the safe but mediocre treatment.
But it is conceivable that there could be a person who is implicitly committed to
MaxiMax. Such a person would consider only what could occur in the best possible
worlds in which some action is performed, believing that the action with the greatest
maximum is what ought to be done, regardless of how likely or unlikely it is that
this maximal outcome will be actual if the action is performed. Again assuming that
information and values are fixed as above, a MaxiMax enthusiast would make the
following judgments:
() a. We ought to either block A or block B.
b. The doctor ought to prescribe the risky experimental treatment.
To see how this is a problem, let’s define “contentful” more precisely. Along with
most of the recent semantic literature on the topic, we have been assuming that the
interpretation of ought in context is sensitive to an information state and a “value”
parameter, either a preference order over worlds w or a value function V. (The latter
is a strictly stronger assumption: a value function naturally determines an agreeing
world-ordering via the rule w w s ⇔V(w ) ≥ V(w ).) Since it’s common ground
between me, Charlow, Carr, and Cariani, let’s assume that information states include
probabilistic information in the form of a measure P(·). We also assume that agents
somehow come to form preferences over objects of a higher type—propositions,
or perhaps actions—and that this is what is utilized directly in judgments about
(e.g.) ought. Depending on the theory, these preferences are represented either by
a preference order s over propositions, or by an expected value function EV (·).8

8 I’ve used w where Lewis and Kratzer would use a symbol closer to  w . I find it more natural to use
the former for “at least as good as,” and the latter for “at most as good as”/“at least as bad as.”

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 daniel lassiter

(Again, the latter is stronger: EV (·) naturally determines an agreeing preference order
via φ s φ ⇔ EV (φ ) ≥ EV (φ ).) For the specific case of ought we can be
fairly noncommittal about how this maps into judgments, except to require that a
cooperative speaker in ideal conditions will endorse ought(φ) only if she judges that φ
is deontically better than all relevant alternatives. “Better” is spelled out in the obvious
way using  s or EV (·), depending on the theory. Now we can define what it means for
a semantics for deontic modals to be “contentful”:
A contentful semantics for deontic modals is a semantics which places constraints
on the kinds of probability measures P(·) and preferences over worlds (w or V)
which can be associated with which preferences over propositions (s or EV (·)).
The Bayesian theory is as contentful as can be: its parameters P(·) and V together fully
determine EV (·). So is the Lewis/Kratzer semantics, where s is fully determined by
w , and P(·) is ignored.
A semantics which is not at all contentful would be one which begins with a P(·)
and w , or a P(·) and V, and allows them to be associated with any s or EV (·)
(as appropriate). Such a semantics would allow, for example, that ought judgments
could be associated with a s ordering generated by a MiniMin rule (you ought
to perform the action with the worst possible worst-case outcome), or a random
shuffling of propositions (RandomChoice), or that the proposition-ordering could be
intransitive or cyclic. While we are focusing on the back-and-forth between maximally
contentful theories (such as the Bayesian one) and very weak theories (such as Carr’s
and Charlow’s), there is clearly a lot of logical space between the extremes: it could be
that the semantics constrains but does not fully determine the relationship between
these components.
I don’t know of any direct empirical evidence of variation among non-theorists in
which choice rules are used in forming ought-judgments, holding information and val-
ues fixed. (As discussed above, “values” must be read here to permit value judgments to
include all of the information about a world, including issues of personal vs impersonal
harm, action vs omission, etc.) A prominent class of potential counter-examples here
involves varying attitudes toward risk in ought-judgments. Such variation is well-
motivated empirically in the case of choice behavior, and it’s reasonable to speculate
that it exists in deontic judgment as well (though direct empirical evidence involving
deontic judgment specifically is needed before we can be confident). However, it is

Note that, in Kratzer () theory w is not a contextual parameter, but is determined indirectly by a
modal base f : w → P(W), an ordering source g : w → P(W), and a rule for constructing w from f and
g: relative to an evaluation world w ,

w   w w ⇔ w , w  ∈ f (w ) ∧ {p ∈ g(w )|w ∈ p} ⊇ {p ∈ g(w )|w ∈ p}.
Any choice of f and g will determine a reflexive and transitive order w . In addition, Lewis () proved
that, for any reflexive and transitive order w over a subset of W, Kratzer can choose some f and g which
deliver exactly that w . The theories are thus equally expressive, and the choice to treat w as a parameter
or derive it using Kratzer’s more elaborate method makes no difference for our purposes.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

bayesian semantics 

generally possible to model moderate risk-seeking and risk-averse choice behavior in

expected utility terms, unless it is at the probability-insensitive extremes of MaxiMin
or MaxiMax (Pratt, ; Arrow, , etc.).
That said, there is considerable disagreement among economists about whether
the way that expected utility models capture moderate risk-averse and risk-seeking
behavior is plausible.9 Arguments against the Pratt-Arrow treatment of attitudes
toward risk could perhaps be converted into a convincing reason to weaken the
Bayesian semantics for ought and other deontic items, if they can be given parallel
empirical motivation for deontic judgments specifically. If so, the response should not
be to adopt a totally noncommittal semantics; instead, the obvious next theoretical
move is to look for inspiration in (for example) Kahneman and Tversky’s ()
Prospect Theory, which makes strong empirical predictions but does not absorb
attitudes toward risk into the value function. Alternatively, a demonstration that some
individuals’ deontic judgments are totally insensitive to non-categorical changes in
probability would do the trick—though I doubt seriously that the type of evidence
which would prove the latter will be found.10
Lacking convincing empirical evidence for inter-individual variation in choice
rules, we are forced to argue from the conceivability of certain kinds of disagreements
to substantive conclusions about the actual semantics, in Moorean fashion. This much
I am willing to concede: it is conceivable that someone could reason deontically using
MaxiMax, MaxiMin, or various other choice rules. Here is my attempt to spell out the
assumptions that would be necessary for this observation to constitute a compelling
argument against any specific semantic proposal. I do this in rather tedious detail,
since there are a number of places where one could object.11
. Sub-argument:
a. If we can conceive of two competent speakers forming contradictory ought-
involving judgments as a result of employing different choice rules—holding
information and values fixed—then it is possible for two competent speakers
to form contradictory ought-involving judgments as a result of employing
different choice rules.
b. We can conceive of two competent speakers employing different choice rules
in forming ought judgments, holding information and values fixed.

9 Rabin () is a prominent critique, but see also Chetty and Szeidl () for evidence that expected
utility models are not as badly off as Rabin’s critique might make it appear.
10 There is interesting work suggesting intra-individual variation in judgments triggered by cognitive
load manipulations (Greene, Morelli, Lowenberg, Nystrom, and Cohen, ). I assume here that we are
modeling reflective judgments formed under minimal task demands. There is also evidence of framing
effects in moral judgments, with logically equivalent scenarios being judged as worse when their negative
aspects are highlighted (Sunstein, ; Kern and Chugh, ). This effect, known as “moral loss aversion,”
is a different phenomenon from risk-aversion, and is usually thought to be a judgmental bias rather than an
effect that should be predicted by a decision theory.
11 Throughout the argument, “people,” “speakers,” etc. are implicitly restricted to non-theorists; otherwise
several of the premises would be quite implausible, for reasons discussed above.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 daniel lassiter

c. So, it is possible [in the sense relevant to (a)] for two competent speakers to
form contradictory ought judgments as a result of employing different choice
rules. [From (a) and (b)]
. Sub-argument:
a. If two competent speakers form contradictory ought judgments as a result of
employing different choice rules, this is a substantive disagreement about an
ought judgment which is generated by the use of different choice rules.
b. In the sense of possible relevant to (a), If p then q entails If it is possible that
p then it is possible that q.
c. So, if it is possible [in the sense relevant to (a)] for two competent speakers
to form contradictory ought-involving judgments as a result of employing
different choice rules, then it is possible [in the sense relevant to (a)] for
there to be a substantive disagreement about an ought judgment which is
generated by the use of different choice rules. [From (a) and (b)]
. So, it is possible [in the sense relevant to (a)] for there to be a substantive
disagreement about what ought to happen which is generated by the use of
different choice rules. [From (c) and (c)]
. Disagreements are either substantive or verbal. Substantive disagreements
require that the parties involved use all of the lexical items with the same meaning
when describing the issue about which they are disagreeing. If there is any
difference in the meanings expressed by the parties to the disagreement when
they describe the issue under contention, the disagreement is merely verbal.
. So, it is possible to vary the choice rule used in the generation of an ought
judgment without varying its meaning. [From () and ()]
The argument appears to be valid, though it falls short of establishing that ought is
not contentful, i.e., that any choice rule is compatible with the meaning of ought. For
that, we would need to assume that, for any definable combination of P(·), w , s —
or P(·), V(·), EV (·), depending on the theory we’re working with—we can conceive
of a disagreement where one party is employing this combination. This would have to
include combinations determined by crazy rules like MiniMin and RandomChoice, as
well as combinations where s is intransitive or cyclic. I find it more or less impossible
to imagine these things, though perhaps this difficulty can be attributed to other
factors (say, to constraints imposed by an intuitive theory of others’ psychology).
The real problems with this argument are broader: a number of the premises are
debatable at best. Consider premise (a), for example. The plausibility of this premise
depends on what sense of “possible” is relevant. Epistemic possibility is a non-starter—
it’s easy to conceive of things that are epistemically impossible, like the moon being
made of green cheese. Metaphysical possibility is a plausible candidate for the intended
sense. If “possible” is resolved in this way, then premise (a) amounts to the assumption
that conceivability entails metaphysical possibility—a hotly disputed point, to put it
mildly. For example, Putnam () argues that it is conceivable that water is not

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

bayesian semantics 

H O, but—given that water is in fact H O—it is not metaphysically possible that it

is not. Likewise, if conceivability entails possibility then we can learn that zombies
are possible, and that a materialist philosophy of mind is false, all without leaving the
armchair. (See Yablo, ; Chalmers, ; and the papers in Gendler and Hawthorne,
. Depending on who you ask, these consequences are either interesting discoveries
or reductios of the assumption that conceivability entails possibility.) Moving away
from heady metaphysical debates and back to homely linguistic questions, the broader
assumption that meaning facts can be reliably discerned by exploring our intuitions
about the conceivability of various kinds of disagreements seems dubious. It assumes
that our introspective abilities in this domain far exceed what seems to hold generally
for our linguistic knowledge and most other aspects of our cognitive lives.
A second problem occurs in (a) and at various other points in the argument, involv-
ing the notion of a “competent” speaker. The hedge “competent” occurs constantly in
the philosophical literature on disagreement, and for good reason: linguistic variation
is a constant presence in our lives, and it is crucial, and very difficult, to factor out its
effects in making judgments involving meaning and disagreement. But it is hard to see
how the vague notion of a “competent” speaker can be cashed out in enough detail to
do the work that is being demanded of it here. Consider again the issue of whether
tomatoes are fruit. Many modern English speakers are of the opinion that this is a
scientific fact—“Tomatoes are not a vegetable, they are really a fruit.” This insistence
clearly implies that people who consider tomatoes a vegetable are incompetent, either
with respect to tomatoes, with respect to fruit, or both. Others deny that tomatoes
are fruit, in full knowledge of the relevant botanical facts. The latter group would
presumably judge the former group as incompetent with respect to the use of one
or both of these expressions. (As a quick web search will reveal, both groups are
well represented and vocal.) It is hard to imagine that we could resolve this issue by
consulting intuitions: each group will intuit that the other is mistaken.
To make matters worse, there is a third possibility: perhaps the two groups are
speaking subtly different languages, and both are correct relative to their own lan-
guages. This is clearly what the Supreme Court thought in its decision in Nix v.
Hedden: the judgment was carefully worded to indicate that both uses were fully
legitimate in their respective realms, and that the only question was which usage
should be employed when interpreting a specific law. But it would surely be difficult
to convince the two parties to the tomato-classification debate that they are not really
disagreeing. In other words, in some cases there probably is simply no fact of the matter
about whether a disagreement is substantive or verbal in nature: the vague notion of
“speaking the same language” can bear only so much theoretical weight. Even if this is
wrong, and there is some hidden fact of the matter in this particular case, it is surely
not one that we as theorists can reliably intuit.
The same holds, I suspect, for the case of the imagined ought-disagreements
between speakers who are implicitly committed to different varieties of decision
theory. If we were to encounter such a disagreement in the flesh, it would not be clear

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 daniel lassiter

how to classify it: perhaps one party is right and the other is wrong, or perhaps the two
parties are speaking languages which differ ever so subtly. Intuitions about this case
are not clear, and we cannot expect to rely on them in adjudicating the issue. Perhaps,
as is plausible in the tomato case, there would not even be a fact of the matter.
The “tomato” and “athlete” examples problematize premises (a) and () of the
argument about ought-involving disagreements as well. Both of these premises pre-
suppose that there is a clear distinction to be drawn between substantive and verbal
disagreements. The assumption that such boundaries exist is independently troubling
in light of the general problematization of the analytic/synthetic distinction since
Quine (). The claim that a disagreement must be either one of meaning or one
of fact presupposes that a such a distinction exists. Worse, the assumption that we
can reliably intuit which is at play in a given case presupposes that we can reliably
distinguish meaning facts from facts about the world. But even if we suppose, contra
Quine, that the analytic/synthetic distinction is real and theoretically useful, the
history of philosophical work which uses or criticizes the concept of analyticity
demonstrates clearly that our intuitions about how to classify particular cases are not
generally reliable.
In sum, the argument from the conceivability of certain kinds of disagreement is
very generic in form, and has several problematic aspects that have nothing particular
to do with the semantics of deontic modals. As a result, this argument does not cause
me to lose sleep. On the other hand, solid empirical evidence of deontic choice-rule
disagreements among non-theorists would constitute a strong argument for relaxing
the semantics to make room for whatever patterns are actually observed. Alternatively,
if there were separable populations of speakers who consistently used ought in different
ways along this parameter, we might be justified in supposing that their languages
differ subtly in the meaning of ought. In my estimation, the most fruitful direction for
research on deontic modals will be oriented toward empirical issues of this type.
This response still leaves important philosophical questions unanswered. What is
the relationship between (e.g.) the Bayesian semantics and the notion of analyticity?
Isn’t the Bayesian semantics implausible because, if correct, it would be an analytic
truth that it is correct, and so obvious and trivial? In the back-and-forth between
philosophers of language and practicing lexical semanticists, there seems to be a
fair bit of confusion on this point at a larger scale. For example, Fodor and Lepore
() criticize Cruse’s () classic, descriptively-oriented lexical semantics text-
book, which takes much time describing relationships among lexical items such as
entailment, antonymy, and synonymy. In their estimation, this work is hopelessly
confused because it confuses the meaning of an expression like hot with the inferences
that people tend to draw from it (e.g., that something that is hot is not cold). They
formulate extensive related criticisms of Pustejovsky’s () Generative Lexicon the-
ory, a richly structured theory of lexical semantics intended to account for systematic
meaning shifts which occur predictably in certain environments and across entire
classes of items. (For example, book, CD, and file can each be interpreted both as a

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

bayesian semantics 

physical object in burn the book/CD/file and as a container of information in read

the book/file and listen to the CD; the latter, but not the former, is compatible with
the object being instantiated as data on an iPad.) Fodor and Lepore argue that, like
Cruse, Pustejovsky makes the grave error of encoding worldly facts such as “books can
contain information” as pieces of information in the lexicon. This is confused because it
is possible to imagine these facts being otherwise, but information that is in the lexicon
could not be otherwise (e.g., Fodor and Lepore, , p. ). Fodor and Lepore argue
that the lexicon should contain nothing at all except statements like these: “cat” denotes
cat, “book” denotes book, and (we can extrapolate) “ought” denotes ought.12
Reading Fodor and Lepore () alongside Pustejovsky’s () reply makes it
clear how deep the misunderstanding goes here: lexical semanticists are simply not
playing the game that (at least some) philosophers think that they are. Pustejovsky
points out that Fodor and Lepore wish to throw the theory out on purely a priori
grounds. They are motivated by the correct observation that “meaning is dirty,” but
have nothing to put in its place: the very real empirical generalizations that Pustejovsky
is trying to capture simply vanish if their counter-proposals are adopted. Meaning is
kept pristine, but we are either re-categorizing these facts as “conceptual” or giving
up on the project of explaining them. Pustejovsky’s argument has a further striking
component: he claims that “the human linguistic capacity is a reflection of our ability
to categorize and represent the world in particular ways” (Pustejovsky, , p. )—
in other words, a theory of language and of the lexicon in particular is intimately
connected with a theory of concepts, reasoning, and cognition at large. No notion of
analyticity is presupposed, and no analytic truths are supposed to follow from a theory
of the lexicon.
This is a conception of meaning that is eminently compatible with Quine’s ()
skepticism about analytic truth. Taking Quine’s arguments seriously does not entail
that we should stop talking about meanings and facts. (Quine himself did plenty of
both.) Rather, it means that we should not expect the distinction between them to
bear theoretical weight: apparently analytic truths are simply those which are relatively
central in our theories of the world, relatively stable across members of a speech
communities, and—for psychological reasons—relatively difficult to imagine being
otherwise. For linguists, then, an appropriately Quinean response to the unclarity
of analytic truth is to keep doing what we have been doing: we go on theorizing
about meaning, keeping in mind that our subjects’ meaning-related beliefs may not

12 Interestingly, Cariani () and Charlow () both give ought more semantic structure than this,
and Cariani endorses a probabilistic semantics for epistemic likely along the lines of Yalcin (, b);
Lassiter (, ). It’s not obvious to me that either of these positions could survive the philosophical
weapon of mass destruction conjured up by the disagreement argument: it’s possible to imagine people
disagreeing about pretty much anything that you might think to call “meaning.” If we take the disagreement
argument to its logical conclusion, I suspect that we will eventually be compelled to Fodor and Lepore’s
bare-bones concept of lexical semantics for all of these expressions.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 daniel lassiter

be sharply distinguished from everything else that they believe about the world.13 As
I understand it, this is how lexical semantics is generally done. On the other hand,
this methodology does not presuppose the non-existence of analyticity either: it could
turn out that some information really is so central that it cannot be doubted and/or
could not turn out to be otherwise. It could even turn out, as Fodor and Lepore argue,
that the boundary between language and the rest of cognition is drawn so narrowly
that all of the interesting work that has been done in lexical semantics is really about
concepts and reasoning. For linguists such as Pustejovsky (, ), arguments like
this, unaccompanied by testable empirical predictions, are rightly met with a shrug.
My attitude toward the debate around deontic modals and decision theoretic
concepts—“How much is in the lexicon, as opposed to being part of a concept of
obligation or a theory of moral reasoning?”—is similar. I’m not convinced that there
are meaningful boundaries to be drawn between meaning, concepts, and reasoning.
But maybe there are; if so, then we cannot expect to locate them by examining
intuitions about the conceivability of various scenarios. What is needed is a predictive
theory of ought that is embedded in a good compositional semantics for the whole
language, and responsive to the theory of syntax, pragmatics, and other aspects of
cognition. We can then judge the result in terms of the ability of the larger whole
to predict the empirical evidence available to us, as well as its overall coherence and
simplicity.
It might well turn out that there is a determinate answer to the question of what
precisely the meaning of ought is, in the heavy-duty, analytic-truth-generating sense.
If so, perhaps ought has a skeletal meaning that is normatively noncommittal, as
Charlow and Carr argue. If so, much of the work that we do under the rubric of “lexical
semantics” is probably better categorized as an investigation into the concept ought,
or into English speakers’ ought-involving reasoning habits, or something else of this
sort. This would be an interesting discovery, though not an especially troubling one
(except, perhaps, when searching for an institutional affiliation). In any case, we will
never know until a testable empirical difference is located.

 Further Objections
Cariani () gives two further arguments against a Bayesian semantics: one involv-
ing attitude ascriptions and one involving dominance reasoning and zero-probability
events. In this section I consider these in turn, arguing again that they are special cases
of very general issues rather than specific problems for the Bayesian semantics. Given
that these puzzles arise independently, and that their general solutions can be expected

13 Obviously linguists working in the Montagovian tradition would be hard pressed to follow Quine’s
verificationism, behaviorism, skepticism about modal notions, etc. What I mean is that this attitude fits
in with Quine’s general philosophy of science, and in particular his belief that philosophy’s job is to aid
scientific inquiry, not to dictate methods to it.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

bayesian semantics 

to extend to the case at hand, they fail to support Cariani’s conclusion that a much more
powerful semantic theory of deontic modality is needed.

. Attitudes
Consider again the MaxiMax enthusiast in the disagreement above; let’s name him
“Carl”. If asked what the doctor in the Medicine scenario ought to do, Carl would
answer that he ought to prescribe the experimental treatment which brings along
with it a very high chance of death. This is because the best possible outcome of
this treatment is a full recovery, which is better than the best possible outcome of the
safe but mediocre treatment. We could move probabilities around as much as we like,
making it .% certain that the patient will die under the experimental treatment,
but crazy Carl won’t budge: unless the experimental treatment leads to death with
probability , his beliefs about what the doctor ought to do are sensitive only to the
most optimistic outcome, the tiny chance that the patient will have a full recovery.
Long-suffering Martha shares Carl’s information and values, but has come to
the reasonable judgment that the doctor ought to prescribe the safe but mediocre
treatment. Having endured Carl’s arguments for several pages now, she might turn
to you and say in exasperation:
() Carl thinks that the doctor ought to prescribe the risky experimental treat-
ment. (But he’s wrong.)
Cariani (, §.) poses an interesting puzzle about how to make sense of Martha’s
statement. Even allowing that Martha’s use of ought is well described by the Bayesian
semantics of §, what are the truth-conditions of () in her mouth? () is no good,
for example.
() Carl thinks that, of the options available to the doctor, the action with the
highest expected moral value is to prescribe the risky experimental treatment.
If Carl knows what expected value is, he will presumably judge The action with the
highest expected moral value is to prescribe the risky experimental treatment to be false—
after all, he and Martha share all of their values and information. So () is false, but
Martha’s statement in () is clearly true; so () is not a good paraphrase of ().
Unfortunately, as Cariani points out, a straightforward, standard operator treatment
of thinks will end up giving us () as the interpretation of ().
What does () mean? A prima facie reasonable interpretation might be ():
() Carl thinks that, of the options available to the doctor, the action whose best
possible outcome is morally best is the risky experimental treatment.
But it would be a mistake to enrich the semantics by allowing thinks to shift some
additional parameter, and then conclude that the problem has been solved. This could
generate (), but it isn’t enough: Carl and Martha would both endorse the content of

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 daniel lassiter

the embedded clause in (), too. The real problem is that Carl and Martha don’t agree
on how ought should be interpreted, and neither () nor () captures this. The full
interpretation of () has a normative, metalinguistic component:
() Carl thinks that, of the options available to the doctor, the action whose best
possible outcome is morally best is the risky experimental treatment, and
that ought should be used to describe the action whose best possible outcome
is morally best.
This kind of interpretative flexibility in attitude reports is needed on independent
grounds. Consider (yet again) the tomato controversy in Nix v. Hedden. Port Con-
troller Hedden (the defendant) could have described the position of the plaintiffs
like this:
() The Nix family think that tomatoes are not fruit. (But they’re wrong: tomatoes
are fruit, and I was correct to make them pay the import duty.)
Let’s allow that, since Hedden classifies tomatoes as fruit, the interpretation function
relevant to interpreting his utterances maps tomatoes to a subset of the things that it
maps fruit to.14 Still, the intended interpretation of () is not “the Nix family think
that everything that is a tomato in Nix’s English is not a fruit in Nix’s English.” Rather,
the issue is about how “fruit” should be interpreted. () means roughly ():
() The Nix family think that, for the purpose of interpreting the Tariff Act,
objects in the extension of “tomatoes” should be treated as being outside the
extension of “fruit”. But in fact they are wrong: for the purpose of interpreting
the Tariff Act, “fruit” should be construed so that everything in the extension
of “tomatoes” falls into its extension as well.
Accounting for this interpretation in the context of a broad compositional semantics
for attitude verbs is an interesting challenge that I won’t try to pursue here. (See
Stalnaker, ; Shan,  for relevant discussion and some directions.) But the
lesson, once again, is that the problems attributed to the Bayesian semantics are
very general problems indeed. A full theory of how () receives the normative,
metalinguistic interpretation paraphrased in () will—I predict—also explain how
() is interpreted as ().
. Dominance Reasoning and Zero-probability Events
Cariani brings up another objection to the Bayesian semantics, drawn from a recent
manuscript by Hájek (ms.) on the foundations of probability. Hájek imagines himself
throwing an infinitely thin dart at the [, ] interval, with a uniform distribution over

14 If you like the alternative interpretation and are convinced that this is metaphysically impossible,
modify the example so that one of the Nixes is the speaker, describing Hedden’s position that tomatoes
are fruit. The argument then goes through as in the main text.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

bayesian semantics 

Table  Outcomes of possible choices in the dart-throwing

example.
Choose Option  Choose Option 

Dart lands on r lousy state world peace

Dart does not land on r lousy state lousy state

the possible landing points. It must land somewhere in this interval, but—for any
particular value x in this interval—the probability that the dart will land on x is .
Indeed, the probability that the dart’s landing point will fall within X is zero for many
large, even infinite, sets X ⊂ [, ]. (For example, X might be the set of all rational
numbers in this interval.) Hájek uses this example to bring out several important
problems for probability theory, such as the ratio definition of conditional probability
and the multiplication definition of independence.
The dart example also brings up an interesting puzzle about Bayesian decision
theory. The latter encodes a certain highly intuitive form of dominance reasoning as
a theorem, but only when possible-but-zero-probability events are excluded. Imagine
that Alan Hájek is about to throw an infinitely thin dart at the [, ] interval. You have to
choose between Option  and Option . If you choose Option , the world will continue
in its current lousy condition no matter where the dart lands. If you choose Option ,
there will be eternal world peace if the dart lands on a certain real value r, and nothing
will happen otherwise.
Intuitively, it is clear that you ought to choose Option . But suppose that world
peace has moral value , ,  and the current state has moral value −. The
expected value of option  is − ×  + , ,  ×  = −. The expected value
of Option  is, well, − ×  + , ,  ×  = −. The zero-probability event of
the dart’s landing on r has no effect on the calculation, and so, against intuition, it is
morally indifferent which option you choose.
Here again, the objection is not really specific to the Bayesian semantics, but rather
is a very general issue for the foundations of Bayesian decision theory, and one which
we may expect optimistically to resolve by adopting whatever patch ends up being the
right one for the theory in general. But we can also say a bit more about the attempt
to use an argument of this form to problematize a proposal about the semantics of
English.
One general objection to this style of argument is that infinitely thin darts are
extremely remote from our everyday experience. If there were an infinitely thin dart
in the room, Hájek wouldn’t be able to find it, or to hold onto it long enough to
throw it. But I personally find it flatly impossible to imagine such a thing: my mental
image is just of a really, really thin dart (which wouldn’t be enough to get the puzzle
going). These observations may sound silly, but they point to a serious issue: we
have no reason to expect that linguistic or moral intuitions that have been forged
in our ordinary lives should be a reliable guide to what would or ought to happen

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 daniel lassiter

in scenarios that are physically impossible, extremely difficult to conceptualize, or

otherwise very remote. Hájek’s example, while doubtless an interesting issue for the
foundations of decision theory, is not obviously relevant to linguistic semantics if the
latter is construed as part of the study of human cognition. (Someone who doesn’t
share the latter assumption might still find the dart puzzle compelling, but then why
care about linguistic intuitions at all?)
I personally find the objection from remoteness compelling, but there is a more
directly linguistic reason to put the issue to the side: English quite generally seems
to enforce a flexible but non-zero granularity in the interpretation of numerical
expressions (cf. Lewis, a; Krifka, ; Sauerland and Stateva, ; Bastiaanse,
). For example, x is  feet  inches tall is an intuitively true statement for many
actual individuals x; but surely no human being is exactly  feet  inches tall if “exactly”
is construed to rule out deviations of a millionth of a nanometer. This means that, to a
rough approximation, non-mathematical English never talks directly about the event
of the dart landing at a real value r. Instead, we talk about events such as the dart
falling within a certain range g >  of r, and—in the example at hand—this is not a
probability zero event. (Compare Lassiter’s (b) response to Yalcin’s () related
objection involving the epistemic adjective possible.)
In general, we have to distinguish carefully between the partially formalized, often
counter-intuitive semantics of the blend of English and math that we use in (e.g.) mea-
sure theory and the semantics of the homely, flexible, imprecise and vague language
that we are modeling when we are trying to give a semantics for English. We can use
the former to model the latter, but we cannot assume that they are equally fine-grained.
Interestingly, this response seems to be consistent with a comment in Hájek ms. stating
the conditions under which the “regularity” assumption that is required for dominance
reasoning to go through can be maintained (emphasis in original):
On the one hand, we can apparently make the set of contents of an agent’s thoughts as big as we like;
on the other hand, we restrict the attitudes that she can bear to these contents—they can achieve
only a certain fineness of grain.

I take this to suggest that dominance reasoning can be upheld if the language being
modeled (non-mathematical English) is restricted so that it is unable to talk about the
troublesome zero-probability sets. It seems plausible that this is the case.

 Conclusion
The Bayesian semantics does a good job of regimenting grammatical and inferential
properties of deontic modals which are troublesome for many theories. In particular, it
automatically encodes many of the information-sensitivity facts which have generated
so much excitement in the recent literature, and gives a clear semantics for widespread
gradability among deontic modals. Since ought is the most widely discussed deontic

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

bayesian semantics 

item in the philosophical literature, we have focused on it here as well, but there is a
great deal to be said about may, must, should, and about good and its modified forms
as good as, (much) better than, etc.: see Lassiter ().
A number of arguments have been made in recent literature to the effect that the
Bayesian semantics is not expressive enough to deal with the interaction of ought with
conditionals and attitudes, with certain kinds of hypothetical disagreements scenarios,
and with situations involving possible events that have zero probability. In each case,
the proposed solution is to weaken so that the predictions of the Bayesian semantics
can emerge as one special case among many.
While it might turn out eventually that there is a good theoretical or empirical
reason to weaken the semantics, I have argued that none of the arguments surveyed
here is compelling. We can explain the problematic data in terms of independent
factors—some formal, some philosophical, and some involving known but unsolved
linguistic problems.15 Until a compelling empirical objection is found, the Bayesian
semantics continues to be a viable theory of the semantics of deontic modals, and—
in my opinion—the best theory available in its balance of empirical coverage and
theoretical restrictiveness.

References
Arrow, Kenneth J () The theory of risk aversion. Helsinki: Yrjö Jahnssonin Säätiö.
Bastiaanse, Harald () The rationality of round interpretation. In Sauerland, U., Nouwen, R.,
van Rooij, R., and Schmitz, H.-C (eds.) Vagueness in Communication. pp. –. Heidelberg:
Springer.
Bierwisch, Manfred () The semantics of gradation. In Bierwisch, M. and Lang, E (eds.)
Dimensional Adjectives: Grammatical Structure and Conceptual Interpretation. pp. –.
Berlin: Springer-Verlag.
Cariani, Fabrizio () The semantics of ‘ought’ and the unity of modal discourse. Ph.D.
dissertation. University of California at Berkeley.
Cariani, Fabrizio () “Ought” and resolution semantics. Noûs. . , pp. –.
Cariani, Fabrizio () Attitudes, deontics and semantic neutrality. Pacific Philosophical Quar-
terly. . , pp. –.
Cariani, Fabrizio () Deontic modals and probabilities: One theory to rule them all? In
Charlow, N. and Chrisman, M (eds.) Deontic Modals. Oxford: Oxford University Press.
Cariani, F., Kaufmann, S., and Kaufmann, S () Deliberative modality under epistemic
uncertainty. Linguistics and Philosophy. , pp. –.
Carr, Jennifer () Deontic modals without decision theory. In Chemla, E., Homer, V., and
Winterstein, G (eds.) Proceedings of Sinn und Bedeutung . pp. –.

15 Note, by the way, that this conclusion should not be construed as an argument against Charlow’s
expressivism. Expressivism is a plausible thesis about deontic and epistemic modals, and probably other
aspects of language as well. I don’t think that being an expressivist requires one to reject a Bayesian semantics,
though. See Yalcin (a) for some relevant discussion.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 daniel lassiter

Chalmers, David () The conscious mind: In search of a fundamental theory. Oxford: Oxford
University Press.
Charlow, Nate () What we know and what to do. Synthese. . , pp. –.
Charlow, Nate () Decision theory: Yes! Truth conditions: No! In Charlow, Nate and
Chrisman, Matthew (eds.) Deontic Modals. Oxford: Oxford University Press.
Chetty, Raj and Szeidl, Adam () Consumption commitments and risk preferences. The
Quarterly Journal of Economics. . , pp. –.
Cruse, D. Alan () Lexical semantics. Cambridge: Cambridge University Press.
Fintel, K. von () The best we can (expect to) get? Challenges to the classic semantics
for deontic modals. Paper presented at the Central meeting of the American Philosophi-
cal Association, February . Available from: http://mit.edu/fintel/fintel--apa-ought.pdf.
[Accessed  Nov .]
Fodor, Jerry A. and Lepore, Ernie () The emptiness of the lexicon: Reflections on James
Pustejovsky’s The Generative Lexicon. Linguistic Inquiry. . , pp. –.
Gendler, Tamar Szabó and Hawthorne, John (eds.) () Conceivability and Possibility. Oxford:
Clarendon Press.
Gibbard, Allan and Harper William L () Counterfactuals and two kinds of expected utility.
In Harper, R., Stalnaker, R., and Pearce, G. (eds.) Ifs: Conditionals, belief, decision, chance, and
time. Dordrecht: D. Reidel.
Goble, Lou () Utilitarian deontic logic. Philosophical Studies. . , pp. –.
Gopnik, Alison and Schultz Laura (eds.) () Causal learning: Psychology, philosophy, and
computation. Oxford: Oxford University Press.
Greene, Joshua D., Morelli, Sylvia A., Lowenberg, Kelly, Nystrom, Leigh E., and Cohen,
Jonathan D () Cognitive load selectively interferes with utilitarian moral judgment.
Cognition. . , pp. –.
Hájek, Alan Staying regular. Manuscript, Australian National University.
Hare, R.M () Some alleged differences between imperatives and indicatives. Mind. . ,
p. .
Hay, Jennifer, Kennedy, Chris, and Levin, Beth () Scalar structure underlies telicity in
‘degree achievements’. In Mathhews, T. and Strolovitch, D. (eds), Semantics and Linguistic
Theory . Ithaca, NY: CLC Publications. pp. –.
Jackson, Frank () Decision-theoretic consequentialism and the nearest and dearest objec-
tion. Ethics. . , pp. –.
Jackson, Frank and Pargetter, Robert () Oughts, options, and actualism. The Philosophical
Review. . , pp. –.
Jeffrey, Richard C (a) Ethics and the logic of decision. Journal of Philosophy. . ,
pp. –.
Jeffrey, Richard C (b) The logic of decision. Chicago, IL: University of Chicago Press.
Kahneman, Daniel and Tversky, Amos () Prospect theory: An analysis of decision under
risk. Econometrica: Journal of the Econometric Society. . , pp. –.
Kamp, Hans () Two theories about adjectives. In Keenan, E. (ed.) Formal semantics of
natural language. pp. –. Cambridge: Cambridge University Press.
Kennedy, Chris () Vagueness and grammar: The semantics of relative and absolute gradable
adjectives. Linguistics and Philosophy. . , pp. –.
Kennedy, Chris and McNally, Louise () Scale structure, degree modification, and the
semantics of gradable predicates. Language. . , pp. –.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

bayesian semantics 

Kern, Mary C. and Chugh, Dolly () Bounded Ethicality. The Perils of Loss Framing.
Psychological Science. . , pp. –.
Kolodny, Niko and MacFarlane, John () Ifs and oughts. Journal of Philosophy. . ,
pp. –.
Krantz, David H., Luce, R. D., Suppes, Patrick, and Tversky, Amos () Foundations of
Measurement. New York, NY: Academic Press.
Kratzer, Angelika () The notional category of modality. In Eikmeyer and Rieser (eds.) Words,
Worlds, and Contexts. pp. –. Berlin: de Gruyter.
Kratzer, Angelika () Modality. In von Stechow and Wunderlich (eds.) Semantics: An inter-
national handbook of contemporary research. Berlin: de Gruyter.
Krifka, Manfred () Approximate interpretation of number words: A case for strategic
communication. In Bouma, G., Krämer, I., and Zwarts, J (eds.) Cognitive foundations of inter-
pretation. pp. –. Amsterdam: Koninklijke Nederlandse Akademie van Wetenschapen.
Kripke, Saul () Naming and necessity. Cambridge, MA: Harvard University Press.
Lassiter, Daniel () Gradable epistemic modals, probability, and scale structure. In Li Nan
and Lutz David (eds.) Semantics & Linguistic Theory (SALT). , pp. –.
Lassiter, Daniel () Measurement and Modality: The Scalar Basis of Modal Semantics. Ph.D
dissertation. New York University.
Lassiter, Daniel (a) Adjectival modification and gradation. In Lappin S. and Fox C (eds.)
Handbook of Contemporary Semantic Theory. Malden: Wiley-Blackwell. nd edition.
Lassiter, Daniel (b) Epistemic comparison, models of uncertainty, and the disjunction
puzzle. Journal of Semantics. . , pp. –.
Lassiter, Daniel (c) Graded modality. Manuscript: Stanford University.
Lassiter, Daniel (d) Modality, scale structure, and scalar reasoning. Pacific Philosophical
Quarterly. . . pp. –.
Lassiter, Daniel () Graded Modality: Qualitative and Quantitative Perspectives. Oxford:
Oxford University Press (in press).
Lewis, David () Counterfactuals. Cambridge MA: Harvard University Press.
Lewis, David () Reply to McMichael. Analysis. . , p. .
Lewis, David () Scorekeeping in a language game. Journal of Philosophical Logic. . ,
pp. –. doi: ./BF.
Lewis, David () Causal decision theory. Australasian Journal of Philosophy. . , pp. –.
Ludlow, Peter () Living words: Meaning underdetermination and the dynamic lexicon.
Oxford: Oxford University Press.
May, Joshua () Moral judgment and deontology: Empirical developments. Philosophy
Compass. . , pp. –.
Meek, Christopher and Glymour, Clark () Conditioning and intervening. The British
Journal for the Philosophy of Science. , pp. –.
Morzycki, Marcin () Adjectival extremeness: Degree modification and contextually
restricted scales. Natural Language & Linguistic Theory. . , pp. –.
Morzycki, Marcin () Modification. Cambridge: Cambridge University Press.
Pearl, Judea () Causality: Models, reasoning and inference. Cambridge: Cambridge Univer-
sity Press.
Plunkett, David and Sundell, Tim () Disagreement and the semantics of normative and
evaluative terms. Philosopher’s Imprint. , pp. –.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 daniel lassiter

Portner, Paul and Rubinstein, Aynat () Extreme and non-extreme deontic modals. In
Chrisman M. and Charlow N. (eds.) Deontic Modals. Oxford: Oxford University Press.
Pratt, John W () Risk aversion in the small and in the large. Econometrica: Journal of the
Econometric Society. . /, pp. –.
Pustejovsky, James () The Generative Lexicon. Cambridge, MA: MIT Press.
Pustejovsky, James () Generativity and explanation in semantics: A reply to Fodor and
Lepore. Linguistic Inquiry. . , pp. –.
Putnam, Hilary () The meaning of “meaning”. In Mind, Language and Reality: Philosophical
papers. Volume . Cambridge: Cambridge University Press.
Quine, Willard van Orman () Two dogmas of empiricism. The Philosophical Review. . ,
pp. –.
Rabin, Matthew () Risk aversion and expected-utility theory: A calibration theorem.
Econometrica. . , pp. –.
Regan, Donald () Utilitarianism and Co-operation. Oxford: Oxford University Press.
Ross, Alf () Imperatives and logic. Philosophy of Science. . , pp. –.
Rotstein, Carmen and Winter, Yoad () Total adjectives vs. partial adjectives: Scale structure
and higher-order modifiers. Natural Language Semantics. . , pp. –.
Sassoon, Galit W () A typology of multidimensional adjectives. Journal of Semantics. . ,
pp. –.
Sauerland, Uli and Stateva, Penka () Scalar vs. epistemic vagueness: Evidence from approx-
imators. In Gibson M. and Friedman T. (eds.) Semantics and Linguistic Theory. . Ithaca, NY:
CLC Publications.
Shan, Chung-chieh () The character of quotation. Linguistics and Philosophy. . ,
pp. –.
Sloman, Steven A () Causal models: How we think about the world and its alternatives.
Oxford: Oxford University Press.
Sloman, Steven A. and Lagnado, David A. () Do we “do”? Cognitive Science. . , pp. –.
Stalnaker, Robert () On considering a possible world as actual. In Aristotelian Society
Supplementary Volume, Vol. , , pp. –.
Sunstein, Cass R () Moral heuristics. Behavioral and brain sciences. . , pp. –.
Wedgwood, Ralph () The Meaning of “Ought”. In Shafer-Landau R., (ed.) Oxford studies
in Metaethics. , pp. –. Oxford: Oxford University Press.
Wedgwood, Ralph () Objective and subjective ought. In Charlow, Nate and Chrisman,
Matthew (eds.) Deontic Modality. Oxford: Oxford University Press.
Yablo, Stephen () Is conceivability a guide to possibility? Philosophy and Phenomenological
Research. . , pp. –.
Yalcin, Seth () Epistemic modals. Mind. . , pp. –.
Yalcin, Seth () Probability Operators. Philosophy Compass. . , pp. –.
Yalcin, Seth (a) Bayesian expressivism. Proceedings of the Aristotelian Society. ,
pp. –.
Yalcin, Seth (b) Context probabilism. In Aloni, M., Kimmelman, V., Roelofsen, F., Sassoon,
G. W., Schulz, K., and Westera M. (eds.) Logic, language and meaning. Vol. , Lecture Notes
in Computer Science. pp. –. Heidelberg: Springer.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i


Contextualism about Deontic
Conditionals
Aaron Bronfman and J. L. Dowell

If you are a semanticist, how best to understand the formal semantics of modal
expressions is an issue that wears its interest on its sleeve. The issue, however, is of
broader interest and importance to those concerned with other debates. One main
task of metaethics, for example, is to understand ordinary moral and, more broadly,
normative and evaluative discourse. Identifying the best semantics and pragmatics of
deontic modal expressions in particular would make an important contribution to
metaethicists’ understanding of such discourse.
Recently, some philosophers of language and linguists have wondered whether there
are any expressions that require a relativist’s distinctive treatment. Contextualists about
some expression E hold that the contribution E makes to the determination of the
truth-conditions of utterances containing E varies from context of use to context
of use. Relativists about E, in contrast, hold that it makes an invariant contribution
to the determination of truth-conditions on any occasion of use. Unlike standard
semantic invariantists, however, relativists hold that the circumstances of evaluation
that determine the truth of utterances containing E are more fine-grained than the
standardly assumed possible worlds. What in addition is needed to determine a truth-
value, for the relativist, depends upon what E is. In the case of deontic modals, some
relativists argue that that addition is a body of information: Deontic modal sentences
are true or false at world, information pairs.1
Assessing the prospects for relativism about deontic modals is crucial to answering
the larger question of whether relativism is a viable research program. Central among
the cases that are thought to motivate relativism are cases involving deontic modals
whose truth requires that they are sensitive to a body of information in some way.2
Parfit’s miners scenario is such a case. A significant point of contention is whether
a contextualist can account for our judgments about deontic modals in that case.

1 MacFarlane (). 2 Kolodny and MacFarlane ().

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 aaron bronfman and j. l. dowell

A challenge for the contextualist, then, is to identify a contextualist account of modal

expressions that fits with those judgments and is independently plausible.3
Here our goal is to help identify the contextualist’s most worthy competitor to
relativism. Recently, some philosophers of language and linguists have argued that
while there are contextualist-friendly semantic theories of deontic modals that fit
with the relativist’s challenge data, the best such theories are not Lewis-Kratzer-style
semantic theories.4 If correct, this would be important: It would show that the theory
that has for many years enjoyed the status of the default view of modals in English and
other languages is in need of revision.
Here we defend the default view by showing how a Kratzer-style semantics is able
to make available readings of the relevant utterances that fit with the pretheoretical
judgments opponents purport it cannot fully capture. Having established this, we turn
to considering the more theoretical grounds proponents have offered for preferring
their rival contextualist views. Here the question is to what extent such grounds
favor semantic over what Korta and Perry call “near-side pragmatic”5 explanations
of our judgments. In particular, we argue that our favored readings figure in near-
side pragmatic explanations of those judgments that possess the methodological and
theoretical advantages of systematicity and unity at least as well as, if not to a greater
extent than, those of opponents who argue for their revised semantic theories on the
basis of these advantages.6 In this way, our discussion is a case-study contribution to
the larger debate among philosophers of language and linguists over when to prefer
semantic over such pragmatic explanations.7

3 Some have argued that the best data for relativism about deontic and epistemic modals is given not by
data at issue here, but by data involving disagreement. For replies to the contention that the contextualist
cannot accommodate the disagreement data, see Dowell (, ).
4 To be clear, we are not suggesting that a Lewis-Kratzer-style formal semantics for modals can only be
given a contextualist construal. For all we say here, there is a relativist interpretation of that semantics that
does as well as the contextualist one we shall defend. Here we aim to assess the claim some contextualists
have defended that a Lewis-Kratzer-style semantics under a contextualist interpretation cannot fit with the
data we discuss here. (For discussion of one way to implement relativism in a Kratzer-style framework, see
Egan ().)
5 “Pragmatics deals with utterances, by which we will mean specific events, the intentional acts of
speakers at times and places, typically involving language. Logic and semantics traditionally deal with
properties of types of expressions, and not with properties that differ from token to token, or use to use. . . The
utterances philosophers usually take as paradigmatic are assertive uses of declarative sentences, where the
speaker says something. Near-side pragmatics is concerned with the nature of certain facts that are relevant
to determining what is said. Far-side pragmatics is focused on what happens beyond saying: what speech
acts are performed in or by saying what is said, or what implicatures. . . are generated by saying what is said.”
(Korta and Perry, : pp. –)
6 Along with Dowell (, , ) and Bronfman and Dowell (forthcoming), this discussion thus
contributes to the larger project of defending a Kratzer-style, flexible contextualist semantics for modal
expressions, supplemented with a near-side pragmatic account of how it is that contexts provide the
parameter values needed to secure appropriate readings.
7 For another such case study, see von Fintel () and Gillies () who each argue for a dynamic
semantic theory of counterfactuals on the grounds that the standard, Lewis-Stalnaker semantics is unable
to explain our judgments about the felicity of Sobel and reverse Sobel sequences. See also Moss (), who
defends the standard semantics by providing a near-side pragmatic explanation of our judgments in such
cases.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

contextualism about deontic conditionals 

Below, we first explain the basic features of Kratzer’s semantics for modal expres-
sions, including conditionals. Then we consider and meet the challenge cases in turn.
Finally, we show how our readings are able to meet any remaining objections and pose
a few of our own to the rival theories thought to be motivated by these challenges to
Kratzer’s canonical view.

 Kratzer-Style Contextualism
On Angelika Kratzer’s canonical semantics, modal expressions are semantically neu-
tral; they make a single contribution to the determination of a proposition on
every occasion of use. What modulates the type of modality expressed—teleological,
bouletic, deontic, epistemic or alethic—is the context of use.8 The plausibility of the
resulting view lies in part in its ability to provide simple and unified explanations
of a wide range of language use. Together with broad cross-linguistic support, the
simplicity of Kratzer’s semantics earns its status as the default view.9
Central to a Kratzer-style semantics is its treatment of modal expressions as
quantifiers over possibilities. Typically, those domains of quantification are restricted.
Restrictions not represented explicitly in the linguistic material are provided as a
function of the context of utterance. The contextual supplementation is twofold. First,
context determines a modal base, f , a function from a world of evaluation, w, to a set
of worlds, ∩f (w), the modal background.
Modal bases may be either epistemic or circumstantial. An epistemic modal base
is a function f that takes a world of evaluation w and returns the set of worlds
consistent with the body of information in w that has some property or properties.
Which properties are relevant is determined by which f is contextually selected; for
example, that function may take the information that has the property of being the
speaker’s at a designated time t in w as an argument and give us the set of worlds
compatible with that information. In principle, context might select any number of
different f s.
A circumstantial modal base is a value for f that takes a world of evaluation
as an argument and delivers a set of worlds circumstantially alike in particular
respects. Here, too, what makes a circumstance among the relevant ones at a world
of evaluation will depend upon which f is contextually selected; for example, a
particular value for f may make circumstances that determine causal relations
between actions and outcomes at the world of evaluation relevant. The modal
background in that case would be the set of worlds alike with respect to those
circumstances.
A second source of contextual supplementation is an ordering source, a function g
from a world of evaluation w to a ranking of worlds in the modal background. Which

8 Kratzer () and ().

9 Some parts of our exposition draw on Bronfman and Dowell (forthcoming).

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 aaron bronfman and j. l. dowell

features of a world w in ∩f (w) give it its relative ranking depends upon the value
for g. For example, g might rank w depending upon how well some salient agent
acts in accordance with the reasons she has or the obligations that apply to her in
w. Or it might rank w in terms of how well it approximates some impartial ideal. The
highest ranked or best such worlds make up the modal’s domain. “Ought,” the modal
of concern here, functions as a universal quantifier over its domain: “ought ϕ” comes
out true at a context-world pair just in case all of the best worlds as determined by that
context and world are ϕ-worlds.10
Since part of what is at issue in the puzzle cases here is the plausibility of a full
Kratzer-style account of deontic conditionals, we’ll need her account of the indicative
conditional on the table. On her semantics for that conditional, the function of the
antecedent is to restrict the domain of a modal in the consequent.11 This is, at least
typically, a covert necessity modal. So, a conditional of the form “if ϕ, then ” where
“” does not itself contain a modal, has the structure if ϕ, must . To see whether
such a conditional is true, we see whether every (relevant) ϕ-world is also a -world.
If so, then the conditional is true. Here we follow Kratzer in assuming that the covert
modal is an epistemic necessity modal: The relevant ϕ-worlds are the ϕ-worlds that
are compatible with some contextually determined body of information12
There are a few options for combining this account of the indicative conditional
with her semantics for modal expressions generally. For all the deontic conditionals
we discuss, we’ll adopt the view that a covert epistemic necessity modal takes scope
over the deontic modal.13 There are a couple of different readings14 of the whole
conditional, if ϕ, must[ought ] that may result, depending upon the context. In
all cases, we assume the antecedent retains its usual semantic function of restricting
the domain of the covert modal. A bit more formally, treating w as the world of
evaluation, the covert modal’s domain will be worlds, w , each of which is a ϕ-world.
To be true, the conditional then requires that the deontic modal is true at each of the
worlds w . To determine this, the deontic modal requires values for f (w ) and g(w ).
These, we’ll argue, are determined flexibly as a function of the context of utterance.
Together they’ll determine a set of worlds w that make up the deontic modal’s
domain.

10 Here we simplify aspects of Kratzer (a) and () to avoid introducing complexities of that
account not at issue here. In particular, we adopt the Limit Assumption, and we ignore the issue of how best
to mark the apparent distinction between “must” and “ought”. Cariani, Kaufmann, and Kaufmann ()
and Carr () also adopt these simplifications in expositing their own views.
11 In some cases not at issue here, the antecedent may restrict a quantifier elsewhere in the sentence, as
in “Always, if a man buys a horse, he pays cash for it” (Kratzer, b).
12 Kratzer, b.
13 For discussion of this type of view and some reasons for adopting it, see, for example, Carr (),
von Fintel (), von Fintel and Iatridou (), Frank (), Geurts (), and Kratzer ().
14 By a “reading,” we simply mean a way a listener might reasonably interpret what’s said in an utterance.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

contextualism about deontic conditionals 

 Miner Variations
. The MINERS objection
The famous miners scenario is one case thought to pose a challenge for a Kratzer-style
framework. Here is Niko Kolodny and John MacFarlane’s characterization (MINERS):
Ten miners are trapped either in shaft A or in shaft B, but we do not know which. Flood waters
threaten to flood the shafts. We have enough sandbags to block one shaft, but not both. If we
block one shaft, all the water will go into the other shaft, killing any miners inside it. If we block
neither shaft, both shafts will fill halfway with water, and just one miner, the lowest in the shaft,
will be killed. (, p. )

In a recent paper, Fabrizio Cariani, Magdalena Kaufmann, and Stefan Kaufmann

argue that there is no way for a Kratzer-style semantics to fit with all of our pretheo-
retical judgments about MINERS.15 Among these judgments are the following truth-
assessments:
NEITHER: We (they) ought to block neither shaft. True.
IF-A: If the miners are in A, we (they) ought to block A. True.
IF-B: If the miners are in B, we (they) ought to block B. True.
Not only must a theory fit with these truth-assessments, they suggest, it must render
them all “true on the deliberative reading of ‘ought’ ” (, p. , footnote ). Hence,
they argue that any Kratzer-friendly readings either fail to accommodate all three
truth-assessments or else do not all qualify as deliberative readings.
In order to evaluate the Cariani, Kaufmann, and Kaufmann claim that a Kratzer-
style semantics cannot fit with our truth-assessments of these sentences under a
“deliberative reading,” we’ll need an understanding of what it takes for a use of “ought”
to get a “deliberative reading” in their sense. They offer a few suggestions. One is by
contrast with an “objective” reading, which seems to be relative to circumstances,
known and unknown (, p. ). Another suggestion: “Deliberative modality” is
“the particular flavor of modality in play in practical deliberations” (, p. ).
Finally, it is the kind “exhibited by”
ARTICLE: We ought to read that article.
ARTICLE, they suggest, is “easily understood as suggesting that reading that article is
the thing to do” (, p. ).
One difficulty for seeing what exactly they have in mind is that the last two
suggestions for how to understand “deliberative modality” or “deliberative reading”
do not contrast with an objective reading in MINERS as their first suggestion holds.
To see this, notice that when NEITHER, IF-A, and IF-B figure in someone’s reasoning
about what to do in MINERS, they will each exhibit deliberative modality in the second

15 See also Charlow ().

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 aaron bronfman and j. l. dowell

sense. This is compatible with the conditionals receiving objective readings. As we’ll
argue below, it’s easy to identify readings on which utterances of all three sentences
are true with the conditionals receiving objective readings. So, if to be deliberative is
to figure in practical reasoning, it will be easy to see how all three can be true under a
“deliberative reading.”
Moreover, insofar as conditionals can suggest that some action is the “thing to do”,
objective readings of the conditionals may also be deliberative in their third sense, e.g.
in contexts in which agents make it clear that what they objectively ought to do settles
the question of what is the thing to do. Since we believe that the contrast with objective
readings is what is most important for understanding the objection to Kratzer that
Cariani, Kaufmann, and Kaufmann seem to have in mind, we will try to improve on
their suggestions in a way that preserves this.

. Deliberative readings: subjective and advisability

As we’ll see, improving on their formulation is a bit tricky. Mark Schroeder’s ()
discussion of the deliberative “sense” of “ought” offers an initial starting point.
According to Schroeder, such an “ought” exhibits five hallmarks: First, the deliberative
sense “matters directly for advice.” While the objective “ought” may figure in good
deliberation about what to do, the deliberative sense settles the question of what it
is advisable to do. Second, the deliberative sense is “the right kind of thing to close
deliberation,” to “settle the question of what to do.” Third, one is “accountable” for
doing as one ought in this sense; failure to do so leaves one open to “legitimate
criticism.” Fourth, it’s the sense of “ought” constrained by what one can do and, finally,
it’s the sense “more closely connected” to the notion of obligation, albeit imperfectly.
While we agree that these are features of “ought”s that figure in practical reasoning,
it will be important for later discussion to note that there are cases in which there is no
single ought-claim that bears all five hallmarks. Instead, there will be distinct ought-
claims, each of which possesses a different, proper subset of these features. In these
cases, acting so as to comply with one such claim is incompatible with complying with
the other.
Eavesdropper scenarios provide good illustrations. Here’s an example from Mac-
Farlane (): Suppose you are deciding whether you ought to bet on Blue Blazer or
Exploder, two horses in an upcoming race. You know that, in the past, Blue Blazer has
proven itself the faster horse. In light of this you conclude,
BLAZER: I ought to bet on Blue Blazer.
Suppose, though, that, unbeknownst to you, I am eavesdropping on your conversation
from behind a bush. Unlike you, I know that today Blue Blazer will be suffering from
the effects of a drug. MacFarlane holds that here “it makes sense for me to think that
you are wrong, and to say,”
EXPLODER: “No, you ought to bet on Exploder” (p. ).

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

contextualism about deontic conditionals 

Let this be a case in which you do not and could not learn that Blue Blazer has
been drugged prior to placing your bet and so you go ahead and bet on that horse.
Have you done as you ought, in Schroeder’s deliberative sense? No doubt you are not
subject to legitimate criticism for betting as you do; in this sense, you have done as
you ought. But you have not done what it would be advisable for you to do. It is
not advisable for you to bet on a drugged horse. Here your utterance bears some of
the hallmarks of Schroeder’s “deliberative sense” of “ought” and my utterance bears
others. MacFarlane has characterized these “ought”s of advice as in between a so-
called “subjective” “ought”, which is tied to information within a deliberating agent’s
epistemic reach, and an objective “ought,” which is not information-sensitive. Such
“in-between” “ought”s are central to his case for relativism. Below we show how all
three readings can be made available within a Kratzer-style framework. Seeing how
this is so will be important for seeing how a Kratzer-style semantics can fit with the
pretheoretical judgments of ordinary speakers for the full range of MINERS cases.
. Kratzer-friendly readings for miner variations
Since our view is contextualist, which reading a deontic modal sentence receives will
be determined as a function of the context of utterance. This means that the best data
for testing theories will be speakers’ judgments about a series of variations on the
basic MINERS scenario, each of which fills out the conversational context in a slightly
different way. As we’ll see, which reading will be most natural for our sentences will
depend upon which version of the scenario is under consideration. One important
dimension along which MINERS scenarios may vary is in whether deliberating agents
expect to receive more information about the location of the miners prior to the time
at which they need to act. Call “EXPECTATION-KNOW” some scenario in which
deliberating agents know they will learn the location of the miners prior to that time.
Call “EXPECTATION-MIGHT” some scenario in which they know they might, but
also might not, learn their location (learning and not learning their location are equally
likely). Let “EXPECT-NOT” be a case in which agents know they won’t learn more.
We consider the following judgments to constitute theoretically neutral data:
NEITHER sounds bad—indeed, clearly false—in EXPECTATION-KNOW, sounds
unwarranted in EXPECTATION-MIGHT, and sounds fine—indeed, clearly true—in
EXPECT-NOT. IF-A and IF-B can each sound fine in any of these cases. We also accept
the Cariani, Kaufmann, and Kaufmann claim that NEITHER, IF-A, and IF-B may all
be true as uttered in the course of a single conversation among agents deliberating
about what to do in MINERS.
One of their central motivations for positing a more complex semantics16 for
deontic modals along with a novel semantic rule for deontic conditionals is their claim
that no Kratzer-style semantics can fit with the full range of this data. Our next task
is to show that this is not so by showing how a Kratzer-style semantics can secure

16 For the sense in which their semantics is more complex, see footnote .

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 aaron bronfman and j. l. dowell

readings that fit with these judgments. We do this in several steps. First, we identify
readings and contexts that make NEITHER, IF-A, and IF-B all true. Then we show
how they can all be assertible in the course of a single piece of deliberation. Recall that
Cariani, Kaufmann, and Kaufmann claim that part of the data is that they are all true
under a “deliberative reading” (, p. , footnote ). As we saw, to test this, we
need to identify what it takes for a reading of a modal to be “deliberative.” We now
have a few overlapping, but distinct senses of “deliberative reading.” Following one of
their suggestions, we have the view that a deliberative reading is any reading of a modal
sentence such that, under that reading, it is properly assertible in the course of a single
piece of practical deliberation. We also, as suggested by Schroeder’s hallmarks, have
two more specific types of deliberative reading: a subjective reading, which is tied to an
agent’s available information, and an advisability reading, which is connected to what
it is advisable for a deliberating agent to do (perhaps tied to an advisor’s information).
Here we show how the Kratzerian can accommodate true readings under each of these
senses of “deliberative reading.”
We begin with NEITHER. Our reading for NEITHER will be the same for all
of our cases, but for the sake of concreteness, we’ll focus on the context in which
it sounds best, EXPECT-NOT. Here we think NEITHER receives what we’ll call a
“subjective” reading. Subjective readings are information-sensitive, where the relevant
information is, very roughly, the information a relevant agent has at the time of
action, t.17 Informally, NEITHER would seem to express the proposition that blocking
neither is, of the actions available to the deliberating agents, deontically ideal, in light
of the information they have at the time that action is necessary.18
More formally, context will select an f that maps a world of evaluation w onto the
set of worlds w like w with respect to the laws and circumstances up until the time of
action t, where we assume that this will hold fixed the options agents have in w at t.
Various issues arise in characterizing an agent’s options. We simply adopt one workable
model with the following features. (i) An option of an agent is represented as the
proposition that she performs some physical or mental action or intentional inaction.
(ii) For something to count as an agent’s option, the agent must know that she is able to
perform it and know how to implement it.19 Thus, for example, “S selects the winning
lottery number” can fail to be one of S’s options, even though, for each number, n,

17 More generally, subjective readings can be sensitive to bodies of information available at a world of
evaluation by some designated time relevant for the ranking of the agent’s options. The information need
not be limited to the information the agent actually has; it may also include information she could or should
have gathered, or information within her “epistemic reach.”
18 Some might see NEITHER as making a claim, not about deontic ideality, but rather about some
particular conception of deontic ideality, such as maximizing expected utility. While we do not rule this
out as a possible reading, we find sentences such as NEITHER to be most naturally understood as making
claims that can be the objects of dispute between, for example, consequentialists and non-consequentialists,
and so we see them as invoking the thinner notion of deontic ideality.
19 Or perhaps an agent’s options are fixed not by what she knows but, more broadly, by what she is in a
position to know she is able to perform by t.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

contextualism about deontic conditionals 

“S selects n” may be one of S’s options. (iii) Options are assumed to be as fine-grained
as possible. Thus “S walks” would not qualify as one of S’s options, but “S walks straight
ahead, slowly, while chewing gum. . . ” could. Because each option specifies the agent’s
behavior so precisely that she does not have any further flexibility in how to act, options
are mutually exclusive.20 For simplicity, we will treat blocking neither as such a fine-
grained option, although doing this is not necessary to our account here. (iv) Options
are assumed to take place over some fixed time period, which may vary with context.
To simplify, we assume the relevant options (e.g. blocking neither) can be performed
instantaneously: perhaps one must now irrevocably decide whether to block a shaft or
none. This helps to put aside complexities such as starting to block one shaft and then
switching to the other.
In the case of NEITHER, context will select a value for g that maps w onto the set of
worlds w in which agents perform that action, ϕ, of their options, O, that is deontically
ideal, given the information our agents have at t in w.21 Since in EXPECT-NOT that
action is blocking neither, NEITHER comes out true. Under this reading, NEITHER
is unwarranted as a conclusion of practical deliberation in EXPECTATION-MIGHT
and false in EXPECTATION-KNOW. This pattern fits precisely with our pretheoretical
judgments about these cases.
Having provided a plausible Kratzer-style reading for NEITHER, we turn now to
identifying plausible readings of IF-A and IF-B, showing how they are assertible in the
course of a single piece of deliberation. So, what should the Kratzerian say about IF-A
and IF-B in our MINERS cases? The issue here is a bit complex as these sentences have
available objective, subjective, and advisability readings in some of these cases.
Start with the objective readings in EXPECT-NOT. There are a couple of ways of
filling out the conversational context of EXPECT-NOT to secure felicitous objective
readings of these conditionals in the course of a single piece of deliberation. Imagine
that agents have not yet arrived at the place in their deliberations in which they realize
they will not learn the location of the miners by t. In that case, they may think that
what they objectively ought to do may settle their deliberative question. IF-A and
IF-B may then represent their thinking about what might objectively be the case.
Alternatively, IF-A and IF-B on an objective reading might each articulate part of their
understanding of the case. They might, for example, play such a role in a conversation
such as the following:
Emma: Ok. Here’s the situation: The miners are all trapped in either shaft A or shaft
B. Only if we use all of our sandbags to block the shaft they’re in, will we save all the
miners.

20 Here we may simplify by assuming options to be as fine-grained as possible because any Kratzer-style
semantics validates Inheritance, so settling which fine-grained options one ought to perform will settle
which coarse-grained options one ought to perform. (The principle of Inheritance holds that if ϕ entails ,
then ought ϕ entails ought .)
21 Recall that we’re trying to keep this reading simple and reader-friendly. For a more general formula-
tion of subjective readings, see footnote .

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 aaron bronfman and j. l. dowell

Lily: I see. So, if they’re all in A, we should block A and if they’re all in B, we should
block B.
Emma: Right. Unfortunately, we’re not going to be able to figure out where they are
by the time we have to decide what to do.
This might be the beginning of a conversation that eventuates in their deliberative
conclusion, expressed by NEITHER.
Finally, each conditional IF-A and IF-B may express a lament about the tragedy of
the situation, even after they have concluded that blocking neither is what they should
do. It would figure naturally, for example, in a conversation such as this one:
Emily: All things considered, we’re going to have to block neither shaft. It’s such a
tragedy that we don’t know where the miners are!
Lily: What’s so bad about that?
Emily: Because, if they’re in A, we should block A. And if they’re in B, we should
block B!
Representing our objective readings for IF-A and IF-B more formally requires
saying a bit about our treatment of deontic conditionals. Recall that we treat such
conditionals as doubly modalized: Each contains a covert, epistemic modal scoped
over the overt, deontic modal. On our objective reading, then, the antecedent of IF-A
restricts the domain of a covert, epistemic modal to the epistemically possible worlds
w in which the miners are in A. The modal background for the overt, deontic modal
will be the set of worlds alike with respect to the relevant circumstances in w , including
that the miners are in A. The value for g(w ) will be the worlds in which agents perform
that action, of those available to them, which is deontically ideal in light of the relevant
circumstances in w . These will all be worlds in which agents block A. So, IF-A comes
out true. Similar considerations will make IF-B come out true.
We have been discussing EXPECT-NOT, in which the conditionals may be felic-
itously uttered, though less practically useful than in EXPECTATION-MIGHT and
EXPECTATION-KNOW. Their practical use, under objective readings, improves in
those latter two cases. This won’t make those readings deliberative in either of the two
senses we identified from Schroeder’s hallmarks, what we’re calling the “subjective
sense” or the “sense of advisability.” But it will suffice to make them deliberative in
the Cariani, Kaufmann, and Kaufmann sense we’re focusing on, namely the sense of
playing a role in a single piece of deliberation.
So far, we’ve focused on Kratzer-friendly objective readings of IF-A and IF-
B that can play a deliberative role. In addition, there’s a second, information-
sensitive, Kratzer-friendly reading available, with which the features of EXPECT-NOT,
EXPECTATION-KNOW, and EXPECTATION-MIGHT are also compatible. As we’ll
see, this reading may be thought of as an advisability reading, one that is “in-between”
the subjective and objective readings. Crucially, we think these cases do not mandate
an advisability reading, rather than an objective reading, of these conditionals; indeed,

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

contextualism about deontic conditionals 

for EXPECTATION-KNOW, subjective readings are also available. Here we merely use
these conditionals to illustrate how advisability readings are available within a Kratzer-
style framework, as well as to show that these conditionals may receive deliberative
readings in one of the senses suggested by Schroeder’s hallmarks. Below we will
consider cases in which subjective and objective readings are unavailable, as evidence
that sometimes modals naturally take advisability readings.22
On an advisability reading of IF-A and IF-B, in addition to its usual semantic
function of updating the value for f for the covert modal, the antecedent serves to
pragmatically indicate a value for the deontic modal’s g parameter. In contrast to the
objective reading, here g(w ) will rank each world w in the deontic modal’s modal
background in terms of the deontic ideality, in light of some body of information,
of the action agents perform in w . Which information is relevant? We suggest it is
a hypothetical body of information consisting of the information deliberating agents
have at t in w together with information specifying where the miners are located in w .
Thus if the miners are located in A at w , the additional information specifies that they
are in A; if they are located in B at w , the additional information specifies that they are
in B. Since the semantic function of the antecedent of IF-A is to restrict the domain of
the covert modal to worlds in which the miners are in A, and these are the worlds at
which the overt deontic modal will be evaluated, the information getting added to the
agent’s information for the purposes of ranking their options is the information that
the miners are in A.
Adding this information to the relevant body, we’re suggesting, is a pragmatic
function of the antecedent. Since the objective reading is also available in our MINERS
cases, we are not suggesting that this reading is forced. Rather, the conversational
context permits the antecedent to play this pragmatic role.
How exactly might it play this role? One idea is that playing this role is suggested by
the relevance of the conditional for conversational purposes. In each of our MINERS
scenarios, agents are deliberating about what to do. Uttering the conditionals is to help
settle this practical question. Agents do not know the location of the miners in any of
these scenarios. In EXPECTATION-KNOW, they know they will come to know their
location. In that case, what would be deontically ideal in light of information agents are
in a position to have by t updated with information about where the miners are located
is highly relevant; indeed, it reflects the epistemic position they expect to be in. This
suffices to make uttering each of the conditionals, on this reading, highly relevant. In
EXPECTATION-MIGHT, agents know they may or may not learn the miners’ location
by t. In this case, too, IF-A and IF-B under the advisability reading will be relevant to
their practical deliberation.
Having discussed NEITHER, IF-A, and IF-B, we now turn to two additional
sentences from Cariani, Kaufmann, and Kaufmann:

22 Here and elsewhere, in offering these readings, we do not claim that no other readings are possible.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 aaron bronfman and j. l. dowell

IF-STILL: If the miners are in A, we (they) still ought to block neither.

EXISTS: We (they) ought to block the shaft the miners are in.23
What are our pretheoretical judgments about utterances of these sentences in our
MINERS cases? Like NEITHER, IF-STILL sounds bad in EXPECTATION-KNOW, not
much better in EXPECTATION-MIGHT, and best in EXPECT-NOT. EXISTS sounds
best in EXPECTATION-KNOW and can sound fine in EXPECTATION-MIGHT. It
may sound false in EXPECT-NOT, unless conversational context makes it a clear
lament (e.g. of their inability to know how to save all ten miners). As with IF-A and
IF-B under objective readings, EXISTS may also serve as an attempt to articulate
a shared understanding of MINERS as a basis for or in the course of deliberation.
This pattern of judgments is fully explained within a Kratzer-style framework with
an objective reading. In EXPECTATION-KNOW, EXISTS is also felicitous under a
subjective reading, where g(w) will rank worlds w in the modal background in terms
of the deontic ideality, in light of the information agents will have at t in w, of the action
performed in w .
What explains our judgments for IF-STILL? We’ve seen that, like NEITHER, IF-
STILL sounds best in EXPECT-NOT and that IF-A and IF-B can also sound fine in
that case. We suggested two available readings for these latter conditionals in all our
MINERS cases: An objective reading and an advisability reading.24 On each of those
readings, the antecedent has a role to play in determining which action is best. If the
antecedent plays such a role for IF-STILL, though, it won’t come out true in EXPECT-
NOT. But it does seem true in that case. Why might the antecedent in IF-STILL not
play such a role in EXPECT-NOT though it does for each of IF-A and IF-B?
The answer rests on the role of “still,” which seems to flag the irrelevance of the
antecedent to the interpretation of the deontic modal.25 We might imagine IF-STILL
uttered in response to an utterance of IF-A. As we’ve seen, IF-A might be uttered in
EXPECT-NOT as part of the agents’ attempt to articulate a shared understanding
of the case. IF-STILL might then flag the irrelevance of the location of the miners
to arriving at their practical conclusion, given that they won’t learn the miners’
location by t.26

23 This is not quite the sentence Cariani, Kaufmann, and Kaufmann consider. Theirs is “there is a shaft
we ought to block” (EXISTS ). We prefer EXISTS on the grounds that, at least in the case of EXPECTATION-
KNOW, it has a clearer action-guiding use than EXISTS .
24 The subjective reading of the conditionals is unwarranted in EXPECTATION-MIGHT and false in
EXPECT-NOT.
25 We seem to find “still” playing this role in other indicative conditionals containing context-sensitive
expressions, e.g.
A: The Sharks might win. A: If he wants to play basketball, he’s not tall.
B: But what if they lose their best player? B: No, even if he wants to play basketball, he’s still tall.
A: Even if they lose their best player, they
still might win.
26 Besides the use of the word “still,” there are other means by which conversational context can signal the
irrelevance of the antecedent to determining which action is best. In such cases it is possible for conditionals

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

contextualism about deontic conditionals 

. Three-shaft version of MINERS

Thinking about how best to understand these deontic conditionals within a Kratzer-
style framework puts us in a position to think about how best to understand a more
difficult type of case. Here’s a three-shaft version of the miners case from Kai von
Fintel:27
Imagine there are three shafts: A, B, and C. We don’t know where the miners are. If we block the
right shaft, all miners are safe. If we do nothing, two miners die. We can blow up Shaft A, which
would of course kill all miners if they are in A, but if they’re not, then blowing up Shaft A and
not blocking either B or C will mean that only one miner dies. So, in our maximally ignorant
information state, we ought to block none of the shafts. In an objective sense, we should block
the shaft the miners are in. Now, consider the following conditional:
[MORE INFORMATION:] If they are not in A, we ought to blow A up.

Von Fintel reports that he can hear this as true. We agree that there are ways
of filling out the conversational context that would make an utterance of MORE
INFORMATION felicitous. A best-case scenario will be one in which agents know
they might learn whether the miners are in A and know they’ll learn nothing else. Call
“EXPECTATION-WHETHER-IN-A-MIGHT” a version of such a case in which agents
believe they have a % chance of learning whether the miners are in A before the time
they need to act. The felicity of utterances such as MORE INFORMATION can seem
puzzling within a Kratzer-style framework, as their felicity can’t be explained by either
a subjective or an objective reading. On the subjective reading we’ve suggested, context
selects a value for f that takes a world of evaluation w to a set of worlds compatible
with facts about the agents’ options in w together with facts about which information
agents in w have at t. Suppose agents don’t learn by t whether the miners are in A.
In that case, the conditional will be false: Blowing A up is not the deontically ideal
action in light of the body of information agents will have at t as, in that case, that
body will leave open the possibility that the miners are in A and blowing up A will
result in all of their deaths. Suppose the miners are in B; in that case the objective
reading comes out false. So what would explain what makes MORE INFORMATION
sound true in EXPECTATION-WHETHER-IN-A-MIGHT? We suggest that MORE
INFORMATION most naturally receives an advisability reading. On such a reading,
recall, the antecedent, in addition to its usual semantic role, pragmatically indicates
an update to the body of information relevant for ranking the worlds in the deontic
modal’s modal background. In MORE INFORMATION, this update adds to the

not containing the word “still” to carry the reading we offer for IF-STILL. For example, if in EXPECT-NOT,
one conversational participant says, “We can’t decide what to do until we know where the miners are, since
if the miners are in A, we should block A,” another may reply, “No, if the miners are in A, we should block
neither shaft. And the same holds if they are in B. We have no way of knowing where the miners are, so
regardless of where they are, we should block neither shaft.” In this case, the conditional “if the miners are
in A, we should block neither shaft” is acceptable because context makes it clear that the antecedent is not
intended as a necessary condition for the assertion of the consequent.
27 Von Fintel ().

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 aaron bronfman and j. l. dowell

information agents have in worlds w in the domain of the covert modal information
about whether the miners are in A. The antecedent guarantees that each such w will
be a world in which the miners are not in A. So the worlds w in the domain of the
deontic modal will be ranked in accordance with whether the agent performs, in w ,
the action, of their options, which is deontically ideal in light of that updated body of
information. In all of the best such worlds, our agents are blowing up A. So, MORE
INFORMATION comes out true.
Von Fintel has offered an alternative explanation of how MORE INFORMA-
TION comes out true. He holds that MORE INFORMATION is “shorthand” or
“enthymematic” for the longer sentence:
“If we learn that they are not in A, we ought to blow A up.” ()
In EXPECTATION-WHETHER-IN-A-MIGHT, this sentence is straightforwardly true
on a subjective reading. In this case, if the agents learn the miners are not in A, they
will learn nothing else. In particular, they will not learn more specific information
pinpointing the location of the miners in B or in C. So if they learn the miners are not
in A, it will be deontically ideal given their information at t for them to blow A up.
Thus if von Fintel is right, there is no need to posit an advisability reading to explain
the truth of MORE INFORMATION.
There is, however, an additional piece of data that our account is better placed to
explain. Begin with the case EXPECTATION-WHETHER-IN-A-MIGHT, in which the
agents have a % chance of learning nothing about the location of the miners, and
a % chance of learning whether or not the miners are in A. Suppose the agents are
aware that if, when the time of action comes, they knowingly allow even one miner to
die unnecessarily, they will be put in jail. In this case, we hear the following sentence
as unwarranted:
If the miners are not in A, we’ll go to jail if we don’t blow A up.
Intuitively, the agents are not warranted in asserting this sentence since there is a %
chance they will remain in their state of complete ignorance about the location of the
miners. If so, they will not be put in jail for failing to blow A up. Similarly, we hear the
following sentence as unwarranted:
MORE INFORMATION JAIL: If the miners are not in A, then we ought to blow A
up and we’ll go to jail if we don’t.
But the proposal that MORE INFORMATION is read enthymematically would tend
to predict that MORE INFORMATION JAIL should be heard as warranted since it
would be read as enthymematic for:
If we learn that the miners are not in A, then we ought to blow A up and we’ll go to
jail if we don’t.
In contrast, our account, on which MORE INFORMATION is not enthymematic, does
not issue such a prediction. This is some evidence in favor of the advisability reading.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

contextualism about deontic conditionals 

 Self-Frustrating Decisions
Jennifer Carr has argued that Kratzer-style contextualism cannot adequately account
for cases of “self-frustrating” decisions. These arise in unusual cases where performing
an action would indicate the existence of reasons against performing that very action.
Carr uses the case DEATH IN DAMASCUS, from Gibbard and Harper (), as an
example:
If you are in the same city as Death tomorrow, then you’ll die. Death has planned to be wherever
he predicts you’ll be, and he’s very reliable in such predictions. Your options are to stay in
Damascus or to go to Aleppo. But, as you know, if you stay in Damascus, then that’s excellent
evidence that Death will already be there. Similarly for going to Aleppo.
(Carr, , p. )

This places you in an unfortunate situation: you expect with high probability that,
whichever decision you make, you will die. Assume you have not made up your mind
about where to go, and you now regard either city as equally likely. Your options
are then symmetric: they offer equally bad prospects. Given this symmetry, we will
assume, with Carr, that both options are permissible: you may go to either city.
Because you may go to Aleppo, we cannot say you should not go to Aleppo. Hence
we have:
ALEPPO: It’s not the case that you should not go to Aleppo.
Consider now the conditional:
IF-ALEPPO: If you go to Aleppo, you should not go to Aleppo.
The reasoning behind IF-ALEPPO goes roughly as follows: If you will in fact go
to Aleppo, then Death is very likely waiting for you in Aleppo, and so you should
not go to Aleppo. IF-ALEPPO uses the antecedent to generate a new set of proba-
bilities for Death’s location, and then evaluates your options in light of those new
probabilities. We will assume, with Carr, that IF-ALEPPO has a true reading along
these lines.
We now consider how Kratzer-style contextualism can account for the truth of
ALEPPO and IF-ALEPPO. Just as for the miners sentences, such a view will hold
that ALEPPO says what it does because context supplies appropriate values for the
parameters f and g. ALEPPO is most plausibly heard as a claim about how it is rational
or reasonable for the agent to act, given her information. It is motivated by the thought
that, given the agent’s information, it is equally reasonable to go to either city. It denies
the claim that not going to Aleppo is the agent’s uniquely most reasonable option. In
other words, it denies the claim that not going to Aleppo is deontically ideal in light of
the agent’s information.
We propose to capture the content of ALEPPO with similar parameter values to
those used for NEITHER. The modal background is circumstantial: f (w) maps a world
to the set of worlds in which the laws and circumstances up through the time of action
t are the same as they are in w. In all these worlds, the agent has the same options as

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 aaron bronfman and j. l. dowell

she does at w: go to Aleppo and stay in Damascus. The only feature that matters to how
a world w in the modal background is ranked is the option the agent chooses in w . In
particular, g(w) ranks worlds w on the basis of whether the option performed in w is
deontically ideal in light of the information that the agent has in w. If all options are
deontically ideal in light of the agent’s information at w, then g(w) treats all worlds in
f (w) as tied-for-best. Otherwise, g(w) divides the worlds into two groups: it ranks as
tied-for-best all w where the option the agent selects in w is deontically ideal in light
of her information at w, and it ranks all other worlds as tied-for-worst.
With these choices for f and g, ALEPPO plausibly comes out true according to
causal decision theory.28 Given that the case and her information are symmetric, she
judges her prospects if she were to go to Aleppo as equivalent to her prospects if she
were to go to Damascus. So a proponent of causal decision theory will hold that the
worlds where the agent stays in Damascus and those where she goes to Aleppo are all
deontically ideal. Hence some of the highest g(w)-ranked worlds in f (w) are worlds
where she goes to Aleppo, and so it is not the case that she should not go to Aleppo.
Just as for IF-A, IF-B, and MORE INFORMATION, we suggest an advisability
reading for IF-ALEPPO.29 Going through our account step by step, we see IF-ALEPPO
as doubly modalized.30 The antecedent you go to Aleppo restricts the higher, covert
epistemic modal, limiting us to what is true in all epistemically possible worlds in
which you go to Aleppo. In itself, this does not do much to help IF-ALEPPO come
out true. If context supplies the same parameter values f and g to the deontic modal
in IF-ALEPPO as it does for ALEPPO, then this conditional will come out false, since
those parameter values make the consequent you should not go to Aleppo false in all
epistemically possible worlds.
This is where the second, pragmatic role for the antecedent comes in. On our view,
the antecedent if you go to Aleppo can indicate to the hearer that different parameter
values are in play. Our suggestion is that the value of f remains unchanged: f (w) still
consists of worlds w that are circumstantially like w. But g is no longer a ranking
of worlds in terms of deontic ideality given the agent’s information. Rather, it is a
ranking of worlds in terms of deontic ideality given a hypothetical body of information
consisting of the agent’s information plus information specifying which city she goes
to in w. In other words, g(w) ranks worlds w on the basis of whether the option
performed in w is deontically ideal in light of the information that the agent has in

28 For the sake of concreteness, we follow Carr in focusing on how a causal decision theorist might
approach this case. Of course, the type of Kratzer-style semantics we’re defending is not committed to the
truth of causal decision theory as opposed to, for example, evidential decision theory.
29 Because the case stipulates only a high likelihood that Death has correctly predicted the agent’s
location, an assertion of IF-ALEPPO would not be warranted on an objective reading. A subjective reading
holds more promise since we might read IF-ALEPPO along the following lines: if the agent will go to Aleppo,
then she will know this at the time of decision, and given this knowledge at that time she subjectively should
not go to Aleppo. But we can put this aside by assuming the agent will make her decision without advance
notice of what she will decide.
30 As Carr () explains, the single-modal view would not predict the truth of IF-ALEPPO, for reasons
derived from the “If p, ought p” problem (Frank, ).

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

contextualism about deontic conditionals 

w plus information specifying which city she goes to in w. Because the agent does not
know where she will go, two kinds of worlds w are epistemically possible for her: those
where this hypothetical body of information includes her going to Aleppo and those
where it includes her staying in Damascus. The antecedent of IF-ALEPPO, however,
restricts the modal background to those epistemically possible worlds where she goes
to Aleppo, and hence to worlds where the hypothetical body of information includes
her going to Aleppo. Given causal decision theory, staying in Damascus is deontically
ideal given such a body of information. The reason is that a body of information that
includes the fact that the agent goes to Aleppo supports with high probability the claim
that Death is in Aleppo, and so assigns the highest causal expected utility to the agent’s
going to Damascus.31
Our analysis of IF-ALEPPO is thus quite similar to our analysis of the advisability
reading of IF-A (and of IF-B and MORE INFORMATION). However, in discussing
the miners case, we noted that IF-A is most clearly relevant to the conversation when
it is possible we will learn the location of the miners before we need to act, since then
it offers advice that may be practically useful. IF-ALEPPO does not appear relevant
in the same way. While IF-A could potentially lead us to block A were we to learn
the miners are in A, IF-ALEPPO cannot lead the agent to stay in Damascus should
she learn that she will go to Aleppo: if the agent genuinely learns that she will go to
Aleppo, then it cannot be true that she will stay in Damascus.
However, we also noted that IF-A can be relevant even if we know we will not
learn the location of the miners before we need to act. It can be part of the process of
articulating our understanding of the situation, or part of a lament about the tragedy
of the situation. IF-ALEPPO appears to have a similar use. In conjunction with “If
you stay in Damascus, you shouldn’t stay in Damascus,” it can help to articulate our
understanding of the facts of the case. It can also express an aspect of the tragedy of
the situation, pointing out that knowledge of either decision would support doing the
opposite.

 Objections: Assessing Semantic versus Near-side

Pragmatic Explanations
We have argued that Kratzer-style semantics can account for the data in MINERS and
in DEATH IN DAMASCUS. Two recent papers, however, appear to claim that this
is impossible. Cariani, Kaufmann, and Kaufmann write that Kratzer-style semantics
incorrectly predicts the falsity of NEITHER in MINERS (, p. ). Similarly, Carr
writes that Kratzer-style semantics, unless modified, cannot “predict or model” the
data about iffy oughts, as exemplified by DEATH IN DAMASCUS (, p. ). On

31 To follow Carr’s discussion and focus on the semantic issue, we put aside some controversy over
whether this is indeed the best interpretation of causal decision theory. See Joyce () for discussion.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 aaron bronfman and j. l. dowell

closer inspection, however, these papers do allow the in-principle possibility of a

Kratzer-style account compatible with the data (, p. , footnote ; , p. ).
They hold, instead, that there are further, more broadly theoretical reasons to reject any
such account. Although neither of these papers considers the kind of solution we have
offered here, it is worth looking at whether any objections they offer raise difficulties
for our proposal.
Cariani, Kaufmann, and Kaufmann argue against one Kratzer-style proposal on the
grounds that it invokes a parameter value that is not “a natural object to be contextually
salient” (, p. ). Similarly, Carr argues against a Kratzer-style proposal on the
grounds that it invokes a parameter value that is “ad hoc” and that may not even be
“coherent” in DEATH IN DAMASCUS (, p. ).
We do not believe the parameter values we have invoked are susceptible to these
objections. In analyzing NEITHER and ALEPPO, we use a circumstantial modal
base f that holds fixed the agent’s options. This is a natural choice in contexts of
deliberation, where the interest is in selecting one option from among those available
to the agent. For the ordering g, we order options by their deontic ideality in light of the
agent’s information. This, too, is salient in deliberative contexts: deliberators care about
whether their action is reasonable in light of the information they have. One might
object that an agent’s primary interest in deliberation is in the goodness or badness
of outcomes, and so hold that the contextually salient ordering must be information
insensitive. Perhaps thinking along these lines, Cariani, Kaufmann, and Kaufmann
consider a Kratzer-style approach that ranks worlds in terms of how many miners are
actually saved. We agree that an information-insensitive ordering can be contextually
salient; indeed, such an ordering is in play in objective ought sentences such as EXISTS.
Nevertheless, when uncertainty is important, it is plausible that deliberative attention
will turn to an information-sensitive ordering.
A similar story holds for the parameter values we have invoked for MORE INFOR-
MATION and IF-ALEPPO. There, we hold that options are ordered in terms of their
deontic ideality given a hypothetical body of information consisting of the agent’s
information plus an additional fact. In accord with our observations above, there
are a variety of contexts where this parameter value may be salient. For example,
if deliberators believe they may learn an additional fact, then planning ahead may
require asking now what their information plus that fact supports doing. Alternatively,
if it would be desirable (though impossible) to know some additional fact, an ordering
that takes this fact into account may become salient, as a way of noting the contrast
between the information we have and the information we wish we had. Such orderings
also express an aspect of the distinctive tragedy of dilemmas such as DEATH IN
DAMASCUS.
Carr raises a more specific problem for a Kratzer-style approach to IF-ALEPPO in
DEATH IN DAMASCUS. She considers a modal background that (i) contains only
worlds where Death is in Aleppo, on the grounds that Death must be in the same
place as you and the antecedent assumes you go to Aleppo, and yet (ii) contains worlds

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

contextualism about deontic conditionals 

where you go to Aleppo and worlds where you go to Damascus. Carr writes that this
proposal “assumes in the very same breath that Death must be in the same place as
you and that he might not be” and concludes such a stipulation may not even be
“coherent” (p. ). Our proposal avoids such an apparent contradiction in a way
similar to the way in which Carr’s own proposal does. We distinguish between the
agent’s options and the agent’s information (actual or hypothetical) about which option
she will choose. The hypothetical information state including the agent’s information
plus the fact that she will go to Aleppo places both the agent and (with high probability)
Death in Aleppo tomorrow. Working with such a hypothetical information state is
consistent with maintaining that the agent has the option to choose either city: in
understanding the conditional, we assume hypothetically that the agent will, out of
her two actual options, choose Aleppo. Of course, the case has the odd feature that the
agent believes that, whichever city she chooses, she will be in the same place as Death,
but also believes that if she were to choose otherwise than she actually does, she would
avoid Death. Still, this feature appears to be coherent and, in any event, is essential to
the description of the case given by Gibbard and Harper.
In arguing against Kolodny and MacFarlane (), Cariani, Kaufmann, and Kauf-
mann point to transparency as a “main advantage” of their own account. They hold
that it is desirable to represent the agents’ priorities “transparently and independently
of the information available to them” (p. ). In MINERS, they represent the priorities
as saving more, rather than fewer, miners; the priorities are thus specified without
reference to the agents’ information. The account we have offered above lacks this
feature: the ranking of worlds given by our parameter value for g is not independent
of the agent’s information.
We do not see this as a disadvantage. Notice first that the ordering of worlds
in play in a given use of a modal expression may depend on many things. For
example, what one must do, legally speaking, may depend on the facts of the case,
the laws themselves, court opinions, conventions of interpretation, and perhaps
on the substance of morality itself. It would be premature to conclude from this
that the semantic value for the legal “must” has argument places for each of these
things.
In the case at hand, it is true that some substantive views about how one should
act are naturally represented as separating the role of information and priorities:
the MaxiMin view invoked by Cariani, Kaufmann, and Kaufmann is an example.
However, other substantive views do not naturally fit this model. For example, a non-
consequentialist view might impose an absolute prohibition on imposing a significant
risk of harm on an innocent person in any circumstance where this can be avoided.
The view allows trivial risks, but rules out risks above a threshold. Such a view, in its
most natural representation, invokes the agent’s information in specifying the relevant
priorities: what the agent should be concerned with is specified partly in terms of the
risks from her point of view. While one could implement such a view consistent with
the letter of transparency (for example, by having most or all of the work done by a

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 aaron bronfman and j. l. dowell

decision rule parameter, with little or no role for priorities), transparency in itself does
not appear desirable here.
Probably the main source of resistance to our proposal will be a broadly theoretical
consideration cited by both Carr and Cariani, Kaufmann, and Kaufmann: systematic-
ity.32 Indeed, Cariani, Kaufmann, and Kaufmann defend transparency at least partly
on the grounds that it secures systematicity: their account derives the ordering of
options via a novel semantic rule that operates on the antecedent of a conditional as
well as information, priorities, and a decision problem. Carr, too, cites systematicity
and derives the ordering of options from a novel semantic rule that operates on the
antecedent of a conditional as well as information and a function from information
to orderings.33 Our account appeals to no such semantic derivation. The parameter
values (e.g. the value for g that refers to deontic ideality in light of an information
state) are supplied by context directly to the consequent of the conditional, rather
than being derived by a semantic rule. In assigning this role to context, rather than
to a novel semantic rule, our account provides a near-side pragmatic explanation for
how the needed truth-conditions get assigned.
Given that, as we show above, a Kratzer-style semantics does make room for
readings that fit with our judgments in the puzzle cases, some more theoretical
consideration is required to decide between the two rival, semantic and pragmatic,
explanations of the cases. Appeal to systematicity would seem to be the right sort of
consideration to play that role. However, a significant worry about the systematicity
argument is whether the rival semantic proposals can, in fact, be developed in a
systematic way. For example, the derivations offered by Cariani, Kaufmann, and
Kaufmann assume the MaxiMin decision rule. While it is easy to generalize the
account to, say, the MaxiMax decision rule, it is not clear how to generalize it to
more sophisticated decision rules (such as expected utility maximization or non-
consequentialist approaches to uncertainty) while preserving the kinds of derivations
cited as evidence of the account’s systematicity. The ability to offer such derivations
only for a small subset of decision rules is not an advantage in systematicity.
But we will argue more directly that considerations of systematicity actually favor
our pragmatic approach. Notice first that, given the variety of readings in principle
available to a modal expression, hearers clearly do have a substantial ability to tell
which parameter values are intended for a given sentence. In particular, readings can
vary in what information or facts are relevant: it can be the information the agent will
in fact have at the time of action, the information she could or should have by that
time, an advisor’s information, or alternatively all the relevant facts (whether known
by anyone or not). A key part of our proposal is to say that the relevant information

32 See also Charlow ().

33 These accounts constitute additions to the Kratzer framework, insofar as the rules and additional
parameters they invoke can be used to recover the Kratzer framework by choosing trivial values for certain
parameters.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

contextualism about deontic conditionals 

may be the agent’s information plus some contextually relevant fact as in MORE
INFORMATION or IF-ALEPPO.
To see that such a proposal isn’t ad hoc, but enjoys independent support, consider
a case from DeRose () regarding “John, who has some symptoms indicative of
cancer, and a ‘filtering’ test which John’s doctor decides to run and which has two
possible results: If the results are ‘negative,’ then cancer is conclusively ruled out; if the
results are ‘positive,’ then John might, but also might not, have cancer: further tests
will have to be run.” (p. ) The test has been run, but the results are not known by
anyone. DeRose notes that Jane, who is familiar with the situation but does not have
the test results, could say:
“I don’t know whether it’s possible that John has cancer.” (p. )
What Jane does not know is, roughly, whether John’s having cancer is compatible with
what she knows combined with the information from the test results. This has a clear
structural similarity to our own account, where the relevant body of information is the
agent’s combined with some additional fact.34
On analogy with DeRose’s case, consider a version of von Fintel’s three-shaft version
of the miners case. The agents in this case know the following: They will not learn
anything about the location of the miners before the time they need to act. There is,
however, a test that can determine whether or not the miners are in A. If they are in
A, it says they are in A. If they are in B or C, it simply says they are not in A. The test
has been run, but the results are not known by anyone. An agent in the case could say:
I don’t know whether we ought to blow A up.
The information relevant to this sentence is not the information the agent will actually
have at the time of action, since that supports doing nothing and so definitively does
not support blowing A up. Nor is it all the possible information about the case, since
this would support blocking whichever shaft the miners are in. Rather it is the agent’s
information about the case combined with information specifying whether or not the
miners are in A.
This example is thus best interpreted with just the kind of parameter value we have
posited. Because “I don’t know whether we ought to blow A up” is not a conditional,
we cannot say here that the parameter value is somehow derived from a semantic rule
operating on an antecedent. Instead, the context, which includes the sentence itself,
supplies the parameter value directly to the modal. Thus, the very mechanism and
type of parameter value we posit for MORE INFORMATION and IF-ALEPPO quite
plausibly operates in this case. Our account of these sentences is systematic insofar as
it simply extends a mechanism we already have reason to accept.
Indeed, our account offers a particularly unified and systematic account of the
following piece of discourse (TEST):

34 See Dowell () for discussion of a version of DeRose’s case in keeping with our present proposal.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 aaron bronfman and j. l. dowell

I don’t know whether we ought to blow A up. If the miners are not in A, we ought
to blow A up. If the miners are in A, we ought not to blow A up.
On our account, context supplies the same parameter values to the modals in all of
these sentences: the ordering g ranks worlds on the basis of what is deontically ideal
in light of the agent’s information plus the fact about whether or not the miners are
in A. Below we will consider in more detail a dilemma that this case raises for our
rivals. But at this point, it is worth noting that the pattern of explanation they offer
for MINERS and DEATH IN DAMASCUS, in which a semantic rule operates on a
contextually supplied parameter value (or values) in conjunction with the antecedent
of a conditional to yield a new ordering, would be overly complex here. On that
pattern, one would end up saying that context supplies one value for g to the first
sentence, and then a different value g  to the two conditionals, but that this value gets
operated on in conjunction with the antecedents to yield a new ordering, possibly the
same as the original g, that gets the truth conditions right.35 Such an account invokes
both a novel semantic rule with additional parameter(s) and an unmotivated context
shift. We posit neither of these, and thus offer a simpler, more systematic account of
this discourse.
Carr offers a final objection to the potential for a Kratzer-style account to handle the
data in DEATH IN DAMASCUS: equivocation. She writes that an adequate account
must allow an “unequivocal treatment of expressions of reasonable decision theories.”
Carr seems to grant here that a Kratzer-style view could accommodate the data with
appropriate parameter values, but maintains that ALEPPO and IF-ALEPPO “should
be compatible at a single context” on the grounds that “they are all the deliverances of a
unified and coherent body of norms: namely, causal decision theory plus the desire to
avoid death” (p. ). Similarly Cariani, Kaufmann, and Kaufmann write, “we believe
(although there is naturally room for further argument) that the data point is that
it is these very conditionals (i.e. [IF-A and IF-B]) that are true on the deliberative
interpretation of ought” (p. , footnote ). This, too, suggests the view that a single
parameter value must be in play for NEITHER, which is clearly deliberative, as well as
IF-A and IF-B.
If there is an intuition that NEITHER, IF-A, and IF-B are true on the very same
readings of the modals, we are not sure just what it is supposed to be. As a point
of comparison, consider a reading of IF-A on which it is false. Cariani, Kaufmann,
and Kaufmann call this the “non-reflecting” reading of the conditional, which they
elucidate as follows:
If the miners are in shaft A, we (still) ought to block neither shaft, for their being in shaft A
doesn’t mean that we know where they are. Indeed, no matter where the miners are, we ought
to block neither shaft. (p. )

35 On Carr’s account, this could be implemented via a context shift in the information parameter.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

contextualism about deontic conditionals 

It does seem plausible that this instance of “if the miners are in shaft A, we ought to
block neither shaft” involves the very same reading of the modal as does NEITHER;
indeed, Cariani, Kaufmann, and Kaufmann “take it to be obvious that the non-
reflecting interpretation is deliberative” (p. , footnote ). Yet, as these authors
acknowledge, there is not quite this feeling of sameness between IF-A when heard
as true and NEITHER (p. , footnote ; p. , footnote ; p. , footnote ).
We agree that there is some feeling of commonality between IF-A, even when heard
as true, and NEITHER. Indeed, our account explains this in line with Carr’s idea that,
in DEATH IN DAMASCUS, sentences like these are “the deliverances of a unified and
coherent body of norms.” On our view, IF-A and NEITHER both invoke an ordering
of worlds in terms of deontic ideality. These sentences do flow from the same norms,
perhaps even norms of causal decision theory. However, there is a slight difference
between them: NEITHER applies these norms to our actual information, while IF-
A (on its advisability reading) applies these norms to our actual information plus
information specifying the location of the miners. Our account therefore explains why
the feeling of commonality is weaker here than it is between IF-A on its non-reflecting
reading and NEITHER. In the latter case, the ordering is identical between the two
sentences, while in the former it is merely very similar.
One might try to press the point about equivocation against our proposal on more
theoretical grounds. Carr, as quoted above, may hold as a desideratum that it should
be possible in English to express the deliverances of causal decision theory without a
change in contextually supplied parameter values. We need not decide whether this
is a legitimate desideratum, since our view satisfies it. On our view, one may express
the deliverances of causal decision theory without a change in contextually supplied
parameter values by using the subjective “ought” in a set of conditionals of the form: “if
your credences and values are . . . , you ought to . . ..” Our view would not, of course,
meet the stronger desideratum that every discourse that expresses the deliverances
of causal decision theory must involve no change in contextually supplied parameter
values. But this stronger desideratum would be unmotivated given the general utility to
conversation of such changes for modals, quantifier domains, and demonstratives.36 , 37
Considerations of whether a set of sentences is intuitively equivocal may ultimately
support our account in a different way. Consider again the piece of discourse men-

36 For example, we easily navigate an unannounced shift from the circumstantial modality of
“hydrangeas can grow here” to the legal modality of “you can plant anything that won’t block visibility
around the corner.” It would be possible, but more cumbersome, to express these thoughts without relying
on context to supply the appropriate parameter values. Dowell (, ) defends the view that our ability
to detect speaker intentions underwrites our competence with contextually supplied parameter values.
37 Jennifer Carr (pc) asks whether our willingness to posit covert modals, combined with contextual
flexibility, might undermine the reasons for accepting the view of if -clauses as restrictors. She suggests that
reading “if p, then q” as must(p ⊃ q), where an epistemic “must” takes scope over a material conditional,
would then offer a simpler semantics. We note that such a view would need to add a condition to avoid
the consequence that “if p, then q” is true whenever must(∼p) is true. Given this, we do not regard the
alternative as simpler than the restrictor view. The viability of the alternative depends in part on whether
it can be developed to deal systematically with the full range of data that have been taken to motivate the
restrictor view (e.g. Lewis, ), a question we do not explore here.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 aaron bronfman and j. l. dowell

tioned above, TEST, for von Fintel’s three-shaft version of the miners case, where we
stipulate the existence of test results that tell whether or not the miners are in A:
I don’t know whether we ought to blow A up. If the miners are not in A, we ought
to blow A up. If the miners are in A, we ought not to blow A up.
These sentences do not appear to be equivocal. This leads to a dilemma for the views
of Carr and Cariani, Kaufmann, and Kaufmann. They may choose to go contrary to
this intuition, and insist that context supplies different parameter values for the first
sentence than for the conditionals. This, we think, would further undermine their
appeal to a sense of commonality between IF-A and NEITHER to argue for sameness
of contextually supplied parameter values there. As mentioned above, it also appears
to be a needlessly complex account of this piece of discourse, one which invokes both
a novel semantic rule with additional parameter(s) and an unmotivated context shift,
in contrast to our straightforward account.
Alternatively, they could hold that context supplies the same parameter values
to all these sentences. In that case, they grant that conditionals such as MORE
INFORMATION, IF-A, and IF-ALEPPO receive, in this and similar cases, their true
readings simply because context supplies the appropriate parameter values, and not
because a novel semantic rule is in play. This puts strong pressure on their accounts
to hold that these conditionals receive their true readings in other cases, such as the
standard MINERS or DEATH IN DAMASCUS cases, for the same reason. Overall,
on this branch, all sides are committed to the core of the Kratzer semantics and to
the existence of a pragmatic mechanism that can assign the kind of parameter values
we have posited in the standard MINERS or DEATH IN DAMASCUS cases. Our
opponents add commitments to a more complicated semantics and to the claim that
the pragmatic mechanism that is common ground cannot operate in those cases. The
additional commitments of the opposing views might be justified if there were strong
independent reason to hold that those cases cannot involve a shift in contextually
supplied parameter values, but our investigation of this issue above has revealed no
such reason. Theoretical virtues would thus appear to favor our account over the
alternatives.

 Conclusion
There has been much recent work in the literature in the philosophy of language,
linguistics, and metaethics over whether a Kratzer-style contextualist semantics for
modal expressions can be made to fit with the full range of data in a series of puzzle
cases. Part of what is at stake in these debates is the viability of relativism as a research
program in the philosophy of language and linguistics; much of the motivation
for relativism is the claim that no contextualist semantics for some expression E is
plausible.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

contextualism about deontic conditionals 

Also at issue is to what extent linguists and philosophers of language should prefer
semantic over near-side pragmatic explanations of ordinary speakers’ judgments that
make up the primary data for semantic and pragmatic theories. As we’ve seen here,
some contextualists, like relativists, defend their novel semantic proposals on the
grounds that no Kratzer-style semantics can be made to fit with the full range of data
in MINERS or in DEATH IN DAMASCUS. Here we have shown how this is not so, by
identifying available Kratzer-friendly readings that fit with the agreed-upon data. This
includes showing both how all of NEITHER, IF-A, and IF-B may be true in a single
piece of practical reasoning in MINERS and how ALEPPO and IF-ALEPPO may both
be true in DEATH IN DAMASCUS, under the assumption that the norms in play are
those of causal decision theory.
The remaining objection to Kratzer-friendly readings in those cases is the claim
that it is “ad hoc” or “unsystematic” to suppose that context is able to secure the
needed parameter values, as our pragmatic explanation requires. But we have seen
how this objection is misplaced; clear features of the context in which those utterances
are felicitous are features that make the needed parameter values highly salient. We
have also, in considering DeRose’s case and a parallel three-shaft MINERS case,
identified independent evidence for the contextual availability of what we are calling
the “advisability” readings that we posit for MORE INFORMATION and for IF-
ALEPPO. This is further evidence that positing such readings is not ad hoc, but
required on grounds independent of the issues raised here.
Finally, we have seen that the charge that the readings we identify are “unsystematic”
is also misplaced. Our theory is able to identify the sense of commonality between
NEITHER and IF-A and IF-B that is pretheoretically plausible and likewise for the
piece of discourse TEST. In contrast, our opponents are faced with a dilemma. In
accounting for TEST, they must either offer a needlessly complex explanation that
undermines their general appeal to the sameness of contextually supplied parameter
values, or else endorse all the materials needed for own account, thus making their
additional commitments appear superfluous.
We conclude that a Kratzer-style, flexible contextualist semantics, supplemented
with a near-side pragmatic account of how it is that contexts supply the needed
parameter values, remains the view to beat.

Acknowledgments
We are grateful to Fabrizio Cariani, Jennifer Carr, Nate Charlow, and Matthew Chrisman for
detailed and helpful comments.

References
Bronfman, A. and Dowell, J. L. (forthcoming) The Language of Reasons and Ought. In Star, D.
(ed.) The Oxford Handbook of Reasons and Normativity. Oxford: Oxford University Press.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 aaron bronfman and j. l. dowell

Cariani, F., Kaufmann, M., and Kaufmann, S. () Deliberative Modality under Epistemic
Uncertainty. Linguistics and Philosophy. ., pp. –.
Carr, J. () The “If p, Ought p” Problem. Pacific Philosophical Quarterly. ., pp. –.
Charlow, N. () What We Know and What to Do. Synthese. ., pp. –.
DeRose, K. () Epistemic Possibilities. Philosophical Review. ., pp. –.
Dowell, J. L. () A Flexible Contextualist Account of Epistemic Modals. Philosophers’ Imprint.
., pp. –.
Dowell, J. L. () Contextualist Solutions to Three Puzzles about Practical Conditionals.
Oxford Studies in Metaethics. , pp. –.
Dowell, J. L. () Flexible Contextualism about Deontic Modals. Inquiry. .–, pp. –.
Egan, A. () Relativism about Epistemic Modals. In Hales, S. D. (ed.) A Companion to
Relativism. pp. –. Malden, MA: Wiley-Blackwell.
Fintel, K. von () Counterfactuals in a Dynamic Context. In Kenstowicz, M. (ed.) Ken Hale:
A Life in Language. pp. –. Cambridge, MA: MIT Press.
Fintel, K. von () Conditionals. In von Heusinger, K. Maienborn, C. and Portner, P. (eds.)
Semantics: An International Handbook of Natural Language Meaning. Vol. . pp. –.
Berlin: de Gruyter.
Fintel, K. von () The Best We Can (Expect to) Get? Ms.
Fintel, K. von and Iatridou, S. () What to Do If You Want to Go to Harlem. Ms.
Frank, A. () Context Dependence in Modal Constructions. Ph.D. dissertation. University of
Stuttgart.
Geurts, B. () On an Ambiguity in Quantified Conditionals. Ms.
Gibbard, A. and Harper, W. L. () Counterfactuals and Two Kinds of Expected Utility. In
Hooker, C., Leach, J., and McClennen, E. (eds.) Foundations and Applications of Decision
Theory. pp. –. Dordrecht: Reidel.
Gillies, T. () Counterfactual Scorekeeping. Linguistics and Philosophy. ., pp. –.
Joyce, J. () Regret and Instability in Causal Decision Theory. Synthese. ., pp. –.
Kolodny, N. and MacFarlane, J. () Ifs and Oughts. The Journal of Philosophy. .,
pp. –.
Korta, K. and Perry, J. () Pragmatics. In Zalta, E. N. (ed.) The Stanford Encyclope-
dia of Philosophy. Winter. [Online] Available from: <http://plato.stanford.edu/archives/
win/entries/pragmatics/>. [Accessed  Nov .]
Kratzer, A. () What “Must” and “Can” Must and Can Mean. Linguistics and Philosophy. .,
pp. –.
Kratzer, A. () The Notional Category of Modality. In Eikmeyer, H. and Rieser, H. (eds.)
Words, Worlds, and Contexts. pp. –. Berlin: de Gruyter.
Kratzer, A. (a) Modality. In von Stechow, A. and Wunderlich, D. (eds.) Semantics: An
International Handbook of Contemporary Research. pp. –. Berlin: de Gruyter.
Kratzer, A. (b) Conditionals. In von Stechow, A. and Wunderlich, D. (eds.) Semantics: An
International Handbook of Contemporary Research. pp. –. Berlin: de Gruyter.
Kratzer, A. () Modals and Conditionals. Oxford: Oxford University Press.
Lewis, D. () Adverbs of Quantification. In Keenan, E. L. (ed.) Formal Semantics of Natural
Language. pp. –. Cambridge: Cambridge University Press.
MacFarlane, J. () Assessment Sensitivity. Oxford: Oxford University Press.
Moss, S. () On the Pragmatics of Counterfactuals. Noûs. ., pp. –.
Schroeder, M. () Ought, Agents, and Actions. Philosophical Review. ., pp. –.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i


Objective and Subjective ‘Ought’
Ralph Wedgwood

Over the years, several philosophers have argued that deontic modals, such as ‘ought’
and ‘should’ in English, and their closest equivalents in other languages, are systemat-
ically polysemous or context-sensitive. On this view, in effect, there are many different
concepts that can be expressed by ‘ought’—as we might call them, many different
‘ “ought”-concepts’—and whenever the term is used, the particular context in which it
is used somehow determines which of these concepts it expresses on that occasion.
More specifically, one way in which these ‘ought’-concepts differ from each other
is that some of them are more ‘objective’, while others are more ‘subjective’ or
‘information-relative’. When ‘ought’ expresses one of these more objective concepts,
what an agent ‘ought’ to do at a given time may be determined by facts that neither the
agent nor any of his friends or advisers either knows or is even in a position to know;
when it expresses one of the more ‘subjective’ concepts, what an agent ‘ought’ to do is
in some way more sensitive to the informational state that the agent (or his advisers
or the like) find themselves in at the conversationally salient time.1
In this chapter, I shall first present some linguistic evidence in favour of this view of
‘ought’. Then I shall propose a precise account of the truth conditions of sentences
involving terms that express these different ‘ought’-concepts. Unfortunately, in the
available space I shall not be able to do much more than simply to propose this
semantic account of these ‘ought’-concepts. In my opinion, the linguistic evidence
makes this account more plausible than any alternative account that metaethicists or
semanticists have devised so far; but I shall only be able to gesture in the direction of
this evidence here.
The general idea of the kind of account that I shall propose is not new. It is basically
akin to the theory of ‘subjective rightness’ that was given by Frank Jackson ()—
since like Jackson’s theory, it gives a starring role to the notions of probability and
of the expected value of a proposition. Nonetheless, my account has several crucial

1 For some philosophers who have advocated distinguishing between the objective and the subjective
‘ought’, see Brandt (, pp. –), Ewing (), Parfit (, p. ), Jackson (), Jackson and Pargetter
(, p. ), and Gibbard (). In a somewhat similar way, Sidgwick (, p. ) distinguished between
objective and subjective rightness and wrongness.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 ralph wedgwood

differentiating features: unlike Jackson’s theory, my account implies that standard

deontic logic is valid for every kind of ‘ought’; it is much more general than Jackson’s
theory, since it is designed to account for all the concepts that can be expressed by
‘ought’ and its equivalents (not just the concept of subjective moral rightness that
Jackson was interested in); and it is also designed to mesh with a quite different account
of how terms such as ‘ought’ interact with conditionals.
The proposal that I shall give here also has affinities with that of Gunnar Björnsson
and Stephen Finlay (), according to which the context-sensitivity of ‘ought’ is
explained by the thesis that uses of ‘ought’ are relativized to bodies of information. In a
somewhat similar way, I shall propose that uses of ‘ought’ are relativized to probability
distributions; and every probability distribution determines a body of information—
namely, the set of propositions that have probability  within that distribution. Still, as
I shall explain in Section  below, my approach differs from theirs in several crucial
ways.
In my view, a full account of the meaning of a term in a natural language would have
to be a fairly complicated story. More precisely, such an account would have to involve
the following components:
a. An account of the syntactic role of the term—that is, of how the term can combine
with other terms to form well-formed grammatical sentences;
b. An account of what it is to understand the term—that is, to be competent in using
the term and in interpreting its use by other speakers;
c. An account of the range of semantic values that the term can have—that is, of
the contributions that the term can make to the truth conditions of sentences in
which it appears;
d. An account of the non-truth-conditional aspects of meaning that the term can
have—for example, of any conventional implicatures or presuppositions that can
be conveyed by the use of the term;
e. An account of how the conversational context in which the term is used deter-
mines its meaning and its semantic value in the particular context in question.
In this essay, I shall focus chiefly on the third of these components, (c), the range of
semantic values that deontic modals such as ‘ought’ and ‘should’ can have—although
I shall also comment briefly on the fifth component, (e), the question of how the
conversational context in which the term is used determines the meaning that it has
in the context in question.
Otherwise, I shall strive to remain neutral about all the other components of the
story. Thus, with respect to the first component (a), I shall not commit myself to any
detailed claims about the underlying syntax or logical form of sentences involving
‘ought’ and ‘should’. In particular, even though I shall claim that ‘ought’ and ‘should’
have different semantic values in different contexts, I shall not commit myself to any
particular view about how these different semantic values arise from the underlying
syntax. Specifically, I shall not commit myself to any view about whether sentences

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

objective and subjective ‘ought’ 

containing ‘ought’ contain hidden variables (or hidden terms of any other kind), so
that the way in which the term’s semantic value shifts between contexts results simply
from different items’ being referred to by these hidden terms, or whether some other
syntactic phenomenon underlies these shifts. I shall not even rule out the idea that the
term ‘ought’ is syntactically simple and unstructured, and simply demands different
semantic interpretations in different contexts.
Similarly, with respect to the second component (b), I shall not here defend any
particular view of what it is to understand or to be a component user of the term. In
fact, I am inclined to favour a certain sort of account of this second component of a
term’s meaning. Specifically, according to an account of this sort, we can explain what
it is to be linguistically competent with a term by appealing to the range of concepts
that the term can be used to express: to be linguistically competent with the term
requires having the ability to use the term to express concepts within that range (in
a way that enables competent hearers to interpret one’s use of the term as expressing
the concept within that range that one intends to express). Then the nature of each of
these concepts can be explained in terms of the conceptual role that the concept plays
in one’s thinking, and in terms of the way in which this conceptual role determines the
object, property, or relation that the concept stands for or refers to.2
However, even though I am attracted to this view of what linguistic competence
consists in, I shall not attempt to defend this view here. Instead, I shall simply give
an account of the range of truth conditions that sentences involving ‘ought’ can have.
To bring out the similarity between the different truth conditions in this range, I shall
put my account in the form of a schema involving three different parameters; as I shall
explain, the different truth conditions that a sentence involving ‘ought’ can have in
different contexts all correspond to different ways of setting these three parameters.
So, in effect, something in the conversational context in which the term ‘ought’ is used
must determine what these parameters are; I shall try to comment, at least in passing,
on what features of the conversational context could do this.

 A Semantic Framework
The general semantic approach that I shall take here is in line with what could be
called the ‘classical’ semantics for deontic logic. According to this approach, ‘ought’
and ‘should’ and their equivalents in other languages are all broadly modal terms, just
like ‘must’, ‘may’, ‘can’, and the like. Every occurrence of ‘ought’ expresses a concept
that functions as a propositional operator—that is, as a concept that operates on a
proposition (the proposition that is expressed by the sentence that is embedded within
the scope of this occurrence of ‘ought’), to yield a further proposition (the proposition
that is expressed by the sentence in which this occurrence of ‘ought’ has largest scope).

2 I have attempted to sketch some parts of this story elsewhere; see especially Wedgwood (,
chapters –).

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 ralph wedgwood

Thus, for example, the occurrence of ‘ought’ in the English sentence ‘This room
ought to be swept’ expresses an ‘ought’-concept that operates on the proposition that is
expressed by the embedded sentence ‘This room is swept’. So the proposition expressed
by the sentence ‘This room ought to be swept’ has the logical form ‘O (This room is
swept)’, where ‘O (. . .)’ is the relevant ‘ought’-concept. In a proposition of the form
‘O(p)’, I shall call the proposition p on which the relevant ‘ought’-concept operates the
‘embedded proposition’.
In general, the conditions under which a sentence expressing such an ‘ought’-
proposition is true at a possible world can be specified as follows. For every such
sentence, and for every possible world w, there is a function that maps possible worlds
onto domains of possible worlds, and a relevant ordering on these worlds, such that
the sentence expressing the ‘ought’-proposition ‘O(p)’ is true at w if and only if, out of
all worlds in the domain that this function assigns to w, the embedded proposition p
is true at all worlds that are not ranked any lower down in this ordering than any other
worlds in this domain.3
If—as will usually be the case—it is possible to express this ordering by means
of words like ‘better’ and ‘worse’, then we can say more simply that the sentence
expressing ‘O(p)’ is true at w if and only if the embedded proposition p is true at all
the optimal worlds in the relevant domain. So, for example, the sentence ‘This room
ought to be swept’ is true at w if and only if the proposition that this room is swept is
true at all the relevantly optimal worlds in the relevant domain.4 So long as there are
always some worlds in the relevant domain that count as optimal in the relevant way,
it turns out that all of the principles of standard deontic logic—in effect, the modal
system KD—will be valid for every ‘ought’-concept.
In this way, this classical approach to the semantics of ‘ought’ involves two param-
eters: a function that maps each possible world onto a domain of possible worlds,
and the relevant ordering on these worlds. As I shall propose in the third section of
this paper, this ordering of worlds can itself be regarded as having an expectational
structure: that is, there is some kind of value, and some probability distribution,
such that this ordering of the worlds is equivalent to an ordering in terms of the
expected value of the worlds, according to this probability distribution. However,
before developing this expectational conception of the relevant ordering, I shall survey
some of different concepts that the term ‘ought’ can express.

3 So, if the proposition ‘O(p)’ is not to be trivial, there must be some worlds that are not ranked any lower
in this ordering than any other worlds in the domain. That is, what David Lewis () called the ‘Limit
Assumption’ must hold. Some philosophers—such as Eric Swanson ()—have denied that the Limit
Assumption must hold for all ‘ought’-concepts. But in my view, there are independent reasons for thinking
that it must hold. Specifically, if ‘ought’ agglomerates over conjunction—including infinite conjunction—
and ‘ought’ implies logical possibility, then it seems that the Limit Assumption must indeed hold: that is, in
effect, there must be a possible world where everything is as it ought to be.
4 This ‘classical’ semantics for deontic operators was defended by such pioneering deontic logicians
as Åqvist () and Lewis (). My defence of this classical semantics is given in Wedgwood (,
chapter ).

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

objective and subjective ‘ought’ 

 The Varieties of ‘Ought’

In earlier work, I have surveyed several of the different concepts that words like ‘ought’
can express.5 As I have argued, some of these ‘ought’-concepts are instances of the
‘practical “ought” ’; some are instances of the ‘purpose-relative “ought” ’, some of the
‘ “ought” of general desirability’, some of the ‘rational “ought” ’, and so on.
For our purposes, the most important point is that each of these kinds of ‘ought’ can
be used in a more or less ‘objective’ or ‘subjective’ way. For example, let us start with
instances of the ‘practical “ought” ’. Suppose that you are on top of a tower, watching
someone trying to escape from a maze on the ground below. Then you might say:
() He has no way of knowing it, but he ought to turn left at this point.
Here what an agent ‘ought’ to do does not depend purely on the information that is
possessed by the agent at the relevant time; so this first example involves the ‘objective’
‘ought’, rather than the ‘information-relative’ ought.
On the other hand, sometimes we use ‘ought’ in such a way that it does depend
purely on the informational state of the relevant agent at the relevant time. Thus, we
might say about the man who is making his way through the maze:
() All the evidence that he has suggests that turning right at this point would be
the best way to escape from the maze, and so that is what he ought to do now.
Here what the agent ‘ought’ to do depends only on the informational state of the
relevant agent at the relevant time. So this second example involves a subjective or
information-relative ‘ought’, not an objective ‘ought’.
In general, many different kinds of ‘ought’ seem to have both an objective and a
subjective or information-relative version. For example, consider the purpose-relative
‘ought’, such as ‘He ought to use a Phillips screwdriver to open that safe’. What makes
this the purpose-relative ‘ought’ is that the truth value of this statement simply depends
on whether or not using a Phillips screwdriver is part of the best way of opening the
safe; the statement takes no stand on whether the person in question ought (in the
all-things-considered practical sense of ‘ought’) to open the safe at all.
It seems clear that this purpose-relative ‘ought’ also comes in both objective and
subjective versions. An objective version of this sort of ‘ought’ might be: ‘He has no
way of knowing it, but he ought to use a Phillips screwdriver to open that safe’. A
subjective or information-relative version of this ‘ought’ might be: ‘Since he doesn’t
know what sort of safe it is, he ought to start with the ordinary screwdriver first’.
In fact, it also seems plausible that other kinds of ‘ought’, such as what I have
elsewhere called the ‘ “ought” of general desirability’ and the ‘rational “ought” ’, also
have both objective and subjective or information-relative versions. In general, it
seems that for each of these kinds of ‘ought’, there must be some systematic connection

5 See especially Wedgwood (, section ., and , section ).

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 ralph wedgwood

between the more objective and the more subjective versions of that kind of ‘ought’.
Moreover, it seems that it must be broadly speaking the same kind of systematic
connection in each case. The next two sections of this chapter will focus on exploring
this connection.
In addition to giving an account of the relationship between the subjective and
objective versions of each of these kinds of ‘ought’, I shall also aim to unify my account
of these phenomena with yet another kind of ‘ought’—specifically, with the so-called
epistemic ‘ought’, as in:
() Tonight’s performance ought to be a lot of fun.
This seems just to mean, roughly, that it is highly probable given the salient body of
evidence that tonight’s performance will be a lot of fun. If this is indeed at least roughly
what the epistemic ‘ought’ means, then it is clear that the ‘salient body of evidence’ need
not include the total evidence available to the speaker at the time of utterance, since it
seems that even if one knows that the orbit of Pluto is not elliptical, it might be true
for one to say:
() The orbit of Pluto ought to be elliptical (although of course it isn’t).
I shall aim to give an account of the semantic values of a range of uses of ‘ought’ that
includes these uses of the term.

 The Expectational Schema

As explained in Section , I am assuming that the truth conditions of sentences that
express ‘ought’-propositions are in line with the classical semantics of standard deontic
logic. The truth conditions of every such sentence involve the following two crucial
elements: first, a function f that maps each possible world w on to a domain of possible
worlds f (w); secondly, an ordering on the worlds in this domain. So, to understand the
semantic value of any ‘ought’-concept, we need to understand these crucial elements.
In this paper, I shall propose a broadly expectational conception of this ordering.
For every use of ‘ought’, the ordering of worlds in the domain is always an ordering in
accordance with the expected value of those worlds. If the ordering of worlds has this
expectational structure, it is itself the result of two more fundamental components: a
probability distribution E; and a value function V, which assigns a value to each of the
worlds within the domain f (w).
There are two ways of interpreting this expectational conception of the orderings
that feature in the truth conditions of these sentences. On the first interpretation,
the analysis of each of these orderings as resulting from a probability function E
and a value function V is built into the semantics of modal terms such as ‘ought’.
On the second interpretation, the semantics just involves these orderings themselves,
without itself containing any such analysis of the orderings; and the analysis is a purely

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

objective and subjective ‘ought’ 

metaphysical thesis about the nature of the orderings in question. In fact, I shall argue
in Section  that there are reasons, concerning the truth conditions of conditional
sentences involving ‘ought’, for interpreting this expectational conception in the first
way, as built into the semantics of terms such as ‘ought’. For the time being, however,
we simply shall leave open the question of which interpretation of this expectational
conception is correct.
In the rest of this section, I shall explain this expectational schema in more detail,
starting with some comments on each of its three elements—the domain function f ,
the probability distribution E, and the value function V.
(i) The first element of any instance of this expectational schema is familiar: it is a
domain function f , which maps every world w on to the relevant domain of possible
worlds f (w). It is this function that identifies the worlds that are, as we might put
it, ‘up for assessment’ by the ‘ought’-concept in question, relative to w. In effect, this
function f fixes what Angelika Kratzer (, chapter ) called the ‘modal base’—
the set of propositions that are true throughout the domain of worlds that are up
for assessment by the ‘ought’-concept, relative to w. We shall explore some specific
examples of such domains of worlds in the next section.
(ii) The second element of any instance of this expectational schema is a proba-
bility distribution E. I shall assume that every probability distribution is a function
that assigns real numbers in the unit interval from  to  to the propositions in
a propositional algebra (that is, a set of propositions that is closed under Boolean
operations like negation, disjunction, and so on). Any function of this sort that obeys
the fundamental axioms of probability theory counts as a probability distribution. So,
in particular, the omniscient probability function—the function that assigns  to every
true proposition and  to every false proposition in the relevant algebra—is itself a
probability distribution.
Another way of thinking of such probability distributions is as defined over a space
of possible worlds, relative to a certain ‘field’ of subsets of this space of worlds. This
‘field’ also constitutes an algebra, in the sense that it is closed under operations such as
complementation, union, and the like; the probability function assigns real numbers
to the sets of worlds in this field.6 This probability function can be thought of as a
measure on the space of worlds: intuitively, it tells us how much of the whole space of
worlds is taken up by each set in this field. (This is why the probability measure has
to obey a basic additivity principle: the proportion of the whole space taken up by
the union of any two disjoint sets of worlds is the sum of the proportions taken up by

6 For technical reasons (see Easwaran ), if the space contains indenumerably many worlds, it may
not be possible to assign a probability to every set of worlds in the space—it may be that only certain sets of
worlds can have a probability assigned to them. This is why the probability distribution is defined over the
worlds only relative to a ‘field’ of sets of worlds—where this field contains all and only those sets of worlds
that correspond to propositions in the relevant algebra. Fortunately, this complication will not matter for
present purposes.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 ralph wedgwood

those sets.) This way of thinking of probability distributions is equivalent to thinking

of them as defined over propositions, so long as for each of the relevant propositions,
the field contains a set of worlds in which that proposition is true. (Indeed, on some
views, each of these propositions is identical to the corresponding set of worlds.)
It seems clear that for some purposes, we will have to consider probability distribu-
tions in which some propositions that are metaphysically necessary but knowable only
empirically—such as the proposition that Hesperus = Phosphorus—have a probability
lower than . If we think of the probability distribution as defined over a space of
worlds, this means that we will have to allow the space to include worlds where
Hesperus = Phosphorus. Such worlds are not metaphysically possible, but they may
still be epistemically possible. So the space of worlds over which the probability
distribution is defined is a space of epistemically possible worlds, not a space of
metaphysically possible worlds.
Although we can make sense of probability distributions in which the proposition
that Hesperus = Phosphorus has a non-zero probability, the sentence embedded inside
a deontic modal term such as ‘ought’ seems to permit the substitution of necessarily
co-referring terms. Since Hesperus is identical to Phosphorus, if you ought to visit
Hesperus, it surely follows that you also ought to visit Phosphorus. To explain this
fact about deontic modals, within the semantic framework that I am assuming here,
the domain of possible worlds f (w) must be a domain, not of epistemically possible
worlds, but of metaphysically possible worlds.
On this picture, then, we have in effect two different spaces of possible worlds—
a domain of metaphysically possible worlds, and a space of epistemically possible
worlds.7 Many different interpretations of these two spaces of possible worlds are
possible, but to fix ideas, I shall propose one such interpretation here. According to
this interpretation, these two spaces of possible worlds correspond to two different
kinds of propositions.
The metaphysically possible worlds correspond to propositions of the ‘Russel-
lian’ kind—structured entities that are composed, by means of operations such as
predication, negation, conjunction and the like, out of entities such as individu-
als, properties, and relations. Metaphysically possible worlds are individuated by

7 Invoking two spaces of possible worlds—a space of epistemically possible worlds and a domain of
metaphysically possible worlds—in this way sets my account apart from most previous accounts of deontic
modals, which have typically sought to explain the semantic value of these modals purely in terms of a single
domain of possible worlds. Thus, for example, the account of Silk () resembles mine to the extent that
it allows the ordering on the worlds to vary with an ‘information state’, but for Silk this information state
is simply a kind of modal base, and so is nothing more than a ‘set of worlds’ (p. ). The main exception
is Jennifer Carr (, p. ) who proposes that the semantic value of deontic modals involves a modal
background, a probability function, and a value parameter (although confusingly she describes the modal
background and the probability function as together constituting an ‘informational parameter’). The main
difference between my account and Carr’s is that her account involves yet another parameter, a ‘decision
rule parameter’, which seems unnecessary to me; she also does not distinguish between metaphysically
and epistemically possible worlds in the way that I regard as important. (For further discussion of Carr’s
proposal, see Section  below.)

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

objective and subjective ‘ought’ 

the Russellian propositions that are true at those worlds: a metaphysically possible
world w is identical to a metaphysically possible world w if and only if exactly
the same Russellian propositions are true at w and w . The Russellian proposition
that you visit Hesperus is composed out of you, the visiting relation, and the planet
Hesperus itself. This proposition is therefore identical to the Russellian proposition
that you visit Phosphorus. Since the propositions that in this way individuate a
possible world must form a logically complete and consistent set, this explains why
there cannot be metaphysically possible worlds in which you visit Hesperus but not
Phosphorus.
By contrast, the epistemically possible worlds are individuated by the Fregean
propositions that are true at those worlds—where Fregean propositions are structured
entities that are composed, by means of operations such as predication and the like, out
of concepts, which are modes of presentation of such entities as individuals, properties,
and relations. An epistemically possible world w is identical to an epistemically
possible world w if and only if exactly the same Fregean propositions are true at
w and w . Since one and the same planet may have several different modes of
presentation—including a ‘Hesperus’ mode of presentation and a ‘Phosphorus’ mode
of presentation—this allows for the existence of an epistemically possible world in
which you visit Hesperus but not Phosphorus.
(iii) Finally, the third element of any instance of this expectational schema is a value
function of a certain kind.
In general, this value function will evaluate a certain set of alternatives—such as a
set of alternative acts, or the like. When we speak of an ‘act’ here, it seems that what we
really mean is a proposition to the effect that the relevant agent performs an act of the
relevant type at the relevant time. So a more general account would involve regarding
this value function as evaluating a certain set of alternative propositions.
To say that these propositions are ‘alternatives’ to each other is to say that they are
mutually exclusive: no more than one of these propositions is true at any world in
the relevant domain of metaphysically possible worlds. I shall also assume that these
propositions are jointly exhaustive: that is, at least one of these proposition is true at
every world in this domain. In other words, this set of propositions forms a partition
of this domain of worlds: at every possible world in this domain, exactly one of these
propositions is true.
Since no more than one of these propositions is true at every world in this domain,
and there is no metaphysically possible world where you visit Hesperus without also
visiting Phosphorus, the proposition that you visit Hesperus cannot be a distinct
member of this set of propositions from the proposition that you visit Phosphorus.
Thus, the propositions in this set must be Russellian propositions (indeed, each such
proposition might simply be identified with a subset of the domain of metaphysically
possible worlds). In effect, every such value function provides a set of Russellian
propositions {A , . . . An } that forms a partition of the relevant domain of worlds, and
assigns a value to each of these propositions.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 ralph wedgwood

We may think of the value that the value function assigns to each Russellian
proposition Ai in this set as a real number V(Ai ) that represents the value of
this proposition Ai . This measure is presumably not unique: the choice of unit
will obviously be arbitrary (just as it is arbitrary whether we measure distance in
miles or kilometres), and the choice of zero point may also be arbitrary as well
(just as it is arbitrary whether we take the zero point on a thermometer to be 
Fahrenheit or  Celsius). But to fix ideas, let us suppose that except in these two
ways, this value function is not arbitrary. Given an arbitrary choice of a unit and
a zero point, this function gives the true measure of the relevant value. In more
technical terms, we are supposing that the value in question can be measured on an
interval scale.8
(iv) In this way, any instance of this expectational schema involves three items:
a function f from each metaphysically possible world to a relevant domain of such
worlds; a probability distribution E; and a value function V defined over a set of
propositions that constitutes a partition of the relevant domain of metaphysically
possible worlds. To represent the fact that a particular instance of the expectational
schema gives an account of the conditions under which a use of a sentence involving
‘ought’ is true, I shall explicitly index this occurrence of ‘ought’ to this trio of items:
‘Ought<f ,E,V> ’.
I have proposed that the value function V is defined over a set of Russellian
propositions that forms a partition of the domain of metaphysically possible worlds.
However, the probability distribution E can assign probabilities to hypotheses about
the value that V assigns to various propositions—where each of these hypotheses
is, in effect, a Fregean proposition. For example, such hypotheses might include:
‘The proposition that I visit Hesperus has value n’, and ‘The proposition that I visit
Phosphorus has value m’—where these two hypotheses are distinct from each other.
In this way, the hypotheses to which E assigns probability refer to Russellian
propositions by means of modes of presentation—where these modes of presentation
of Russellian propositions are, in effect, Fregean propositions. It seems that just as the
relevant set of Russellian propositions forms a partition of the domain of metaphysi-
cally possible worlds, the corresponding set of Fregean propositions forms a partition
of the space of epistemically possible worlds. Since each of these hypotheses involves
a Fregean proposition A E (as a mode of presentation of a Russellian proposition AM ),
we may think of each of these hypotheses as having the form ‘V(AE ) = n’.
We can now give a definition of the EV-expected value of a Fregean proposition
AE , in the following way. Consider a collection of hypotheses {h , . . . hk }, where each
hypothesis hi has the form ‘V(AE ) = ni ’. Suppose that this collection of hypotheses
also forms a partition, in the sense that it is epistemically certain that exactly one of

8 Strictly speaking, to accommodate incommensurability, we need to consider a set of such value

functions, rather than a unique value function. But I shall ignore this complication for the purposes of
the present discussion.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

objective and subjective ‘ought’ 

these hypotheses is true; and suppose that E assigns a probability to each of these
hypotheses. Then the EV-expected value of AE is the probability-weighted sum of
the values of AE according to these hypotheses, where the value of AE according to
each hypothesis is weighted by the probability of that hypothesis. Symbolically, the
EV-expected value of AE is:

i ni E(V(AE ) = ni ).
9

Since the set of Fregean propositions that feature in these hypotheses forms a parti-
tion of the epistemically possible worlds, the epistemically possible worlds themselves
can be ordered in terms of the EV-expected value of the proposition in this set that is
true at each world. Let us say that the epistemically possible worlds that are not ranked
lower down in this ordering than any other such worlds have ‘maximal EV-value’.
For each of these epistemically possible worlds, we need to find the metaphysically
possible worlds that in the relevant way ‘correspond to’ that epistemically possible
world. In the simple cases, a metaphysically possible world wM corresponds to an
epistemically possible world wE if and only if all the Fregean propositions true at
wE are true at wM (that is, these Fregean propositions are modes of presentation
of Russellian propositions that are true at wM ). In more complex cases (such as
epistemically possible worlds in which you visit Hesperus but not Phosphorus), there
are no metaphysically possible worlds where all these Fregean propositions are true. In
these cases, we would have to give a different account of what it is for a metaphysically
possible world wM to ‘correspond to’ an epistemically possible wE ; for example, we
could say that that the metaphysically possible worlds that correspond to wE are those
at which a maximal subset of the Fregean propositions that are true at wE are true.
We can now define a selection function S<f , E, V> over the metaphysically possible
worlds that will pick out the metaphysically possible worlds that correspond to
the epistemically possible worlds with maximal EV-value: for any metaphysically
possible world wM , wM belongs to the subset of f (w) picked out by this selection
function S<f , E, V> (f (w)) if and only if for some epistemically possible world wE , wM
corresponds to wE , and wE has maximal EV-value.
The truth conditions of sentences of the form ‘Ought<f ,E,V> (p)’ can now be
specified in terms of this selection function S<f , E, V> :
‘Ought<f ,E,V>(p)’ is true at w if and only if p is true at every world w  ∈
S<f , E, V>(f (w)).
Let us illustrate this proposal by considering the example of Frank Jackson’s ()
‘three drug’ case. In this case, the speakers using ‘ought’ are focussing on the practical

9 An alternative approach would be to understand the ‘expected value’ of a Fregean proposition A as

E
defined in terms of the conditional probability of such hypotheses—conditional on the proposition AE that
is in question. To put it symbolically, the appropriate expectation might be: i ni E(V(AE ) = ni | AE ).
Unfortunately, I cannot take the time here to explore whether this alternative approach might give a better
interpretation of certain uses of ‘ought’.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 ralph wedgwood

situation of a certain agent x at a time t; in this situation, there are three options
available to x at t—giving the patient drug , giving the patient drug , and giving the
patient drug . The agent x knows that drug  is second best. Unfortunately, although
x knows that either drug  is best or drug  is best, x does not know which—and x
knows that if drug  is best, drug  will be disastrous, while if drug  is best, drug
 will be disastrous. The speakers are considering what x ought to do at t given the
informational state that x is in at t. Then the three parameters f , E, and V will be
something like the following:
• f (w*) is the set of metaphysically possible worlds that are practically available to
x at t (so in these worlds, everything that x cannot change by x’s actions at t is
exactly as it is in w*).
• E is a probability distribution that in the appropriate way corresponds to x’s
informational state at t.
• V is a value function that assigns values to the three Russellian propositions, A ,
A , and A —the propositions that at t, x gives the patient drug , drug , and drug
, respectively—where these three propositions form a partition of the domain of
worlds f (w*).
E assigns probabilities to various hypotheses—including hypotheses about the value
that V assigns to A , A , and A . In referring to these Russellian propositions A , A ,
and A , these hypotheses use modes of presentation of these propositions—and we are
assuming that these modes of presentation of Russellian propositions are themselves
Fregean propositions. To keep things simple, however, let us suppose that E puts the
relevant Fregean propositions into a one-to-one correspondence with the Russellian
propositions. (That is, for each of these Russellian propositions, there is exactly one
Fregean proposition that is a mode of presentation of that Russellian proposition
such that E attaches non-zero probability to any hypotheses involving that Fregean
proposition.) Thus, there is also a corresponding set of Fregean propositions forming
a partition of the epistemically possible worlds—AE. , AE. , and AE. —corresponding
to A , A , and A .
Assume that for each of these Fregean propositions AE.i , there are two hypotheses
hi. and hi. of the form ‘V(AE.i ) = n’ to which E assigns non-zero probability:
• h. is ‘V(AE. ) = ’; and h. is ‘V(AE. ) = ’.
• h. is ‘V(AE. ) = ’; and h. is ‘V(AE. ) = ’.
• h. is ‘V(AE. ) = ’; and h. is ‘V(AE. ) = ’.

Suppose that for all i, E(hi. ) = E(hi. ) = .. Then the EV-expected value of each of
these three Fregean propositions AE. , AE. , and AE. is as follows:
• EV(AE. ) = E(h. ) ×  + E(h. ) ×  = .
• EV(AE. ) = E(h. ) ×  + E(h. ) ×  = .
• EV(AE. ) = E(h. ) ×  + E(h. ) ×  = .

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

objective and subjective ‘ought’ 

Thus, the epistemically possible worlds that have maximal EV-value are all and only
the worlds at which AE. is true, and since the metaphysically possible worlds in f (w*)
corresponding to the epistemically possible worlds at which AE. is true are all and only
the worlds where A is true, S<f , E, V> f (w*) contains precisely these metaphysically
possible worlds. It follows that ‘Ought<f , E, V> (A )’ is true at w*.
As I explained above, this proposal is simply an account of conditions under which
‘ought’-sentences are true. I am not proposing that there are hidden variables referring
to these parameters f , E, and V in the actual syntax of these sentences. I am also
not claiming that linguistic competence with ‘ought’ involves some kind of implicit
knowledge or grasp of this semantic account; this semantic account does not by itself
settle the question of how best to account for our competence with ‘ought’.
However, I shall argue in Section  that all of these three parameters—f , E, and
V—are part of the semantics of ‘ought’, in the sense that they must be included in
any systematic account of the truth conditions of the full range of sentences involving
‘ought’. So, in normal contexts when ‘ought’ is used, something must determine what
these three parameters are. Presumably, this will involve the speakers in the context
actually thinking of something that somehow determines these parameters. I shall not
take a definite stand on what exactly the speakers in the context must be focussing on
in this way. (No doubt, few actual speakers employ the formal mathematical concept
of probability to think of a probability distribution!) For example, the probability
distribution E might be determined by the speakers’ in some way thinking of or
imagining a certain epistemic perspective—where as a matter of fact, this perspective
can be modelled by the probability distribution E.
As I shall put it, in the context in question, each of these three parameters f , E,
and V is ‘contextually salient’ (although—as I have said—I shall remain neutral about
what exactly is involved in these parameters’ being contextually salient in this way).
In the next section, I shall show how different settings of these three parameters f , E,
and V can yield intuitively plausible truth conditions for each of the kinds of ‘ought’
considered in Section .

 Instances of the Expectational Schema

The schema set out in the previous section offers a simple way of understanding the
maximally objective kinds of ‘ought’. With these kinds of ‘ought’, E is the omniscient
probability distribution—the function that assigns probability  to every truth and
probability  to every falsehood.
The differences between the semantic values of various objective kinds of ‘ought’
are reflected, not in the probability distribution E, but in the different settings of the
other two parameters—the function f that fixes the relevant domain of metaphysically
possible worlds, and the value function V that measures the value of the worlds in each
domain.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 ralph wedgwood

It seems plausible that the semantic value of every instance of the practical ‘ought’
is focussed on the situation of a particular agent x at a particular time t. (It is this that
has tempted many philosophers—including Mark Schroeder ()—to argue that the
practical ‘ought’ actually stands for a relation between an agent and an act-type.) So it
seems that the semantic value of this use of ‘ought’ will involve a function f that maps
each world w onto the worlds that are ‘practically available’ from the situation that the
agent x is in at the time t in w—in effect, the worlds that the agent x can realize through
the acts that she performs at t in w.
This semantic value will also involve a function V that measures the value of the
various acts that the agent performs at any of these available possible worlds. For
example, more specifically, V might be a measure of the choiceworthiness of the act
that the agent performs in this situation within each of these worlds. On this view,
then, if the relevant ‘ought’ is the objective practical ‘ought’, focussed on the situation
of an agent x at a time t, then ‘Ought (p)’ is true at a world w if and only if p is true
in all the worlds that are practically available from the situation that x is in at t in w
where x does one of the maximally choiceworthy acts available at that time t.
With the more subjective forms of the practical ‘ought’, V and f are exactly as they
are with the objective practical ‘ought’, and E is some less omniscient probability
distribution—that is, it is a probability distribution that encodes a significant degree of
ignorance and uncertainty about the world. For example, in many contexts we might
use a practical ‘ought’ in such a way that its semantic value involves a probability
distribution that corresponds to the system of credences that would be ideally rational
for a thinker to have if their experiences, background beliefs, and other mental states
were exactly like those of the agent x at t.
This, however, is not the only concept that a subjective practical ‘ought’ can express.
If the speakers have pertinent information that is not yet available to the agent who
is under discussion, it will often be natural for the speakers to use an ‘ought’-concept
whose semantic value involves a probability distribution that reflects this information.
Moreover, if the agent herself also thinks that there is some available information that
she has not yet acquired, it will be very natural for the agent to use an ‘ought’-concept
that in this way involves a probability distribution that incorporates this information
that the agent hopes to acquire.10
In general, a probability distribution is in effect a way of representing a certain epis-
temic perspective; and an epistemic perspective can become conversationally salient
for many reasons. For example, as we have noted, many probability distributions
correspond to the systems of credences that an ideally rational thinker would come to
have in response to certain experiences, given a certain set of background beliefs and
other mental states. If this collection of experiences and other mental states is precisely
the collection of experiences and states that a conversationally salient agent has at
a conversationally salient time, this can explain why the corresponding epistemic

10 This is how I would aim to answer the objections of Kolodny and MacFarlane ().

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

objective and subjective ‘ought’ 

perspective will be salient in the conversational context in question. There are many
factors that can explain why a certain agent and time are salient in a conversational
context. For example, in many contexts, the salient time will often be the time of action,
rather than the time of utterance; and the salient agent may be an adviser or observer
of the agent on whom this occurrence of the practical ‘ought’ is focussed, rather than
that agent herself.
This idea of relativizing ‘ought’-concepts to probability distributions is clearly akin
to the idea of Björnsson and Finlay () that occurrences of ‘ought’ are relativized
to bodies of information, conceived of simply as sets of propositions. However,
there are a number of crucial differences. First, although every probability function
determines a body of information (consisting of the propositions to which the function
assigns probability ), the converse does not hold: there are many different probability
distributions in which exactly the same propositions have probability . In this way,
probability distributions contain more structure than mere bodies of information.
Secondly, my proposal is not committed to their view that every occurrence of ‘ought’
is relativized to an ‘end’ or ‘standard’ that can be understood in wholly non-normative
terms. Finally, my proposal is easier to integrate with some of the classical theories
in this area: unlike their account, my proposal entails standard deontic logic; and it
clearly yields the right verdicts in contexts where it is assumed that the agent ought to
maximize some kind of expectation of some kind of value.
We can make sense of objective and subjective versions of many kinds of ‘ought’.
For example, this point seems to hold, not just of the practical ‘ought’, but of the
purpose-relative ‘ought’, the ‘ought’ of general desirability, and the rational ‘ought’ as
well. In each case, the objective and the subjective ‘ought’ differ only with respect to
the relevant probability distribution E: with the objective ‘ought’, E is the omniscient
probability distribution, whereas with the more subjective ‘ought’, E is a probability
distribution that corresponds to the credence function of a possible thinker who
(although perfectly rational) is significantly more ignorant and uncertain about the
world.
It would be intrinsically interesting to explore exactly how this schema can be
worked out in detail for each of these other kinds of ‘ought’; but to save space, I shall
here only explain how it would work for the purpose-relative ‘ought’. So far as I can
see, the purpose-relative ‘ought’ resembles the practical ‘ought’ in that they are both
implicitly focussed on the situation of a particular agent x at a particular time t. So the
relevant function f from worlds to domains of worlds is again the function that maps
each world w onto the worlds that are ‘practically available’ from the situation that the
agent x is in at the time t in w.
The only respect in which the purpose-relative ‘ought’ differs from the practical
‘ought’ is in involving a different value function V. For the purpose-relative ‘ought’,
there is some purpose P that is contextually salient, and the value function V ranks the
various acts that the agent performs at any of the worlds that are practically available
to the agent at the time in question, not in terms of their overall choiceworthiness, but

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 ralph wedgwood

purely in terms of how good these acts are as a means to accomplishing that purpose
P. Otherwise, the two kinds of ‘ought’ work in more or less the same way.
As I remarked in Section  above, it would be preferable if our account of ‘ought’
could also encompass the other kinds of ‘ought’ that I considered in that section—
including the epistemic ‘ought’ (as in ‘Tonight’s performance ought to be a lot of fun’,
which as I said seems roughly equivalent to saying that the embedded proposition that
tonight’s performance will be a lot of fun is highly probable given the salient evidence).
The schema that I proposed in the previous section may be able to capture the
epistemic ‘ought’, in something like the following way. For the epistemic ‘ought’, the
three parameters may be the following. First, f can simply be the function that maps
each world onto the set of all possible worlds that are compatible with everything
that is known for certain in the context. Secondly, E can be a probability distribution
modelling some possible epistemic perspective. (Again, this could be pretty well any
perspective; the participants in a conversation will just have to interpret the contextual
clues in order to discern which perspective is contextually salient in the relevant way.)
Finally, V could simply be a function that ranks answers to a certain question, which
we can think of as a partition of alternative answers to the question, by ranking the true
answer to the question above all the false answers—say, by assigning a value of  to the
true answer and  to false answers. Now, as is well known, probabilities are themselves
simply expectations of truth values. So the ranking of answers to this question in
terms of their EV-expected value is identical to the ranking in terms of these answers’
probability according to E; and this ranking determines a corresponding ordering
of worlds in accordance with the probability of each world’s answer to the question.
So, for example, if the rival answers to the question are simply p and ‘¬p’, then the
sentence ‘It ought to be that p’, involving this epistemic ‘ought’, will be true just in
case p is more than probable than ‘¬p’ (according to the probability distribution that
corresponds to E).
One might wonder whether p’s being barely more probable than ‘¬p’ is enough to
make it true to say ‘It ought to be that p’, using this epistemic ‘ought’. At least, if we
were considering a fair lottery with  numbered tickets, we would not typically say
such things as ‘The winning ticket ought to be one of the  tickets numbered between
 and ’.
However, the reason for this may be that the question that we normally have in mind
is not simply whether or not the embedded proposition is true, but whether or not
some more general explanatory picture of the world is true. If this general explanatory
picture is more than % probable, and the proposition p follows from this explanatory
picture, then it will be true to say ‘It ought to be that p’ (since p will be true in all the
worlds within the domain where this explanatory picture is true). A proposition p that
follows from a general explanatory picture of this sort will typically be significantly
more probable than that general picture itself.
This simple account of the value function V, in terms of the truth value of answers
to a certain question, may turn out not to be completely defensible in the end; a more

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

objective and subjective ‘ought’ 

complicated account of this value function may be required. But at all events, to
capture the range of ways in which we use the epistemic ‘ought’, we have to allow that
many different probability distributions (or spaces of epistemically possible worlds)
can be involved. In particular, when a speaker asserts a proposition involving an
epistemic ‘ought’-concept of this sort, the probability distribution E involved in this
concept’s semantic value does not have to correspond to the information that is
actually available to the speaker. It may be a different probability distribution.
For example, even if the speaker knows perfectly well that the orbit of Pluto is not
elliptical, the relevant probability distribution E does not have to assign a probability
of  to the proposition that the orbit of Pluto is elliptical; it may be a probability
distribution that corresponds to the credences that it would be rational to have given
a body of information that is different from the speaker’s actual total evidence but
contextually salient for other reasons. So this approach has no difficulty handling such
puzzling instances of the epistemic ‘ought’ as ‘The orbit of Pluto ought to be elliptical
(though of course it isn’t)’.11

 ‘Ought’ and Conditionals

In this section, I shall comment on what this expectational account of ‘ought’ implies
about how ‘ought’ interacts with conditionals. It is here that we shall see why the
probability distribution E needs to be part of the semantics of ‘ought’.
The general idea is familiar from such classic discussions of conditionals as that of
Angelika Kratzer (, chapter ). According to Kratzer, quite generally, the effect
of conditionals is to restrict some domain of possible worlds that is involved in the
semantics of a modal operator that appears (at least implicitly) as the dominant
operator of the consequent of the conditional—by restricting this domain of worlds to
that subset of the domain where the antecedent of the conditional is true.
As I mentioned in the previous section, we can think of each probability distribution
as itself a space of possible worlds—where a ‘space’ of worlds is more than a mere set
of worlds. A space of worlds involves not just a set of worlds but also a measure on
subsets of this space. That is, there is a certain ‘field’ of subsets of the space such that
we can make sense of ratios between the proportions of the whole space that are taken
up by these subsets. For example, we can make sense of the idea that one subset takes
up twice as large a proportion of the whole space as another. So we can in effect view
the probability distribution E as a structured measurable space of worlds of this sort.

11 This view of the epistemic ‘ought’ also helps to explain why it has such different truth conditions from
the epistemic ‘must’—even though both modals are broadly speaking necessity operators. For ‘must’, the
ordering on the possible worlds makes no difference to the sentence’s truth conditions; and according to my
proposal, the only relevance of the probability distribution E is to generate the ordering of possible worlds
in terms of their EV-expected value. So the truth conditions of ‘Must (p)’ depend purely on whether p is
true throughout f (w), and is unaffected by what E and V are in the relevant context.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 ralph wedgwood

Once we have the idea of a space of possible worlds—as opposed to a mere domain
or set of worlds—it is natural to reinterpret this ‘restricting’ function of conditionals.
Instead of simply replacing the domain of possible worlds with the subset of the
original domain where the conditional’s antecedent is true, we may conceive of the
conditional as replacing the original space of possible worlds with the sub-region of
the space where the conditional’s antecedent is true.
Where the space of worlds has no more structure than a simple set of worlds,
the sub-region of the original space will simply be the subset where the antecedent
is true—just as on Kratzer’s original proposal. However, where the space of worlds
has the structure of a probability distribution, replacing the space with the sub-
region where the antecedent is true is equivalent to replacing the original probability
distribution with the result of conditionalizing it on the antecedent.
According the account that I have proposed here, the semantics of ‘ought’ involves
two spaces or domains of possible worlds—the domain of metaphysically possible
worlds that is fixed by the function f , and the space of epistemically possible worlds E.
The antecedent of the conditional will restrict one of these spaces of worlds; but it may
be up to the particular conversational context to determine which of these two spaces
is restricted in this way.
So, some conditionals will restrict the domain of metaphysically possible worlds
f (w) to the subset of that domain where the antecedent is true; but other conditionals
will restrict the space of epistemically possible worlds E to the sub-region of that
space where the antecedent is true. Just to give them labels, I shall call the first sort
of conditional ‘ought’ the ‘metaphysical conditional’, and I shall call the second sort of
conditional the ‘epistemic conditional’.
The truth conditions of these two kinds of conditionals can be specified as follows:
. Metaphysical: For any two propositions p and q: ‘[If p] q’ is true at w iff q [f / f  ] is
true at w—where q [f / f  ] is the result of uniformly replacing f in q with f  , which
is the function from any possible world w to the subset of f (w ) where p is true.
. Epistemic: For any two propositions p and q: ‘[If p] q’ is true at w iff q [E/ E  ]
is true at w—where q [E/ E  ] is the result of uniformly replacing E in q with E  ,
which is the sub-region of E where p is true.
It is clear that the clause for this second epistemic conditional requires that the space
of possible worlds E must itself be part of the semantics of the sentence that expresses
the proposition q. It is only if E is part of the semantics that the effect of embedding
this sentence within a conditional can be to restrict this space E to the sub-region of
the space where the antecedent proposition p is true.12

12 It is at this point that my account differs most clearly from the account of Bronfman and Dowell (,
chapter ). In their account, the ordering source g (interpreted as a function from worlds of evaluation to
orderings on the worlds in the modal base) is a single parameter, rather than a pair of parameters E and V.
(Following Dowell (), they propose that with the subjective ‘ought’, but not the objective ‘ought’, the
ordering source is sensitive to the contents of some salient body of information at the world of evaluation.)

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

objective and subjective ‘ought’ 

The truth conditions that I have assigned here to the metaphysical conditionals
involving ‘ought’ are in effect the same as those that were assigned to the so-called
dyadic ‘ought’-operator by the classical deontic logicians such as Åqvist () and
Lewis (). On the other hand, the truth conditions that I have assigned to the
epistemic conditionals involving ‘ought’ have the effect of replacing the probability
distribution E that would be involved in the semantic value of the consequent of
the conditional if it appeared unembedded with the result of conditionalizing that
probability distribution on the antecedent.
For an example of the metaphysical conditional, consider the familiar examples of
the ‘second best “ought” ’ that have been used to illustrate the dyadic ‘ought’-operator.
Suppose that an adviser is remonstrating with a recalcitrant advisee. First, the adviser
says ‘You ought not to shoot up heroin’, and then when the advisee indicates that he
may not follow this advice, the adviser continues, ‘And if you do shoot up heroin, you
ought to shoot up with clean needles’.
If these statements involve the practical ‘ought’, focussed on the advisee’s situation
at the time of the utterance, then the adviser’s first statement is true because out of all
the worlds that are practically available to the advisee at the relevant time, the worlds
where the advisee acts in a maximally choiceworthy way are all ones where he does
not shoot up heroin. The second statement is true because out of all the worlds that
are practically available to the advisee at the relevant time and the advisee does shoot
up heroin, the worlds where the advisee acts in a maximally choiceworthy way are all
worlds where he shoots up with clean needles.
For an example of the epistemic conditional, consider the following variant of Frank
Jackson’s () three-drug case—specifically, a four-drug case. There are two drugs, 
and , such that it is known for certain that one of these two drugs will completely
cure the patient while the other drug will kill him, but unfortunately it is unknown
which of the two drugs will cure the patient and which will kill him. In addition, there
are two other drugs,  and , each of which will effect a partial cure, but one of which
will have an unpleasant side-effect—though it is not yet known which drug will have
that side-effect. Suppose that the patient is about to have a test: it is known that if the
test is negative, it is drug  that will have the unpleasant side-effect, while if the test is
positive, drug  will have the unpleasant side-effect. Then it is true to say ‘If the test
result is positive, we should give the patient drug .’
This statement is true because we give drug  in all possible worlds in the relevant
domain in which we take the course of action that maximizes expected choicewor-
thiness, according to the probability distribution that results from our current system

They interpret deontic conditionals as involving two modal operators, with the conditional’s antecedent
restricting the domain of an operator that has wide scope, and the deontic modal operator having narrow
scope, embedded inside the scope of the other operator. While I find my account marginally more plausible
than theirs, I cannot compare the advantages of the two approaches here; further work is clearly needed on
this point.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 ralph wedgwood

of credences by conditionalizing on the proposition that the test result is positive.13

This seems to give the intuitively correct truth conditions for this sentence—which
supports the semantic proposal that I am making here.
This is not to say that every conditional with an ‘ought’ in the consequent conforms
to one of these two patterns. Suppose for example that we are considering another
agent—call her Alice—and wondering which of two courses of action, A and B, it is
most rational for her to take. Then we might say: ‘I’m not sure which of these two
courses of action Alice regards as preferable. But if she thinks that A is preferable to B,
she should do A’.14 This seems to me to be an ordinary indicative conditional—to be
explained in the same way as all other indicative conditionals (whatever that way is). In
this case, the local context in which this occurrence of ‘should’ occurs makes a certain
possible epistemic perspective E salient—specifically, the perspective that Alice would
have if she thinks that A is preferable to B. The semantic value of this occurrence of
‘ought’ is to be analysed in terms of this epistemic perspective E.

 Objections and Replies

There are two main things that I have done in this chapter. First, I have set out a related
family of truth conditions—in effect, the truth conditions that the sentences expressing
a family of concepts might have. These truth conditions naturally divide into those that
belong to more ‘objective’ concepts and those that belong to more ‘subjective’ concepts,
depending on whether the probability distribution involved in these truth conditions
is the omniscient probability distribution, or a probability distribution that in some
way reflects a greater degree of ignorance and uncertainty.
Secondly, I have suggested that these truth conditions capture the semantic values of
uses of the English deontic modals such as ‘ought’ and ‘should’. In the space available,
this suggestion could not be defended in detail. The suggestion seems plausible to me,
but I concede that much more empirical evidence about the semantic intuitions of
competent speakers of English would have to be considered to give a full defence of
this suggestion. If this suggestion seems less plausible to some readers than it does
to me, then the concepts that I have highlighted—even if they are not expressed in
English by deontic modals such as ‘ought’ and ‘should’—may at least turn out to be
useful for various theoretical purposes.

13 This interpretation of these epistemic deontic conditionals seems to me to avoid the problems for rival
accounts that are canvassed by Nate Charlow (). Those rival accounts all represent the relevant body of
information by means of the ‘modal base’—that is, the propositions that are true throughout the relevant
domain of worlds f (w); my account represents this body of information in a fundamentally different way—
by means of the probability distribution E that, together with the relevant value V, determines the ordering
of the worlds in this domain. In this way, my account agrees with Charlow’s central point, that a good
semantic account must make provision for conditionalizing, not only the modal base, but also the relevant
ordering of the worlds. It is precisely for this reason that I propose that there are two different kinds of
deontic conditionals.
14 I owe this example to Alex Silk.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

objective and subjective ‘ought’ 

In this final section, I shall offer a brief reply to an objection that might be raised
against my suggestion that this family of truth conditions really captures the semantic
values of uses of ‘ought’. Some readers may worry that my account seems to build
in some controversial assumptions about rational choice into the very semantics of
‘ought’. In some unpublished notes quoted by Kai von Fintel (, p. ), Kratzer
objects to accounts that do this, asking rhetorically: ‘Why pack information about
rational decision making into the meaning of modals?’
Strictly speaking, however, I have not in fact said anything about rational decision-
making here. Admittedly, my account makes use of the general idea of the expected
value of a proposition, which is an idea that is also invoked in many theories of rational
choice—for example, by those theories that imply that a rational choice must maximize
expected utility. My account has in fact made absolutely no mention of utility at all.
(There may be contexts where the value function V involved in the semantic value of
an occurrence of ‘ought’ is a utility function; I take no stand on the issue.) Still, it may
seem that the mere fact that I have made use of the general idea of the expected value
of a proposition brings my account too close to ‘packing information about rational
decision making into the meaning of modals’.
There are two main problems that might be alleged to affect accounts of the
semantics of ‘ought’ that appeal to the idea of the expected value of a proposition. First,
one might think that this idea is too controversial and too technical to be implicit in
the linguistic competence of ordinary speakers. Secondly, one might think that there
are some specific cases that cannot be handled in an intuitively acceptable way by any
such account.
My account is not vulnerable to the first problem, since I have explicitly distanced
myself from any attempt to explain linguistic competence in terms of an implicit
grasp of the truth conditions that I have described. It is undeniably an important
question what linguistic competence consists in, but unfortunately I cannot address
that question here. At all events, it is far from obvious that the truth conditions
that I have described are incompatible with any plausible account of linguistic
competence.
The second problem is potentially more serious. For example, consider an agent—
call him John—who harbours grave doubts about all views according to which one
should choose options that maximize some kind of probabilistic expectation of some
kind of value. Instead, John is attracted to a rival theory of rationality, such as the
maximin theory—according to which in every choice situation, one should choose
one of the options whose worst possible outcomes are at least no worse than the worst
possible outcomes of the available alternatives. Suppose that the most plausible version
of the expected-value theory would favour John’s choosing act A, and the maximin
theory would favour his choosing act B. It would seem true to say ‘For all John knows,
he ought to choose B’. Can we really handle cases of this sort in a satisfactory manner
if the notion of maximizing expected value is built into the semantics of ‘ought’ as
I propose?

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 ralph wedgwood

These cases seem hardly typical of the normal use of deontic modals, since they
concern the use of these terms by theorists or philosophers in talking about other
theorists. It is questionable whether such esoteric uses are the most reliable evidence
for a theory of the meaning of words in a natural language. Nonetheless, a number of
recent writers seem to have been moved by cases of this sort. For example, Jennifer
Carr () has proposed that there should be a separate parameter in the semantics
of deontic modals for a decision rule: in some contexts, this decision rule might be
maximizing expected utility, but in other contexts, it might be some other decision
rule, such as the maximin rule.
In my view, however, this manoeuvre greatly complicates the semantic account
of deontic modals, in a way that is far from obviously warranted by the linguistic
evidence. Statements of the form ‘For all John knows, p’ are hard to interpret. In some
contexts, it seems that it could be true to say such things as ‘For all Pythagoras knew,
there are only finitely many prime numbers’. To explain why this sentence is true, it is
surely not necessary to argue that there are some possible worlds where there are only
finitely many prime numbers! In a similar way, it should be possible to explain why
the sentence ‘For all John knows, he ought to choose B’ without supposing that there
is any context such that the notion of maximizing expected value plays no role in the
correct account of the truth conditions that an ‘ought’-sentence has in that context.
A similar objection is raised by Fabrizio Cariani (, p. ), who focusses on cases
where a sentence involving ‘ought’ is embedded inside a larger sentence, such as ‘John
believes that he ought to choose B’, which ascribes a belief to a heterodox theorist like
John. Cariani argues that an account like mine will have difficulties explaining why
this belief-ascription is true.
In fact, however, my account has no difficulty providing such an explanation. It
seems most promising to link my account of the semantics of ‘ought’ with a relational
analysis of belief-ascriptions. According to this relation analysis, the belief-ascription
is true because John stands in the belief-relation to a content of the appropriate kind
that can be expressed in this context by the embedded sentence ‘he ought to choose B’.
According to Cariani (ibid.), ‘the appeal to the relational analysis is merely evasive,
unless it is complemented by an account of what content is expressed by a deontic
sentence in a given context’. But it is surely not obvious that in order to defend my
account of the semantics of ‘ought’, I need to commit myself to a full account of
the semantics of belief-ascriptions here. It is enough if I can make it plausible that
it is possible in principle to give an illuminating analysis of belief-ascriptions that
harmonizes with my account of ‘ought’.
So, to fix ideas, I shall suggest a possible analysis of this sort. I am not firmly
committed to all the details of this suggestion; the suggestion is included here
only to respond to Cariani’s objection. According to this suggestion, in this context
the embedded sentence ‘he ought to choose B’ expresses a Fregean proposition—
presumably, a Fregean proposition that John could express in an appropriate context
by uttering the sentence ‘I ought to choose B’.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

objective and subjective ‘ought’ 

It seems clear that this Fregean proposition is capable of being true or false; in
that sense, this proposition has truth conditions. Presumably, to make this sugges-
tion compatible with my account, this Fregean proposition must have has the same
(extensional) truth conditions that my account assigns to the ‘ought’-sentence in this
context. However, there is no reason to think that John himself must entertain this
Fregean proposition by means of explicitly thinking of these truth conditions. John
must latch onto this proposition somehow, but it is not necessary for him to latch onto
the proposition by means of an implicit grasp of the most systematic account of the
truth conditions of sentences that express this proposition. Instead, I suggest, John
latches onto this proposition by deploying some mode of presentation of the property
that, according to my account, the embedded proposition that John chooses B would
have to possess for the whole ‘ought’-proposition to be true. The content of John’s
belief is a Fregean proposition that applies this mode of presentation to this embedded
proposition. Exactly how John grasps this mode of presentation of this property is a
delicate question, but it seems possible that he could grasp this mode of presentation
without having any awareness of how the property is analysable in terms of a domain
function f , a probability distribution E, and a value function V.
A further concern that Cariani raises is whether my account will make it the case
that John’s beliefs are ‘logically inconsistent’. In principle, there are many views that
philosophers have defended that are inconsistent with the correct semantics for some
natural-language expressions. For example, some philosophers have defended the view
that there are deontic dilemmas—cases in which it is simultaneously true that you
ought to do A and also that you ought not to do A. According to almost all the accounts
of ‘ought’ that semanticists have proposed, these philosophers’ views are inconsistent
with the correct semantics for ‘ought’. In principle, I accept that cases could be devised
in which the beliefs of John the maximin theorist would be similarly ‘inconsistent’.
However, since beliefs can be in this sense ‘inconsistent’ in highly non-obvious ways,
I do not see how this counts as any sort of objection to my account.
In general, cases where an ‘ought’-sentence of this kind is embedded within a
hyperintensional context such as ‘John believes that. . . ’ or ‘For all Barbara knows. . . ’
raise so many problems of their own that they seem not to provide firm grounds for
objecting to my account. Moreover, so far as I can see, there is no clear case where
we have the intuition that a sentence that has ‘ought’ as the dominant operator—
for example, a sentence of the form ‘Barbara ought to do A’—is true, in a way that
clearly cannot be handled by the account that I have proposed.15 In short, the linguistic
evidence does not clearly undermine my account of the semantics of ‘ought’.
Even though my account is unified in that the notion of maximization features in
my account of the semantic value of every occurrence of ‘ought’, it is in other ways

15 Indeed, I suspect that in any case where we are tempted to assert a sentence of the form ‘Barbara ought
to do A’, on the grounds that A is what is recommended by Barbara’s non-maximizing theory, our assertion
is either false, or else true only when this occurrence of ‘ought’ is understood as the purpose-relative ‘ought’,
relativized to the goal of conforming to the non-maximizing theory in question.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 ralph wedgwood

an immensely broad and flexible account of the term. Many other philosophers of
language and metaethicists have proposed much narrower interpretations of ‘ought’,
which dramatically under-predict many of the readings of ‘ought’ that seem genuinely
available.16 By contrast, the range of truth conditions that I have identified in Sections
– above is much wider. So my suggestion—that all the truth conditions identified
here belong to concepts that can be expressed by ‘ought’ in ordinary English—implies
that these deontic modals, such as ‘ought’ and ‘should’, are capable of expressing this
wide range of concepts, depending on the particular context in which they are used.
In this way, my suggestion clearly runs the opposite danger—that of over-predicting
the readings of ‘ought’ that are available. For example, the schema that I outlined in
Section  seems to predict that there is a practical ‘ought’-concept that is indexed to the
situation that I am in right now, and to the space of epistemically possible worlds that
corresponds to Julius Caesar’s state of information on that fateful morning of  March
 BC. It is doubtful, to say the least, whether there is any way of using such terms as
‘ought’ in English that will express this concept.
It does not seem clear to me that this point grounds any decisive objection to my
approach. We should concede, it seems to me, that this concept really exists, but that we
have no natural way of expressing it in English (or in Latin, or in any natural human
language), largely because of the very limited interest that this concept would have
for us. Admittedly, the suggestions that I have made in this paper would need to be
supplemented in order to explain why there is no natural way of using our natural-
language terms to express many of these concepts. But I see no reason to think that
such supplementation will prove impossible.
In general, of the two dangers that face such interpretations of natural-language
expressions, the danger of over-predicting the readings that are available seems less
grave than the danger of under-predicting such readings, since it will often be possible
to supplement an account that looks likely to over-predict the available readings of a
term with some further account that explains why those readings will not in fact be
available in any real conversational context. An account that underestimates the range
of concepts that a term can express, on the other hand, seems to admit no way of being
supplemented in order to rectify this deficiency. So there are some general reasons to
be optimistic that the sort of approach that I have sketched here will help us to achieve
a better understanding of these deontic modals such as ‘ought’ and ‘should’.

Acknowledgments
This paper was originally written and posted on my website in the summer of . In the spring
and summer of , it was presented as talks at Berkeley and at Edinburgh; I am grateful to
the members of those audiences for helpful comments. Finally, in revising the paper in ,

16 For an example of an interpretation of ‘ought’ that is dramatically narrower than mine, see Judith
Thomson ().

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

objective and subjective ‘ought’ 

I benefited greatly from some highly illuminating comments from Nate Charlow, Matthew
Chrisman, Alex Silk, and Malte Willer.

References
Åqvist, Lennart () Good Samaritans, Contrary-to-Duty Imperatives, and Epistemic Obli-
gations. Nous. . pp. –.
Björnsson, Gunnar and Finlay, Stephen () Metaethical Contextualism Defended. Ethics.
. . pp. –.
Brandt, R. B. () Ethical Theory. Englewood Cliffs, NJ: Prentice Hall.
Bronfman, Aaron and Dowell, J. L. () Contextualism about Deontic Conditionals. In
Charlow, Nate and Chrisman, Matthew (eds.) Deontic Modality. Oxford: Oxford University
Press.
Cariani, Fabrizio () Deontic modals and probabilities: One theory to rule them all? In
Charlow, Nate and Chrisman, Matthew (eds.) Deontic Modality. Oxford: Oxford University
Press.
Carr, Jennifer () Deontic modals without decision theory, Proceedings of Sinn und Bedeu-
tung , Paris, September –. pp. –.
Charlow, Nate () What we know and what we do. Synthese. . pp. –.
Dowell, J. L. () Flexible Contextualism about Deontic Modals: A Puzzle about Information-
sensitivity. Inquiry. . –. pp. –.
Easwaran, Kenny () Regularity and Hyperreal Credences. Philosophical Review. . .
pp. –.
Ewing, A. C. () The Definition of Good. New York, NY: MacMillan.
Fintel, Kai von () The best we can (expect to) get? Challenges to the classic semantics for
deontic modals. Presented at the Central Meeting of the American Philosophical Association.
February , . [Online] Available from: http://mit.edu/fintel/fintel--apa-ought.pdf
[Accessed  November .]
Gibbard, Allan () Truth and Correct Belief. Philosophical Issues. . pp. –.
Jackson, Frank () A probabilistic approach to moral responsibility. In Barcan Marcus, R.,
Dorn, George, and Weingartner, Paul (eds.) Logic, Methodology, and Philosophy of Science
VII. pp. –. Amsterdam: North-Holland.
Jackson, Frank () Decision-Theoretic Consequentialism and the Nearest and Dearest Objec-
tion. Ethics. . . pp. –.
Jackson, Frank and Pargetter, Robert () Oughts, options, and actualism. Philosophical
Review. . pp. –.
Kolodny, Niko and MacFarlane, John () Ifs and Oughts. Journal of Philosophy. .
pp. –.
Kratzer, Angelika () Modals and Conditionals: New and Revised Perspectives. Oxford: Oxford
University Press.
Lewis, David K. () Counterfactuals. Oxford: Blackwell.
Parfit, Derek () Reasons and Persons. Oxford: Clarendon Press.
Schroeder, Mark () Ought, Agents, and Actions. Philosophical Review. . . pp. –.
Sidgwick, Henry () The Methods of Ethics th edition. London: Macmillan.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 ralph wedgwood

Silk, Alex () Evidence-Sensitivity in Weak Necessity Deontic Modals. Journal of Philosoph-
ical Logic . . pp. –.
Swanson, Eric () Ordering Supervaluationism, Counterpart Theory, and Ersatz Fundamen-
tality. Journal of Philosophy. . . pp. –.
Thomson, J. J. () Normativity. Chicago, IL: Open Court.
Wedgwood, Ralph () The Nature of Normativity. Oxford: Clarendon Press.
Wedgwood Ralph () The ‘Good’ and the ‘Right’ Revisited. Philosophical Perspectives. .
pp. –.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i


‘Ought’: Out of Order
Stephen Finlay

The semantics of ‘ought’ and related modal verbs in natural language is not for the faint
of heart. A central desideratum is a unifying theory accommodating the many diverse
uses, both normative (moral, instrumental, rational, legal, etc.) and nonnormative
(logical, metaphysical, nomological, epistemic, dispositional, etc.) This places it at
the intersection of many issues in different subfields of philosophy, linguistics, and
logic, at least, and so presumably no single scholar could acquire all relevant expertise.
Angelika Kratzer’s (David Lewis-influenced) ordering semantics, which today is widely
regarded as orthodoxy, is therefore a remarkable and audacious achievement. Whereas
the Lewis–Kratzer semantics was developed from a primary focus on counterfactuals
and then extended to normative and other uses, this chapter offers a metaethicist’s per-
spective, taking as its primary cues the behavior of English modal verbs in normative
sentences.
When I first developed my own theory of the meaning of ‘ought’ in – (initially
published as “Oughts and Ends” in ), like other metaethicists I was ignorant
of Kratzer’s seminal work on modals in linguistics (, , ). This defect in
metaethics has since been corrected, and today philosophical work on normative
modals typically adopts the Lewis–Kratzer framework. But this new attention has
uncovered serious difficulties for its application to normative sentences, stimulating
an explosion of work at the intersection of metaethics and linguistics. Whereas these
efforts at repairing or replacing the Lewis–Kratzer framework invariably propose
additional semantic complexity, I’ll argue here that these difficulties are all avoided
more straightforwardly and naturally by the simpler semantics I’ve advanced in
“Oughts and Ends” and subsequent work (a, b, ), a version of the dyadic
semantics which the Lewis–Kratzer approach supplanted. I’ll argue that Kratzer’s
signature innovation of an ordering source parameter, though ingenious, introduced
unnecessary complexity into the semantics of modals that we’re better off without, at
least for English auxiliaries such as ‘ought’, ‘must’, and ‘may’. The simpler semantics has
the resources to address the issues motivating ordering semantics, without having its
weaknesses.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 stephen finlay

After sketching the central features and motivations of the ordering semantics in
Section , I’ll programmatically investigate problems arising from (i) instrumental
conditionals (Section )—which motivate my rival approach to normative modality
(Section ), (ii) gradability and “weak necessity” (Section ), (iii) information-
sensitivity (Section ), and (iv) conflicts (Section ). I will argue that a simpler dyadic
approach addresses these problems at least as well as the ordering semantics, given
three moves: (a) an end-relational analysis of normative modality, (b) an analysis of
‘ought’ or “weak necessity” in terms of less-than-universal quantification (most rather
than all), and (c) appeal to the same resources of conversational pragmatics as utilized
by Kratzer. I conclude (Section ) by addressing metasemantic objections to my claims
about simplicity, with observations about what we should want from a semantics for
‘ought’.

 Ordering Semantics: Dyadic to Polyadic

By orthodoxy and with great success, modal auxiliaries are analyzed as quantifiers
over possibilities. Necessity verbs such as ‘must’ and ‘have to’ are universal quantifiers
meaning roughly in all possibilities, while possibility verbs such as ‘may’, ‘might’, ‘can’,
and ‘could’ are existential quantifiers meaning roughly in some possibilities. Different
kinds of modality result from defining the relevant domain differently: by consistency
with physical laws (nomological), bodies of evidence (epistemic), codes of laws or rules
(deontic), etc.
Dyadic semantics classically captures this variability by postulating that in addition
to their scope, modal verbs take a second argument for a domain restrictor. Following
Kratzer, this modal base is identified as a conversational background, usually implicit
in the context but articulable with a ‘given. . . ’ clause. Formally, a modal base is a
function f from a context w to a “premise-set” of propositions f (w). What is necessary
in w given f is what logically follows from f (w), and what is possible in w given f is what
is logically consistent with f (w). On a simple dyadic approach, this argument-place is
undiscriminating. It can take backgrounds that are (i) “realistic” or factual, like given
the circumstances at time t or given what subject s knows, (ii) “nonrealistic”—including
normative backgrounds—like given s’s desired ends, given what is prescribed by the code
of laws L, or given what s believes, or (iii) mixed. But Kratzer insists (beginning with
her ) that “realistic and normative backgrounds need to be kept separate” (,
p. ). She offers two reasons.
First, her “most important argument” is the need to accommodate the gradability
of modality, as in degrees of deontic ideality, epistemic likelihood, and counterfactual
closeness. A metaethically important case is the modal force of ‘ought’ (or ‘should’),
which is stronger than that of possibility modals such as ‘may’, but weaker than that of
necessity modals such as ‘must’, and is therefore labeled ‘weak necessity’.
Second, normative and other nonrealistic modalities often involve inconsistency:
ideals can conflict either with each other (e.g. incompatible goals or contradictory

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

‘ought’: out of order 

laws), or with salient circumstances (e.g. what one ought to do given violation of some
norm, or “contrary-to-duty obligations”). Since anything follows from an inconsistent
premise-set, a dyadic semantics predicts that relative to such backgrounds must p is
true for any arbitrary p, and, since nothing is consistent with an inconsistent set, that
may p is false for any arbitrary p. But inconsistent ideals don’t seem to render normative
claims trivially true or false.
Kratzer’s ingenious solution to both problems is to suggest that modal verbs take a
second conversational background, an ordering source. This is to move from a dyadic
to a polyadic semantics (although ‘dyadic’ is often used in the literature, confusingly,
where ‘polyadic’ is meant). This ordering source is also a variable function, g, from a
context w to a premise-set of propositions g(w), comprised of goals, ideals, laws, etc.
But rather than restricting the domain, g orders it, roughly according to how closely
the possibilities approximate the premise-set. The modal force of a verb consists in
how it quantifies over a subset of possibilities picked from the ordering by a selection
function. Simplifying slightly, must p says, relative to the f and g in w, that p is true
in all possibilities consistent with f (w) that are highest-ranked by g(w), while may p
says that p is true in some possibilities consistent with f (w) that are highest-ranked by
g(w). In this paper I use the term ordering semantics narrowly to refer only to theories
positing an ordering source parameter.
These two kinds of background are utilized in different ways to account for different
modal flavors. The modal base is stipulated always to be realistic, on Kratzer’s division
of labor. Normative or deontic flavors are generated by normative ordering sources,
whereas bare alethic (e.g. logical, metaphysical) modals have empty ordering sources,
so that all possibilities in the domain are equally ranked. Other modalities (epistemic,
counterfactual, etc.) are distinguished by particular combinations of backgrounds.
We’ll see that normative modalities pose significant difficulties for this shift from
dyadic semantics to polyadic ordering semantics.

 Instrumental Conditionals
Consider instrumental conditionals,1 such as,
() If you want to go to Harlem, you must take the A train.
() If you are (going) to go to Harlem, you must take the A train.
() (In order) to go to Harlem, you must take the A train.

1 (Some of) these are commonly called ‘anankastic conditionals’ by linguists and ‘hypothetical imper-
atives’ by philosophers; both labels are misnomers. (a) ‘Anankastic’ signifies necessity/compulsion, but
necessity is neither necessary (e.g. “existential anankastics” such as ‘If you want to go to Harlem, you could
take the A train’) nor sufficient (e.g. ‘If you are in Harlem, then you must be in New York’). (b) These
sentences have declarative rather than imperative mood.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 stephen finlay

(I begin with ‘must’, as ‘ought’ introduces additional complications.) Surface differ-

ences notwithstanding, these are widely agreed to have instrumental readings as
saying, approximately, that taking the A train is a necessary means to going to Harlem,
and to have the normative flavor of a teleological modal, with the force of hypothetical
advice.
The Lewis–Kratzer semantics has difficulties accounting for these conditionals,
as first observed by Kjell Johan Sæbø (). This might seem a minor nuisance
posed by an obscure subclass of sentences; Sæbø suggests they have “attracted little
attention in linguistics and philosophy” (, p. ). I believe this impression is
mistaken, for the reason that instrumental conditionals are key to analyzing normative
modality quite generally. On a controversial though perennially popular view (e.g.
Foot ), every normative or deontic ‘ought’ (‘must’, etc.) sentence is at least implicitly
instrumental in some way, even moral oughts. If this view is correct, then a satis-
factory treatment of instrumental conditionals is essential for a successful semantics
for normative modals. I’ll first sketch the problems these sentences pose for the
ordering semantics, and then briefly explain how I’ve argued, in previous work, that
these can be naturally resolved by returning to a dyadic semantics. I won’t attempt
here to establish the success of this proposal; interested or skeptical readers may
examine my earlier work. Rather, I shall make the case that this simpler semantics
motivated by the need for a satisfactory treatment of instrumental conditionals
also compares favorably with the ordering semantics with respect to the further
features of gradability, conflicts, and information-sensitivity—despite (two of) these
being Kratzer’s primary motivations for adopting ordering semantics over a dyadic
framework.
Focus first on ‘If s wants . . .’ sentences such as (). By Kratzer’s own influential
“restrictor” semantics for conditionals, ‘if p’ functions to update the modal base f
with the antecedent p, yielding the restricted domain f +p(w). This suggests reading
() as saying, for salient f and g, that in all possibilities consistent with f in which you
want to go to Harlem that are highest-ranked by g, you take the A train. However,
there’s no guarantee that the salient ordering source g will rank go-to-Harlem worlds
highest, so restricting to worlds where you have this desire doesn’t entail that in
all the best remaining worlds you take the A train, even if it’s the only way to get
to Harlem. Intuitively, the goal you go to Harlem should instead be determining
which possibilities are best. This presents the ordering semantics with (at least) two
puzzles:
(i) How can the antecedent clause influence the ordering source?
(ii) How does the reference to desire or purpose contribute to sentences like ()—
since the relevant goal is the “internal antecedent”, (you) go to Harlem?
For reasons of space I won’t focus on the second puzzle here. There are abundant
grounds for thinking that the reference to desires doesn’t make a straightforwardly

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

‘ought’: out of order 

compositional contribution to truth conditions,2 as most writers have concluded and

I will here assume.3
Sæbø responds to the first puzzle by proposing, radically, that ‘if ’ is ambiguous, and
can update either the modal base or the ordering source. On its instrumental reading,
the antecedent of () would update the background ordering source g with the internal
antecedent, that e : you go to Harlem. This fix isn’t sufficient, however. Merely updating
an ordering source with the goal e doesn’t guarantee that all the g + e -best possibilities
will be go-to-Harlem worlds, because of a problem from inconsistent goals: e could be
incompatible with other goals in the salient g, like going to Hoboken (perhaps you
have a grandmother who lives in each place).4 A sentence such as () could then be
true in such circumstances:
() #To go to Harlem, you may go to Hoboken instead.
The “designated goal” in the antecedent clearly needs to take priority over any
other goals or ideals. One suggestion is that it interacts with the preliminary ordering
source, g , by eliminating anything inconsistent with it, thereby guaranteeing that the
g+e-best possibilities will all be go-to-Harlem worlds.5 However, this encounters a
problem from consistent goals. Suppose you happen to have the desire or goal to kiss
the footballer Ruud van Nistelrooy, who will be on the A train. Then it’s predicted that
() may be true even if there are other, equally reliable ways to go to Harlem (Huitink
)—and worse, that () may be true (Nissenbaum );
() #To go to Harlem, you have to kiss Ruud van Nistelrooy.
To solve these two problems of inconsistent and consistent goals, Janneke Huitink
() proposes that the antecedent functions to constitute the ordering source, rather
than merely update it. Since ordering sources tolerate inconsistencies by design, and
conditionals normally function to update backgrounds, this proposal seems ad hoc
and anomalous.6 It also throws the baby out with the bathwater, since instrumental

2 First, the contrapositive of () is apparently ‘If you don’t take the A train, you can’t go to Harlem’, rather
than ‘If you don’t take the A train, you can’t want to go to Harlem’ (Sæbø , p. ).
Secondly, () seems roughly equivalent to () and (). While these other constructions can sometimes
indicate an agent’s purpose, this doesn’t seem to be their essential function here (Finlay b). Nissenbaum
() proposes the gloss: you must take the A train with the purpose of going to Harlem; von Fintel and
Iatridou () offer convincing objections.
Thirdly, these sentences exhibit anomalous detaching behavior, resisting modus ponens. One can coher-
ently assert ‘If Henry wants to be a famous mass murderer, then he has to kill a lot of people’, while refusing
to assert ‘Henry has to kill a lot of people’ upon learning that Henry has this disturbing desire.
3 Condoravdi and Lauer () offer a dissenting view, turning on a reading of ‘want’ that I would argue
is artificial.
4 von Fintel and Iatridou (), von Stechow et al. ().
5 Proposed in an early version of von Fintel and Iatridou ().
6 As von Stechow et al. () point out, it is also (almost) functionally equivalent to eliminating ordering
sources altogether, as I propose below.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 stephen finlay

conditionals do sometimes exhibit sensitivity to other goals or ideals, at least in the

case of ‘ought’ variants such as (), though arguably not for ‘must’.7
() To go to Harlem, you ought to take the A train. It’s cheapest.
The favored solution to these problems, proposed by Kai von Fintel and Sabine
Iatridou (), introduces further semantic complexity by adopting Kratzer’s idea
of multiple ordering sources. They suggest that ‘ought’ differs from ‘must’ in taking
a secondary ordering source, having the form ought f,g,g (p) with roughly the modal
force: in all possibilities consistent with f that are highest-ranked by g  that are highest-
ranked by g  . For instrumental conditionals, the designated goal determines the
primary ordering source g , with other salient goals or ideals providing the secondary
ordering source g . () therefore says that in all possibilities highest-ranked with
regard to going to Harlem that are also highest-ranked by the secondary ideals (e.g.
cost, comfort), you take the A train. This analysis solves many problems, though
still predicts the truth of () in the van Nistelrooy scenario, which they suggest—
implausibly, I think—is infelicitous rather than false:
() #To go to Harlem, you ought to kiss Ruud van Nistelrooy.
However, a problem of unattainable goals remains for any analysis that assigns
the designated goal to an ordering source.8 By design, ordering sources needn’t be
consistent with modal bases. Suppose there are no possibilities consistent with f in
which you go to Harlem; e.g. you’re locked up in jail. The singleton set of ideals {you
go to Harlem} still generates an ordering: all possibilities in the modal base will be
equally best. So the ordering semantics seems to predict incorrectly that the following
sentences would be true in this context:
() #To go to Harlem, you must not go to Harlem.
() #To go to Harlem, you must stay in jail.
Observe that these sentences seem false rather than merely infelicitous, unlike their
purported translation,
() In all accessible possibilities which are closest to the ideal that you go to
Harlem, you don’t go to Harlem/ you stay in jail.
My claim here is not that these problems are fatal for ordering semantics. One
might for example postulate a (controversial but popular) “diversity condition” in the

7 This is contested; see discussion of the “Chinese train” example in von Stechow et al. (). The correct
response, I believe, appeals to ellipsis in the antecedent (von Fintel and Iatridou , p. ). The existence
of inferior means justifies challenges to and retractions of ‘must’ claims; e.g. A: “No, you don’t have to take
the Chinese train to go to Vladivostok; you could take the Russian train.”; B: “Yes, but you do have to take
the Chinese train to go to Vladivostok comfortably.”
8 Cf. von Stechow et al. (, p. ).

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

‘ought’: out of order 

semantics,9 requiring that the scope of a modal operator be neither entailed nor ruled
out by f whenever g isn’t empty. But I wish to point out how much more simply
and naturally a dyadic semantic analysis of instrumental conditionals avoids these
problems.
Notice that the challenge confronting ordering semantics here is, effectively, to
explain how a goal in an ordering source could behave as if it were in the modal
base instead. The truth conditions of instrumental conditionals, such as ()–(), seem
sensitive only to possibilities consistent with achieving the designated goal. So suppose
we simply allow the conditional to function in its ordinary way, unambiguously adding
the antecedent’s designated goal to the modal base instead.10 I have argued (especially
Finlay, b, ) that a fully unifying, compositional, and conservative analysis then
comes into view at least for sentences like () with ‘if s is (going) to. . . ’ antecedents,
which also supports promising analyses of other grammatical variants such as () and
().11 Briefly:
The prospective aspect in these antecedent clauses encourages a transparently and
compositionally temporal analysis: to represent an event e as “[going] to” happen is to
project it into the relative future. A sentence such as () therefore grammatically rep-
resents the modality in the consequent clause as being (temporally or metaphysically)
prior to the event e in the antecedent; i.e. the necessity of taking the A train is located
prior to your going to Harlem. This reverses the order we read into conditionals by
default;12 for example, the non-instrumental conditional () is naturally interpreted
as representing the necessity of taking the A train as (temporally) consequent upon
going to Harlem:
() If you go to Harlem, then you must take the A train.
Whereas Kratzer doesn’t herself analyse how modality interacts with temporality, I
maintain that by virtue of the temporal structure of ()–(), ‘must’ there expresses
the necessity of (temporal or metaphysical) pre-conditions for the antecedent’s event
obtaining, and thereby an instrumental (or, roughly, means-end) relationship.
This simple analysis avoids the ordering semantics’ problem from inconsistent goals
(since it requires consistency with the designated goal) and also the problem from
consistent goals (since e.g. kissing van Nistelrooy isn’t a necessary condition for going
to Harlem), while avoiding any ambiguity in the semantics for ‘if ’. Of course, we also

9 E.g. Condoravdi ().

10 Finlay (); see also discussion in von Stechow et al. (, p. ).
11 ‘In order to’ is plausibly an idiomatic device for representing the same reversed temporal relationship
described below. I argue for an analysis of ‘if s wants. . . ’ variants as relevance/instrumental conditional
hybrids (Finlay b, , pp. –), which explains the anomalies in note . Relevance analyses are also
tentatively suggested in von Fintel and Iatridou (), von Stechow et al. ().
12 Sæbø () analyzes this as placing the consequent’s event (the means) prior to the antecedent’s event
(the end), which fails to distinguish instrumental conditionals from many epistemic conditionals; e.g ‘If the
ground is wet, then it has to have been raining.’ In Finlay (b) I endorsed these as jointly necessary and
sufficient conditions for an instrumental conditional; in Finlay () I rather reject Sæbø’s condition.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 stephen finlay

need an account of the contrast between ‘must’ and ‘ought’, and of inconsistencies (e.g.
to address the problem of unattainable goals); here I beg the reader’s patience until
Sections  and , respectively. But if I am correct, then instrumental conditionals are
best analyzed without appeal to any ordering source.

 The Source of Normativity

Can’t the champion of ordering semantics just concede this result, and say that at
least some instrumental conditionals involve an empty ordering source? The problem
is that instrumental conditionals are almost universally agreed to have an at least
minimally normative character—which is largely why moral philosophers puzzle over
“hypothetical imperatives” and Kratzerian analyses have almost universally assumed
that the designated goal must go into an ordering source. So such a concession would
apparently acknowledge that at least some normative modals are best analyzed without
appeal to ordering sources. I’ll now explain why this poses a serious challenge to the
ordering semantics.
Kratzer intends her semantics to be neutral toward the metaethical question of what
makes an occurrence of a modal verb normative. Answering that question, she says
(p.c.), is the philosopher’s job. But the ordering semantics is metaethically committal
nonetheless, in a potentially problematic way. The stipulation that modal bases are
realistic imposes a condition that normativity must enter through an ordering source.
Although sometimes writing as if nonempty ordering sources are also sufficient for
normative modality (e.g. distinguishing ordering sources as “normative backgrounds”
from modal bases as “factual backgrounds”), Kratzer also posits nonnormative order-
ing sources at least in analyzing some epistemic, dispositional, and counterfactual
modals. Rather, a normative flavor supposedly arises from ordering sources with
normative content or selected in a normative way.
The dyadic analysis of instrumental conditionals supports a conflicting view at
least of the source of an instrumentally normative flavor: that it reduces to modality
restricted by some state of affairs in the relative future, or what I call an end,13 as
included in the modal base. A reasonable hypothesis is that our sense for normative
flavor is pragmatically triggered by a sentence’s suitability for practical purposes
such as giving advice or guiding deliberation: uses for which these end-relational
or teleological sentences are especially apt whenever the end is relevantly valued or
desired. What could be more apt for purposes of advice than telling someone what is
necessary for achieving their desired or intended goal?
This leaves open the possibility that Kratzer’s approach is correct for non-
instrumentally normative modality, such as the deontic modalities of morality, law,

13 Paul Portner labels these ‘priorities’; I prefer ‘ends’ both for continuity with ancient philosophical
tradition, and because ‘priority’ suggests rather the importance or weight of an end/norm (e.g. as used in
Horty ).

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

‘ought’: out of order 

and all-things-considered practical reason. Deontic claims certainly differ in impor-

tant ways from ordinary “hypothetical imperatives”. But plausibly, normativity is
one genus of which these are different species. My work in metaethics argues that
normative modality is just end-relational modality, concerning what is necessary,
possible, etc., conditional on the obtaining of some end.14 This would vindicate the
classic view in metaethics that the “categorical imperatives” of morality are really just
special kinds of “hypothetical imperatives”. If this end-relational theory is correct, then
the dyadic analysis of instrumental conditionals suggests that all normative modalities,
even the moral/deontic, might be best analysed without ordering sources.15 , 16
Semanticists may be understandably reluctant to commit to controversial philo-
sophical positions, but metaethical neutrality is here an unattainable luxury. The
Lewis–Kratzer semantics appears incompatible not only with my end-relational the-
ory, but also with popular rival metaethical views. Contrary to analysis of the moral
‘ought’ as relative to ordering sources “represent[ing] the content of a body of laws
or regulations” (Kratzer , p. ), some philosophers argue that the content of
moral laws can’t even be described or conceived except as laws about what we ought
to do, where this ‘ought’ expresses a primitive or irreducible normative relation (e.g.
Boghossian , pp. f). If this is correct, then attempts to accommodate the moral
‘ought’ with a unifying relational semantics will lead to vicious circularity: what we
morally ought to do is whatever is highest-ranked by the laws concerning what we
morally ought to do.
Metaethicists of this stripe also view the extent of moral disagreement as refuting
claims that moral ‘ought’s are relativized to particular laws or ideals. Speakers sub-
scribing to different moral laws seem genuinely to disagree when making moral ‘ought’
claims, which most metaethicists believe wouldn’t be the case if each were just speaking
about what is highest-ranked by their respectively favored laws.17 Other metaethicists
deny that normative claims quantify over possibilities or express propositions at

14 I analyze the prescriptive, “categorical” character of moral utterances as rhetorical, arising from
suppressing reference to the end made salient by the speaker’s rather than the agent’s desires; see especially
Finlay (, chapter ), and compare Harman ().
15 Why is this theory dyadic, if we can distinguish between the end e, and the “realistic” part of the modal
base, f-e? Answer: the modal force of ‘ought’ (i.e. the operation it performs on its inputs) doesn’t differentiate
between e and anything else in f . The end therefore needn’t be identified as a separate parameter; see further
discussion in Section .
16 Two further supporting considerations: () the distinctive deontic logic shared by teleological and
deontic modals reduces to ordinary modal logic given only the assumption of conditionalization on a
hypothetical outcome (the “Kanger–Anderson reduction”). () Teleological and deontic uses of modal verbs
share grammatical features of tense which distinguish them from epistemic and other uses. I argue (b,
, chapter ) that the end-relational analysis explains these compositionally, and suggest that in deontic
uses they function as grammatical indicators of suppressed conditionalization on an end, which triggers
our sense for normativity.
17 In unpublished work, Janice Dowell proposes analyzing moral ‘ought’ with ordering sources invoking
another normative concept, such as: in view of the most important standards. This solution depends on these
further concepts not being themselves relativistic, but I argue (, pp. –) for a parallel treatment. I
address the problem of disagreement in Björnsson and Finlay (), Finlay (, chapter ).

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 stephen finlay

all. Despite the aspiration to neutrality, the ordering semantics may therefore be
compatible with only a narrow range of metaethical theories. Since taking sides
in these debates is unavoidable in pursuing a unified modal semantics, we should
boldly go wherever the linguistic evidence leads—which, I’ll now argue in earnest,
is consistently in the direction of a dyadic semantics and the end-relational theory.

 Gradability and “Weak Necessity”

Kratzer identifies the need to accommodate gradable modality, or degrees of possibility
(goodness, likelihood, typicality, closeness) as her primary argument for ordering
sources. I think it is at least unclear whether such a need exists for English modal
auxiliaries such as ‘must’, ‘may’, and ‘ought’, which don’t naturally take degree modifiers
such as ‘-er’, ‘-est’, ‘most’, ‘slightly’, ‘very’, etc. (Portner ).18 But there may be a
special case for a gradable semantics for ‘ought’ and ‘should’, which in both epistemic
and normative uses lie somewhere between ‘must’ and ‘may’ in logical strength. Must
p entails ought p, while ought p entails may p, and not vice versa. I’ll now investigate
whether ordering sources are needed to account for this distinctive “weak necessity”
of ‘ought’.
From the outset, ‘ought’ also presents a prima facie problem for ordering semantics.
Intuitively, ‘ought’ is to ‘must’ as ‘best’ is to ‘only’ (Sloman ), and so ought p is
commonly glossed as it’s best that p. You must take the A train to go to Harlem if it is
the only available way, and ought to take the A train to go to Harlem if it is the best
available way. But because the ordering semantics analyzes ‘must’ as meaning roughly
in all the best possibilities, finding logical space for ‘ought’ presents a difficulty.
Whereas Kratzer doesn’t ever propose an analysis of ‘ought’ herself,19 various pro-
posals have recently been advanced on behalf of ordering semantics. These all involve
introducing additional semantic complexity—such as multiple ordering sources (von
Fintel and Iatridou ), “merging” operations on these (Charlow ; Katz, Portner
and Rubinstein ), counterfactual conditions (Silk ), additional contrast-set
parameters (Cariani, Kaufmann and Kaufmann , Charlow , Carr ) and
special rules of use (Rubinstein ). I believe these proposals all face serious
obstacles, though I can’t demonstrate that here.20 Instead, I’ll show how the dyadic

18 Kratzer responds that this is a “language-specific fact” and that auxiliaries in some other languages
and some English modal adjectives do take such modifiers (, p. )—which I don’t think justifies
attributing a gradable semantics to English auxiliaries. Admittedly, there are some ways to qualify these
with degree modifiers (see Portner and Rubinstein ), but these seem to apply indiscriminately, even to
uncontroversially nongradable terms such as ‘dead’ and ‘guitar’ (e.g. ‘X is more a guitar than Y is.’).
19 P.c. She does offer an analysis of “weak necessity” (, p. ).
20 For problems for multiple ordering sources, see Rubinstein (). Silk faces the problem that ‘ought’
isn’t synonymous with ‘would have to’, and a nonarbitrary kind of counterfactual condition is yet to be
identified. Rubinstein’s suggestion that ‘ought’ indicates a controversial ordering source conflicts with the
Theseus/miner scenarios below, in which relevant ideals are uncontroversial, and with moral uses of ‘must’,
which are often controversial. See Section  for some objections to other accounts.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

‘ought’: out of order 

framework provides a natural, simpler way of accommodating weak necessity and

degrees of possibility.
Given the equivalence,
‘must’ : ‘ought’ : ‘may’ :: all : ??? : some,
the naïve and obvious hypothesis is that the modal force of ‘ought’ (and ‘should’)
is roughly that of most, which has appropriately intermediate logical strength.21
(This suggestion is precisified below.) This applies intuitively to the epistemic ‘ought’;
e.g. that
() It ought to rain today
says roughly that it rains today in most possibilities consistent with the background
evidence. On the classical theory of probability as a proportion or measure of
possibility-space, () would then be equivalent to saying that it probably rains today
given the evidence, as many writers have accepted.22 “Weak necessity” would therefore
turn out to be greatest probability.23
This simple analysis is usually raised only to be summarily dismissed. The primary
objection offered is the classical theory’s difficulties with infinity (e.g. Portner , pp.
–). On the Kripkean possible world semantics in which the Lewis–Kratzer theory is
modeled, there are infinitely many possible worlds in any typical domain, so the notion
of a count or proportion of possibilities is undefined. But this is merely a choice of
formal model, which, as Kratzer insists (, p. ), shouldn’t itself dictate the content
of semantic theory. Alternative models quantify over possibilities of a more coarsely
individuated nature, corresponding to partitions on sets of possible worlds, which are
more plausibly countable or measurable.24 Note also that the ordering semantics faces
its own problems with infinity. To avoid the artificial Limit Assumption that there
always are best worlds, Kratzer advances definitions for the simplest auxiliaries such
as ‘must’ and ‘may’ that she herself admits are “more complicated than might seem
necessary” (, p. ).25

21 This is called a “common intuition” in Portner (, p. ), and the “traditional view” in Copley
(, p. ), but I’m unaware of anyone else defending it in print. Copley attributes it to Larry Horn, who
reports (p.c.) maintaining the view outside of print. Previously I emphasized probably, tentatively analyzed
in terms of most (a, p. ; , p. ); see also Wheeler (, ).
22 Some common objections: (i) Ought p but not p is acceptable, unlike probably p but not p (e.g. Copley
). Reply: unlike tenseless modal adverbs, auxiliaries such as ‘ought’ needn’t be relativized to full present
evidence (Thomson ; Finlay , p. ; Wedgwood ). (ii) Ought p seems bad when p is only
incrementally more likely than some relevant alternatives. Reply: it may just be infelicitous; compare most
likely, p.
23 Compare Kratzer (, p. ).
24 Finlay (, p. ). (Some writers, such as Cariani, Kaufmann and Kaufmann ; Dowell ;
Charlow , induce such partitions in a possible worlds framework, to solve problems for ordering
semantics; see section ). Another option is to adopt a measure function defined for infinity.
25 Formally:

[[must p]]f ,g = T iff for all u ∈ ∩f (w) there is a v ∈ ∩f (w) such that v ≤g(w) u, and for all z ∈ ∩f (w),
if z ≤g(w) v, then z → p.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 stephen finlay

Reluctance to commit to a controversial theory of probability is understandable,

but this may be another place where modal semantics can’t afford the luxury of
neutrality. Observing that the semantics for ‘ought’ needs to be sensitive to proba-
bilities, multiple writers propose replacing the modal base with an argument-place for
a quantitative probability distribution over possible worlds, usually also replacing the
ordering source with a value function.26 On the classical theory such a parameter is
unnecessary, as a modal base is sufficient to determine a unique probability function.27
I myself am optimistic that this approach can be vindicated,28 but how can semantic
theorizing proceed when faced with such difficulties?
I suggest we bear in mind that our subject is the meaning of natural language.
Whatever philosophical challenges it faces, the classical theory of probability has
excellent claim to be the naïve theory of common sense, which is good reason
not to build anything more sophisticated into our semantics, “beyond the limits of
what the faculty of language provides for everyone” (Kratzer , p. ).29 I find
it implausible that ordinary modal speech and thought quantifies over maximally
fine-grained possible worlds. Given only the information that a group of miners is
trapped in one of two shafts, for example, it’s extremely natural to say that there are
two relevant possibilities, and that relative to this information each therefore has the
probability /. (It also seems natural to represent probabilities geometrically rather
than arithmetically, treating most as a measure of possibility-space rather than as
a count of discrete possibilities.) At the least, I suggest that the theoretical virtues
demonstrated below for the most analysis warrant a sympathetic reconsideration of
these difficulties.
A second objection is that this simple hypothesis can’t be extended to the normative
uses of ‘ought’: what a person ought to do isn’t what they probably do (e.g. Portner
, pp. –). But I argue (in Finlay ) that it actually yields a promising analysis
of normative ‘ought’s if we adopt a contrastivist semantics. This interprets ‘ought’ claims
relative to a contrast set d(w) of mutually incompatible propositions, or options, {r ,
r , . . . , rn }, which we can treat as generated by a conversational background d. (While
this adds a second parameter, it is specific to ‘ought’, for which ordering semantics

A formalization of a basic multiple ordering source semantics for ‘ought’ that avoids the Limit
Assumption:
[[ought p]]f ,g,g = T iff for all u ∈ ∩f (w) there is a v ∈ ∩f (w) such that v ≤g(w) u, and for all z ∈ ∩f (w),
if z ≤g(w) v, then there is a q ∈ ∩f (w) such that q ≤g(w) z and q ≤g(w) z, and for
all r ∈ ∩f (w), if r ≤g(w) q and r ≤g(w) q, then r → p.
Swanson () presents a scenario that requires even further complications to the ordering semantics.
26 For example, Goble (), Wedgwood (, ), Cariani ().
27 Compare Williamson (), Kolodny and MacFarlane (). Probabilistic information can be
included in modal bases where necessary.
28 For an amateur attempt, see Finlay (, pp. –). One constraint on an acceptable model is that
relevant possibilities needn’t be known.
29 This claim is defended in Section . Ordering semantics are often defended against quantitative
frameworks on similar grounds (e.g. Kratzer , p. ).

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

‘ought’: out of order 

generally introduce third or more parameters.30 For this reason, I think it is reasonable
to continue describing this as a move within the dyadic semantics framework.) For the
practical ‘ought’ used in advice and deliberation, this background d is articulated with
a phrase like ‘rather than any other option in the agent’s power to perform at t’. My
proposal is that ‘ought’ has the modal force of comparative probability, meaning in
more possibilities (or: more likely) than any alternative, ri , in d(w).31
To see the application to normative modalities, consider a teleological ‘ought’ like
() To evade arrest, Max ought to mingle with the crowd.
Following the blueprint of the end-relational analysis of ‘must’ above, we assume a
preliminary modal base f of circumstances or evidence, which is then updated with
the end e: that Max subsequently evades arrest. As with the teleological ‘must’ and
‘may’ (and also ability modals, following Kratzer (, p. )), we exclude from f any
facts or information about the agent’s psychological dispositions to choose any one
option over any other. Suppose Max is disposed not to mingle with the crowd because
he falsely believes it wouldn’t succeed; we wouldn’t on those grounds dissent with the
assertion of an instrumental possibility sentence like ():
() To evade arrest, Max could mingle with the crowd.
Relative to a preliminary modal base of this kind, every option is equally represented
in the possibility-space; I call this symmetry of choice.32
Updating this preliminary background f , characterized by symmetry of choice, with
the end e yields the following result: the option chosen in most of the remaining
possibility-space f +e(w) is the option on the choice of which e eventuates in more
of the possibility-space than it does on the choice of any alternative, ri . In simpler
language, the option most likely chosen is that which, if chosen, would make the end
most likely. Formally, given symmetry of choice, Equivalence holds:
Equivalence: ∀ri ∈ d(w): pr(p|f +e) > pr(ri |f +e) iff pr(e|f +p) > pr(e|f +ri ).
Equivalence might seem too complex to be playing any role in ordinary normative
thought. But represented geometrically it’s highly intuitive:

30 We could dispense with this parameter by treating d(w) as fixed by p plus context, but a contrast
parameter for ‘ought’ is independently well-motivated, embraced for numerous reasons; see Sloman (),
Jackson (), Cariani (a, b), Snedegar (), Finlay and Snedegar ().
31 My simple dyadic semantics can be formally expressed as:

[[must p]]f = T iff for ALL u ∈ ∩f (w): u → p.

[[ought p]]f ,d = T iff for MOSTd u ∈ ∩f (w): u → p; or:
[[ought p]]f ,d = T iff for all r ∈ d(w): |{u ∈ ∩f (w): u → p}| > |{v ∈ ∩f (w): v → r}|.
(Compare note .) Whereas I advanced this analysis of ‘ought’ primarily to get the right truth-conditions for
normative uses, note that Yalcin () has since advanced a very similar comparative analysis for ‘probably’.
32 A naturalized version of Kant’s assumption of transcendental freedom: to deliberate, one must regard
oneself as equally free to choose any available option.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 stephen finlay

r5 p

e
r4 r1

r3 r2

As is obvious at a glance, once we restrict to the e-space (shaded), more of the

remainder is p-space than ri -space, for any i. Informally, this analysis says that whereas
what must be done in order to achieve some end is whatever is necessary for the end,
what ought to be done in order to achieve the end is just whatever is the surest or most
reliable way to achieve it.33
Promisingly, this captures the difference between must and ought in many ordinary
scenarios. But can it be extended to an analysis of all normative ‘ought’ sentences?
An obvious objection is that what ought to be done, even instrumentally, isn’t always
just what makes some primary end more likely, but frequently depends on multiple,
sometimes competing ends or desiderata, such as comfort, decency, cost, or safety.
Here we confront Kratzer’s other reason for introducing ordering sources. While I
won’t address the issue of multiple ends or conflicts until Section , I’ll now show that
this semantics at least offers a superior treatment of simple cases, involving single ends.
Consider the following scenario:34
Single-Minded Theseus: Theseus is rushing through the Labyrinth with a single goal in mind: to
find and kill the minotaur before it slaughters his compatriots. He reaches a room with three
doors, with the following probabilities of success: door A=., door B=., door C=.

My dyadic semantics generates the following, intuitive verdict:

() To save his compatriots, Theseus ought to choose A, though he doesn’t have to,
as he could choose B, but he must choose one of A or B, and mustn’t choose C.

33 Cariani (b, p. ) offers two objections to symmetry of choice. First, “it is plausible to assume
that the contextually supplied probability function might be either some salient credence. . . or an evidential
probability function.” Reply: this is to reject without argument my claim (and Kant’s) that thought about
whether S ought to do A essentially ignores any evidence that S will do A, by virtue of its very nature
as normative. Second, “there is no guarantee that a set of alternatives equiprobable relative to an initial
background will remain equiprobable after the background is updated”, e.g. by a conditional such as if it’s
snowing outside. Reply: symmetry of choice is always (re-)applied after any circumstantial updating, for
the reason that the assumption of transcendental freedom applies precisely in the circumstance of choice.
Another objection to the most analysis, pressed by Cariani and Ralph Wedgwood (p.c.), is that it entails
failure of Agglomeration: Op & Oq → O(p & q). I believe this may rather be another virtue; see Jackson
().
34 Finlay (a, p. n; , p. ).

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

‘ought’: out of order 

The ordering semantics is unable to yield these results so simply or naturally. Perhaps
the most promising strategy is to treat reliability as a secondary ideal for the ‘ought’
claim, utilizing multiple ordering sources (von Fintel and Iatridou ). But reliability
isn’t normatively optional in the way secondary ideals are. Even if the only thing that
matters is saving his compatriots, Theseus ought to take door A. Reliability might be
thought a secondary ideal because it can be traded off against cost or comfort, etc.,
but this is really to trade off the importance of success, i.e. of the end itself, against
conflicting ideals.
This end-relational semantics has an additional advantage. It accommodates
degrees of possibility through the (classical) probabilistic structure of possibility-
spaces, which in effect directly induces an ordering of the options in the contrast
set. Importantly, this differs from the gradability provided by ordering sources in the
Lewis-Kratzer semantics, which instead directly induce an order on possible worlds
according to their approximation to the ideal. So while the most analysis simply
identifies the “best” option as the most likely/reliable option, the ordering semantics
rather identifies an option as “best” just in case (roughly) it is true in all the best
worlds. This indirectness is problematic: what is true in the best worlds might also
be true in the very worst worlds, and it seems arbitrary to determine an option’s value
by looking exclusively at best worlds and ignoring worst worlds.35 As we’ll see below,
some recent work in ordering semantics therefore seeks to induce orderings on sets
of possible worlds, partitioned by options—by introducing a contrast-set parameter
(e.g. Cariani a), a special selection function (e.g. Charlow ), or a special kind
of ordering source (e.g. Dowell ). The most analysis delivers these desired results
for free.

 Information-sensitivity
Further difficulties for the ordering semantics are observed in how normative modals
are sometimes sensitive to subjects’ information. Whereas Kratzer originally analyzes
normative modals as always taking circumstantial modal bases, picking out premise-
sets of facts objectively, some normative claims are subjective or information-relative.
A simple amendment is to allow modalities with epistemic modal bases, picking out
sets of facts that constitute what the relevant subject knows (“information-states”).36
But this may not be enough, as there are also more subjective normative uses of modal
verbs, such as the ‘ought’ of practical reason, which is sensitive to what is merely
believed rather than known.
Kratzer proposes (p.c.) to accommodate these subjective uses with a circumstantial
modal base of facts about the agent’s beliefs, and an ordering source consisting of

35 Compare Jackson (). This selection function is also what forces the Limit Assumption, which the
most analysis avoids without fuss.
36 Finlay (, p. n); Kolodny and MacFarlane (, p. ).

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 stephen finlay

norms of rationality, which I’ll call a subjectivized end strategy. While I agree that we
sometimes make claims about what an agent ought rationally to do given his beliefs,
this fits uncomfortably with other ordinary subjective ‘ought’ claims. First, we often
refer to ordinary, objective ends when making subjective ‘ought’ claims. Suppose for
example that Theseus’s beliefs favor choosing door A, although this option is in fact
incompatible with achieving his goal. We might then say,
() Given what he believes, to save his compatriots Theseus ought to take door A.
Secondly, it makes wrong predictions about appropriate modal force. There is
plausibly no separation between what is rationally necessary and what is rationally
optimal. In order to act rationally, Theseus must take door A. It seems infelicitously
weak to say that to act rationally he ought to take door A, and simply false to say that
to act rationally he could take door B. By contrast it’s natural to say that given what he
believes, to save his compatriots Theseus ought to take door A but doesn’t have to, and
could take door B though he oughtn’t.
Once Kratzer’s realism constraint is abandoned, we can simply appeal to a modal
base consisting of the (nonrealistic) contents of the subject’s beliefs (“given what you
believe”) to account for these subjective normative claims.37 The realism constraint
seems gratuitous, in any case. While the modal base has the function of domain restric-
tor in the semantics, sometimes conversationally relevant modalities are restricted
rather than merely ordered by nonfactual information.38 Instead of appealing to norms
of rationality, the dyadic semantics allows us simply to appeal to the same objective
ends (e.g. to save his compatriots), and account for the subjective ‘must’, ‘may’, and
‘ought’ as concerning what is respectively necessary, possible, and most likely relative
to the subject’s information i and the end e. Consider a subjectivized version of
Single-Minded Theseus, where the probabilities are subjective (i.e. relative to Theseus’s
beliefs); my semantics then predicts the following, intuitive verdict, parallel to ():
() Given what he believes, to save his compatriots Theseus ought to choose A,
though he doesn’t have to, as he could choose B (albeit foolishly!), but he must
choose either A or B, and mustn’t choose C.
Unlike Kratzer’s proposed solution, this semantics generates the appropriate judg-
ments and respects the relativization to objective ends. It requires no appeal to
controversial norms of rationality (indeed, I think it might rather explain them), and
directly explains why what one ought to do given complete accurate beliefs converges
on what one ought to do given the facts: these different functions pick out identical
premise-sets.

37 Finlay (a, p. ). With inconsistent beliefs the modal base must be restricted to a consistent subset
(Section ).
38 E.g. exocentric epistemic claims. Consider a variation of von Fintel and Gillies’ mastermind scenario
where a player accepts false information which she realizes entails that there are two reds. A nondeceived
observer can correctly say, “She knows there must be two reds”; i.e. must given her (false) information.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

‘ought’: out of order 

A second, more widely discussed problem about information-sensitivity involves

the way that strengthening information can change what an agent ought to do, by
changing which options are best. Niko Kolodny and John MacFarlane () illustrate
this with the following scenario:39
Miners: Ten miners are trapped together in one of two shafts, A and B, in danger of drowning in
impending floodwaters. We can block either shaft, but not both, which will save all the miners
if they’re in the blocked shaft but drown them otherwise. Or we can block neither shaft, which
is certain to save nine miners and drown one.

In a context of this information, the following sentences are all typically judged true:
(a) If the miners are all in shaft A, then we ought to block A.
(b) If the miners are all in shaft B, then we ought to block B.
(c) Either the miners are all in A or they are all in B.
(d) We ought to block neither shaft.
This presents a puzzle. The conditional clauses of (a) and (b) apparently just
strengthen the information, updating the modal base with the antecedent. In all the
best worlds where the miners are in A we block A, and in all the best worlds where
they are in B we block B. But since all accessible worlds are either in-A or in-B worlds,
it follows that all the best accessible worlds are either block-A or block-B worlds, so
none can be block-neither worlds, which falsifies (d) on the ordering semantics.40
The fundamental issue here concerns nonmonotonicity, as Nate Charlow ()
observes.41 Note first that simple necessity or ‘all’ conditionals are (left downwards)
monotonic; i.e. if p then necessarily q entails if r then necessarily q for any r that entails
p. For example, from () we can safely infer ():
() If Garfield is a cat, then he must be a mammal.
() If Garfield is a yellow cat, then he must be a mammal.
The ordering semantics does allow for nonmonotonicity, but in only the following
way: eliminating the p worlds which were initially ranked highest can result in the
highest-ranked worlds that remain being not-p worlds. However, this explanation is
unavailable for the miner puzzle. No worlds are eliminated in the state of ignorance,
and apparently all the best worlds are block-A or block-B worlds. What the ordering
semantics needs to explain is how adding information can change the ordering of
possible worlds: what Kolodny and MacFarlane label ‘serious information depen-
dence’. This is a serious challenge, because in the Lewis–Kratzer framework, adding

39 Due to Regan ().

40 For objections to simple solutions, see Kolodny and MacFarlane ().
41 Kolodny and MacFarlane attribute it to the failure of modus ponens, but Charlow proves that the same
problem arises for theories of conditionals that invalidate modus ponens.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 stephen finlay

information merely eliminates possible worlds from the domain, which cannot by
itself change the relative order of any two worlds within the domain.
The various responses recently offered on behalf of ordering semantics fall into
two categories, which both encounter serious difficulties. One (conservative) strategy
appeals to an information-sensitive (or subjectivized) ordering source, such as: in view
of what maximizes expected value.42 Plausibly speakers can and do sometimes make
normative claims about what ought to be done to maximize some salient kind of
expected value, and this would indeed generate different orderings as information
strengthens. But as a general interpretation of information-sensitive claims, this
analysis is problematic. It’s natural to suppose that the salient ends or ideals for these
claims remain the objective ends (e.g. of saving miner , miner , etc.), rather than
higher-order or subjectivized ends.43
This is easier seen in simple cases involving a single end, like Theseus’s above.
(Or equivalently, suppose blocking neither shaft instead provides a % chance of
saving all ten miners.) Suppose Theseus acquires additional information reducing the
likelihood of success if he selects door A from . to .. Given this strengthening of his
information, he ought to choose door B (. probability of success) rather than door A.
But it’s natural to say,
() In order to save his compatriots, Theseus ought to choose door B
just as it’s also natural to say, relative to his earlier information, that in order to save
his compatriots, he ought to choose A. Here we have serious information dependence
relative to a constant, objective end.
Further evidence comes from data about appropriate modal force. In the
(unstrengthened) subjective Theseus scenario, the following is true:
() In order to maximize expected value, Theseus must choose door A, and can’t
choose B.
But it’s intuitive to say rather that he merely ought to choose A, and could choose B, as in
() above, indicating that these information-sensitive ‘ought’ claims aren’t relativized
to a subjectivized end like to maximize expected value.
Relatedly, Charlow (, p. ) observes that serious information-dependence
seems to arise uniquely for “weak necessity” verbs such as ‘ought’, but not for ‘must’,
discovering intuitions that the following set of sentences aren’t consistent:44
(a) If the miners are all in A, then we must block A.
(b) If the miners are all in B, then we must block B.
(c) The miners are either all in A, or they are all in B.
(d) We must block neither shaft.

42 von Fintel (); Dowell (); Silk (). 43 Cf. Charlow (), Silk ().
44 These intuitions are disputed (e.g. Carr ), I think because in the original miner puzzle it’s too easy
to accommodate context shifts in ends/ideals; see Section .

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

‘ought’: out of order 

The conservative strategy isn’t subtle enough to accommodate this difference;

strengthening information changes what must be done in order to maximize expected
value. This evidence tells against an information-sensitive ordering source.
The second strategy proposes, more radically, to solve this problem for ordering
semantics by appealing to additional (third, fourth . . . ) parameters, and a wider
palette of modal operations on these. These solutions are complicated and varied, but
encounter problems of their own.
Observing the limitation of serious information-dependence to “weak necessity”,
Charlow () follows von Fintel and Iatridou in adopting a secondary ordering
source, and proposes that ‘ought’ semantically allows a “merging” operation between
the two ordering sources g and g . As he implements this idea to address the miner
puzzle, the higher-order ideals in g (w) select from the primary ideals in g (w) a set of
“practical ends” which are the strongest actionable weakenings of the primary ideals.
An ideal is “actionable” just in case there is an action guaranteeing that outcome
relative to the information. In the original miner scenario, the strongest actionable
weakening of the ideal saving all ten miners is the practical end, saving at least nine
miners. In view of this practical end, blocking neither shaft is the best option given the
limited information.45 Serious information-dependence is explained by the possibility
that strengthening information causes additional or stronger practical ends to become
actionable.
Charlow’s proposal creates new difficulties. First, it falsely predicts that in the
original miner scenario, the following sentences are both consistent (and true) without
any shift in context:
() We ought to block neither shaft.
() #We must block either shaft A or shaft B.46
Secondly, absent a special, as yet untold story about the semantics of the antecedent,
it would seem to predict that () is true, since saving nine is the strongest actionable
weakening of the designated primary ideal of saving all ten:
() #(If we want) to save all ten miners, we ought to block neither shaft.
But blocking neither is the only option we may not choose in order to save all ten.
Fabrizio Cariani, Magda Kaufmann and Stefan Kaufmann () propose an addi-
tional contrast-set or “decision problem” parameter d—as in my own semantics for
‘ought’—instead of a secondary ordering source, consisting in mutually exclusive

45 A problem: not all worlds where at least nine miners are saved are block-neither worlds. Charlow
suggests one solution (, pp. –n): the set of practical ends generated by the merging also includes
the negation of any non-actionable primary ideals; i.e. not (all ten miners are saved). However, suppose that
if we block neither shaft there is a % chance the tenth miner survives, while another option guarantees
saving only nine. This analysis then predicts we ought not to block neither shaft—because we might thereby
save all ten miners.
46 Charlow suggests this is merely infelicitous, but it seems correct to say, “It’s not true that you must
block either A or B; you could and should block neither.”

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 stephen finlay

“chooseable” options.47 This partitions the f −accessible worlds into sets defined by
chosen option. The ordering source g ranks each partition by just the ends true
throughout it, i.e. those ends guaranteed by the option. Since saving nine miners
is the only relevant end guaranteed by any option, worlds are ranked only by their
closeness to this end. Serious information-dependence is explained by the possibility
that strengthening information may lead to additional ends being guaranteed by some
option (e.g. given that the miners are in A, blocking A is guaranteed to save ten). This
avoids the problems observed for Charlow, but shares a further problem: appeal to
“guaranteed” or “actionable” outcomes is artificially restrictive, because sometimes we
ought to pursue a chance of an outcome. Even if blocking neither shaft provides only a
% chance of saving nine miners, it plausibly ought to be chosen over a % chance
of saving everyone (Carr ).48
A variation from Graham Katz, Paul Portner and Aynat Rubinstein () offers a
fix for this problem. They propose separate epistemic (gE ) and deontic (gD ) ordering
sources, and a merging operation on these. As applied to the miner puzzle, the idea is
that valued outcomes are first ordered into tiers by how likely (actionable) they are, by
gE , and then are ordered within each tier by their closeness to the deontic ideal gD . An
option ought to be chosen iff it corresponds to the deontically best of the epistemically
most actionable outcomes. In Jennifer Carr’s probabilized miner scenario, this is the
option of blocking neither shaft, since the most actionable end (at % if one blocks
neither shaft) is saving nine. But this fails to accommodate normative judgments that
trade low probability for high value. Suppose blocking neither shaft will certainly
save exactly one miner, while blocking a shaft gives a % chance of saving all ten.
Presumably we ought to block a shaft, although saving exactly one is the uniquely
most actionable outcome.
I have not shown that serious information dependence is a fatal problem for
the ordering semantics. Its champions may reply, as Charlow does (p.c.), that my
objections merely target contingent details of the way they have proposed imple-
menting these more general strategies to address particular variations of the miner
puzzle. I think this itself reflects a problem: it remains to be shown that any of these
strategies can be developed into a concrete theory that systematically generates correct
results when applied to arbitrary cases, rather than on an ad hoc basis. In one of the
most sophisticated attempts, Cariani () concludes from these problems that the
semantics for ‘ought’ must be probabilistic, which he implements by modifying his
earlier view, above, replacing the modal base with a pair consisting of an information-
state and a probability function <i, Pr>, and assigning probabilistic content to the
premises in the ordering source. I can’t do justice to this proposal here, but one
problem it faces is explaining why ‘must’ and ‘may’ aren’t similarly probabilistic, once

47 See also Charlow (); Carr ().

48 Single-Minded Theseus shows that appeal to a secondary normative ordering source can’t be sufficient
to explain serious information-dependence, since this is a feature also of single-end scenarios.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

‘ought’: out of order 

we’ve probabilized ordering sources; Cariani suggests a further semantic rule that
these terms ignore such premises.
The point I want to make is that here, also, the challenge for ordering semantics
is to show that it can generate the same results that my simpler semantics delivers
already, without further elaboration. For the end-relational semantics provides a
straightforward explanation of serious information-dependence. Nonmonotonicity is
a basic characteristic of probability and most conditionals: strengthening information
often switches most p from true to false, unlike all p. While most philosophers already
born are now dead, for example, most philosophers already born after  are not
now dead. Similarly, if the information is strengthened in a way that eliminates the
possibilities where Theseus both chooses door A and saves his compatriots, then in
most of the remaining possibility-space where his compatriots are saved, he takes door
B. (Similarly for the simplified miner case where blocking neither shaft is % likely
to save all the miners.)
Unlike the kind of nonmonotonicity provided by ordering sources in the basic
Lewis–Kratzer semantics, the most analysis explains serious information-dependence.
It also avoids all the problems observed above for versions of ordering semantics.
It invokes the same (objective) ends as objective normative claims, accommodating
the use of restrictors like in order to save his compatriots, and it makes the right
predictions about appropriate modal force. It correctly predicts nonmonotonicity
uniquely for “weak necessity” verbs like ‘ought’ and ‘should’, and not for ‘must’. It
accommodates probabilities with ease. As already observed, it doesn’t imply, coun-
terintuitively, that worlds where agents act as they subjectively ought are always the
“best” worlds: the worlds where Theseus saves his compatriots are best, even if they
aren’t those in which he acts as he ought. Finally, it predicts the right verdicts in at
least these simple single-end cases, because so long as the end is held fixed, changes
in expected value are just changes in the conditional probability of the end. The
information-sensitivity of ‘ought’ therefore provides further support for abolishing
ordering sources.
However, it may fairly be objected that celebration is premature, as I haven’t yet
shown that my semantics yields the right results for the original miner puzzle. A full
solution to this requires a solution for cases involving multiple ends/ideals, to which
I now turn.

 Conflicts
One argument for an ordering source parameter remains: the need to accommodate
inconsistent premise-sets, or conflicts either between incompatible ends/ideals or
between ideals and facts. Ordering sources provide a way of reaching a consistent
domain from inconsistent inputs. But despite this advertised virtue, here too the
ordering semantics encounters a problem: the Lewis–Kratzer ordering mechanism

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 stephen finlay

seems too crude to handle all but the simplest of normative scenarios. An ordering
source g ranks a possible world w as better than a world w iff w satisfies all the
ideals in the premise set g(w) that w does, plus at least one more. This presents some
obvious difficulties when we consider ordinary use of normative modals.
First, it doesn’t allow comparisons between two worlds if each satisfies some ideal
the other doesn’t, although this is an ordinary situation where it is often perfectly clear
that one possibility is better than the other. Suppose you have two desired ends, (e )
attending your daughter’s wedding, and (e ) watching the football match live; here
() is surely true:
() You ought to attend your daughter’s wedding rather than watch football.
It is obviously because of the greater importance of attending the wedding that it ought
to be chosen over watching football. But ordering sources don’t encode the degrees
of importance (or normative “weight”) of different ideals, treating every ideal in the
premise-set g(w) on a par. A second problem arises from the interaction of importance
with uncertainty, as the ordering semantics doesn’t allow the intrinsic importance of
ends in g(w) to be balanced against the probability or expectation of achieving them,
as modeled in decision theory.
Many creative solutions have been proposed on behalf of ordering semantics. We
might try capturing differences in importance by assigning ends to higher- or lower-
ranked ordering sources. However, this can’t capture a properly decision-theoretic
framework, in which less important ends can outweigh more important ends either by
combining their normative weight or by their greater expected (probability-weighted)
value. In Section  we observed proposals to “merge” separate ordering sources (either
normative and epistemic, or first- and second- order normative), and for probability-
sensitive ordering sources. But no general account is offered of how to generate
decision-theoretic verdicts systematically, balancing importance against probability
in intuitive ways. I think these strategies are unlikely to succeed, for the reason that
ordinal scales can never fully reproduce cardinality.
There is a way the ordering semantics can generate any desired ordering of worlds,
however.49 Suppose there are three possible outcomes, to be ranked in the order o , o ,
o . This result can be generated by the premise set {o , (o ∨o ), (o ∨o ∨o )}. But this
technical fix is problematic as an account of natural language semantics. In the Lewis–
Kratzer framework, the ordering ≤g(w) is generated by a semantically fixed rule from
the premise-set g(w), which is in turn generated by the conversational background g,
a function from the context determined by the speaker’s intentions. So the speaker’s
intended set of ideals is in the driver’s seat. By contrast, this solution reverse-engineers
a premise-set from the desired ordering, which can’t be generated from any intuitive
inputs such as degrees of importance and probability. The only kind of conversational
background that could plausibly generate such a premise-set is something like: given

49 E.g. Swanson ().

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

‘ought’: out of order 

the set of ideals that by the semantic rule ≤ yields the ordering corresponding to expected
value by the salient probability and value functions. But it is implausible that this is the
kind of conversational background operative in ordinary normative thought. Whereas
Lewis–Kratzer ordering sources are usually described in highly intuitive ways (e.g. in
view of what you desire), on this fix they and their derivative premise-sets bear little
resemblance to intuitive inputs for normative thought, even in mundane cases.
In view of such problems, many writers reject ordering semantics in favor of
quantitative (“fully Bayesian”) semantics that instead have parameters for at least (i)
a probability distribution over worlds, and (ii) a value function assigning normative
weights to outcomes, and give ‘ought’ the modal force of a decision rule on expected
value.50 This has the virtue of systematically generating the verdicts of decision theory,
but there are problems here too. First, whereas an ability to accommodate decision-
theoretic judgments is a desideratum, such judgments shouldn’t be guaranteed true by
semantics alone, since competent speakers can vary in their tolerance of risk.51 Carr
() therefore proposes an additional (third) parameter for a decision rule. Secondly,
it isn’t clear how to apply these semantics to ordinary epistemic or circumstantial uses
of ‘ought’,52 which don’t seem to involve value functions or decision rules at all.
I believe that the simple dyadic semantics has been abandoned too quickly in the
face of this issue, because it supports an attractive solution which makes the extra
semantic machinery of ordering and quantitative semantics gratuitous. This solution
appeals only to resources that play a central role in Kratzer’s own application of the
ordering semantics to different uses of modal verbs: i.e. to the pragmatics (or “pre-
semantics”) of how the appropriate conversational backgrounds are fixed and premise-
sets constructed:
The way we understand a particular occurrence of a modal can be at least partly explained by
an interaction of independently motivated semantic and pragmatic principles.
(Kratzer , pp. –)

In particular, it endorses Kratzer’s appeal to the Rule of Accommodation:

If the utterance of an expression requires a complement of a certain kind to be correct, and the
context just before the utterance does not provide it, then ceteris paribus. . . a complement of the
required kind comes into existence. (, p. )

This solution (explored in greater detail in Finlay , chapter ) is as follows.

First, given that modal verbs function logically as quantifiers over possibilities,
modal claims relative to an inconsistent modal base will generally be pointless.
Informally, a modal base provides a partial sketch of some way the world could be,
but the world couldn’t be such that p and ∼p are both true. On a dyadic semantics,

50 E.g. Goble (); Wedgwood (); Lassiter ().

51 Kratzer (, p. ); Carr ().
52 For the distinction between epistemic and circumstantial ought, see Copley ().

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 stephen finlay

competent interpreters can therefore be expected to accommodate modal utterances

using pragmatic cues to identify the speaker’s intended, consistent, modal base f .
Second, on the end-relational theory, normative modals are relativized to a single
end or outcome e in the modal base. When this end isn’t articulated explicitly, we can
expect it to be the most salient end in the context, which will generally be the relevantly
most preferred / important end of the person(s) whose perspective is salient—whether
this is the agent, the speaker, the audience, or some further party.
Even putting aside conflicts of ends, this proposal might seem ill-equipped to
handle all-things-considered practical contexts involving multiple ends. In the original
miner scenario, deliberation can’t be concerned with any single end (like saving all
the miners, or saving the most we can, etc.), but rather with a set of ends such as
{saving miner , saving miner , . . . , saving miner }.53 But there is a simple, intuitive
solution. In practical contexts (of deliberation or advice), the most salient end can
be identified (perhaps only de dicto) as the overall outcome that is the relevantly
preferred object of pursuit. Call such an outcome the contextually preferred end. If the
dyadic semantics is correct, then an ordinary normative utterance of ‘s must/ought
to/may do A’, simpliciter, can therefore be interpreted by default as asserting that given
f: (it is the case that i, and it is going to be the case that e), in all/most/some of the
remaining possibilities, s does A—where i is the salient circumstances or information
(with symmetry of choice) and e is the contextually preferred end. This provides
intuitive resolutions to the puzzles about conflicts.
First, in the case of contrary-to-duty obligations, any end or ideal that is incompati-
ble with the salient circumstances or information i will not be (part of) the contextually
preferred end e, since we don’t seriously pursue ends we’re sure are unattainable.
Therefore, unattainable ends would typically be omitted from the modal base, just as
ability modals ignore information about agents’ psychological dispositions in Kratzer’s
analysis (, p. ). When judging whether you ought to apologize for causing
offense, for example, any possibilities where you never violated offense norms are
irrelevant. On the other hand, if we tell somebody locked in jail that if he wants to go to
Harlem he could take the A train, he will immediately assume that his incarceration is
being excluded from the modal base; this resolves our case of unattainable goals from
Section .
Second, whenever a person believes that two ends e and e are mutually
incompatible, her contextually preferred end won’t incorporate both, and for the
same reason: we don’t pursue unattainable outcomes. Instead, a speaker’s contextually
preferred end will be roughly the most preferred combination of ends she believes
attainable. If you desire to visit your grandmother in Harlem, and your grandmother in
Hoboken, but can’t do both, then we’d typically advise you best by telling you what you
ought to do in order to visit the grandmother you’d most contextually prefer to visit.

53 In Kratzer’s (borrowed) slogan, we need commas rather than ampersands (, p. ).

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

‘ought’: out of order 

Third, in cases where multiple ends are weighed under uncertainty, an agent’s
overall motivation will balance intrinsic preference for outcomes (importance) against
their expectability (probability). The contextually preferred end therefore needn’t be
intrinsically preferred, or what the agent most desires. In the miner scenario, for
example, our intuitive judgments plausibly reflect a contextual preference for saving
at least nine (whether at % or % odds) over saving ten (at % odds).54 This
accounts for the truth of (d): if we’re going to save at least nine miners, it is most
likely that we block neither shaft (“We ought to block neither”). In support of this
analysis, observe that if asked, “Why ought we to block neither shaft?” it’s natural to
reply, “In order to save at least nine miners”, but not “In order to save all the miners”.55
On this dyadic semantics, the complex interactions between importance and prob-
ability are assigned to the psychology of normative thought rather than to semantics.
While more work is needed to establish that these pragmatic resources are sufficient to
address every issue arising from conflicts,56 this explanation has noteworthy advan-
tages over both the ordering semantics and its decision-theoretic rivals. It respects
linguists’ concerns about the artificiality of building quantitative measures into natural
language semantics (e.g. Kratzer , p. ), yet also easily accommodates decision-
theoretic verdicts without ad hoc semantic maneuvers. This is accomplished without
either building a decision rule into semantic competence (thereby respecting the
semantic desideratum of normative neutrality), or relativizing ordinary normative
speech to a decision rule as an additional conversational background.57 It therefore
employs only the same simple semantic resources needed for epistemic and circum-
stantial uses, not postulating empty ordering sources or other redundant argument-
places. Surprisingly, a dyadic semantics might even provide the best natural language
framework for handling cases of conflicts.

54 In complex scenarios an agent’s contextual preference is plausibly a conjunction of probability-

weighted outcomes; e.g. to have a % chance of e and a % chance of e . . . ; see Evers (); Finlay (,
p. ).
55 Superficially this resembles Charlow’s appeal to the “practical end” of saving nine miners and no more,
but the most analysis needn’t exclude worlds where all ten are saved, as not among the relevantly “best”
worlds.
56 Perhaps most pressing is the difference between ‘must’ and ‘ought’ in these cases. Sometimes we ought
to but don’t have to choose an option because it promotes an end of greater importance, rather than making
any single end more likely. Suppose six miners are in A and four are in B; arguably we then ought to block
A but don’t have to, as we could block B. But to achieve the (contextually preferred) end of saving six, we
must block A, and cannot block B.
Hypothesis: in basic (single-end) cases, using ‘ought’ will indicate that p is merely the most preferred of
possibly multiple acceptable options, whereas ‘must’ indicates that p is the only acceptable option (Finlay
, p. ). Use of these words may therefore carry generalized conversational implicatures to this effect,
which would explain the preference to use ‘ought’ whenever the end is only marginally preferred over others,
and reserving ‘must’ for communicating that other ends are unacceptable. Use of ‘may’ or ‘could’ would
pragmatically indicate an end that is merely accepted rather than preferred. I also argue (, pp. –)
that this explains the puzzling impropriety of ‘You must, but you won’t’ (e.g. Copley ).
57 In contrast to Carr (), for example. One objection to this pragmatics-heavy approach is that
because it liberally postulates context shifts, it fails to account for many disagreements and valid inferences.
I believe any contextualist semantics faces this issue, and that it is amenable to pragmatic resolution (Finlay
 Chapter ).

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 stephen finlay

 Metasemantic Issues
My objections to ordering semantics may be thought to stem from an idiosyncratically
narrow conception of semantics. I’ve argued that a dyadic semantics is sufficient in part
because certain peculiarities of our use of ‘ought’ can be attributed to “pragmatics”.
But whereas I’ve assumed a narrow conception of semantics as investigating the
conventional meanings of words (etc.), others understand “semantics” more broadly, as
investigating the truth-conditions of sentences as uttered in contexts, and understand
“pragmatics” more narrowly, as concerning what is merely implicated by utterances.
Kratzer might seem to endorse this broader conception of semantics by beginning her
textbook on compositional semantics by writing that “a theory of meaning . . . pairs
sentences with their truth-conditions” (Heim and Kratzer , p. ). Such a semantics
for ‘ought’ would aim not just to identify the meaning of the word ‘ought’ itself, but
rather to give a complete theory of the truth conditions of ‘ought’ sentences as uttered
in contexts.
My claim to have offered a simpler semantics may therefore be thought confused,
on the grounds that I’ve really just relabeled what is functionally still an ordering
source as part of the “pragmatics”. The ordering source’s function in the Lewis–Kratzer
semantics is nothing other than to generate consistent premise-sets from inconsistent
inputs, and it is therefore just a formal implementation of the same kinds of contextual
pressures I’ve described in Section . Then I haven’t really shown that we can do
without ordering sources. Since the ordering-source mechanism is flexible enough to
yield any ordering we might want, my objections to it as crude also fail.
A theory of the meaning of ‘ought’ should provide more than this, however. If
the task of semantics is just to give a formal statement of the truth conditions of
sentences, then semantic theories would be nonexclusive. Kratzer provides one way
of modeling the truth conditions of ‘ought’ sentences, but these could be modeled
with equal legitimacy either by (i) more fine-grained semantic theories with three or
more parameters, which separate different kinds of inputs to orderings such as utilities,
probabilities, and decision-rules (as in Carr ), or by (ii) coarser semantic theories
with just one parameter, identifying the background f with an often complex function
from context which itself eliminated inconsistencies—as in the dyadic semantics I’ve
championed.58
This fails to provide other things that semanticists, including Kratzer, clearly want
from a semantic theory. First, it has nothing to say about whether a term is lexically
ambiguous—if it doesn’t eliminate the possibility of lexical ambiguity altogether.59
Capturing truth conditions with a single complex formula provides no evidence that
a word has a single common meaning—especially once we begin positing parameters
commonly left empty. A single rule can be provided even for paradigms of ambiguity
like ‘bank’ and ‘mole’, by positing a parameter that toggles to different values. There

58 Hall (unpublished) makes this point forcefully. 59 Compare Kratzer (, p. ).

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

‘ought’: out of order 

would be no difference between saying that ‘must’ has a unified meaning involving two
parameters including an ordering source which is empty in alethic uses, and saying
that ‘must’ has two meanings, being either a simple dyadic operator, or a polyadic
operator taking ordering sources.
Secondly, it needn’t reveal anything about the semantic competence of ordinary
speakers, and so wouldn’t realize Kratzer’s stated aim of identifying the “abilities of a
person who has a complete grasp of the modal system” of some natural language (,
p. ). Unity is a desideratum of lexical semantics mainly because of the need to explain
how speakers are able to learn the language, and construct and interpret arbitrary new
sentences. A theory of meaning for ‘ought’ therefore owes us a compositional account
of the contribution that competent speakers implicitly understand the word ‘ought’
itself as making to the truth conditions of a sentence, alongside the contributions of
other words and contextual features.
The modal base and ordering source aren’t merely inputs to truth–conditions
in the Lewis-Kratzer semantics, but conversational backgrounds: functions fixed by
what the speaker has in mind, and made explicit by expressions like ‘given what we
know’ and ‘in view of what the law prescribes’. The modal force of ‘ought’ is the
rule that competent speakers apply in identifying the truth conditions of an ‘ought’
sentence relative to these backgrounds. Different semantic theories are therefore not
just formal or notational variants, and it is a serious challenge to the feasibility of the
ordering semantics if, as I’ve argued, it requires “conversational backgrounds” that
don’t capture what ordinary speakers have in mind, or posits a semantic rule that
doesn’t reflect ordinary modal or normative thought. If I am right about how ordinary
users of normative ‘ought’ select the relevant domain—by a complex function f picking
out both circumstances/information under symmetry of choice and a contextually
preferred end—then the ordering semantics are wrong.
Recent semantic theorizing about modals has fixated on the ideal of unity, but has
largely ignored the ideal of simplicity, which is also important with respect to the
acquisition, use, and interpretation of expressions in natural language. For simple tasks
we can expect use of simple tools; to butter a slice of bread we reach for a simple butter
knife, not for a Swiss Army knife. We should therefore be skeptical of a theory—no
matter how unifying—which claims that when someone says (e.g.) “If this apple costs
less than a quarter, then it must cost less than a dollar”, by ‘it must cost less than a
dollar’ she means (without the usual simplifying gloss making the Limit Assumption)
that for every accessible world u there is an accessible world v that comes at least as close
to the ideal determined by the (empty) ordering source g, and for every accessible world
z, if z is at least as close to that ideal as v, then this apple costs less than a dollar in z
(Kratzer , p. ).
If some modal judgments require more complex semantic resources then we should
expect to find special terms to have developed for those purposes, rather than all our
modal vocabulary developing extra machinery left unused in simpler applications, like

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 stephen finlay

empty ordering sources or modal bases. However, we shouldn’t be surprised if instead

we find people using simple but flexible tools in creative ways to accomplish more
complicated tasks, as I suggested above that pragmatic pressures identify consistent
premise-sets in cases of conflicts. (Butter knives can be put to a wide variety of uses!)
The dyadic semantics respects this constraint. It doesn’t posit any kind of argument
for modal verbs which is ever left empty (except in bare logical modalities), and rather
than building complex but inflexible rules for reaching consistent premise-sets into the
semantics itself, it merely extends Kratzer’s appeal to “the interaction of independently
motivated semantic and pragmatic principles” (, p. ).
The fact that normative modalities are lexically the most promiscuous is therefore
further evidence in support of the simpler semantics. In addition to dedicated nor-
mative verbs, such as ‘needs to’, we can get normative readings of virtually any modal
auxiliary, including epistemic verbs such as ‘might’,60 and ability/dispositional verbs
such as ‘can’ and ‘could’. If the ordering semantics were correct, we should rather expect
normative modalities to be among the most lexically specialized and discriminating,
seeing as they require the greatest number of distinct parameters. This promiscuity is
however predicted by my end-relational theory that a normative modality is generated
simply by including an end in the modal base.61

 Conclusion
The difficulties for the Lewis–Kratzer ordering semantics from the behavior of nor-
mative modals, despite stimulating ever-more-complicated fixes in the literature, can
all be avoided in simple and intuitive ways by instead giving up on ordering sources
and returning to a simpler dyadic framework. Instrumental conditionals, the “weak
necessity” and serious information dependence of ‘ought’, the information-sensitivity
of subjective normative claims, and even judgments weighing conflicting ends under
uncertainty all seem more amenable to the simpler dyadic semantics. While more
work is needed to show that the dyadic approach is viable, these results provide
compelling reasons for believing that the move to ordering semantics was a mistake.
Of course, given the desideratum of a fully unifying theory, the data to which semantic
theorizing must answer extends far beyond the behavior of English modal auxiliaries
in normative uses that we’ve examined here. But my suspicion is that other kinds of
use (e.g. epistemic, counterfactual) raise parallel problems for ordering semantics and
will yield to parallel solutions in a dyadic framework.

60 The existence of such readings of ‘might’ is often denied (e.g. Portner , p. ), but consider the
normative flavor of suggestive advice in ‘To go to Harlem, you might take the A train’.
61 Our lack of grammatical devices for distinguishing conversational backgrounds as modal bases or
ordering sources (‘given. . . ’, ‘in view of. . . ’, ‘in light of. . . ’ etc. are all neutral) is also suggestive, as we should
expect linguistic resources to have developed to mark such differences in logical role.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

‘ought’: out of order 

Acknowledgments
I’m indebted to many people for their help, but especially Angelika Kratzer, Matthew Chrisman,
Nate Charlow, Bridget Copley, Fabrizio Cariani, Mark Schroeder, an anonymous reviewer, the
participants of my Fall  graduate seminar, and the audience at the Arché Workshop on
Normative Language at St. Andrews. Remaining errors are not theirs.

References
Björnsson, Gunnar and Finlay, Stephen () Metaethical Contextualism Defended. Ethics.
.. pp. –.
Boghossian, Paul () What is Relativism? In Greenough, P. and Lynch, M. (eds.) Truth and
Relativism. pp. –. Oxford: Clarendon Press.
Cariani, Fabrizio (a) Ought and Resolution Semantics. Noûs. .. pp. –.
Cariani, Fabrizio (b) Epistemic and Deontic Should. Thought. .. pp. –.
Cariani, Fabrizio () Deontic Modals and Probabilities: One Theory to Rule Them All.
In Charlow, N. and Christman, M. (eds.) Deontic Modality. pp. –. Oxford: Oxford
University Press.
Cariani, Fabrizio, Kaufmann, Magda and Kaufmann, Stefan () Deliberative modality under
epistemic uncertainty. Linguistics and Philosophy. .. pp. –.
Carr, Jennifer () Subjective Ought. Ergo: An Open Access Journal of Philiosophy. ..
Charlow, Nate () What We Know and What To Do. Synthese. .. pp. –.
Condoravdi, Cleo () Temporal Interpretation of Modals. In Beaver, D., Kaufmann, S., Clark,
B., Casillas-Martinez, L. (eds.) The Construction of Meaning. pp. –. Stanford CA: CSLI
Publications.
Condoravdi, Cleo and Lauer, Sven () Preference-Conditioned Necessities: Detachment and
Practical Reasoning. Pacific Philosophical Quarterly. .. pp. –.
Copley, Bridget () What Should Should Mean? [Online] Available from: https://hal.
archives-ouvertes.fr/halshs-/ [Accessed  Oct .]
Dowell, Janice () Flexible Contextualism about Deontic Modals. Inquiry. .–. pp. –.
Evers, Daan () Weight for Stephen Finlay. Philosophical Studies. .. pp. –.
Finlay, Stephen () Oughts and Ends. Philosophical Studies. .. pp. –.
Finlay, Stephen (a) What Ought Probably Means, and Why You Can’t Detach It. Synthese.
.. pp. –.
Finlay, Stephen (b) Normativity, Necessity, and Tense. In Shafer-Landau, R. (ed.) Oxford
Studies in Metaethics. . pp. –. New York, NY: Oxford University Press.
Finlay, Stephen () Confusion of Tongues. New York, NY: Oxford University Press.
Finlay, Stephen and Snedegar, Justin () One Ought Too Many. Philosophy and Phenomeno-
logical Research. .. pp. –.
Fintel, Kai von () The Best We Can (Expect to) Get? [Online] Available from:
http://web.mit.edu/fintel/fintel--apa-ought.pdf. [Accessed  Oct .]
Fintel, Kai von and Iatridou, Sabine () What to Do If You Want to Get to Harlem. [Online]
Available from: http://mit.edu/fintel/www/harlem-rutgers.pdf. [Accessed  Oct .]

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 stephen finlay

Fintel, Kai von and Iatridou, Sabine () How to Say Ought in Foreign. In Guéron J. and
Lecarme J. (eds.) Time and Modality. pp. –. Dordrecht: Springer.
Foot, Philippa () Morality as a System of Hypothetical Imperatives. Philosophical Review.
.. pp. –.
Goble, Lou () Utilitarian Deontic Logic. Philosophical Studies. .. pp. –.
Hall, Keith (unpublished) Normative Neutrality and Kratzer’s Semantics for Deontic Modals.
Manuscript.
Harman, Gilbert () Moral Relativism. In Harman, G. and Thomson, J. (eds.) Moral Rela-
tivism and Moral Objectivity. pp. –. Cambridge, MA: Blackwell.
Heim, Irene and Kratzer, Angelika () Semantics in Generative Grammar. Malden, MA:
Blackwell.
Horty, John F. () Reasons as Defaults. New York, NY: Oxford University Press.
Huitink, Janneke () Anankastic Conditionals and Salient Goals. In Maier, E., Bary, C., and
Huitink, J. (eds.) Proceedings of Sinn und Bedeutung . pp. –.
Jackson, Frank () On the Semantics and Logic of Obligation. Mind. .. pp. –.
Katz, Graham, Portner, Paul and Rubinstein, Aynat () Ordering Combination for Modal
Comparison. In Chereches. A. (ed.), Proceedings of SALT. . pp. –.
Kolodny, Niko and MacFarlane, John () Ifs and Oughts. Journal of Philosophy. ..
pp. –.
Kratzer, Angelika () What ‘Must’ and ‘Can’ Must and Can Mean. Linguistics and Philosophy.
.. pp. –.
Kratzer, Angelika () The Notional Category of Modality. In Eikmeyer, H. J. and Rieser, H.
(eds.) Words, Worlds, and Contexts. pp. –. Berlin: De Gruyter.
Kratzer, Angelika () Modality. In von Stechow, A. and Wunderlich, D. (eds.) Semantics: An
international handbook of contemporary research. pp. –. Berlin: De Gruyter.
Kratzer, Angelika () Conditionals and Modals. New York, NY: Oxford University Press.
Lassiter, Daniel () Measurement and Modality. Ph.D. Thesis NYU.
Nissenbaum, Jon () Kissing Pedro Martinez: (existential) anankastic conditionals and
rationale clauses. In Georgala, E. and Howell, J. (eds.) Proceedings of SALT. . pp. –.
Portner, Paul () Modality. New York, NY: Oxford University Press.
Portner, Paul and Rubinstein, Aynat () Extreme and Non-Extreme Deontic Models. In
Charlow, N. and Chrisman, M. Deontic Modality. Oxford: Oxford University Press.
Regan, Donald () Utilitarianism and Cooperation. New York, NY: Oxford University Press.
Rubinstein, Aynat () Figuring out what we ought to do: the challenge of delineating
priorities. U. Penn Working Papers in Linguistics. .. pp. –.
Sæbø, Kjell Johan () Necessary Conditions in a Natural Language. In Féry, C. and
Sternefeld, W. (eds.) Audiatur Vox Sapientiae: A Festschrift for Arnim von Stechow. pp. –.
Berlin: Akademie.
Silk, Alex () ‘Ought’ and ‘Must’: Some Philosophical Therapy. In Chereches, A. (ed.)
Proceedings of SALT. . pp. –.
Silk, Alex () Evidence Sensitivity in Weak Necessity Deontic Modals. Journal of Philosoph-
ical Logic. .. pp. –.
Sloman, Aaron () Ought and Better. Mind. .. pp. –.
Snedegar, Justin () Contrastive Semantics for Deontic Modals. In Blaauw, M. (ed.) Con-
trastivism in Philosophy: New Perspectives. pp. –. New York, NY: Routledge.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

‘ought’: out of order 

Stechow, Arnim von, Krasikova, Sveta and Penka, Doris () Anankastic Conditionals Again.
In Solstad, T., Gronn, A. and Haug, D. (eds.) Festschrift for Kjell Johann Sæbø. Oslo: University
of Oslo.
Swanson, Eric () Modality in Language. Philosophy Compass. .. pp. –.
Swanson, Eric () On the treatment of incomparability in ordering semantics and premise
semantics. Journal of Philosophical Logic. . pp. –.
Thomson, Judith Jarvis () Normativity. Chicago, IL: Open Court.
Wedgwood, Ralph () The Meaning of ‘Ought’. In Shafer-Landau, R. (ed.) Oxford Studies in
Metaethics. . pp. –. New York, NY: Oxford University Press.
Wedgwood, Ralph () Objective and Subjective ‘Ought’. In Charlow, N. and Chrisman, M.
(eds.) Deontic Modality. Oxford: Oxford University Press.
Wheeler, Samuel C. III () Inference and the Logical ‘Ought’. Noûs. .. pp. –.
Wheeler, Samuel C. III () Neo-Davidsonian Metaphysics. New York, NY: Routledge.
Williamson, Timothy () Knowledge and Its Limits. New York, NY: Oxford University Press.
Yalcin, Seth () Probability Operators. Philosophy Compass. .. pp. –.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i


On a Shared Property of Deontic
and Epistemic Modals
Jessica Rett

 Introduction
Modals such as the English must and may can receive an epistemic or deontic inter-
pretation to the exclusion of some other interpretations (e.g. an ability interpretation,
Hacquard, ). The same is true for some modals in other languages: mesti (‘must’)
in Malay (Drubig, ), Egyptian Arabic laazim (‘must’), and Tamil modal suffixes
(Palmer, ).1
Despite this lexical tendency to treat deontic and epistemic modality as a natural
class, it’s relatively hard to find ways in which these two flavors behave similarly seman-
tically to the exclusion of others. Instead, epistemic modals seem to behave differently
from all other ‘root’ modal flavors, including deontic modality. In particular, epistemic
modals tend to scope high while root modals tend to scope low, as is evident in their
interaction with other quantifiers and in their temporal anchoring (Ippolito, ;
Hacquard, , ). Given this, from a semantic point of view, it’s surprising to
find a lexical kinship between deontic and epistemic modality across languages.
Meanwhile, epistemic modals display some semantic properties that seem best
encoded in the lexical entries of modals such as must: an anaphoricity requirement
(Stone, ); an indirect evidence requirement (von Fintel and Gillies, ); and a
flexibility in terms of whose knowledge is reflected in the modal base (von Fintel and
Gillies, ). Because must can additionally receive a deontic interpretation, these
proposals inadvertently predict that deontic must displays similar behavior. At the
moment, it is unclear whether these predictions are borne out.

1 I will focus on the contrast between deontic and epistemic interpretations on the one hand and ability
and future interpretations on the other. There are a few additional modal flavors distinguished in the
descriptive literature which are available for must: teleological (e.g. In view of his goals, John must pass the
test) and bouletic (e.g. In view of his desires, John must pass the test). I will set these aside for now, as I am
unaware of any significant way in which they differ from deontic modality. I will however briefly return to
bouletic modals (and should and ought) in §.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

on a shared property of deontic and epistemic modals 

I will argue that we can reconcile the linguistic evidence for a kinship between
deontic and epistemic modality with at least some recognized properties of epistemic
modality if we assume that modals such as must impose what I’ll call an ‘inference
requirement’: a requirement that the prejacent be inferred from some premises, a set
of propositions reflecting contingent and logical assumptions about the context at the
time of utterance. This proposal is effectively an extension of the epistemic modal
account in Stone () to deontic modals, following suggestions in Glass (). In the
case of epistemics, the consequence is an inferential evidence requirement (similar to
the one described in von Fintel and Gillies, ). In the case of deontics, as discussed
in Lance and Little (), this same requirement is manifested as a normative claim
that admits of exceptions.2
My evidence for treating deontic and epistemic modality as a natural class in this
way comes from the unacceptability of must (and to a large extent may) in inversion
exclamatives, exemplified in ().3
() a. *(Wow,) Must/May Sue be the murderer! epistemic
b. *(Wow,) Must/May Sue complete the assignment on time! deontic
In contrast, inversion exclamatives can be headed by modals with other interpreta-
tions, as illustrated in ().
() a. (Wow,) Can Sue dance! ability
b. (Wow,) Will Sue be mad! future indicative
c. (Wow,) Would Sue like to win the race! future subjunctive
I will argue that exclamations are unacceptable in contexts in which the speaker
has inferential evidence for the prejacent; that this is what explains the contrast in
acceptability between () and (), and gives credence to an account in which deontic
and epistemic modality both indicate that the prejacent was inferred from a salient set
of premises.
In this paper, I will be focusing on the so-called ‘strong’ necessity modal must (and
its possibility counterpart may), leaving aside the so-called ‘weak’ necessity modals
ought and should. While these weak necessity modals arguably receive epistemic in
addition to deontic interpretations (e.g. It ought to be raining by now), they differ in
substantial enough ways from strong necessity modals to warrant putting them aside,
at least for now (von Fintel and Iatridou, ). It is, however, worthwhile noting that
these weak necessity modals are in general unacceptable in inversion exclamatives.

2 I am indebted to Nate Charlow for pointing out the relevance of this work.
3 This empirical claim is sensitive to certain intonational and stress patterns; for instance, the inversion
exclamative *Man, must Sue complete the assignment! has a distinct interpretation from a superficially
similar utterance with no intonation break after man: MAN must Sue complete the assignment! (McCready,
). I will address these subtleties in §.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 jessica rett

 A Semantics for Deontic and Epistemic Must

In this discussion, I’ll focus on matrix (read: unembedded) sentences that lack other
scoping elements. I’ll begin by discussing the evidential restriction encoded by epis-
temic modals (§.) and one strength-based account of it; I’ll then present the analysis
of epistemic modals as anaphoric to generalizations in Stone  (§.). In §., I
extend a version of this analysis to deontic modals, allowing for a unified account of
the deontic and epistemic interpretations of modals such as must.

. The evidential requirement on epistemic modals

It’s widely recognized that matrix sentences containing bare epistemic modals M, of
the form M(φ), are infelicitous in contexts in which the speaker has direct evidence for
φ (Karttunen, , though see De Haan, , for a dissenting view). This is typically
illustrated with the pair in ():
() a. It is raining.
b. It must be raining.
In a context in which the speaker is looking out the window, watching the rain fall, (a)
is acceptable, but (b) is not. In contrast, (b) is acceptable and natural in a context in
which the speaker has only seen someone enter the building with a wet umbrella.
As von Fintel and Gillies () point out, this difference complicates discussion
of ‘Karttunen’s Problem’: the question of which is logically stronger: φ or must(φ).
Whereas many have concluded that must(φ) is weaker (Groenendijk and Stokhof,
; Lyons, ; Kratzer, ), von Fintel and Gillies conclude, based on the
evidential complications, that this is a false dichotomy (, p. ): ‘weakness and
indirectness are not two sides of a single coin at all. They are just different. Any
arguments for a weak semantics need to be more than just reminders that must carries
an indirect evidential signal.’
There is an important question of exactly what sort of evidence must requires (I’ll
discuss epistemic possibility modals at the end of the section). I will agree with the
characterization, in von Fintel and Gillies (), that must requires ‘the presence of
an indirect inference or deduction rather than of a direct observation’ (p. ). I will
however treat this requirement as an inferential evidence requirement, as opposed to
an indirect evidence requirement. This distinction will be made clear in what follows.
Von Fintel and Gillies, following Willett (), endorse a particular typology
of evidence based on distinctions lexically encoded in evidential markers across
languages, as shown in Figure  (see also De Haan, ; Aikhenvald, ). As is
appropriate for evidential languages, this typology differentiates based on the type of
the evidence received, not on the strength of the evidence received.
A speaker has direct evidence for a proposition that references an eventuality iff the
speaker has seen, heard or otherwise sensorily perceived that eventuality. Examples
include seeing it rain; hearing the bell ring; and smelling the pie burning. A speaker

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

on a shared property of deontic and epistemic modals 

Evidence type

Direct Indirect

Visual Auditory Other sensory Reported Inference

Results Reasoning

Secondhand Thirdhand Folklore

Figure 

has reported evidence for a proposition φ if the speaker has learned that φ from some
third party, or if φ is a matter of common cultural knowledge (folklore), etc.
Finally, there are two ways, according to this typology, in which a speaker can
have inferential evidence for a proposition φ: the speaker can infer that φ from some
physical evidence that she interprets as an indicator of φ (the ‘results’ reading); or she
can infer that φ from other things she knows (the ‘reasoning’ reading). An example of
the former is inferring that John is home from seeing his shoes and bag in the hallway;
an example of the latter is inferring that John is home from knowing what time it is
coupled with knowing John’s schedule (and that he is reliable and punctual, etc.).
These evidential distinctions are demonstrated below with data from Tsafiki, a
Barbacoan language spoken in Ecuador (Dickinson, , pp. –). (-a) demon-
strates the direct-evidence interpretation that Tsafiki sentences which lack an overt
evidential marker receive. (-b) illustrates the Tsafiki reportative evidential. Tsafiki
has two distinct inferential evidentials: the one in (-c) is used when the speaker has
inferred the prejacent from sensory evidence, while the one in (-d) is used when the
speaker has inferred the prejacent from information already in her knowledge base.
Notably, evidential languages mark the strongest evidence in a situation in which the
speaker has more than one type of evidence for a proposition.
() a. Manuel ano fi-e.
Manuel food ate-decl
‘Manuel ate.’ (The speaker saw him.)
b. Manuel ano fi-nuti-e.
Manuel ano ate-ev -decl
‘Manuel has eaten.’ (They said so.)
c. Manuel ano fi-nu-e.
Manuel food ate-ev -decl
‘Manuel ate.’ (The speaker sees the dirty dishes.)

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 jessica rett

d. Manuel ano fi-nki-e.

Manuel food ate-ev -decl
‘Manuel ate.’ (He always eats at :; it’s now :.)
I will argue here that the evidential requirement encoded in epistemic modals such
as must is an inferential requirement: a requirement that the speaker have inferential
evidence (of either the ‘results’ or ‘reasoning’ variety) for the prejacent.4 This char-
acterization is importantly different from a prohibition against direct evidence (or
a requirement that the speaker have indirect evidence) because, as Figure  shows,
the two make different predictions with respect to reported evidence. It is also dif-
ferent from a prohibition against reliable evidence (a strength-based characterization),
because inferential evidence can be reliable or indefeasible. These different predictions
have been blurred in recent accounts, so I will tease them apart here.
I agree with the observation that utterances of must(φ) ‘signal that the prejacent
was reached through an inference rather than on the basis of direct observation or
trustworthy reports’ (von Fintel and Gillies, , p. ). But von Fintel and Gillies
elsewhere characterize the evidential restriction encoded by must and other epistemic
modals as a prohibition against direct evidence (p. ): ‘the modal signals that [the
speaker’s] information isn’t direct’.
If the evidential restriction encoded by must was really just a prohibition against
direct evidence (instead of a requirement for inferential evidence), we would predict
that must(φ) is acceptable in cases of reported or hearsay evidence. But it is not.
Imagine a situation in which John is listening to the radio using earphones in a
room with Sue. John hears the news report that the President has been shot. In this
scenario, he can relay this news to Sue by saying, The President’s been shot or They
say the President’s been shot, the latter being a typical use of a reportative evidential
in evidential languages. But it is infelicitous, in this context, for John to report The
President must have been shot.5
Furthermore, the evidential restriction encoded by epistemic modals needs to be
a restriction about evidence type, not a restriction about the strength of the evidence
(although Matthewson () presents a dissenting view). In other words, an utterance
of must(φ) requires that the speaker have inferred that φ from some set of premises,
but it doesn’t require that the speaker’s evidence for φ be defeasible or unreliable. While

4 Although see Winans (to appear) for evidence that will is even more picky than this, at least in certain
syntactic contexts.
5 In a scenario in which John hears from Bill that it is raining outside, and Sue then asks him if it’s raining
outside, John can felicitously respond, Bill said so, so it MUST be raining outside. As far as I can tell, these
utterances are different from standard uses of reportative evidentials because they generally require focal
stress on must; they require that the evidence source be made explicit; and they are more natural if the
prejacent already have been introduced in the discourse. In these cases, the utterance also seems to reflect a
certain authority on behalf of the source. In light of the unacceptability of must in more canonical reported
evidence cases, such as the radio example above, I will tentatively suggest that these acceptable cases involve
a third-person report in addition to an inference from the source’s authority, making it a more natural fit for
inferential evidentials. This is in direct contrast to the conclusions drawn in Matthewson ().

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

on a shared property of deontic and epistemic modals 

there might be a tendency for direct evidence to be reliable and for inferential evidence
to be unreliable (you might have more confidence that John is home if you see him at
home than if you infer that he’s home from knowledge of his schedule), evidence type
does not necessarily correlate with evidence strength.
Davis et al. () address this point in great detail. Two brief examples will
suffice to illustrate the point: a speaker could have direct evidence for φ but might
be hallucinating, in which case that evidence is unreliable. Or a speaker could have
inferential evidence from nondefeasible, non-contingent premises, as in the case of
a mathematical proof, the conclusions of which are notoriously quite natural with
epistemic necessity modals, as in Therefore, x must equal .
Despite this, the account in von Fintel and Gillies () is formalized in terms of a
prohibition against reliable or strong evidence rather than a requirement in terms of
the type of evidence. Informally, it requires that the speaker’s evidence for the prejacent
must ‘fail to directly settle whether’ φ (p. ). This might, as they suggest, appropri-
ately prohibit must in cases in which the evidence source is a trustworthy report, but
it does not make the right predictions in scenarios like the two discussed above when
the strength of evidence does not correlate in expected ways with the type of evidence.
More formally, von Fintel and Gillies first define what it means for something to
count as evidence for a proposition: they use the term ‘kernel’—represented as a set of
propositions K—for privileged information that counts as evidence in the context.
They then characterize must as a prohibition on evidence that ‘directly settles’ the
prejacent () in a context of evaluation c and a world of evaluation w.
() Fix a c-relevant kernel K:
i. [[must φ]]c,w is defined only if K does not directly settle [[φ]]c
ii. If defined, [[must φ]]c,w =  iff BK ⊆ [[φ]]c
The first clause of () encodes their version of the evidential restriction as part of the
definedness conditions of a must(φ) statement.6
What it means for a kernel to ‘directly settle’ the prejacent φ is clearly central here;
von Fintel and Gillies say (p. ), ‘The basic intuition is that K can fail to directly
settle whether P even though K entails whether P; epistemic modals carry an evidential
signal that exploits that gap’. As discussed above, this doesn’t necessarily rule out direct
evidence because sensory evidence from a hallucinating individual doesn’t entail the
prejacent; it also doesn’t succeed in allowing all types of inferential evidence, because
inferential evidence from non-defeasible premises does entail the prejacent in the
relevant sense.
I’ll end this section by discussing the scope of the restriction presented here. I’ve
argued that epistemic must encodes an evidential restriction, and that this restriction

6 von Fintel and Gillies (, p. ) characterize the evidential restriction of must as a presupposition
by a process of elimination; they reject the idea that evidentiality is encoded in a conventional implicature
in terms of specific details of the account of conventional implicature in Potts (). But because
evidentials generally introduce new information, the evidential restriction is arguably better thought of
as non-truth-conditional, not-at-issue content (Murray, ).

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 jessica rett

is best characterized as a requirement that the speaker’s evidence be inferential, rather

than a restriction against weak or unreliable evidence. I thus differ with von Fintel and
Gillies on exactly how this restriction should be formulated, but I agree with them that
the restriction isn’t isolated to must. It is arguably a property of epistemic possibility
modals like may and might as well as epistemic ought.
von Fintel and Gillies emphasize that ‘Our claim isn’t that must uniquely carries
this extra evidential component over and above its quantificational oomph. Our claim
is that epistemic modals in general carry this signal’ (p. ). Their lexical entry for
the epistemic possibility modal might (given a context of evaluation c and a world of
evaluation w) is in ().
() Fix a c-relevant kernel K:
i. [[might φ]]c,w is defined only if K does not directly settle [[φ]]c
ii. If defined, [[might φ]]c,w =  iff BK ∩ [[φ]]c = ∅
In (), might has an evidential requirement but encodes a weaker, existential
requirement in its truth conditions. Because might(φ) requires only that there be
some world in the modal base compatible with φ, it is markedly harder to detect the
presence of an evidential restriction of a bare modal. But, because presupposition
and other types of not-at-issue content project outside of negation, we can detect
an evidential requirement when might (and other epistemic possibility modals) are
embedded under negation. () is their example involving epistemic can (von Fintel
and Gillies, , p. ).
() [Billy, seeing brilliant sunshine]
a. It’s not raining.
b. ??It can’t be raining.
In contrast, these sentences are acceptable in a context in which the speaker has only
indirect evidence that it isn’t raining; von Fintel and Gillies label (-b) acceptable in a
context in which Billy sees people enter a windowless room from outside wearing ‘sun
gear’ such as hats, sunglasses and parasols.
It remains to be seen whether this inferential evidential restriction is a necessary
property of epistemic modals across languages (although there’s some evidence it
could be; Matthewson et al., ); for now, I will agree with von Fintel and Gillies
that it is a property of the English epistemic modals must, may, might, and ought. In
the next section, I present an account of this inferential restriction that allows for an
extension to the deontic counterparts of these modals.

. The inferential evidence requirement

In his  paper, Matthew Stone proposed a semantic analysis for epistemic must
formalizing Palmer’s (, p. ) intuition that ‘it is the notion of deduction or
inference from known facts that is the essential feature of must, not just the strength of

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

on a shared property of deontic and epistemic modals 

commitment by the speaker. For must does not have the same kind of meaning as the
adverbs certainly, definitely, etc., which are, indeed, indications simply of the speaker’s
confidence or commitment.’ His is an account of the evidential restriction based on
evidence type rather than evidence strength, and I will adopt it and extend it in what
follows.
Part of Stone’s argument against characterizing the evidential restriction of must in
terms of reliability or validity comes from the frequent use of must to mark conclusions
in mathematical proofs, ‘where every step of reasoning meets the strictest criteria
of validity’ (p. ). Informally, Stone’s analysis of must treats it as anaphoric to an
argument that justifies the conclusion of the prejacent. Doing so, he suggests, ‘serves
[speakers’] communicative intentions to make the dependence of claims on evidence
particularly clear’ (p. ). His informal description of the meaning of must(φ) is in ().
() Some particular collection of facts A, salient in the common ground, provide
(or have provided) a decisive reason to adopt the belief that φ.
Decisive reasons for the belief that φ can be logical (i.e., can entail φ) but can also be
defeasible.
To implement his account, Stone draws on the notion of (potentially defeasible)
generalizations from accounts of default reasoning in Reiter (), Harman (),
and Simari and Loui () (later developed in Horty, ). This is in part because
his account ‘depends on a model of reasoning in which defeasible conclusions may be
adopted on the basis of packages of evidence natural enough that speakers can refer
to them in conversation’ (p. ).
Stone first establishes the nature of the set of possible premises for an inference
(p. ): ‘The formalization starts from a context κ consisting of a set K of established
propositions consisting of ground formulas KC and (logical) rules KN and an addi-
tional set  of defeasible rules.’ He defines what constitutes an argument for a given
proposition: effectively, a set of propositions T (a subset of established propositions K)
counts as an argument for a proposition h in a context κ–, i.e. T, hκ – if h is entailed
by any context that includes T and standard common-ground assumptions in that
context.
Stone also provides a definition of what it means for an argument T to justify a
proposition h in a context κ (notated as κ  T, hκ ). Informally, in Stone’s account,
arguments are constructed step by step from subarguments, and an argument justifies
a proposition if ‘after a certain point in this induction, no further evidence against it
comes to light’ (p. ). This notion of a justified argument is invoked in the definition
of truth in a context κ, given in (). The end result is the semantics for epistemic must in
(), characterized in terms of two semantic arguments: a sentence S and a contextually
salient argument A.
() [[S]] is true in κ if and only if κ  [[S]]) (i.e., if [[S]] is justified in κ).
() [[Must S(A)]] is true in κ if and only if κ  A, [[S]])κ .

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 jessica rett

According to (), must(φ) is true iff there is a salient argument A that justifies φ.
The analysis therefore invokes anaphora to a salient argument to account for the
inferential evidence restriction encoded by epistemic must. This innovation makes a
particularly useful prediction: must(φ) can be false while φ is true. Stone’s example of
such a situation is a context in which two potentially conflicting arguments are taken
for granted (i.e. are in K), namely:
() a. A : A (recently) struck match is a hot match (unless it was wet when
struck).
b. A : Something that has (recently) been boiled is hot and wet.
In a context in which it’s known that a match has been struck, but also that it’s been
boiled, the proposition ‘The match is hot’ is true. But the sentence The match must be
hot is either true or false, depending on whether A or A is a (or the most) contextually
salient argument for must. This is illustrated in the following (felicitous) exchange:
() A: The match was struck, so it must be hot.
B: Well, no. It is hot because it was boiled. It didn’t light.
This example highlights the same point made in von Fintel and Gillies (), namely,
that determining the relative strength of φ and must(φ) is significantly complicated by
the evidential restriction of must.
The account in Stone () reinforces the claim from §. that the semantics of
must (and, arguably, other epistemic modals) should require that the prejacent follow
from an argument or generalization. The above example highlights the need to ensure
that these arguments can (but need not) be defeasible. In the next section, I’ll argue
that this inferential requirement can be used to explain the apparent affinity between
deontic and epistemic modality.

. A common account for deontic and epistemic modals

I claim here that what deontic and epistemic must have in common is that they both
require that the prejacent follow from some argument or generalization. This claim
involves, in effect, extending the semantics of Stone’s () epistemic must to its
polysemous deontic use, following Glass ().
The idea that deontic must statements also require inferential evidence is echoed in
some areas of ethical theory. Lance and Little () draw an explicit parallel between
epistemic and deontic reasoning (in contrast to e.g. reasoning about aesthetics).
According to them, epistemic and deontic reasoning have in common that they are
inferences from a set of potentially defeasible premises (Aristotle’s hos epi to polu,
‘for the most part’ generalizations). Their characterization of the type of inference
involved relies on a notion of a defeasible normative generalization, which itself
‘involves two parts: understanding what happens in circumstances that are in some
sense privileged, and, second, understanding . . . which deviations [are] acceptable’

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

on a shared property of deontic and epistemic modals 

(p. ). In the case of moral reasoning, this amounts to a) knowing whether an
action is wrong in an idealized situation, and b) knowing which (if any) circumstances
could nevertheless justify that action (a context-sensitivity referred to as ‘switch
valence’). Examples include (pp. –) causing pain (acceptable when it’s ‘constitutive
of athletic challenge’); lying (acceptable when done to Nazis, or as part of a game); and
not heeding the express wishes of competent agents (acceptable ‘in the S&M room’).
Effectively, the claim is that ‘[m]oral understanding, while drenched in exception,
is understanding of a structure, not merely a series of instances’ (Lance and Little,
, p. ). This understanding of a structure is the natural result of knowledge of
these normative generalizations along with an ability to infer from them based on
the particulars of a given context. Horty () embraces this parallel, as well, in his
formal treatment of reasons as defaults (parallel to Stone’s  treatment of epistemic
modals): he argues that defeasible generalizations in ethics and epistemology help
account for the generalizations, used in natural language, that are ‘useful for our
ordinary reasoning’ (p. ). However, I take my claim that deontic reasoning involves
inferential evidence to be compatible with the view that prejacents are inferred
from moral or ethical intuitions instead of (or in addition to) general principles
(e.g. Dancy, ).
I am not in a position to evaluate these ethical theories, but I find them encouraging
for a semantic theory that attempts to explain the morphological connection between
epistemic and deontic modals via an inferential evidence restriction. Important for
the present discussion is that deontic and epistemic judgments have in common an
inference from some premises, not—at least for my purposes—that these premises be
defeasible. As Stone () already pointed out, epistemic must is perfectly acceptable
in cases in which the prejacent φ is inferred from a logical truth (or a series of logical
truths), in e.g. mathematical proofs. Lance and Little () argue that, in the case of
moral reasoning, the relevant premises are by nature defeasible, but I will take a weaker
stance, namely that the relevant premises in both cases are potentially defeasible.
There is, however, one important difference between moral judgments and epis-
temic judgments: I can make an epistemic judgment (i.e. decide that φ is true)
without inferring from a set of premises, but (as Lance and Little and others have
argued) I cannot make a deontic judgment without inferring from a set of premises
or an intuition. In other words, deontic judgments depend on deontic reasoning, but
epistemic judgments may or may not depend on epistemic reasoning. In the epistemic
realm, the difference corresponds to the difference between having direct (or reported)
evidence for φ versus having inferential evidence for φ.
The consequence is that a definition of must that requires that the prejacent be
inferred from some premises prohibits the use of epistemic must in certain contexts
(ones not involving inference), but does not prohibit the use of deontic must, because
inference is always involved in cases of moral judgment.
As a result of this difference, an account of deontic and epistemic must in which
it carries an inference requirement—i.e., an extension of Stone’s () account to

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 jessica rett

the deontic domain—predicts that deontic modals will not display something parallel
to an inferential evidence restriction: a commitment that φ is inferred from certain
premises will rule out some uses of epistemic must, but not of deontic must. We will
thus have to look elsewhere for evidence that deontic and epistemic modals have in
common this inferential restriction. This is the topic of §.
Below is a modal-semantic proposal for a common account of deontic and epistemic
must. It represents a basic extension of the inferential analysis of epistemic must in
Stone () to deontic modals. It characterizes must in both cases as acceptable only
if the prejacent is justified according to some salient set of premises K in the context
of utterance. I will therefore refer to these modals as ‘inferential modals’.
This approach was anticipated by Glass (), who argues that deontic and epis-
temic modals have in common that they ‘invoke a body of rules of some sort: for
deontic must, these are normative rules such as “do not litter,” and for epistemic must,
descriptive rules about how things tend to unfold, such as “wet umbrellas indicate
rain” ’ (p. ). Her focus was on addressing Karttunen’s Problem, and she encodes the
inferential restriction in the ordering source of the modal.
I will maintain Stone’s use of the variable K to range over a set of established
propositions, including logical rules and contingent (defeasible) premises.
() For some accessibility relation R and ordering source g salient in c:
a. [[must φ]]c,w is defined iff there is some salient set of propositions K such
that K justifies φ in c.
b. If defined, [[must φ]]c,w is true iff ∀w [wRw → φ(w )].
Like the analysis in von Fintel and Gillies (), () encodes the inferential restriction
as a precondition on the truth or falsity of must(φ). This means that a sentence of the
form must(φ) will be undefined if there is no salient K in c that justifies φ. Unlike the
analysis in Stone (), I will leave unspecified what it means for a set of premises
to justify a proposition, although it is entirely plausible that Stone’s account or some
equivalent can be extended to handle the deontic cases as well as his original epistemic
cases.
Instead of defining the relationship between K and φ in terms of indirectness or K’s
inability to ‘directly settle’ whether φ, () requires there be a set of propositions K that
justifies φ in c. This accounts for must’s apparent reliance on inference while prohibit-
ing neither logical entailments nor defeasible generalizations. Along with others in the
evidential literature, I will assume that some pragmatic mechanism (possibly similar to
scalar implicature) explains why elements that encode a requirement for a weak type
of evidence are generally interpreted as prohibiting stronger types of evidence (e.g.
direct evidence). Alternatively, the restriction in (-a) could be strengthened from an
existential requirement to a requirement about the strongest evidence salient in the
context.
I’ll discuss an epistemic and a deontic example that are correctly predicted to be
acceptable. Imagine that A and B are in a windowless room, watching a group of

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

on a shared property of deontic and epistemic modals 

people enter from outside with wet umbrellas. In this scenario, A can felicitously and
truthfully utter to B, It must be raining outside. () predicts that this utterance is
felicitous because a set of propositions K—containing the proposition that people are
entering from outside with wet umbrellas—is salient in the context of utterance and
also justifies the prejacent under normal circumstances. The sentence is true because
it holds in all the nearby worlds accessible from the world of evaluation by the salient
epistemic modal base.
() makes several further predictions about the above situation. First, that the
utterance is infelicitous or undefined if the relevant premise (that people are entering
with wet umbrellas) is not salient in the context of utterance. Imagine a different
scenario in which A sees people entering in from outside with wet umbrellas, then
walks down the hall to B’s office. In this scenario, A’s utterance It must be raining outside
is infelicitous, unless it is followed by an explicit introduction of the relevant evidence
(...I just saw some people come in with wet umbrellas).
Secondly it predicts that this reference to a set of premises—as an explicit part of
the non-truth-conditional content—can be indirectly denied in discourse, like other
not-at-issue content (Potts, ). This is demonstrated in ().
() A: (Watching people enter with wet umbrellas) It must be raining outside.
B: Hey, wait a minute, they’re washing the roof right now. So you can’t
conclude that for sure.
Thirdly, ()—by virtue of its anaphora to a salient K—preserves the prediction from
Stone’s () account that φ can be true while must(φ) is false, depending on which
K is salient. Stone’s example is repeated below from ():
() A: The match was struck, so it must be hot.
B: Well, no. It is hot because it was boiled. It didn’t light.
A’s utterance is anaphoric to a particular set of premises: the inference from a match
being struck to its being lit to its being hot. B’s response challenges the validity of
this inference while endorsing the conclusion that φ. If these predictions are right,
they support the treatment of the evidential restriction of must as anaphora to a
salient generalization, as opposed to encoding it in e.g. the ordering source (as Glass
() does).
Finally, () predicts that epistemic must(φ) is undefined in a context in which the
speaker does not have inferential evidence for φ (or has better than inferential evidence
for φ). This distinction is made in evidential systems across languages, e.g. the Tsafiki
data in (). Inferential evidence is a type of evidence; it is not characterized relative
to direct and reported evidence by its weakness. In a situation in which the speaker
looks out the window and sees it raining, her conclusion that it is raining comes from
a salient perceptual event, not from inference. In a situation in which the speaker
hears from a colleague that it’s raining, her conclusion comes from a salient speech
event. In these cases, an utterance of It must be raining outside is predicted by ()

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 jessica rett

to be undefined because the conclusion is not justified by inference (or because it is

justified by stronger evidence than inference).
There is a deontic parallel to these epistemic cases. Imagine that A and B know that
Sue and John are driving to Las Vegas tomorrow, and they learn that John doesn’t have
a driver’s license. In this context, A can felicitously and truthfully utter to B, Sue must
do all the driving tomorrow. () predicts the utterance to be felicitous because there
is a salient generalization in the context that justifies the prejacent: namely, that John
doesn’t have a driver’s license. Notice that this information is arguably distinct from
the modal base invoked by the modal, presumably one modeling the rules of the road
in the world of evaluation. The sentence is true in this scenario because it holds in all
of the nearby accessible worlds.
As in the epistemic example, the salient information from which the prejacent is
inferred must be salient in the discourse. In the above scenario, if A and B both know
that Sue and John are driving to Las Vegas, but only A knows that John doesn’t have a
license, the statement Sue must do all the driving tomorrow would be odd unless it was
followed by A’s explanation of her evidence for the prejacent.
By virtue of its salience, K can also be challenged in discourse. Imagine again that A
and B learn together that John doesn’t have a driver’s license, but B additionally knows
that Bill is joining them on their trip.
() A: Sue must do all the driving tomorrow.
B: Hey, wait a minute, Bill has a driver’s license. So you can’t conclude that.
We can also establish a scenario, inspired by the Stone example, that makes it
possible for φ to be true while at least some utterance of must(φ) is false. In a scenario
in which A and B know that Sue and John are driving to Las Vegas tomorrow, the
following exchange is felicitous:
() A: John doesn’t have a license, so Sue must do all the driving tomorrow.
B: Well, no. Sue must do/is doing all the driving tomorrow because she’s the
only one insured on her car.
As in the epistemic cases, this discussion about the inference A is making in her
utterance of must(φ) seems independent of the modal base and ordering source of
the modal: intuitively, the exchange in () holds fixed a circumstantial modal base in
which Sue abides by the laws of the road and an ordering source prioritizing worlds
that are as similar as possible to the world of evaluation.
One final comment on the analysis: unlike the other accounts discussed, () doesn’t
characterize must as lexically restricting the modal bases or accessibility relations
available to must. Stone () and von Fintel and Gillies () do so because their
definitions of must are explicitly restricted to its epistemic uses; Glass () proposes
two separate but parallel lexical entries for deontic and epistemic must. It remains to
be seen whether such a restriction is necessary; it’s possible that deontic and epistemic
modals are the only ones that carry an inferential requirement.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

on a shared property of deontic and epistemic modals 

Ability modals don’t need to involve inference, as () shows (Austin, ; Brennan,
; Hackl, ).
() John can ride horses/a horse. ability
() is acceptable in a context in which the speaker has seen John ride a horse before,
which arguably counts as direct evidence for the prejacent. It also seems acceptable in
a context in which the speaker has been informed that John rides horses. This suggests,
in contrast to deontic and epistemic must, that ability can does not encode an inference
requirement, or isn’t an inference modal.7
Future modals, as well, don’t seem to encode an inferential restriction, although this
is less clear.8 The example in () suggests that will can be interpreted with respect to
an epistemic modal base (although this use might be restricted to generic statements;
Haegeman, ).9
() As far as I know, oil will float on the water.
It is relatively hard to conceive of what would count as direct, perception-based
evidence for a claim about the future (crystal balls aside). But we can conclude that
will is not an inferential modal because it is compatible with even the strongest, most
reliable reported evidence for an eventuality. Imagine that Mary’s daughter Sue is
graduating from high school and is deciding which university to attend. On the day
she is required to make a decision, Sue announces to Mary that she will attend UCLA.
Mary can then report to her friend:
() Sue will attend UCLA in the fall.
In this scenario, Mary has reported evidence for the prejacent, resulting in a felicitous
utterance. This suggests that the future will, unlike deontic and epistemic must, does
not carry an inferential restriction.

7 It’s interesting to note that can also receives deontic or epistemic interpretation, depending on context.
((-b) is an example of the former; as an example of the latter, Hackl () gives: John can be married to
his cousin, according to law.) These interpretations, like the deontic and epistemic interpretations of must,
seem to have an inferential requirement. I’ve suggested here that () is evidence that can—in contrast to
must—doesn’t lexicalize an inferential requirement; if this is the case, it remains to be seen why the deontic
and epistemic interpretations of can are restricted to inferential evidence.
8 See Klecha () for very compelling arguments that will is a modal. His analysis, in (i), treats will as
encoding universal quantification over worlds in the modal base M as well as existential quantification over
future times j.
(i) [[will]]w,i = λp∀v ∈ M(w, i)[∃j > i [p(j)(v)]]
Klecha offers several empirical arguments in favor of a modal analysis of will and its subjunctive counterpart
would. Primary among them is the ability of will to interact with if -clauses in conditionals and to participate
in modal subordination. Winans (to appear) presents an account of will as an evidential modal that derives
its use as a future marker from its evidential restriction, which she characterizes as non abductive.
9 There is no clear consensus on this issue, however. Giannakidou and Mari () argue that the future
does not encode epistemic knowledge but rather direct or reported partial knowledge. They argue that this
difference corresponds to a difference in speaker confidence for the prejacent (high confidence in the case of
the future, relatively low in the case of epistemics). Palmer () and De Haan () claim that the future
is not an evidential, and Winans (to appear) argues that will and must are both epistemic modals but differ
in their evidential restrictions.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 jessica rett

In presenting this account of deontic and epistemic must, I’ve made the following
claims: Epistemic must carries an evidential restriction requiring that the speaker have
inferential evidence for the prejacent. This evidential restriction is best characterized,
as in Stone (), as anaphora to a set of premises from which the speaker is justified in
concluding the prejacent. Deontic and epistemic must have this evidential restriction
in common (Glass, ); this correctly predicts that the salient generalization behind
both the epistemic and deontic generalizations can be denied in discourse. And it has
the potential to offer insight into why deontic and epistemic modalities are often co-
lexicalized across languages.
As I’ve suggested, there is some additional linguistic evidence in favor of this analy-
sis. There is independent reason to think that exclamations, a type of expressive speech
act, are incompatible with inferential evidence. And it seems as though exclamations—
or at least exclamations in which the modal is prominent in a certain way—are
unacceptable with deontic and epistemic modals (but not with other modals).

 Modals and Inversion Exclamatives

In this section, I argue that exclamations are incompatible with inferential evidence,
and therefore that they can function as a test for the analysis above; namely, the claim
that deontic and epistemic modals carry an inferential restriction. I begin by review-
ing exclamations generally; I then argue that the data from inversion exclamatives
present at least some compelling reasons to think that deontic modals—like epistemic
modals—carry an inferential requirement.
An exclamation is a type of speech act in which the speaker expresses that the
content of the exclamation is unexpected. I use the term exclamative to refer to a
subset of exclamations formed from syntactic objects other than declarative sentences.
Examples of each subtype of exclamation are in () (from Rett, ).
() a. Robin baked a blueberry pie! sentence exclamation
b. What a pie Robin baked! wh-exclamative
c. (Oh,) The pie Robin baked! nominal exclamative
d. (Boy,) Did Robin bake a pie! inversion exclamative
I use the term ‘express’ in the Kaplan () sense: as content that ‘displays something
which either is or is not the case.’ This unexpectedness is typically speaker-oriented
(Harris and Potts, ), and typically manifests itself as something like surprise. They
can also be uttered insincerely (Searle, ; Rett, ). So a speaker’s utterance of
the wh-exclamative What a beautiful apartment! represents the apartment as having
exceeded her expectations whether or not the representation is accurate. In previous
work, I’ve characterized the illocutionary force of exclamation as a function from a
proposition to a speech act uttered by a speaker s in a context C, as in () (Rett, ).
This definition will be supplemented later in this section.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

on a shared property of deontic and epistemic modals 

() E-Force(p), uttered by sC , is appropriate in a context C if p is salient and true

in wC . When appropriate, E-Force(p) counts as an expression that sC had not
expected that p.
Exclamatives are exclamations formed from strings other than declarative sen-
tences. Inversion exclamatives, which receive disproportionate attention in what
follows, are formed with sentences that display subject-auxiliary inversion. When a
declarative sentence contains no auxiliary, its tense is inverted (and the verb do is
inserted to support it morphologically), as in ().10
() a. Robin baked a pie. declarative sentence
b. Did Robin bake a pie! inversion exclamative
In other work, I’ve argued that exclamatives (but not sentence exclamations) have
in common that they denote degree properties at some level (Rett, , ). Among
other things, this explains why wh-exclamatives cannot be headed by e.g. who (*Wow,
who she met the other day!), and why exclamatives like What scholars we met yesterday!
can be used to exclaim about the degree to which the scholars were scholarly but not
to express that the scholars met yesterday were unexpected.
. Inversion exclamatives and inferential evidence
In this section, I focus on inversion exclamatives in particular. I return in §. to
discuss the extent to which the claims made here extend to exclamations generally.
Inversion exclamatives are infelicitous in contexts in which the speaker’s strongest
evidence for the content of the exclamative is inferential. I will motivate this claim
using inversion exclamatives that do not contain modal auxiliaries; I will then discuss
the consequences of this claim for the class of inferential modals.
The inversion exclamative in () is most naturally interpreted as expressing the
speaker’s surprise at how fast the car goes. In what follows, I’ll assume that () is
uttered in a context in which the car’s reference is clear, the car’s speed is in fact high,
and both are salient.
() (Wow,) Does that car go fast!
() is felicitous in this type of context if the speaker has just witnessed the car going
fast. This is a direct evidence context, where the speaker has visual evidence of the
content of the exclamative. It is also felicitous in a context in which the speaker has
just heard from an interlocutor that the car goes up to mph. This is a hearsay or
reportative evidence context.

10 Subject-auxiliary inversion is most commonly associated with yes/no questions in English, but
there are many ways in which inversion exclamatives differ from yes/no questions (see McCawley, ),
suggesting that the inversion in exclamatives comes about for other reasons, perhaps akin to the inversion
triggered by only (Progovac, ).
(i) Only last night *(I did)/(did I) eat pizza for the first time.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 jessica rett

In contrast, () is infelicitous in a context in which the speaker has only inferred
the content of the exclamative. It is infelicitous in a context in which the speaker has
examined the engine under the hood (but not e.g. seen the car perform), and it is
infelicitous in a context in which the speaker knows that John only drives incredibly
fast cars, and sees a picture of John in the car in question. This is in contrast to the
sentence That car must go fast, which (as we’ve seen) is compatible with inferential
evidence.
I interpret this as showing that inversion exclamatives carry an evidential restric-
tion: they are infelicitous if the speaker’s best evidence for their content is inferential.
Inversion exclamatives thus provide a useful test for which modal auxiliaries require
inferential evidence and which do not.
Recall that inversion exclamatives can be headed by modals as well as by the
auxiliary verb do, as in () (repeated in ()). But as McCawley () observed, they
are unacceptable with deontic and epistemic modals, as in () (repeated in ()).
() a. (Wow,) Can Sue dance! ability
b. (Wow,) Will Sue be mad! future indicative
c. (Wow,) Would Sue like to win the race! future subjunctive
() a. *(Wow,) Must/May Sue be the murderer! epistemic
b. *(Wow,) Must/May Sue complete the assignment on time! deontic
The judgments in () and () reflect the utterance of these strings as exclamatives,
which (in the case of inversion exclamatives) means that they receive an intonation
pattern distinct from that of yes/no questions, in particular a high-level intonation with
emphasis, typically manifested in lengthening effects (Bartels, ). This intonation
is brought out by such particles as wow, but these particles are not obligatory.11
We can see that the contrast in () and () tracks the speaker’s type of evidence.
Imagine that Mary is a detective investigating a murder, and she has just discovered
incontrovertible proof that Sue perpetrated the crime (say, she received the results of a
critical DNA test). In this scenario, her conclusion that Sue is the murderer is based on
an inference from a set of premises, which licenses an epistemic modal in the assertion

11 Some particles, for instance man and boy, can cause an inversion sentence to receive a reading slightly
different from an exclamative interpretation. The two interpretations are exemplified in (i), and discussed
at length in McCready ().
(i) a. Man, does Robin like cake!
b. MAN does Robin like cake!
The particle in (i-a) receives what McCready refers to as ‘comma intonation’, and corresponds to the
inversion exclamatives addressed here. In contrast, (i-b) receives McCready’s ‘integrated intonation’, and
receives an interpretation similar to but arguably distinct from exclamatives. In what follows, I will make
claims only about exclamations like (i-a), and will stick to the particle wow to attempt to bring out this
intonation unambiguously.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

on a shared property of deontic and epistemic modals 

in (-a). But in this scenario, despite the relevance and truth of the proposition, and
despite Mary’s having recently learned it, an utterance of (-b) is unacceptable. It
cannot be felicitously used to express surprise at the proposition that Sue must be the
murderer (or even at the extent to which she must be the murderer).
() a. Sue must be the murderer.
b. *Wow, must Sue be the murderer!
The same can be done with deontic modals. Imagine that Bill is Sue’s father and
helps keep track of Sue’s schoolwork. Sue’s teacher imposes relatively modest penalties
for late assignments, but Sue still isn’t doing well in the class. Bill has just learned,
however, that the teacher will not be accepting this particular homework assignment
late; he has also just learned that Sue needs to pass the assignment in order to pass
the class. In this scenario, his utterance of (-a) is appropriate and felicitous; he has
inferred the prejacent from a set of premises (including the teacher’s new policy and
Sue’s current grade in the class). However, he cannot felicitously utter (-b); it cannot
be used to express surprise at the proposition that Sue must complete the assignment
on time (or even at the extent to which she must).
() a. Sue must complete the assignment on time.
b. *Wow, must Sue complete the assignment on time!
In contrast is the ability interpretation of e.g. can. Imagine a scenario in which Beth
is at a club, watching Sue dance the samba extremely well. In this situation, it’s felicitous
for her to utter (-a) to Joe. It is also felicitous for her to utter (-b). In this context,
Beth’s utterance of (-b) counts as an expression of surprise that Sue can dance (or,
more accurately, about the great extent to which she can; Rett, ).
() a. Sue can dance.
b. Wow, can Sue dance!
As discussed in §., this is arguably because ability modals are not inferential
modals; there is no sense in which a claim about John’s abilities requires inference
(although it is compatible with inference). In fact, Hackl () uses compatibility with
direct evidence as a test for the difference between epistemic uses of can and its uses
as a circumstantial, opportunity or ability modal (p. ).
() shows that the modals will and would pattern with can in this respect, suggest-
ing that they are not inferential modals. As I argued in §., this is plausibly because
these modals also do not encode an inferential restriction. Imagine a scenario in
which Mary hears from her meteorologist friend that it will rain a substantial amount
tomorrow. In such a context, she can felicitously utter either (-a) or (-b).
() a. It will rain tomorrow.
b. Wow, will it rain tomorrow!

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 jessica rett

It therefore seems that inversion exclamatives differentiates between inferential and

non-inferential modals, which is what we expect given that exclamations seem to be
incompatible with inferential evidence.
. Alternative explanations
In this section, I discuss and reject several alternative explanations for the distinction
in () and (). First, the unacceptability of deontic and epistemic modals in inversion
exclamatives cannot be attributed to the non-assertoric force of inversion exclama-
tives. Yes/no questions also lack assertoric force, but license deontic and epistemic
modals.
() a. Can Sue dance? ability
b. Will/Would Sue be mad? future
c. Might Sue be the murderer? epistemic
d. Must Sue complete the assignment on time? deontic
In contrast, auxiliary modals are unacceptable as the heads of imperatives (compare
*Can dance! for Be able to dance!), which suggests that the distinction in () and ()
is exclamative-specific.
Secondly, it doesn’t appear as though the distinction in () and () reflects
independent syntactic differences between must/may and the other modals.12 Must
and may are generally unacceptable with cliticized negation, in contrast to other
modals, as () shows. When available, the negation in mustn’t takes narrow scope, in
contrast to the negation in e.g. can’t. The difference between these classes of modals has
been attributed to certain syntactic restrictions on the movement of these auxiliaries
(Zwicky and Pullum, ).
() a. John can’t/won’t/wouldn’t go to the party.
b. John *mayn’t/?mustn’t go to the party.
The weak necessity modal should is instructive here: it can occur with cliticized
negation (e.g. shouldn’t) but it is nevertheless unacceptable in inversion exclamatives
(e.g. *(Wow,) Should Sue complete the assignment on time!). While I have put weak
necessity modals aside as a more complicated case than strong necessity modals,
I take these data to show that an account that reduces the data in () and () to
syntactic differences will not be sufficient. Certainly, all of these modals can participate
in subject/auxiliary inversion in yes/no questions. On a related note, it also seems as
though an account of must as a positive polarity item (PPI; Iatridou and Zeijlstra, ;
Homer, ) will not suffice to account for (), as may is not characterized as a PPI
in these theories.
Thirdly, exclamatives are scalar; I’ve argued that they denote degree properties at
some level (Rett, , ). Yet inferential modals don’t differ from other modals

12 Thanks to Lauren Winans (p.c.) for help with this point.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

on a shared property of deontic and epistemic modals 

in their gradability.13 A typical test for gradability is the ability to be modified by an

intensifier such as very or to form a comparative. None of the auxiliary modals in ()
and () pass these tests; (-b) is from Lassiter ().
() a. Sam can/will (*more than Bill) dance (*more than Bill). ability/future
b. Sam must/may (*more than Bill) be at home (*more than Bill). epistemic/
deontic
There is of course an acceptable reading of these sentences in which they mean
something like, ‘Sam is capable of dancing more often than Bill’ or ‘It must be the
case that Sam is home more often than Bill.’ This is a frequency reading, in which the
comparative targets the event variable of the verb, not the modal; see Doetjes ()
for details.
Fourthly, there are a number of ways in which epistemic modals have been observed
to differ from other modals—in particular, root modals—but in these phenomena,
deontic modals pattern with other root modals, not epistemics. The differences can
be best summarized in terms of scope: epistemic modals tend to be interpreted high;
root modals tend to be interpreted low (Hacquard, , ). This is evident in their
interaction with other quantifiers; epistemic modals tend to scope above quantifiers
such as every, and root modals below (Brennan, ).
() a. #Every radio can get Chicago stations and no radio can get Chicago
stations. ability
b. #(By law,) Every radio may get Chicago stations and no radio may get
Chicago stations. deontic
c. Every radio may get Chicago stations and no radio may get Chicago
stations. epistemic
The epistemic interpretation of the modal in (-c) is not reported to be contradictory:
the sentence is interpreted as reporting an epistemic possibility that every radio get
Chicago stations, along with an epistemic possibility that none do. The explanation
is that this lack of contradiction is due to the fact that epistemic modals—but not
root modals, including deontic ones—scope outside of individual quantifiers. But
because both epistemic and deontic modals are prohibited from appearing in inversion
exclamatives—to the exclusion of e.g. ability modals—we cannot attribute the differ-
ence between () and () to known scoping differences between epistemic and root
modals.
To sum up: inversion exclamatives—even those without modals—are incompatible
with inferential evidence. It turns out they are also unacceptable with the epistemic

13 While modal auxiliaries don’t seem to be gradable, this does not mean, as Lassiter () and Klecha
() have argued, that their modal bases are intrinsically non-scalar (cf. It’s more that Sam is at home
than that Bill is at home). For our purposes, the relevant conclusion is this: in looking for an explanation
of why deontic and epistemic modals are unacceptable in inversion exclamatives, we cannot appeal to their
lack of gradability, because standard tests for gradability characterize can, will, must and may as equally
non-gradable.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 jessica rett

and deontic interpretations of the modals must and may, which I characterized in §
as ‘inferential modals,’ lexically encoding an inferential evidence requirement.
I have restricted my claim about inferential modals to must and may, but weak
necessity modals such as should and ought are also unacceptable in inversion excla-
matives, as () shows.
() a. *(Wow,) Should he shut his mouth!
b. *(Wow,) Ought John to turn in his homework!
Despite this, I will continue to put weak necessity modals aside; they differ from
strong necessity modals in a number of ways I cannot control for here (Iatridou and
Zeijlstra, ).14 I will also remain agnostic on the status of possible bouletic or
gnomic interpretations of these modals.
. Inferential evidence and other types of exclamation
I’ve argued that inversion exclamatives are unacceptable when the speaker has only
inferential evidence for their content. There is some evidence that this prohibition is
attributable to the nature of mirativity or speaker surprise; there is other evidence that
it is not. For the purpose of this paper, I will stipulate the inferential restriction for
inversion exclamatives. It remains to be seen how general this restriction is, and why
it exists. However, it does seem that the more tolerant an exclamation construction is
of inferential evidence, the more tolerant it is of inferential modals.
Wh-exclamatives are relatively unacceptable in inferential evidence contexts. Hold-
ing fixed the relevance of the President’s tallness (and assuming he is tall in the world
of evaluation), () is acceptable in a context in which the speaker has just seen the
President (a direct evidence context), or read online that he is   (a reportative
context).
() How very tall the President is!
But it seems odd in a context in which the speaker has seen evidence from which she
can (reliably) infer that the President is tall: if she sees the height of the teleprompter at
a speech or if she sees a bathtub at the White House that was custom-made for him. In
these contexts, an epistemic modal seems much more appropriate: The President must
be tall.
And, as expected, possibility inferential modals are unacceptable in
wh-exclamatives (), while non-inferential modals are acceptable ().
() a. *How very tall the President may/might be! epistemic
b. ?What a big car she may drive! deontic
() a. What a big mountain she can climb! ability
b. What a smart student she will be! future

14 As Nate Charlow (p.c.) points out, should seems to be scalar in the way that other modal auxiliaries
aren’t: it’s fine to say Sam should wash his hands after work more than Bill should.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

on a shared property of deontic and epistemic modals 

But the acceptability of deontic and epistemic wh-exclamatives is improved with a

stronger modal force (see also Rett, ).
() a. How very tall the President must be! epistemic
b. What a big car she must drive! deontic
Sentence exclamations are relatively compatible with inferential evidence. () is
acceptable in a context in which the speaker has seen Sue in her new car (i.e. when there
is direct evidence of the proposition). It is also acceptable in a context in which the
speaker has just heard from a friend that Sue has bought a new car (reported evidence
of the proposition).
() (Wow,) Sue bought a car!
It seems relatively acceptable in inferential evidence contexts as well. Suppose that
John knows that Sue has always wanted to buy a car, and not having one is one of
her biggest regrets. Then imagine that John sees a large withdrawal from Sue’s bank
account amounting to the cost of a new car. In this context—an inferential evidence
context—() seems felicitous.
Both inferential and non-inferential modals are acceptable in sentence exclama-
tions.
() a. Wow, John must/might be the murderer! epistemic
b. Wow, John must get a permit for his sailboat! deontic necessity
c. Wow, John may have a ferret for a pet! deontic possibility
d. Wow, Sue can dance! ability
e. Wow, Sue will graduate! future
So it seems as though we cannot attribute the inferential evidence prohibition
in inversion exclamatives to an exclamation-general prohibition against inferential
evidence; in other words, the restriction on inferential evidence demonstrated in §.
does not seem to be attributable to the illocutionary mood of exclamation generally.
However, the differences between inversion exclamatives and the data above suggest
that the prohibition against inference exhibited so strongly in inversion exclamatives
might be sensitive to givenness, and the prohibition against inferential modals might
be sensitive to discourse prominence.
First, givenness: exclamatives with inferential modals are improved in contexts in
which the modal component of the exclamation is already salient in the discourse.
For example, (-b) is acceptable in a context in which it’s been established that Sue is
allowed to drive a particular car, which turns out to be big (cf. Sue is allowed to drive
one car in the lot. And what a big car she may drive!). In these contexts, the modal
proposition is taken for granted, and the exclamative is used to express surprise about
a related or correlating degree.
Second, and related: it seems plausible that inferential modals are more unaccept-
able in inversion exclamatives because in these constructions the modal is syntactically

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 jessica rett

more prominent than in the other types of exclamation. In other words, the syntax
of the inversion construction might prohibit the hearer from ignoring the role of
inference in the speaker’s expression of surprise, something that might be easier in the
other types of exclamation. The relative acceptability of the examples in ()—which
contain but are not headed by inferential modals—speak in favor of an explanation
along these lines.
() a. Wow, does Mary have some debt she must pay back!
b. Wow, can John work out who the murderer must be!
If this is right, then it might be possible to consider the prohibition against infer-
ential modals in inversion exclamatives to be the result of a general incompatibility of
exclamation and inferential evidence; we might explain the differences within the class
of exclamation constructions with some notion of structured content that differenti-
ates between, among other things, the content encoded in an inversion exclamative
headed by must and the examples in (). This is an encouraging possibility, but I am
unable to offer such a refinement here. I’ll begin the final section by summarizing the
arguments so far and offering brief discussion of one potential area for expanding the
discussion here: non-specific indefinites.

 Conclusions and Extensions

There are a few reasons to look for a semantic kinship between deontic and epis-
temic modals. First, there are modals in several languages that lexicalize deontic
and epistemic interpretations to the exclusion of other modal bases or accessibility
relations (Palmer, ). Second, epistemic modals have long been argued to carry an
evidential restriction—described here as a requirement that the speaker have inferred
the prejacent from some set of premises—which several authors have been tempted to
encode in lexical entries of e.g. must (Stone, ; von Fintel and Gillies, ). If must
can receive either an epistemic or a deontic interpretation, and if it lexically encodes
an inferential requirement, we would expect deontic and epistemic must to pattern
similarly in at least some semantic respects.
I’ve argued here that deontic and epistemic must (and may) do share an inferential
restriction (hence the label ‘inferential modals’). In particular, I’ve argued that sen-
tences with must have in common that they require the speaker have inferred the
prejacent from some set of (potentially defeasible) premises, in contrast to having
arrived at the prejacent from direct or reported evidence. This parallel has its roots
in similar parallels drawn in work on moral reasoning (e.g. Lance and Little, )
and in those implementing particular logics involving potentially defeasible inference
rules or default reasoning (e.g. Horty, ). My formal account of this, similar to
the proposal in Glass (), amounted to a simplified extension of the treatment
of epistemic modality in Stone () to deontic interpretations of modals like must
as well.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

on a shared property of deontic and epistemic modals 

If deontic and epistemic modality have in common an inferential requirement, we

would predict that modals across languages either could or could not lexicalize this
requirement. When they do, as with English must and may, the modal is consequently
restricted to only deontic and epistemic interpretations (and the epistemic interpre-
tation would have an evidential restriction like the one observed in §.). If they do
not, for instance English can, the modal might impose neither a restriction on the
type of interpretation it can receive (the modal ‘flavor’) nor an evidential restriction.
They could, alternatively, lexicalize other things; Matthewson () argues that while
modals in English lexicalize modal force but not flavor, modals in St’át’imcets (a Salish
language spoken in Western North America) lexicalize modal flavor but not force.
There is still a significant amount of cross-linguistic work to be done on this topic.
§ explored a way in which this difference between deontic and epistemic modals
on the one hand and other modal flavors on the other—which I’m claiming amounts
to the difference between inferential and non-inferential modals—could be tested for
explicitly. I’ve argued that inversion exclamatives are unacceptable in situations in
which the speaker has only inferential evidence for the content of the exclamation.
This correctly predicts that inferential modals cannot head inversion exclamatives,
despite the fact that they do not differ from other modal auxiliaries in their scalarity or
compatibility with non-assertoric illocutionary force, etc. But there is clearly more to
be said on the topic of inferential evidence and exclamation generally: only possibility
inferential modals are prohibited from wh-exclamatives, and all inferential modals are
acceptable in sentence exclamations.
Nevertheless, the fact that deontic and epistemic modals form a natural class in
their unacceptability in inversion exclamatives is an encouraging fact for those looking
for a semantic corollary to the lexical kinship of deontic and epistemic modality.
Especially in the absence of a better explanation, the fact that both epistemic modals
and exclamations are picky about the sort of evidence the speaker has for the relevant
proposition supports a connection between the two (and its extension to deontic
modality).
Curiously, in addition to inferential modals, inversion exclamatives are incompat-
ible with non-specific indefinites (McCawley, ). (To quote Farkas (), ‘The
notion of specificity in linguistics is notoriously non-specific.’ I will follow Groenendijk
and Stokhof () and Jayez and Tovena () in characterizing an indefinite as
specific when it is used in a context in which the speaker can identify its referent, and as
non-specific otherwise. I’ll refer to this property as ‘speaker identifiability’.) I consider
this prohibition to be at least plausibly related to the prohibition against inferential
modals, and will end by discussing a possible theoretical parallel.
Original accounts of indefinites treated them as existential quantifiers, similar to
DPs headed by the quantifier some (Russell, ; Ludlow and Neale, ); in contrast,
Strawson () proposed that indefinites are referring expressions. These accounts
both failed to differentiate between specific and non-specific indefinites. Recent
theories tend to blend these proposals, analyzing non-specific indefinites as existential

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 jessica rett

quantifiers and specific indefinites as referring expressions. Many such approaches

(Kamp, ; Fodor and Sag, ; Heim, ; Abusch, ) treat indefinites as
denoting (restricted) variables either valued by an assignment function (for specific
indefinites) or bound by a quantifier, existential (for non-specific indefinites) or other.
This is true even for dynamic treatments, e.g. Groenendijk and Stokhof (), which
employ multiple variable assignments to model the indeterminacy associated with
non-specific indefinites.
The English determiners some and a can receive specific or non-specific interpre-
tations, depending on context. The word certain can generally be used to bring out
specific interpretations of indefinites; non-specific interpretations can be brought out
by context or by explicit claims of ignorance ().
() a. Mary met a certain man for dinner. specific
b. Mary met a man for dinner and I need to know who. non-specific
Imagine a scenario in which Mary learns that Joe Schmo will win an extremely
valuable jackpot in the lottery, and that he is from New York. In this context, Mary
can felicitously utter () to exclaim about the amount of money Joe will win.
() (Wow,) Will someone from New York win the lottery!
In contrast, imagine Mary learns that New York’s lottery has hit a particularly high
jackpot, and that the winning ticket will be chosen tonight. In this scenario—in which
all Mary knows about the lottery is that the jackpot is exceptionally high, and that it
will go to someone in New York—it is infelicitous to utter () to exclaim about the
amount of money that was won.15
This contrast between specific and non-specific indefinites is more stark in excla-
matives formed with unambiguous versions of the indefinites, as the contrast between
() and () shows.
() a. (Boy,) Can a certain someone bake a cake! (after eating a slice of cake Mary
baked)
b. (Man,) Did a certain someone win the lottery! Joe will be rolling in dough!
() a. *(Boy,) Can someone or other bake a cake! (after eating a slice of cake at a
restaurant)
b. *(Man,) Did someone or other win the lottery!
It’s worthwhile noting that the specific indefinite exclamatives in ()—in contrast to
versions with proper names—are most natural in situations in which the speaker has

15 Nate Charlow (p.c.) points out that, in this context, the free relative Wow, is whoever purchased that
ticket going to win the lottery! is acceptable. This is a compelling observation, although at least one prominent
analysis of -ever free relatives (von Fintel, ) characterizes the contribution of ever in terms of uncertainty
rather than non-specificity.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

on a shared property of deontic and epistemic modals 

reason to obscure the identity of his intended referent (e.g., the speaker in (-a) knows
Mary baked the cake, but does not want to reveal that he knows).
This intolerance of non-specific indefinites is a property of exclamatives generally,
as the wh-exclamatives in () demonstrate (I use ‘§’ to mark ‘specific interpretation
only’). Imagine a context in which Sue overhears a concerto played in an elevator, but
does not know who wrote it. In this context, she cannot utter the sentences in () to
express surprise at how talented its composer is.
() a. §(Wow,) What a composer someone is!
b. §(Wow,) How very talented someone is!
Like the examples in (), these exclamatives are acceptable only in a situation in which
it’s clear that Sue knows who the composer or talented individual is.
A typical example of the specific/non-specific contrast is in (), from
Von Heusinger ().
() A student in Syntax  cheated on the exam.
a. His name is John. specific
b. We are all trying to figure out who it was. non-specific
Assuming, following Groenendijk and Stokhof () and Jayez and Tovena (),
that non-specific indefinites lack the property of speaker identifiability (that is, that
an indefinite is non-specific iff the speaker cannot identify the individual that satisfies
the description), then non-specific indefinites are licensed only when the speaker has
inferred the existence of an individual that satisfies the description. In other words, if
speaker identifiability is the right way to characterize non-specificity, we might be able
to cash out non-specificity in terms of inferential evidence.
The non-specific interpretation of () requires a context in which the speaker
has evidence that is informative enough to tell them someone cheated, but not so
informative that they can tell who; for example, a context in which the professor’s
answer key was discovered stolen. In such a context, the speaker must infer from this
indirect evidence that someone has cheated. It is an inference based on descriptive
generalizations about other worlds in which answer keys are stolen; it comes about
neither from direct evidence nor from reported evidence.
The non-specific interpretation of the indefinite in (), repeated below, involved a
similar pattern of inference.
() §(Boy,) Will someone from New York win the lottery!
The non-specific interpretation was natural in a context in which the speaker knew
only that New York’s lottery had hit a particularly high jackpot. The leap from this
knowledge to the claim in () involves the premise that someone will buy the
winning ticket, and that that person will be from New York. This, too, represents a
(defeasible) inference based on descriptive generalizations about the world or model of
evaluation.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 jessica rett

I have not done this topic or these data enough justice to make a strong claim; I
only mention them in passing as another phenomenon, that is barred from inversion
exclamatives, to which considerations of inferential evidence could be brought to
bear. If non-specific indefinites denote existential individual quantifiers that carry an
indirectness requirement, they are clear counterparts to deontic and epistemic modals,
and we are closer to understanding their unacceptability in inversion exclamatives.
The plausibility of this comparison is, I believe, bolstered to some small degree
by work in Rullmann et al. (), in which St’át’imcets (Lillooet Salish) modals
are argued to be ‘akin to specific indefinites in the nominal domain’ (p. ) and
are analyzed accordingly with choice functions. If language encodes an inferential
restriction in one domain, it’s reasonable to expect that it might mark inference in
another domain.

Acknowledgments
Thanks to Nate Charlow, Matt Chrisman, Sam Cumming, Lelia Glass, and Lauren Winans for
comments and suggestions, and to my Winter  semantics class for judgments.

References
Abusch, D. () The scope of indefinites. Natural Language Semantics. . pp. –.
Aikhenvald, A. () Evidentiality. Oxford: Oxford University Press.
Austin, J. () Ifs and cans. In Unnson, J. and Warnock, G. (eds.), J.L. Austin: Philosophical
Papers, Part . pp. –. Oxford: Oxford University Press.
Bartels, C. () The intonation of English statements and questions: A compositional interpre-
tation. New York, NY: Garland.
Brennan, V. () Root and epistemic modal auxiliary verbs. Ph.D. thesis. University of Mas-
sachusetts, Amherst.
Dancy, J. () Ethical particularism and morally relevant properties. Mind. . pp. –.
Davis, C., Potts, C. and Speas, M. () The pragmatic values of evidential sentences. In Gibson,
M. and Friedman, T. (eds.), SALT . pp. –. CLC Publications.
De Haan, F. () Evidentiality and epistemic modality: setting boundaries. Southwest Journal
of Linguistics. . pp. –.
Dickinson, C. () Mirativity in Tsafiki. Studies in Language. . pp. –.
Doetjes, J. () Adverbs and quantification: degrees versus frequency. Lingua. . pp. –.
Drubig, H. () On the syntactic form of epistemic modality. Manscript. University of
Tübingen.
Farkas, D. () Specificity distinctions. Journal of Semantics. . pp. –.
Fintel, K. von () Whatever. In Jackson, B. and Matthews, T. (eds.) Proceedings of Semantic
and Linguistic Theory . CLC Publications.
Fintel, K. von and Gillies, A. () Must stay strong. Natural Language Semantics. . pp. –.
Fintel, K. von and Gillies, A. () Might made right. In Egan, A. and Weatherson, B. (eds.)
Epistemic modality. Oxford: Oxford University Press.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

on a shared property of deontic and epistemic modals 

Fintel, K. von and Iatridou, S. () How to say ought in foreign: the composition of weak
necessity modals. In Guéron, J. and Lecarme, J. (eds.) Time and Modality. pp. –. New
York, NY: Springer.
Fodor, J. and Sag, I. () Referential and quantificational indefinites. Linguistics and Philoso-
phy. . pp. –.
Giannakidou, A. and Mari, A. () The future of Greek and Italian: an epistemic analysis. In
Proceedings of the Chicago Linguistic Society. . pp. –.
Glass, L. () Deriving indirectness and questioning entailment for epistemic must. Annual
Meeting of the Linguistic Society of America.
Groenendijk, J. and Stokhof, M. () Modality and conversational information. Theoretical
Linguistics. . pp. –.
Groenendijk, J. and Stokhof, M. () A pragmatic analysis of specificity. In Heny, F. (ed.)
Ambiguities in intensional contexts. Boston, MA: Reidel.
Groenendijk, J. and Stokhof, M. () Dynamic predicate logic. Linguistics and Philosophy. ..
pp. –.
Hackl, M. () On the semantics of ability attributions. Manscrupt. MIT.
Hacquard, V. () On the interaction of aspect and modal auxiliaries. Linguistics and Philos-
ophy. . pp. –.
Hacquard, V. () Modality. In Maienborn, C., von Heusinger, K., and Portner, P. (eds.),
Semantics: An International Handbook of Natural Language Meaning. pp. –. Berlin:
Mouton De Gruyter.
Haegeman, L. () The semantics of will in present-day English: a unified account. Brussels:
Royal Academy of Belgium.
Harman, G. () Change in view: principles of reasoning. Cambridge, MA: MIT Press.
Harris, J. and Potts, C. () Perspective-shifting with appositives and expressives. Linguistics
and Philosophy. . pp. –.
Heim, I. () The semantics of definite and indefinite noun phrases. Ph.D. thesis. University of
Massachusetts, Amherst.
Homer, V. () Neg-raising and positive polarity: the view from modals. Semantics and
Pragmatics. .. pp. –.
Horty, J. () Reasons as defaults. Oxford: Oxford University Press.
Iatridou, S. and Zeijlstra, H. () On the scopal interaction of negation and deontic modals. In
Aloni, M. (ed.) Logic, Language and Meaning Proceedings of the th Amsterdam Colloquium.
In Croum, C., Smith-Stark, T. C., and Weiser, A. (eds.) pp. –. Berlin: Springer-Verlag.
Ippolito, M. () The time of possibilities. Ph.D. thesis. MIT.
Jayez, J. and Tovena, L. () Epistemic determiners. Journal of Semantics. . pp. –.
Kamp, H. () A theory of truth and semantic representation. In Groenendijk, J., Janssen, T.,
and Stokhof, M. (eds.), Formal methods in the study of language. pp. –. Amsterdam:
Mathematical Centre.
Kaplan, D. () The meaning of ‘ouch’ and ‘oops’. In Cornell Conference on Context Depen-
dency. Ithaca, NY: Comell University,  March.
Karttunen, L. () Possible and must. In Kimball, J. (ed.) Syntax and Semantics. Volume .
pp. –. New York, NY: Academic Press.
Klecha, P. () Positive and conditional semantics for gradable modals. In Guevara,
A. A., Chernilovskaya, A., and Nouwen, R. (eds.) Proceedings of Sinn und Bedeutung ,
pp. –. MITWPL.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 jessica rett

Klecha, P. () Diagnosing modality in predictive expressions. Journal of Semantics. . ..

pp. –.
Kratzer, A. () Modality. In von Stechow, A. and Wunderlich, D. (eds.) Semantics: an
International Handbook of Contemporary Research. pp. –. Berlin: De Gruyter.
Lance, M. N. and Little, M. O. () Defending moral particularism. In Dreier, J. (ed.)
Contemporary debates in moral theory. pp. –. Malden, MA: Blackwell.
Lassiter, D. () Measurement and modality: the scalar basis of modal semantics. Ph.D. thesis.
NYU.
Ludlow, P. and Neale, S. () Indefinite descriptions: in defense of Russell. Linguistics and
Philosophy. . pp. –.
Lyons, J. () Semantics . Cambridge: Cambridge University Press.
Matthewson, L. () Cross-linguistic variation in modality systems: the role of mood. Seman-
tics and Pragmatics. . pp. –.
Matthewson, L. () Evidential restrictions on epistemic modals. In Epistemic indefinites.
Oxford: Oxford University Press.
Matthewson, L., Davis, H., and Rullmann, H. () Evidentials are epistemic modals: evidence
from St’át’imcets. Linguistic variation yearbook. . pp. –.
McCawley, N. () Boy! Is syntax easy. In Corum, C., Smith Stark, T. C., and Weiser, A. (eds.)
Papers from the Ninth Regional Meeting of the Chicago Linguistic Society. pp. –.
McCready, E. () What man does. Linguistics and Philosophy. . pp. –.
Murray, S. () Evidentiality and the structure of speech acts. Ph.D. thesis. Rutgers University.
Palmer, F. () Mood and modality. Cambridge: Cambridge University Press.
Potts, C. () The logic of conventional implicatures. Oxford: Oxford University Press.
Potts, C. () Conventional implicature and expressive content. In Maienborn, C., von
Heusinger, K., and Portner, P. (eds.) Semantics: An International Handbook of Natural
Language Meaning. Volume . pp. –. Berlin: Mouton de Gruyter.
Progovac, L. () Negative polarity: entailment and binding. Linguistics and Philosophy. .
pp. –.
Reiter, R. () A logic for default reasoning. Artificial Intelligence. . pp. –.
Rett, J. () Degree modification in natural language. Ph.D. thesis. Rutgers University.
Rett, J. () A degree account of exclamatives. In Proceedings of SALT . CLC Publications.
SALT data: Friedman, T. and Ito, S. (eds.). pp. –.
Rett, J. () Exclamatives, degrees, and speech acts. Linguistics and Philosophy. . pp. –.
Rett, J. () On modal subjectivity. In Paperno, D. (ed.) UCLA Working Papers in Linguistics.
Volume . pp. –. Los Angeles, CA: UCLA.
Rullmann, H., Matthewson, L., and Davis, H. () Modals as distributive indefinites. Natural
Language Semantics. . pp. –.
Russell, B. () On denoting. Mind. . pp. –.
Searle, J. () Speech Acts. Cambridge: Cambridge University Press.
Simari, G. and Loui, R. () A mathematical treatment of defeasible reasoning and its
implementation. Artificial intelligence. . pp. –.
Stone, M. () The reference argument of epistemic must. In Proceedings of the First Interna-
tional Workshop on Computational Semantics. pp. –. ITK.
Strawson, P. () Introduction to logical theory. London: Methuen.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

on a shared property of deontic and epistemic modals 

Von Heusinger, K. () Specificity and definiteness in discourse structure. Journal of Seman-
tics. . pp. –.
Willett, T. () A cross-linguistic survey of the grammaticalization of evidentiality. Studies in
Language. . pp. –.
Winans, L. (to appear). Predictions and Inferences. Ph.D. thesis. University of California, Los
Angeles.
Zwicky, A. and Pullum, G. () Cliticization vs. inflection: English n’t. Language. .
pp. –.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i


Modalities of Normality
Seth Yalcin

 Introduction
The modals ought and should are widely thought to take both deontic and epistemic
readings. At some informal level, this claim is hardly questionable. Considering a
sentence like
() Noam ought to be in his office.
It is evident that there are at least two ways to take this sentence. One of these ways
has a deontic ring, and seems to concern what is normatively called for in some sense;
the other has an epistemic ring, and seems more tied to what it is reasonable to expect
to be true in some sense.
But more could be, and usually is, meant by the idea that these modals can take
both deontic and epistemic readings. On the stronger interpretation, what is meant
is that ought and should each exhibit two kinds of reading (two “flavors” of modality)
which are common to other modals, such that they stand in certain nontrivial logical
relations with those other modals (relative to a choice of flavor). If we say in this sense
that ought and should can be epistemic, what we mean is that they share a certain
epistemically-flavored reading with other modals such as must, have to, may and might,
such that holding fixed this flavor (and relevant features of context), the corresponding
modal sentences exhibit some nontrivial logical interaction or pattern of entailment—
in particular, they can be ordered by logical strength:
must φ, have to φ  ought φ, should φ  may φ, might φ
Indeed, ought and should are often called “weak” necessity modals because they are
thought to be logically weaker than the corresponding “strong” necessity modals such
as must and have to, when we hold fixed a flavor of these modals as (e.g.) epistemic or
deontic. When I talk about the thesis that ought and should possess both deontic and
epistemic readings, I have this stronger sense of “reading” in mind.
This chapter has two main objectives. First, I make a case against the idea that
ought and should are capable of true epistemic readings. Anticipated in certain respects

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

modalities of normality 

by Copley (, ), we will see that these modals admit of a certain reading
distinct from, but easily confused with, the epistemic flavor of modality, which I will
provisionally call the pseudo-epistemic reading. Once we recognize the existence of
the pseudo-epistemic reading, we face the question whether there is any reason to
think that, in addition to it, ought and should can also take a true epistemic reading,
and function as true epistemic modals. I ultimately find no reason to postulate a true
epistemic reading: the pseudo-epistemic reading suffices to explain the data that have
prompted theorists to suppose that there is an epistemic reading.
My second objective is to explore the nature of the pseudo-epistemic reading. I will
suggest that this reading has to do with what is normally the case (compare von
Fintel and latridou, ), or with what it would be reasonable to expect to be true.
I ultimately end up calling the pseudo-epistemic reading the normality reading of
these modals. Inspired by Veltman (), I consider some connections between these
modals and default reasoning. I close by noting some connections between the ground
covered here and the default logic-based approach to deontic modals explored by
Horty ().
To be explicit about it: this paper is not directly about deontic should and ought. It
is about their close pseudo-epistemic cousins. But part of what animates the project
is the hope to shed some indirect light on deontic should and ought by first trying to
get clearer about the most common non-deontic reading of these modals. A relatively
unified theoretical understanding of these two sorts of reading would be desirable if
we could get it. Seeking a unified understanding, we expect there to be some structural
parallels between the pseudo-epistemic modals and their deontic counterparts.

 Pseudo-epistemics
Consider a case which many would, at least initially, take as drawing out the putative
epistemic reading of the English modals ought and should. Suppose Jones is in a
crowded office building when a severe earthquake hits. The building topples. By sheer
accident, nothing falls upon Jones; the building just happens to crumble in such a way
as not to touch the place where he is standing. He emerges from the rubble as the only
survivor. Talking to the media, Jones says in wonderment one or other of the following:
() I should be dead right now.
() I ought to be dead right now.
In a similar vein, Jones’s sister says things like this:
() It’s incredible! That quake was massive. He should be/ought to be dead. We’re
so lucky he survived.
Obviously Jones is not saying that in deontically or bouletically preferred situations, he
is dead; nor is his sister saying that. That is, we don’t have here any (normal) deontic

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 seth yalcin

reading of the weak necessity modals. Instead it seems we have a paradigm case of
the putative epistemic reading. But the modals are not functioning here like epistemic
modals, for observe that no modal which is uncontroversially epistemic works in these
sentences. Notably it would be bizarre for Jones to say:
() # I am probably dead right now.
Further, observe that if () and () were true epistemics, they would on any account
obviously entail the epistemic readings of:
() # I might be dead right now.
() # I may be dead right now.
Generally, sentences that entail defective sentences are defective themselves; () and
() are not defective; so plausibly () and () do not entail ()/(); so plausibly they do
not contain true epistemic modals. To have a temporary label, let’s call the reading that
the weak necessity modals are taking in these sentences the pseudo-epistemic reading.
The first point is that it is easy to mistake the pseudo-epistemic reading for a true
epistemic reading.
Does there exist, in addition to the pseudo-epistemic reading of the weak necessity
modals, a true epistemic modal reading of these modals? We investigate this possibility
below.
An unpublished paper by Bridget Copley (Copley, ) anticipates these observa-
tions. She calls attention to the following contrast:
() # The beer must be cold by now, but it isn’t.
() # The beer may be cold by now, but it isn’t.
() The beer should be cold by now, but it isn’t.
Copley correctly notes that these data present a problem for the idea that the should
in () is epistemic. Of course, again, this does not yet show that should cannot be
epistemic. It shows that there is reading that looks very epistemic but is not.1, 2

1 On her ultimate analysis, Copley allows for an epistemic reading of should. As noted, I will push against
this view.
2 Swanson (), building on Copley, gives another example of the contrast:

(i) They left an hour ago, and there isn’t any traffic. So they should be here by now. But they’re not.
(ii) # They left an hour ago, and there isn’t any traffic. So they must be here by now. But they’re not.
In explanation of the contrast, Swanson writes that with “They must be here by now”, the speaker aims
to add to the common ground the proposition that they are here—a discourse move in tension with the
subsequent remark “But they’re not”—whereas with “They should be here by now”, the speaker does not
attempt to change the common ground in this way (Swanson, , p. ). But this difference cannot be
the full story, since (as effectively noted by Copley ()) it would not account for the contrast between (i)
and the following:
(iii) # They left an hour ago, and there isn’t any traffic. So they might be here by now. But they’re not.
(iv) # They left an hour ago, and there isn’t any traffic. So they probably are here by now. But they’re not.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

modalities of normality 

 Teasing Apart
Closely related data are also noted by Thomson (). Considering the idea that
there is a reading of ought tantamount to the epistemic modal probably, she raises
the following objection:
Consider Rasputin. He was hard to kill. First his assassins poisoned him, then they shot him,
then they finally drowned him. Let us imagine that we were there. Let us suppose that the
assassins fed him pastries dosed with a powerful, fast-acting poison, and then left him alone for
a while, telling him they would be back in half an hour. Half an hour later, one of the assassins
said to the others, confidently, “He ought to be dead by now.” The others agreed, and they went
to look. Rasputin opened his eyes and glared at them. “He ought to be dead by now!” they said,
astonished. It might be thought that when they first said the words, they meant that it was then
probable that he was dead. Not so when they second said the words. By the time they second
said the words, they knew perfectly well that he wasn’t dead. (pp. –)

(The assassins could not have replaced their second remark with “He is probably dead
right now!”) The first use of ought by the assassins would be reckoned by most everyone
as the putative epistemic reading. The second use corresponds to what I have called
the pseudo-epistemic reading. But is there any good reason to insist that there are two
distinct readings of the modal here, rather than a single (pseudo-epistemic) reading?
In response to her case, Thomson recommends the following analysis:
. . . what it calls for is simply that we distinguish: if I say “The car keys ought to be on the hall
table,” then I assert different propositions, according as my state of knowledge is different. If (i)
I don’t know that the car keys are, or that they aren’t, on the hall table, then if I say “They ought
to be on the hall table,” what I mean is that it is probable that they are there. If (ii) I know that
they aren’t there, then if I say “They ought to be on the hall table,” what I mean is that it was
probable that they would be there. (p. )

The thought is that on what I have called the quasi-epistemic reading, ought (and,
I take it, should), unlike probably, can optionally be evaluated relative to a past state
of information, even when superficially appearing in a present tense construction.
Roughly, the thought is that ought and should can mean probably, but they can also
mean was probable that it would be.3
Thomson’s idea is a natural one. However, it is subject to the following counter-
examples. Suppose an urn has five marbles, one black, four white. A marble is selected
at random. We observe it is black. We can say:
() It was probable that the marble selected would be white.

It is not plausible that “They might be here by now” or “They are probably here by now” serve as a means
of adding the proposition that they are here to the common ground; yet (iii) and (iv) are marked. A natural
conclusion to draw is that the should in (i) is not a true epistemic; hence it does not entail the corresponding
epistemic might claim.
3 Thomson later hints that we might need to change this to: ought and should can mean very probably,
but they can also mean was very probable that it would be.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 seth yalcin

But it would be odd to complain:

() ? The marble selected ought to be white.
() ? The marble selected should be white.
Indeed, note it is odd to say these even before the marble is revealed.
Here is a tentative step towards diagnosis (not far from some of the suggestions in
Copley (, ); see also von Fintel and Iatridou (, p. )): to say () or
(), it seems one has to take oneself to be in a situation where the normal way for the
things to unfold is for the marble to be white. And that obviously is in tension with the
stipulations of the case. The probability of white was of course notably higher than that
of black; but this doesn’t imply that the selection of the black marble was abnormal. In
normal situations of this type—marble selection from the urn described—the marble
selected sometimes is black.4
For another example, suppose we are playing craps. Alas, you roll snake eyes (a pair
of ones). You would be correct to believe:
() It was likely that the dice would not come up snake eyes.
However, it would be odd to say, and a mistake to believe, either of the following:
() ? The dice should not have come up snake eyes.
() ? The dice ought not to have come up snake eyes.
A cheater might whisper one of these to her conspirator, right after loading the dice
against snake eyes. Outside this kind of thing, however, it rings false.
Another example, switching to the future tense. You buy a lottery ticket in a fair
lottery. Compare:
() You will very likely lose.
() ? You should lose.
() ? You ought to lose.
The weak necessity modals seem clearly dispreferred, at least when we try to read them
in some epistemically flavored way. Some report getting more of a deontically-flavored
reading than an epistemic-like reading here. Of course, that itself is puzzling, on the
hypothesis that there is a true epistemic reading. Were there such a reading, we would
naturally expect it to shine through in just this kind of example, where the putative
epistemic reading would be far more context-appropriate than the deontic reading.5

4 Another way to put it might be in terms of what one would be entitled to expect. If () or () were true,
one would be entitled to expect, in some sense, that the marble selected would be white. But it is plainly not
true in this case that one would be entitled to expect that. (I am indebted here to conversation with Judy
Thomson.)
5 Notice further that the judgments flip if the situation is that you bought every lottery ticket except one.
In that case,
You should/ought to win.
seems true. Observe that the probability that you win in this scenario is the same as the probability you

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

modalities of normality 

Games of chance supply easy cases where ought and should pull apart from what is or
was epistemically likely or unlikely. But we can find such examples without them. Most
Bostonians are Americans. Hannah just married Henry, a Bostonian. Is that enough
to say: Henry ought to be an American? That feels comparatively worse than: Henry
is probably American.6 Or again, suppose I am struggling to summarize a colleague’s
objection. In the course of doing so, I might say:
() I am probably misunderstanding you.
() It’s very likely I am misunderstanding you.
() I must be misunderstanding you.
By contrast, it would be queer to say:
() ? I ought to be misunderstanding you.
() ? I should be misunderstanding you.
Finlay () adopts the view that what I am calling pseudo-epistemic ought is
basically akin to probably, except:
Whereas ‘probably’ is tied total present evidence, ‘ought’ like other auxiliaries can be relativized
to any background, such as evidence at some previous time t, some subset of evidence at t, or
what some agent s believes, and so on. ()

What we have been observing, however, is that this level of context-sensitivity is prof-
ligate. Such a view would suggest that ought should be fine pretty much everywhere
probably is, and in more places besides. That is not the case. A view like Finlay’s
overgenerates. Overgeneration is what usually happens when the level of contextual
flexibility postulated is out of proportion with the facts.
The earthquake and Rasputin examples were cases where modals which are uncon-
troversially epistemic are marked, but where should and ought can take a perfectly
acceptable reading. The above examples highlight cases where epistemics such as
probably, very likely, and even must are fine, but where should and ought are marked.
There seem to be two possibilities. One is that there is no epistemic reading of should
and ought; there is only what I have called the pseudo-epistemic reading, a reading
whose nature it remains to explain. A second possibility is that should and ought are
yet capable of a true epistemic reading (in addition to the pseudo-epistemic reading),
but it is not one that is even loosely equivalent to something like probably or very
probable.
Either way, these results tell against the idea, in Kratzer (), of classifying
probably/likely semantically with ought and should as a “weak necessity” modal. It

lose in the scenario where you bought only one ticket. Nevertheless, the truth values of the corresponding
should and ought sentences diverge. This highlights the way in which oughts and shoulds come apart from
what is likely.
6 I owe this example to Judy Thomson (p.c.).

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 seth yalcin

is unlikely these operators have structurally analogous semantics. While the kind of
model structure appropriate to probability operators is currently a matter of active
investigation (Yalcin, , b; Lassiter, ; Holliday and Icard, ), the exam-
ples above make it appear doubtful that ought and should possess a reading that is to
be articulated in terms of this kind of structure, as would naturally be expected if they
possessed true epistemic readings. The pseudo-epistemic reading of weak necessity
modals appears to involve some different kind of structure.
I take it that we should not posit possible readings of modals beyond necessity; and
I am unaware of data showing that, above and beyond the pseudo-epistemic reading
of weak necessity modals, we must also recognize a true epistemic reading. I do not
claim to have shown that this further reading does not exist. But the burden of proof is
on those who would wish to recognize such a reading. Meanwhile it seems worthwhile
to understand how far one can get without supposing there is such a reading.
Here I should briefly pause to defend against one kind of objection. Kratzer ()
suggested that it is possible to explicitly control the restriction of a modal with an in
view of -phrase. If one has this view, one will think it is trivial to show that there can be
an epistemic reading of the weak necessity modals. One merely has to cite an example
like this:
() In view of the evidence, Bob ought to be in his office.
But I deny that () settles anything, for I deny it is generally true that in view of -
phrases systematically semantically control the interpretation of modals. The connec-
tion between the interpretation of modals and in view of -phrases is loose and indirect
at best, as the following examples illustrate:
() In view of what we know, you can’t be parked here. (deontic reading of the
modal clearly available)
() In view of what the tribal laws are, the guy performing the ceremony must be
the chief. (epistemic reading of the modal clearly available)
() In view of what the laws say, we should protest them. (deontic reading
preferred, but not one relative to what the laws require)
Examples could be multiplied. It is not at all obvious that there is any canonical overt
realization of the restriction of a modal. So there is no real obstacle here to the view
that the modal in a construction like () is never a true epistemic.
Let me also defend against another sort of objection. Building on Horn (),
Copley () observes the following contrast:
() a. #Xander must be there, in fact, he should be.
b. Xander should be there, in fact, he must be.
One might think that this is evidence that should has a reading on which it is strictly
weaker than epistemic must. But that would be much too fast. Observe:

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

modalities of normality 

() a. #Xander must be there, in fact, he normally is.

b. Xander is normally there, in fact, he must be there.
() a. #Xander must be there, in fact, he usually is.
b. Xander is usually there, in fact, he must be there.
From () it would obviously be a mistake to conclude that normally is on a scale
with, and is strictly weaker than, epistemic must. Similarly for usually. A better
explanation for the contrasts here would be to say that the should/normally/usually-
claims serve to partially address some question under discussion, a question which is
fully answered by the must claim. If we lead with a full answer (with the must claim),
following this immediately with information only partially addressing the question is
pragmatically odd.
Henceforth I proceed under the assumption that there is after all no epistemic
reading of the weak necessity modals. The problem is now to further clarify the nature
of the pseudo-epistemic reading. How should we characterize the examples so far
reviewed? A rough step towards analysis, hinted at already above, would be to say this:
they are cases where normality and probability come apart. The pseudo-epistemic use
of weak necessity modals tracks, not probability, but something more like normality.
Examples like the earthquake and Rasputin cases are ones where what ought to be
the case lines up with what would normally have been the case rather than with what
probably is the case. The probability that Jones is dead is zero, but as he observes, he
ought to be dead. Examples like the marbles case are ones where what ought to be the
case again lines up with what would normally have been the case rather than with either
what probably is the case or with what was likely to have been the case.

 Regrouping
I have recommended against Kratzer’s grouping of probably with the weak necessity
modals, and against the idea of allowing that there is an epistemic reading of these
modals. Still, some might suppose that the apparatus of Kratzer (, ) yet has the
resources to classify the pseudo-epistemic reading. Let me consider one possibility in
this vein. First I briefly review the main features of Kratzer’s approach.
On Kratzer’s system, modals induce quantification over some partially ordered,
restricted class of worlds. The relevant partially ordered, restricted class of worlds is
generally fixed by two ingredients (what Kratzer calls conversational backgrounds): a
modal base and an ordering source. Formally these are both functions from worlds to
sets of propositions. The modal base fixes the restricted class of worlds quantified over,
as a function of the evaluation world: this will just be the set of worlds making all the
propositions delivered by the modal base true (the intersection of the propositions
given by the modal base). The ordering source is then used to induce an order on the
worlds in this domain.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 seth yalcin

Kratzer originally hypothesized that there were two principal kinds of modal base
(corresponding roughly to the traditional linguistic distinction between epistemic and
“root” modals): epistemic and circumstantial. Epistemic modals have epistemic modal
bases. These deliver, relative to an evaluation world, a set of propositions known. Non-
epistemic modals have circumstantial modal bases. As I discuss further below, these
are somewhat harder to give a generic characterization of; but the basic idea is that
given a world, they deliver a set of true propositions characterizing some contextually
determined set of relevant circumstances at that world.
Kratzer further hypothesized that the two kinds of modal base are apt to be ordered
by characteristically different kinds of ordering sources, and that these differences give
rise to the various possible “flavors” of modal operators we observe. Circumstantial
modal bases, when ordered at all, can combine with deontic, teleological (or goal-
oriented), or bouletic ordering sources, and perhaps others. These ordering sources
yield sets of propositions characterizing what is required, what is aimed for, what is
wished for, and so on; and they induce a ranking of the circumstantial modal base
worlds according to how well the worlds conform to what is required, aimed for,
wished for, etc. Give or take some lexical idiosyncrasies, root modals are assumed to
potentially take any of these flavors. Kratzer proposed that epistemic modal bases, by
contrast, were apt to combine with either doxastic or stereotypical ordering sources,
ranking the epistemically accessible worlds according to how well they conform
to certain (contextually fixed) beliefs or to stereotypical situations, respectively. In
particular, a stereotypical conversational background yields, relative to a world, a set
of propositions characterizing the normal course of events in that world.
I set aside the idea that there exists a specifically doxastic reading of epistemic
modals, since there appears to be no evidence for this idea.7 Stereotypical ordering
sources play the more dominant role in the discussion of epistemic modals in Kratzer
(). On a natural reading of her account, stereotypicality is what primarily orders
the epistemic domain (the worlds quantified over by epistemic modals): properties of
stereotypical ordering sources are what grounds the differences between the grades of
epistemic modality (must versus probably, and probably versus might).
Against this, I want to suggest that we work under the hypothesis that stereotypical
ordering sources never play a role in the semantics of epistemic modals. Epistemic

7 The idea that there exists a doxastic reading of epistemic modals is mentioned only once by Kratzer
(), and in passing, without evidence. Portner (), following Kratzer, more explicitly maintains that
there is such a reading of epistemic modals. He supplies a number of examples which, he suggests, illustrate
modals interpreted with an epistemic modal base and a doxastic ordering source. But, for these examples, it
appears possible to hold either that the relevant epistemic modal base is not ordered, or that it orders itself,
or that the order is a primitive feature of the relevant information state. So these examples do not motivate
the idea of a doxastic reading.
(Of course, epistemic modals under explicit belief operators arguably quantify over belief worlds
(Stephenson, , Yalcin, ); but that is orthogonal to the present issue, which is about what kinds
of information can order the worlds quantified over. Moreover, we are talking about what is possible for the
interpretation of modals in abstraction from explicit semantic shifting.)

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

modalities of normality 

modals are not sensitive in any special way to facts of normality or stereotypicality.
These modals are indeed sensitive to bodies of information, and a body of information
may well be modeled by a set of worlds together possibly with some kind of ordering
on those worlds. But the relevant ordering is not well-interpreted as anything like an
ordering by stereotypicality or normality. The relevant notion is instead likelihood,
broadly construed. I take it this suggestion is already motivated by the examples given
above. The epistemic modality primarily concerns what might actually be the case,
and what is likely or unlikely to be the case. It does not concern what is normally
the case.
Of course, truths about what is normally the case may affect what you think
might be the case. But truths about anything can affect what you think might be
the case. The point is that epistemic modality does not have a special semantic
connection to the notion of normality. Epistemic modals are not sensitive to normality
orderings.8
(To some, this will seem like hairsplitting about how to use the technical term
“epistemic modal.” It is not. Epistemic modals are semantically distinctive in ways that
set them apart from other modals in significant respects (see for instance Groenendijk
et al., ; von Fintel and Iatridou, ; Yalcin, , ; Anand and Hacquard,
; Bledin, ). The evidence strongly favors the thesis that the epistemic modals
form a natural class.9 The thesis that ought and should do not admit of epistemic
readings—readings that belong in this class—is substantive and nonterminological.)
Still, Kratzer was not wrong to suggest that some modals can be sensitive to facts
of normality or stereotypicality. This was an important insight. Evidently, the weak
necessity modals can be understood in this way: that is what I have already suggested
about the pseudo-epistemic reading. But for reasons already reviewed, if we stay within
the Kratzerian approach to modality, we do not want to interpret this reading of
the modals as involving quantification over the epistemically accessible worlds (the
worlds quantified over by epistemic modals); that is, we do not want to say that these
modals take an epistemic modal base, and thus are epistemic modals. Again, this
is because we do not want ought φ/should φ to have a reading on which it entails
(epistemic) might φ, or is entailed by (epistemic) must φ. For that runs counter to
our examples.
We have noted that Kratzer bifurcates the space of modal bases into epistemic
and circumstantial varieties. So if we want to stay within the Kratzerian paradigm,
we should say that that the pseudo-epistemic reading of weak necessity modals
corresponds to the combination of a circumstantial modal base with a stereotypical
ordering source.

8 Though there may be a connection between normality and epistemic modality at the level of default
inference. See §.
9 I favor the thesis that the puzzles discussed in Yalcin (, ) are diagnostic of epistemic modals.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 seth yalcin

So for example, the quantification introduced by Jones’s modal operators is

restricted to worlds that hold fixed certain circumstances obtaining in the actual
world, circumstances like the following: a certain earthquake happened, Jones was
in a building at the time, the building collapsed, etc. But we do not hold fixed all the
circumstances known to obtain. In particular, we do not hold fixed the circumstance
that Jones survived. Jones’s pseudo-epistemic modals must be allowed to quantify
over a domain of worlds which include worlds where he was killed. By contrast, true
epistemic modals do not allow this kind of thing. They do not allow this kind of
selective deletion of items of information possessed in context. If it is a part of our
information that Jones—against all appropriate expectations—survived, then it is just
not the case that Jones might be dead.10

 Difficulties
But the “circumstantial modal base/stereotypical ordering source”-analysis faces some
difficulties. Relevant for assessing it is the question whether and what other modals
can take the pseudo-epistemic reading. Consider first the strong necessity modals.
Suppose Jones had instead said:
() ? I must be dead right now.
() ? I have to be dead right now.
The epistemic and deontic readings of these modals are false, owing to the facts of the
context; so we might naturally expect the pseudo-epistemic readings of these modals
to shine through here. Surely Jones is dead in all the relevant normal (circumstan-
tially accessible) worlds—not just, say, those that are especially normal according to
the relevant normality ordering (as Kratzerians might naturally analyze ought and
should11 ). But these constructions are plainly marked. There are two possibilities: (a)
must and have to cannot take the pseudo-epistemic reading; or (b) on the pseudo-
epistemic reading, the strong necessity modals entail their prejacents.

10 Can’t we say something like, “In light of the information in the report, Jones must be dead. But he
isn’t—he’s right there!”—thereby selectively deleting whatever items of information entailed that Jones is
dead? But I take it that an operator like In light of the information in the report is a (hyper)intensional
environment which semantically shifts the state of information relevant for the epistemic modal under it
(see, e.g., Hacquard, ; Yalcin, ; Stephenson, ; Anand and Hacquard, ). The claim made
in the main text concerns bare epistemic modals sentences where no compositionally induced semantic
shifting takes place. There is no tension here.
11 See for instance von Fintel and Iatridou (), in which weak necessity modals are construed as having
a strictly smaller domain of quantification than the corresponding strong necessity modals: “strong necessity
modals say that the prejacent is true in all of the favored worlds, while weak necessity modals say that the
prejacent is true in all of the very best (by some additional measure) among the favored worlds” (p. ).
They implement this idea in a Kratzerian framework by hypothesizing that weak necessity modals generally
involve two ordering sources rather than one (with the second serving to impose a further restriction not
present in the case of the corresponding strong necessity modals). See also Rubinstein ().

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

modalities of normality 

Either proposal seems unexpected from the point of view of the Kratzerian analysis
just described. From a theoretical point of view, (b) is not motivated. If the set of
worlds that pseudo-epistemic have to and must quantify over is some set of worlds
that are normal relative to the evaluation world, there is no conceptual reason that the
evaluation world needs to be in this set. The world can be a very abnormal place, even
relative to its own standards of normality.
One might propose that the strong necessity modals simply universally quantify
over the circumstantial modal base worlds, and not over any further restriction of that
modal base by normality or stereotypicality. Since the actual world must be in this set
(the circumstantial worlds), this would indeed compel the strong necessity modals to
entail their prejacents. Note that this is, in effect, to accept (a): while weak necessity
modals can get a special pseudo-epistemic normality reading, strong necessity modals
would only get a pure circumstantial reading.
Option (a) is also prima facie unattractive from the point of view of the basic
Kratzerian framework. In the context of this account, there is no principled reason
why there should not be a pseudo-epistemic reading of the strong necessity modals—
in the way that the system predicts that there are, for example, deontic readings of
both weak and strong necessity modals. It is a familiar point that often the most
important evidence for linguistic theory is negative evidence—the unavailability of
certain readings or constructions. As we have been observing throughout, modals are
in fact highly restricted in the readings that they can take. The absence of a pseudo-
epistemic reading for the strong necessity modals requires explanation in the context
of Kratzer’s theory.
There are further sources of discomfort with option (a). We are considering the
hypothesis that the pseudo-epistemic reading corresponds to a circumstantial modal
base and a stereotypical ordering source. Other things being equal, then, a pure
circumstantial reading of a strong necessity modal should entail a sentence with a
pseudo-epistemic weak necessity modal. But the intuitive evidence for this entailment
is slim. Consider one of Kratzer’s examples of a case of a strong necessity modal getting
a pure circumstantial reading:
() I must sneeze. (pure circumstantial)
() I have to sneeze. (pure circumstantial)
Neither of these intuitively entails either of:
() ? I ought to sneeze. (pseudo-epistemic)
() ? I should sneeze. (pseudo-epistemic)
One might reply that the pseudo-epistemic weak necessity sentences presuppose (or
entail) that their prejacents are false in some of the relevant circumstantial worlds,
and this blocks the entailment. Whether or not this suggestion is artificial depends
on whether this kind of constraint applies also to other readings of weak necessity

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 seth yalcin

modals. It seems not to be widely assumed in the deontic case, where strong necessity
modals are generally thought to entail the weak ones (indeed, that is where the jargon
of “strong” versus “weak” comes from).
In any case, a further problem is that it is just strange to say these in the relevant kind
of context. Why? Suppose you sense a sneeze coming on, but you are not convinced
it is inevitable. Why don’t () and () seem like natural words of warning, slightly
weaker than the warning conveyed by () and ()? (As deontic ought is thought
to be weaker than deontic must.) We have the intuition that the flavor of modality
is qualitatively different—not just weaker—when we move from the strong necessity
modals here to the weak ones. This is surprising. Even if the pseudo-epistemic readings
of () and () are marked without some additional setup in this kind of scenario, we
naively might have thought that there should be a pure circumstantial reading of ought
and should available, such that () and () can be appropriate when you feel a sneeze
approaching, in the way () and () are. But such a reading seems not to be available.
This requires explanation.12
The evidence so far reviewed suggests—tentatively—the following: strong necessity
modals do not have a pseudo-epistemic reading, and weak necessity modals do not
have a pure circumstantial reading. Further, it is not clear whether circumstantial
strong necessity modals entail the pseudo-epistemic weak necessity modals.
We should wrap up this discussion of the way that circumstantial modality might
interact with the pseudo-epistemic modality by briefly considering the situation with
possibility modals. Kratzer () suggests that can can take a pure circumstantial
reading. If the pseudo-epistemics involve a circumstantial modal base, it would
therefore be natural to expect pseudo-epistemic modal sentences to entail sentences
with pure circumstantial possibility modals. So suppose Jones says:
() ? I can be dead right now.
This sounds like an strangely worded offer, from Jones, to kill himself.
Though odd owing to the facts of the particular example, () arguably does have
a reading on which it is strictly speaking true (and on which the entailment from the
corresponding pseudo-epistemic weak necessity claim goes through). The modal here
takes an ability-like reading, which some would run together with, or closely tie to, the
circumstantial reading. Interestingly, however, this possibility modal (like may) does
not seem capable of being used to express the (true, contextually appropriate) thought

12 The pseudo-epistemic readings of () and () also do not come very naturally, but they do at least
seem clearly available. It is easier to hear if we fill in some special circumstances—for example, that the
speaker has just inhaled some sneezing powder.
Seemingly also relevant to the availability of the pseudo-epistemic reading is the aspectual structure of
these constructions. Notably, in the presence of a telic verb, the pseudo-epistemic reading is rather more
natural with a progressive infinitive, or with some explicit temporal marking:
() I ought to be sneezing.
() I should sneeze any minute now.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

modalities of normality 

that among the relevant normal worlds is a world where Jones is dead. Many speakers
can achieve something like that reading with:
() I could be dead right now. (pseudo-epistemic?)
This tends as a default to get a pure epistemic reading (on which it is marked), but
(especially with the right tone of amazement) it can receive the pseudo-epistemic
reading, or something intuitively close to it.13 This reading is heard as equivalent to
something along the lines of:
() I could have been dead right now.
() I might have been dead right now.
—where these are not expressing the philosopher’s mere metaphysical possibility,
but rather seem instead to effect existential quantification over some nearby normal
worlds.14
Where does this leave us? The idea that the pseudo-epistemic reading involves a
circumstantial modal base and a stereotypical ordering source might yet be massaged
into working. I have raised some questions about it, but I have not shown that it cannot
be done. A successful development would presumably say more about the character
of circumstantial modal bases. It is fair to say that there is not a tremendous amount
of clarity about the nature of circumstantial readings in the literature. Kratzer herself
now seems ambivalent, if not skeptical, about clearly defining this category from a
semantic point of view (see e.g. Kratzer, , p. ). Still, it may be possible to do, for
all I have said.
Nevertheless, in the following sections, I wish to explore a different line of analysis.

 Normality
In light of the examples discussed above, I want to consider the idea that on the
pseudo-epistemic reading, ought φ and should φ are rather directly connected to
something about the way things normally unfold. As a first attempt, let us consider
the following idea:
It should/ought to be that φ ≈ Normally, φ
We could also put the idea in a way more aligned with the fact that these modals
generally take infinitival clauses:

13 Select speakers can hear () in this way as well.

14 Though it should be acknowledged that many speakers would tend to reach instead for a construction
with a telic predicate, as in:
() I could have been killed.
() I might have been killed.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 seth yalcin

first pass
α should/ought to F ≈ It is normal for α to F.
(It is of course highly doubtful that we will be able to exactly paraphrase pseudo-
epistemic ought and should somehow in terms of the word normal. Still, exploring
how close we can get may teach us something.)
The following examples seem to illustrate a prima facie mutual entailment, and
hence motivate first pass:
() Normally, Bob is in his office right now.
It is normal for Bob to be in his office right now.
() Bob should be/ought to be in his office right now.
A difficulty for the first pass is that Rasputin-like cases does not fit this paradigm
as nicely:
() Rasputin should be/ought to be dead now.
() ? Normally, Rasputin is dead now.
() ? It is normal for Rasputin to be dead now.
Normally φ seems to suggest that φ corresponds to some kind of repeatable event, or
a situation of a repeatable type; but, at least out of context, Rasputin’s death does not
strike us as such an event or situation. Yet the pseudo-epistemics are (as already noted)
fine here, even without any special setup. Similarly for the earthquake case:
() ? It is normal for me to be dead now.
The sentences are improved if we combine a normality operator with the sort of
morphology usual to counterfactuals. We can say:
() Normally, Rasputin would be dead now.
() It would be normal for Rasputin to be dead now.
() It would be normal for me to be dead now.
We might therefore adjust the proposal: α ought to F and α should F, on their pseudo-
epistemic readings, mean something roughly equivalent to:
second pass
α should/ought to F ≈ It would be normal for α to F.
(This brings to mind von Fintel and Iatridou’s () observation that crosslinguisti-
cally, ought is often expressed by combining a strong necessity modal with counterfac-
tual morphology. It also brings to mind Swanson’s suggestion that this kind of should
concerns what “one would naturally expect” to be the case (Swanson, , p. ).)
The following seems to be a problem for this approach. On a Monday at the office,
I can say:

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

modalities of normality 

() Bob normally doesn’t come into the office on Mondays; he stays home. So it
would be normal for Bob to be home now. But I see that his car is parked
outside. So he ought to be in his office.
But according to the second pass, there should be a problem here: the sentence It would
be normal for Bob to be home now and the sentence He ought to be in his office should
be in conflict. But evidently they are not. We can affirm them both.
What to say? Perhaps the following: in (), the normal is talking about: normal
for a Monday. By contrast, I tentatively conjecture that the pseudo-epistemic modality
involves a notion of normality which is in some sense “all relevant things considered”.
The ought in () concerns what is normally true on a Monday where Bob’s car
is parked outside—and perhaps factoring in further relevant contextually supplied
information, too. Although Bob is not normally in his office Mondays, it may yet be
true that Bob is normally in his office on those Mondays in which his car is parked
outside. Perhaps this latter fact, or something in the near vicinity, explains why we
accept “He ought to be in his office” in ().
This leaves us with something like:
third pass
α should/ought to F ≈ It would be normal, all relevant things considered, for α to F.
Of course, we should like to say much more about what makes for normality in this
sense. This seems to be a difficult problem, the surface of which I have barely scratched.
The only point I wish to make here is that there seems to be no reason why this kind
of reading of ought and should could not be captured in terms of simple universal
quantification over a domain of normal worlds—the worlds where, roughly, things
in fact unfold as they (all relevant things considered) normally should. Tentatively
embracing this possibility, I will now switch to calling the pseudo-epistemic reading
the normality reading. I take it this reading has a normative quality. Roughly it
expresses something about what one is, or would be, entitled to expect to be true.

 Defaults
Pseudo-epistemic modality now distinguished from epistemic modality, I will try to
take some steps towards clarifying what exactly the relationship is between these
modalities. In the next two sections, I explore two possible points of contact. First,
it may be that the domain of normal worlds is fixed as a function of the set of worlds
relevant for epistemic modals. Second, there may be important relations of default
inference from the pseudo-epistemic modality to the epistemic modality.
Start with the first idea. Even if ought and should, on their normality readings, are
taken to quantify universally over a domain of normal worlds, it may yet be that
this domain of normal worlds is one that is somehow determined as a function of
the set of worlds that epistemic modals quantify over. We could probe this question

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 seth yalcin

by looking to see whether semantically shifting the body of information relevant

for epistemic modals affects judgments about the truth of oughts and shoulds of
normality. Plausibly, the indicative conditional is a device for semantically shifting this
body of information (for arguments, see Gillies, ; Yalcin, , c; Kolodny
and MacFarlane, ; also Stalnaker, ; Heim, ). Crudely, with an indicative
φ → ψ, we consider temporarily updating our information with φ, and check to see
whether the information corresponding to ψ is incorporated therein. It appears that
normality modals are indeed sensitive to such shifts, as it seems we can consistently
affirm both of the following on the normality reading:
() Noam ought to be in his office.
() If Noam was arrested on the way to work, he ought to be in jail.
It would be a mistake to draw the modus tollens inference here, concluding that Noam
was not arrested. Similarly with:
() My car should be parked on the street outside.
() If my car was stolen last night, it should be in a chop shop by now.
We find a structurally analogous situation with the deontic reading of ought and
should—or so argue Kolodny and MacFarlane (), considering parallel examples
for the deontic reading. Such examples suggest that ought φ and should φ are not
persistent in the following sense: if ought/should φ is accepted relative to some state
of information, it does not necessarily follow that it is accepted relative to any strictly
stronger state of information. You are of the view that Noam ought to be in his office.
But if you strictly grew your state of information, keeping all your factual information
and adding to it the proposition that Noam was arrested on the way to work, it may
well no longer be the case that you accept that Noam ought to be in his office. In at
least this way, normality oughts and shoulds seem not to be straightforwardly factual.
In thinking about the non-persistent character of modalities of normality, we do
well to study Veltman (). Veltman explores the possibility of a semantics for
normally which licenses default inferences from claims about what is normally true
to claims about (inter alia) what is presumably true. Following in his path, we could
ask analogous questions concerning our target modals. For example, the following
kind of inference seems to be licensed as a default:
Noam ought to be in his office. (normality reading)
⇒ Presumably, Noam is in his office.
That is: if the only relevant information you have is that Noam ought to be in his office,
then you are in a position to judge: presumably, Noam is in his office. But this conclusion
is licensed only as a default. If one acquires further information of the appropriate
sort, this can defeat the default, rendering the conclusion no longer derivable. Thus
the following pattern is not valid, even though it strictly extends the premises of the
above argument:

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

modalities of normality 

Noam ought to be in his office. (normality reading)

Noam is not in his office.
 Presumably, Noam is in his office.
For this we would obviously require a nonmonotonic consequence relation.15
A default semantics for normality modals in the style of Veltman () could enable
us to predict these patterns of entailment. Let me illustrate, using the tools Veltman
provides to extend the kind of semantic system developed in Yalcin () for treating
epistemic modals.16
We have an intensional semantics with points of evaluation which are pairs of worlds
and what I will call expectation-laden information states. The extensions of expressions
are relativized to such points. An expectation-laden information state is modeled as
a pair of a set of worlds s (the information state) and an expectation pattern . The
information state component reflects a possible stance on what the facts are. The
expectation pattern component reflects a view on what is normal. It is a preorder
(reflexive transitive order) on the domain W of possible worlds, intuitively readable
as is as at least as normal as. We will think of the preorder as induced by a finite set of
propositions N (the normality propositions) taken to characterize a view about what
is normally the case: w  v iff every proposition in N true at v is true also at w.17
(Thus strictly we should write N , but we generally leave the reference to the normality
propositions tacit.) We assume N always contains the logical truth W. We will call a

world max normal if it is in N; i.e.:
Definition . A world w is max normal relative to an expectation pattern , or -
normal, iff for all w ∈ W, w  w .
We lay down the following definitions:
Definition . The normal set relative to an expectation pattern , n , is the set of -
normal worlds. (The normal set relative to an expectation-laden information state i,
ni , is just the normal set relative to i’s expectation pattern.)
Definition . An expectation pattern  is coherent iff n = ∅.
(We assume that in the non-defective case, expectation patterns are coherent.) Now
let O symbolize ought and should on their normality readings. Then we could consider
the idea that Oφ expresses a constraint on the expectation pattern of an information
state, namely that the prejacent φ be true throughout the normal set it determines:

15 A relation R between a set of sentences  and a sentence φ is monotonic iff whenever  ⊆  , {φ :

 
R( , φ)} ⊆ {φ : R( , φ)}; else it is nonmonotonic.
16 Veltman () deploys his tools in the analysis of normally. I will apply many of his formal ideas to
the analysis of normality ought and should instead. I will assume nothing about the correct semantics for
normally.
17 According to von Fintel and Iatridou (), “epistemic ought differs from epistemic must/have to in
being sensitive not just to the hard and fast evidence available in a situation but also to a set of propositions
that describe what is normally the case” (p. ). While I agree that it is useful to appeal to a set of normality
propositions, the idea of taking the domain of quantification for pseudo-epistemics to be a subset of the
epistemically accessible worlds is problematic, for reasons already reviewed.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 seth yalcin


[[Oφ]]w,i = iff ∀w ∈ ni : [[φ]]w ,i = 
It is instructive to compare this to the clauses for the epistemic possibility () and
necessity () modals (defended in Yalcin ()):

[[φ]]w,i = iff ∃w ∈ si : [[φ]]w ,i = 

[[φ]]w,i = iff ∀w ∈ si : [[φ]]w ,i = 
These modals are sensitive to the information-state component, but not the expecta-
tion pattern. We can observe the failure of Oφ to entail φ by defining an appropriate
notion of consequence:18
Definition . φ is incorporated (or accepted, or supported) in an expectation-laden
information state i iff for all w in si : [[φ]]w,i = .
Definition . ψ is an informational consequence of a set of sentences ,   ψ, just
in case: for all information states i that incorporate every φ ∈ , ψ is incorporated
in i.
It is not hard to see that Oφ  φ. For Oφ can be incorporated in an expectation-
laden information state that rules out φ as a candidate for actuality (as an epistemic
possibility). One can of course have the view that the world one inhabits is not among
the most normal possibilities. An expectation-laden information state reflects such a
view when the set of worlds left open by the information state is disjoint from the
normal set determined by its expectation pattern. In this kind of case, the information
one has rules out the possibilities that are max normal.
Still, within any given information state, there will generally be differences in
normality among the worlds left open. It is useful to define the class of worlds that are
most normal within the set of worlds an expectation-laden information state leaves
open:
Definition . A world w is optimal relative to an information state i, or i-optimal, iff
w ∈ si and there is no w ∈ si such that w > w .
Definition . The optimal set relative to an information state i, oi , is the set of i-
optimal worlds.
Using this notion of optimality, we can give a semantics for presumably (P) following
in Veltman’s footsteps:19

[[Pφ]]w,i = iff ∀w ∈ oi : [[φ]]w ,i = 

18 For further discussion of this notion of consequence, see Bledin ().

19 This semantics is surely only an approximation. The restricted embedding potential of presumably
suggests that it is semantically more like a discourse hedge or parenthetical than an ordinary sentential
operator. We are following Veltman in abstracting from this.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

modalities of normality 

Observe that Pφ  φ, since the optimal set will be a nonempty subset of any
nonempty information state.20 Since as recently noted, Oφ  φ, we know Oφ  Pφ.
That is the desired result. We would not approve of Jones’s reasoning like this:
I should be dead right now.
⇒ Presumably, I am dead now.
Still, as noted above, the movement from a normality should/ought-claim to a claim
about what is presumably true plausibly has default license. The intuitive problem with
the above reasoning is that in Jones’s context, the epistemic possibility that he is dead
is eliminated; consequently the corresponding presumably-claim is not licensed. To
consider what patterns of inference are default licensed, we should define a notion of
consequence that restricts to states of information incorporating the premises and the
premises only—states that are informationally minimal. We may do this as follows:
Definition . The minimal states i incorporating a set of sentences  are the states
i such that i incorporates every element of , and there is no i such that (a) i
incorporates every element of , and (b) si ⊂ si .
Definition . ψ is a default consequence of a set of sentences ,  D ψ, just in case
all the minimal states incorporating  incorporate ψ.
Now we can observe that although Oφ  Pφ, Oφ D Pφ. For example, return to:
Noam ought to be in his office.
⇒ Presumably, Noam is in his office.
The minimal states incorporating the premise are those where the information state
component is just W. These are all the states i such that the proposition that Noam is
in his office is true throughout ni . At these states ni coincides with oi , so the inference
is licensed.
Observe next that although Oφ D Pφ, {Oφ, ¬φ} D Pφ. (So D is nonmono-
tonic.) For example, the following is not default licensed:
Noam ought to be in his office.
Noam is not in his office.
⇒ Presumably, Noam is in his office.
The minimal states incorporating the premises are those where the information state
component is just the set of worlds where Noam is not in his office. This means that
at each of these states i, oi is some subset of these worlds. The conclusion drawn is not

20 The right result, intuitively. For example:

() Presumably, Noam is in his office.
⇒ Noam might be in his office.
() ? Noam is not in his office, but presumably he is in his office.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 seth yalcin

incorporated by any of these states. (On the contrary, they all incorporate “Presumably,
Noam is not in his office”—intuitively a correct result.)
In this chapter so far we have seen many reasons to doubt that the following is a
valid inference:
Noam ought to be in his office.
⇒ Noam is probably in his office.
But now we can ask a different question. Is this inference default valid? That idea is
considerably more plausible. If the only relevant information one has is the premise,
the conclusion does seem like an acceptable one to (tentatively) draw. Vindicating this
pattern as a default inference would, it seems, do much to accommodate the intuition
that there is some nontrivial connection between what normally ought or should be
and what is likely. (And correspondingly, it would further subtract from whatever
motivation remains for recognizing a bona fide epistemic reading of ought and should.)
There are various ways in which one might vindicate this pattern as a default
inference. It depends on what semantics for probability operators one assumes. One
possibility is to think of information states as equipped with a probability measure
Pr conditionalized on the state (as discussed for instance in Yalcin (c)). Then one
could require that in minimal states i incorporating only information about normality,
Pri (ni ) > . (or whatever threshold is thought to be appropriate to license probably-
claims in general).21 This would be to say that in abstraction from factual information,
it is likely that the world is normal.

 Information-sensitivity, Nonfactualism
About Normality
We have yet to say what is going on with:
() Noam ought to be in his office.
() If Noam was arrested on the way to work, he ought to be in jail.
Above we noted some theories on which indicative if-clauses are understood to shift
the information state relevant for evaluating the consequent of the conditional. Let
me give a basic example of this kind of semantics. I will restrict attention to indicative
conditionals with nonmodal antecedents.
Definition . For nonmodal sentences φ, the proposition expressed by φ, [φ], is {w :
[[φ]]w,i = } (for arbitrary choice of i).
Definition . For nonmodal φ, the nearest state to i = s,  incorporating φ, i + φ,
is the pair sφ , , where sφ is s ∩ [φ].

21 For further discussion of the appropriate threshold, see Yalcin ().

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

modalities of normality 

Then a simple information-sensitive semantics for the indicative conditional → would

be:
[[φ → ψ]]w,i =  iff i + φ incorporates ψ
We can briefly note that this helps us to model the information sensitivity of epistemic
modals. For example:
() It is not the case that Noam must be in his office.
() If Noam is in the building, he must be in his office.
One can of course rationally accept both of these (and without also accepting the
seeming modus tollens conclusion that Noam is not in his office). In the present
context what we can say is that the if-clause shifts the sets of worlds that the epistemic
necessity modal in () quantifies over to a strictly stronger state of information, one
incorporating the antecedent information. See Yalcin (, c); Gillies () for
further discussion.
Now the thing to observe is that on the semantics so far provided, normality oughts
and shoulds are not information-state sensitive at all; rather, they are expectation-
pattern sensitive. In the context of this account, the putative information-shifting fea-
ture of indicative if-clauses therefore does not help us explain the data in () and ().
There are at least two possibilities. One possibility is that the expectation pattern
of an expectation-laden information state is in fact determined as a function of
the information state. On this view, if we strictly grow the information state, the
expectation pattern systematically changes as result, and (hence) so do the oughts and
shoulds. (Compare Kolodny and MacFarlane, ; Cariani, ; Charlow, ; Silk,
 on deontic modals.) Another possibility is that the if-clause shifts the expectation
pattern “directly”, and not indirectly via its effect on the information state. On this
approach, we should change the semantics for the indicative by showing how to shift
from a given  to a new preorder φ , one which minimally adjusts  somehow as a
function of the antecedent information [φ].
Let me give one abstract, nonlinguistic reason for preferring the former approach. In
modeling modals of normality, we are indirectly modeling states of mind that consist
partly in a view about what the normal course of events consists in, about what ought
to be or should be so in roughly the sense of what one would be entitled to expect
to be true. Plausibly (but not uncontroversially), a view about what ought to be so in
this sense does not reduce to a purely factual view about how things are: that is why
we appeal to expectation patterns over and above information states. Nevertheless,
our views about what (normally) ought to be or should be the case do not float free
from what we take the facts to be. We take there to be a relation of supervenience
between the pattern of expectation we endorse and the way we take the world to be.
We take our views about what we are entitled to expect to be fixed by the facts. One’s
expectations, we take it, should not change unless there are corresponding changes

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 seth yalcin

in how one takes the world to be. (Compare (Gibbard, , chapter ), discussing
normative discourse.)
We could put it like this: one component of one’s state of mind determines an expec-
tation function f , which maps information states to expectation patterns. This reflects
one’s conception of how the information one possesses governs the expectations it
would be appropriate to have. The important feature of the expectation function is that
it is a function: given an information state, we do not take there to be multiple, equally
legitimate expectation patterns. Disagreements between agents about what (normally)
ought to or should be the case may be traceable to differences in the ways that each
agent takes the world to be (their information); but they may also be traceable to a
more fundamental difference in the expectation function each agent endorses.
Taking this approach, we would do well to model expectation-laden information
states as pairs s, f  of an information state and an expectation function. The expec-
tation pattern of an expectation-laden information state i is then just f (s). This will
determine the normal set for the state, and the semantics will be as above. In principle,
this approach can capture the shiftiness apparent in (), as the expectation pattern will
now shift with the information.
It remains to explore what formal constraints (if any) govern f . The matter deserves
separate exploration. But I note that one constraint we surely do not want is the
requirement that the normal worlds determined by a given state s, f  be a subset of s.
This would make it the case that Oφ  φ—the incorrect result, as we have repeatedly
observed.

 Expressing Defaults, Expressing Norms

It is about time to say something about the deontic readings of ought and should.
A natural and simplifying thought would be that the deontic and pseudo-epistemic
readings of ought and should have some nontrivial amount of underlying formal
structure in common. That after all would help explain why the same linguistic
expressions are used for both, and in many languages. In that spirit, the preceding
two sections, if on the right track, would lead us to suspect that the concepts of
default logic would be of some use in theorizing about deontic modals and deontic
modality.
Indeed they seem to be. Horty () in particular provides some illuminating
analyses of deontic oughts using the resources of default logic. On his theory (as I
would describe it), deontic oughts express constraints on default theories. Structurally,
default theories are akin to expectation-laden information states, except that (roughly)
the expectation pattern component is replaced by a set of default rules together with
a preordering of the rules by their relative priority. To a first approximation, a default
rule X → Y is the sort of thing that recommends the transition to proposition Y from
proposition X, as a default. Default rules can be used to encode what conclusions are
epistemically default licensed given some information. (Rough example: the default

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

modalities of normality 

rule corresponding to Birds fly default recommends the conclusion that x flies, given
the information that x is a bird.) Such rules can also be used to capture which actions
are default licensed given some information. (Rough example: the default rule corre-
sponding to Keep your promises recommends that I show up, given that I promised
to.) Horty uses the latter application to clarify deontic oughts. To say φ ought to be the
case is to say that the realization of φ is default recommended, in a sense formalized
with the notion of a default theory.22 Together with the work above, this would seem
to point in the direction of a unified treatment of deontic and pseudo-epistemic oughts
as devices for expressing defaults. I don’t suggest I have achieved that unification here,
of course; I suggest only that it emerges as a promising avenue to pursue.
While Horty’s discussion of oughts is trained chiefly on the deontic reading and
on surrounding issues about practical reasons and deliberation, it is plain he intends
his style of analysis to apply to the notion of an epistemic reason and to theoretical
deliberation, and to what we have been calling pseudo-epistemic oughts. (Not surpris-
ing, since perhaps the standard application of default logic is to theoretical reasoning.)
From a linguistic point of view, it would be worthwhile to investigate the connections
between a semantics for pseudo-epistemic oughts of the sort sketched above and one
based on default theories of Horty’s variety—to see which differences are notional and
which substantive, and where the data point in the cases where they come apart. I leave
this for future work.
Stepping back, I would be inclined to take Horty’s analysis, and the analysis of this
chapter, in what I would call an expressivistic direction, in the style of Yalcin (a). If
it is right that we need something like expectation-laden information states or default
theories to model the language of oughts and shoulds, it needn’t follow that this talk
serves to literally describe expectation-laden states of mind, or the default we embrace,
or our habits of belief or action. Unembedded, these kinds of sentences do not describe
the defaults we embrace; rather, they serve to express those defaults. Yalcin (a)
offers a recipe for making precise of this sense of “express".
Defaults, as formalized here or by Horty, might also sensibly be called norms.
Characteristic of norms, after all, is their defeasibility. Switching terminology, then,
the picture that emerges is that ought and should correspond, in a relatively technical
sense, to modalities of norms, on both their deontic and pseudo-epistemic readings.

Acknowledgments
Thanks to Nate Charlow, Matthew Chrisman and Sophie Dandelet for detailed comments
on an earlier draft. Thanks also to Fabrizio Cariani, Wes Holliday, Daniel Rothschild, Aynat
Rubinstein, Judy Thomson and to participants in my minicourse on modality and dynamics at
the Hebrew University of Jerusalem in May of  for helpful discussion.

22 Actually, he formalizes several possible senses, discussing the prospects for each.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 seth yalcin

References
Anand, Pranav and Hacquard, Valentine () Epistemics and attitudes. Semantics and Prag-
matics. . .
Bledin, Justin () Logic informed. Mind. . . pp. –.
Cariani, Fabrizio () ‘Ought’ and resolution semantics. Noûs. . . pp. –.
Charlow, Nate () What we know and what to do. Synthese. . . pp. –.
Copley, Bridget () So-called epistemic should. Snippets. . pp. –.
Copley, Bridget () What should ‘should’ mean? Unpublished manuscript. [Online]
Available from: http://halshs.archives-ouvertes.fr/docs/////PDF/copley.should.pdf.
[Accessed  Nov .]
Finlay, Stephen () Confusion of Tongues: A Theory of Normative Language. New York, NY:
Oxford University Press.
Fintel, Kai von and Iatridou, Sabine () Epistemic containment. Linguistic Inquiry. . .
pp. –.
Fintel, Kai von and Iatridou, Sabine () How to say ought in Foreign: The composition
of weak necessity modals. In Guéron, Jacqueline and Lecarme, Jacqueline (ed.) Time and
Modality,. pp. –. Springer.
Gibbard, Allan () Thinking How to Live. Cambridge, MA: Harvard University Press.
Gillies, Anthony () Epistemic conditionals and conditional epistemics. Noûs. . .
pp. –.
Gillies, Anthony () Iffiness. Semantics and Pragmatics. . . pp. –.
Groenendijk, Jeroen, Stokhof, Martin, and Veltman, Frank (). Coreference and modality.
In Lappin, Shalom (ed.) Handbook of Contemporary Semantic Theory. pp. –. Malden,
MA: Blackwell.
Hacquard, Valentine () Aspects of Modality. PhD thesis. MIT.
Heim, Irene () On the projection problem for presuppositions. In Barlow, Michael,
Flickinger, Daniel P., and Wescoat, Michael T. (eds.) Proceedings of the Second West Coast
Conference on Formal Linguistics. pp. –. Stanford, CA: Stanford Linguistics Association.
Holliday, Wesley and Icard, Thomas () Measure semantics and qualitative semantics for
epistemic modals. In Snider, T.(ed.) Proceedings of Semantics and Linguistic Theory .
University of California, Santa Cruz, CA, – May. pp. –. CLC Publications.
Horn, Laurence () A Natural History of Negation. Chicago, IL: Chicago University Press.
Horty, John () Reasons as Defaults. New York, NY: Oxford University Press.
Kolodny, Niko and MacFarlane, John () Ifs and oughts. Journal of Philosophy. CVII. .
pp. –.
Kratzer, Angelika () What ‘must’ and ‘can’ must and can mean. Linguistics and Philosophy.
. pp. –.
Kratzer, Angelika () The notional category of modality. In Eikmeyer, Hans-Jürgen and
Rieser, Hannes (eds.) Words, Worlds, and Contexts, New Approaches to Word Semantics.
pp. –. Berlin: de Gruyter.
Kratzer, Angelika () Modality. In von Stechow, Arnim and Wunderlich, Dieter (eds.)
Semantics: An International Handbook of Contemporary Research. pp. –. Berlin: de
Gruyter.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

modalities of normality 

Kratzer, Angelika () Modals and Conditionals: New and Revised Perspectives. Oxford: Oxford
University Press.
Lassiter, Daniel () Measurement and Modality: The Scalar Basis of Modal Semantics. Ph.D.
thesis. New York University.
Portner, Paul () Modality. New York, NY: Oxford University Press.
Rubinstein, Aynat () Roots of Modality. Ph.D. thesis. University of Massachusetts, Amherst.
Silk () Evidence sensitivity in weak necessity deontic modals. Journal of Philosophical Logic.
. . pp. –.
Stalnaker, Robert () Indicative conditionals. Philosophia. . . pp. –.
Stephenson, Tamina C. () Judge dependence, epistemic modals, and predicates of personal
taste. Linguistics and Philosophy. . . pp. –.
Swanson, Eric () Modality in language. Philosophy Compass. . . pp. –.
Thomson, Judith Jarvis () Normativity. Peru, Illinois: Open Court Publishing Company.
Veltman, Frank () Defaults in update semantics. Journal of Philosophical Logic. . .
pp. –.
Yalcin, Seth () Epistemic modals. Mind. . . pp. –.
Yalcin, Seth () Probability operators. Philosophy Compass. . . pp. –.
Yalcin, Seth (a) Bayesian expressivism. Proceedings of the Aristotelian Society, CXII. .
pp. –.
Yalcin, Seth (b) Context probabilism. In Aloni, M, Kimmelman, V., Roelofsen, F., Weidman
Sassoon, G., Schulz, K., and Westera, M. (eds.) Logic, Language, and Meaning: Proceedings of
the th Amsterdam Colloquium. University of Amsterdam, – Ddecember . pp. –.
Berlin: Springer.
Yalcin, Seth (c) A counterexample to Modus Tollens. Journal of Philosophical Logic. . .
pp. –.
Yalcin, Seth () Epistemic modality de re. Ergo. . . pp. –.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i


Extreme and Non-extreme
Deontic Modals
Paul Portner and Aynat Rubinstein

 Introduction
When we argue that our priorities militate in favor of an action, we can do so in a way
that allows for competing alternatives, or in a way which does not. Suppose that A and
B agree that no policy can both provide insurance for uninsured citizens and allow
people to make their own choices:
() A: It is important that our uninsured citizens get insurance.
B: It’s also important that people make their own choices.
A: So how do we balance these things?
() A: It is crucial that our uninsured citizens get insurance.
B: And it’s crucial that we allow people to make their own choices.
A: So we’re stuck.
() A: Our uninsured citizens should get insurance.
B: People should also make their own choices.
A: So how do we balance these things?
() A: Our uninsured citizens must get insurance.
B: And yet people must make their own choices.
A: What can we do?!
We refer to operators such as important, crucial, should and must as deontic necessity
modals.1 Our goal in this chapter is to contribute to the understanding of the

1 This use of the term ‘necessity’ is common in linguistics but is more inclusive than is usual in
philosophy. It reflects a theoretical assumption that the operators with weaker force (should, important)
have the same quantificational core as clearer examples of necessity modals such as must. In a similar way,
linguists often assume that attitude verbs such as believe and want and upper-range probability operators
such as likely and probable have a necessity semantics (Hintikka, ; Kratzer, , ).
The modals in question are not strictly limited to deontic interpretations, but rather have the full range
of priority readings (deontic, teleological, and buletic; see Portner ). These subtypes are similar by
linguistic criteria, and the use of a particular modal in context is often indeterminate as to its subtype.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

extreme and non-extreme deontic modals 

differences between the two types of deontic necessity statements that they produce.
What is it about must and crucial that leads to an impasse, while should and important
do not?
This chapter will focus on the semantic and pragmatic analysis of the modal
elements in ()–(). We situate our work with respect to two broader issues in the
semantics of modality. The first concerns the nature of differences in strength among
necessity modals. In (), we see strong necessity modals paired with their weak
necessity counterparts:
() (a) You must call Barbara. (entails You should . . .)
(b) It is crucial to talk to Barbara. (entails It is important . . .)
(c) It is certain that Barbara will win. (entails It is likely . . .)
• Question : Why are there two strength levels of modal necessity, and how do
they differ?
The recent literature on this topic is diverse, and includes important work by, among
others, von Fintel and Iatridou (); Finlay (, ); Kolodny and MacFarlane
(); Rubinstein ().
The second broader issue is based on the observation that many modal operators
have scalar semantics, and are gradable in the same way as concrete, non-modal adjec-
tives are. The gradability properties of a number of modal operators are exemplified
in (). We see that should, important, and likely are gradable, similarly to a concrete
adjective such as big.
() a. You should call Barbara more than (you should call) Alice.2 (cf. bigger)
b. It is very important to talk to Barbara. (cf. very big)
c. It is just as likely that Barbara will win as it is that Alice will. (cf. as big)
• Question : How can the gradability properties of modal operators be composi-
tionally modeled within a general framework of gradability?
The issue of the gradability of modal expressions has received attention only fairly
recently, by Portner (); Yalcin (, ); Lassiter (, ); Klecha (,
); and Katz et al. (), among others.
In this chapter, we contribute a novel linguistic characterization of the difference
between weak and strong deontic necessity modals. Among non-modal adjectives,
the literature has identified a class known as the extreme adjectives; many relative
adjectives have extreme counterparts, for example big has the extreme counterpart
huge and smart has brilliant (Bolinger, ; Paradis, ; Rett, ; Morzycki,

2 Note that (a) has an irrelevant reading where more compares frequency. We are focusing on the
reading where it compares degrees of importance or priority. One naturally occurring example which
arguably has this reading is the following: A fillet more than a whole fish should have that beautiful
fresh smell (http://www.smh.com.au/lifestyle/homestyle/tried-and-tasted/how-to-cook-the-perfect-piece-
of-fish--mkn). Accessed from GloWbE. All corpus data cited in this chapter were produced by
the BYU corpus project (Davies, , , ).

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 paul portner and aynat rubinstein

). We will show that must, crucial, certain are also extreme elements, contrasting
with non-extreme should, important, likely. Since pairs of extreme vs. non-extreme
adjectives are so common, it is to be expected that gradable modal operators would
come in extreme and non-extreme versions as well. The existing literature on extreme
adjectives then points us to a formal analysis of their gradability properties.3
The analysis here will focus on deontic and other priority modals, giving less
attention to epistemic examples like (c). Although we discuss epistemic modals in
reviewing the relevant previous literature, our formal analysis will only apply to the
priority types. We limit our attention in this way not only because deontic modality is
the focus of the present volume, but also because priority modals show a logical profile
distinct from epistemic and ability/opportunity modals.4 , 5 We also limit our attention
to unembedded occurrences of modals. In other work, we have begun to discuss how
the same pragmatic distinction can be applied to cases in which the modal is embeded
in an attitude context (Portner and Rubinstein, ).
The Chapter is organized as follows: section  reviews previous work on modal
semantics, focusing in particular on the motivation for developing a theory with links
to the scale-based semantics of adjectives, and then pushes this perspective further by
incorporating the class of extreme expressions. We begin in . by reviewing the stan-
dard premise semantics classification and theory of modality within formal semantics
due to Kratzer. We then provide in section . some background on semantic classes of
adjectives, discussing how different types of gradable adjectives can be characterized in
terms of a scale-based semantics. Based on this background, we review and assess work
which looks at modals from a similar perspective, such as Lassiter () and Klecha
(). A prominent idea in this literature is that many modal operators require a scale-
based semantics, and that subclasses of modal operators should be identified with
subclasses of gradable adjectives. In section . we use a variety of tests to establish a
parallel between strong and weak necessity modals and the extreme and non-extreme
classes of adjectives exemplified by huge vs. big. We show that all of the strong necessity
modals are in fact extreme words and point out that this perspective reconciles some
of the theoretical tension in previous work on gradable modals.
In section , we develop a semantic and pragmatic analysis of extreme and non-
extreme deontic and other priority modals. In this section, we propose that the scales
of these modals are constructed from the same types of modal parameters which
are familiar from the standard Kratzerian premise semantics for modality, that is

3 Lassiter () notes the similarity between certain strong necessity modals (in particular, deontics) and
what he calls ‘high degree adjectives’ such as huge. He does not, however, note their unique grammatical
properties or connect them to the literature on extreme adjectives.
4 For example, while priority modals show the Union Property pattern, epistemic ones do not (Halpern,
; Yalcin, ; Lassiter, ): φ ≥ ψ, φ ≥ χ entail φ ≥ (ψ ∨ χ ). The fact that epistemics don’t show
the Union Property is an important puzzle for the standard Kratzerian (, ) analysis.
5 Dynamic modals do not show the same possibility/necessity distinction as other types. This can be
seen in the fact that can is stronger than a modal logic existential  (can(φ) does not entail can(φ ∧ ψ) ∨
can(φ ∧ ¬ψ), compare Kenny, , Portner, , section .) and the fact that it has no obvious necessity
counterpart.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

extreme and non-extreme deontic modals 

conversational backgrounds which function as modal bases and ordering sources;

this foundation in premise semantics allows us to draw connections to recent work
on weak and strong necessity (von Fintel and Iatridou, ; Rubinstein, ) and to
understand better the pragmatic difference between weak and strong—which is to say,
non-extreme and extreme—deontic necessity statements. We summarize our results
and discuss the prospects for future research in section .

 Semantic Classes of Modals

. Standard classification of modals
Modal words have long, and correctly, been classified according to two major dimen-
sions of meaning: their force and their judgment type (or simply type, often informally
called ‘flavor’). Modal logic traditionally distinguishes two levels of force, possibility
and necessity, whereas linguists have noted further levels, such as weak necessity and
slight possibility. Kratzer (, ) famously gives an analysis of several of these
levels within a quantificational framework.
Turning to the (judgment) types, there is a wider range. Some well-known types
include epistemic modality, deontic modality and related priority types, ability and
circumstantial modality. But this is a not a closed class, since each of the named types
can occur in an indefinite number of subvarieties; for example, among the deontics
we can distinguish God’s law, national and international law, professional ethics, tribal
custom, and classroom etiquette. Within buletic modality, we have different subtypes
for each attitude holder (Alice’s desires, Bob’s desires, etc.), and possibly also different
kinds of desire (things we want versus things we hope for). Kratzer (, , )
denies that these distinctions point to any lexical ambiguity; instead, she treats the
modals as dependent on parameters of context which serve to narrow down the set of
relevant possible worlds over which the modal quantifies.
This framework has been the basis for studying many different kinds of modals
in different languages, and has proven fruitful in understanding their linguistic and
logical properties as well as crosslinguistic variation. However, a number of important
puzzles and problems for the quantificational approach have accumulated over the
years. One of these is one of the main issues addressed in this paper: the fact that modal
expressions are often gradable. Kratzer’s treatment of levels of modal force provided
a partial solution to this problem, but it fell short of a full, compositional analysis
of natural language modal expressions. For example, the theory does not compose
the meaning of more likely than using a plausible derivation of the comparative
construction more generally.

. Scalar classification of modals

It is natural to think of gradable predicates as subclassifying objects into groups by
measuring them along a lexically encoded dimension: for example, long classifies

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 paul portner and aynat rubinstein

objects according to how they measure in length. There are competing views about
how to formally implement gradability, some of which assume degrees as part of the
semantic ontology, and others which do not. In this paper we adopt a degree-based
approach to gradability, following Kennedy and McNally (); Kennedy () and
others. We do so for concreteness, and under the assumption that this choice does
not affect our main claims about the distinction between extreme and non-extreme
predicates, and modals in particular. As an entry point to the extensive literature on
adjective meaning and gradability, including the debate about the semantic status of
scales and degrees and the cognitive basis of these notions, we refer the reader to recent
overview articles on these topics (Beck, ; Demonte, ).
Some of the most common reflexes of gradability crosslinguistically, exemplified in
(), are the ability of a predicate to occur in comparative constructions, to be modified
by degree modifiers such as almost, very or slightly, and to appear in degree how-
questions.
() a. The green vase is prettier than the brown vase.
b. The glass is almost full.
c. How long was the Beagle?
Languages may also allow modification of certain gradable predicates by measure
phrases. Long in English allows this type of modification, whether in its comparative
form (a), or in its morphologically unmarked form (b), known as the positive form.
() a. The Mayflower was (about)  meters longer than the Beagle.
b. The Beagle was  meters long.
Gradability has been most closely studied with respect to adjectives, and indeed
most of the modal expressions we focus on in this paper belong to this category.
It is well known, however, that other grammatical categories admit of gradability
as well (Bolinger, ; Doetjes, ; Kennedy and McNally, ; Kennedy and
Levin, ; Sassoon, , ). The verbal elements should and ought to attest
the cross-categorial nature of gradability in the modal domain (see (a) above and
data below).
It is noteworthy that not all adjectives pass the tests for gradability, or at least that
some pass them more easily than others (Sapir, ; Burnett, ). Good candidates
for non-gradability are adjectives such as true and pregnant and privative adjectives
such as fake:
() a. ?Your sister is more pregnant than this woman.
b. ?It is more true that crows are black than it is that apples are red.
While these examples are not ungrammatical, they require contextual support to
be interpreted. As is often assumed, we take it that coercion (a semantically or
pragmatically induced shift in meaning) is involved (see, for example, Kennedy ;
Burnett , ), although in this domain we do not find explicit theories of the

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

extreme and non-extreme deontic modals 

coercion process on a par with those discussed in the aspectual domain (Moens and
Steedman, ; de Swart, ).
The literature is somewhat ambivalent about the gradability status of one particular
class of adjective, particularly important here, the extreme adjectives.
() a. ?Godzilla is more gigantic than Mothra. (Morzycki, , p. )
b. ?The Tesla is more excellent than the Volt. (Compare Paradis, )
Both Morzycki and Paradis rate extreme adjectives as questionable in the comparative,
while Rett () assumes that they are grammatical. (As we’ll see below, Morzycki
discusses the fact that they are more acceptable in some contexts than others.)
Among the modal predicates we focus on, we observe that the strong necessity
modals (must, crucial, certain) are somewhat degraded in the comparative, unlike their
weak counterparts:
() a. You ought/?have to call Zoe more than (you ought/?have to call) Barbara.
You should/?must call Zoe more than (you should/?must call) Barbara.6
b. It is more important/?crucial to talk to Zoe than it is to meet her face to
face.
c. It is more likely/?certain that the sun will set tonight than it is that it will
rise tomorrow.
Quantitative data bears out our feeling that strong necessity modals are less natural
than weak ones in the comparative: important and likely occur in the comparative
much more frequently than crucial and certain, and this pattern is comparable to that
observed with non-modal extreme/non-extreme pairs.7

.. background on semantic classes of adjectives

Following Kennedy and McNally (, p. ), we assume degrees as abstract repre-
sentations of measurement; a scale S is a pair < D, R > consisting of a set of degrees
D that are organized by a total order R along a dimension . Gradable predicates
relate the members of their domain (objects or, in the case of modals, propositions)
to degrees on their lexically associated scale. The denotation of the predicate includes
a measure function, μ, whose role is to measure objects on the relevant dimension.
Long, for example, denotes a relation between objects x that have a length and degrees
of length d.

6 We note that some speakers find the examples with modal auxiliaries worse on the whole than those
with semi-modals. The relative pattern of acceptability judgments reported in the text is stable across these
types, however.
7 The percent of occurrence in the comparative and mutual information scores between target items
and the comparative in the British National Corpus (Davies, ) are as follows: important=.%
(PMI=.), crucial=.% (.), likely=.% (.), certain=.% (.). As examples of non-modal
adjectives: big=.% (.), gigantic=.% (.), pretty=.% (.), gorgeous=.% (.).

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 paul portner and aynat rubinstein

() a. μlength = [λx : x has a length . the degree d such that x measures d in
length]
b. [[ long ]] = [λxλd . μlength (x) = d] (informally = λxλd . x is d-long)
Comparative morphemes and degree modifiers place restrictions on the degree
returned by the measure function, typically by relating it to another degree (Kennedy,
, p. ). These morphemes are essentially the glue that derives appropriate truth
conditions for sentences in which gradable predicates occur. In (), assuming a
simple-minded syntax, we see how the denotation for a phrasal comparative mor-
pheme -er/more works:8
() a. [[ -er/more ]]= λGλyλx.∃d∃d [d > d ∧ G(x)(d) ∧ G(y)(d )]
b. The Mayflower was longer than the Beagle.
c. [[ (b) ]] = [[ [Mayflower [[-er long] than Beagle]] ]] = True iff
∃d∃d [d > d ∧ μlength (Mayflower) = d ∧ μlength (Beagle) = d ]
In order to account for gradable adjectives in their simplest, positive form, the
grammar needs to provide semantic glue, parallel to -er/more . . . than, which can relate
the measured degree to another degree not explicitly mentioned, the standard. For
some adjectives (namely, relative adjectives), the standard is contextually determined,
while for others (absolute adjectives), the standard is lexically determined by the
properties of the scale (Unger, ; Rotstein and Winter, ; Kennedy and McNally,
). A null morpheme pos (for ‘positive’) introduces the standard and expresses that
the measured degree is at least as high (Cresswell, ; Bartsch and Vennemann, ;
Bierwisch, ; von Stechow, ; Kennedy, ).
() [[ pos ]]= λGλx.∃d[d ≥ std(G) ∧ G(x)(d)]
Example (a) could be judged true in the context of a conversation about old
explorer ships, but it would likely be judged false in a discussion about the evolution
of non-military ships in the past two centuries. Thus, long has a contextual standard
in this case; std([[ long ]] ) is given by context. In contrast, (b) has a lexical standard,
and is true if the ship was completely full; hence std([[ full ]] ) is the maximum degree
of fullness.9
() a. The Beagle was long.
b. The Beagle was full.

8 The phrasal analysis of the comparative is coupled with a straightforward semantic interpretation
(Heim, ), but it is unclear whether it is to be preferred over a more elaborate clausal analysis involving
ellipsis (Lechner, ). See Bhatt and Takahashi () and Beck () for an overview of the debate.
The particular formalizations given in (b), (a), and () are not as simple as they could be, but are
given in these ways to make them compatible with the presentation of Morzycki’s ideas below.
9 Normal issues of granularity mean that, in different contexts, there can be variation in what counts as
reaching the maximum or minimum on the scale. For example, a glass may count as full even if it could
hold slightly more liquid. These uses are known as ‘loose’ or ‘imprecise’ (Kennedy and McNally, ).

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

extreme and non-extreme deontic modals 

() [[ (a) ]] = [[ [Beagle [pos long]] ]] = True iff

∃d[d ≥ std([[ long ]] ) ∧ μlength (Beagle) = d]
Beyond the basic toolset of dimensions, degrees, orderings, and standards of com-
parison, Rotstein and Winter () and Kennedy and McNally () argue that a
linguistically relevant property of scales is whether or not they contain maximal and
minimal elements. The four scale types which result are totally open, totally closed,
upper closed, and lower closed.
Patterns of modification and entailment have been used to argue that all four types
of scales are represented among gradable adjectives. For example, compatibility with
proportional modifiers (half, mostly, most of the way) is used as a diagnostic of totally
closed scales.10
() Totally closed scale
a. The glass is half/mostly full.
b. Her eyes were half/most of the way closed.
(Kennedy and McNally, , p. )
Adjectives that do not accept proportional modifiers can be further subclassified using
endpoint modifiers. Some of these are said to diagnose the existence of a maximum
degree (e.g., %, completely, fully), while others have been used to diagnose the exis-
tence of a minimal degree (e.g., barely, slightly). Assuming in addition that antonym
pairs have the same set of degrees but reverse orderings of these degrees, we can
identify adjectives with all four scales:
() a. Totally open scale
i. Her brother is completely ??tall/??short.
ii. Her brother is barely/slightly ??tall/??short
b. Lower closed scale
i. The pipe is fully ??bent/straight.
ii. The pipe is barely bent/??straight.
c. Upper closed scale
i. The treatment is completely safe/??dangerous.
ii. The treatment is slightly ??safe/dangerous.
[(i) examples from Kennedy and McNally , p. ]
As for the positive form, it has been argued that, when an adjective has a lexical
standard, this lexical standard is always identified with an endpoint of a closed scale
(Kennedy and McNally, ; Kennedy, ). Thus, tall, with its open scale, must
use a contextual standard. In contrast, full, with its totally closed scale, takes a lexical
standard given as the (upper) endpoint of its scale.

10 See Rotstein and Winter (); Kennedy and McNally () for additional tests.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 paul portner and aynat rubinstein

.. the scalar approach to modals

Many modals—both adjectives and auxiliaries—are grammatical in constructions
that diagnose gradability (Portner, ; Yalcin, ; Lassiter, ; Klecha, ).
Gradable interpretations affect the full range of modality (judgment) types mentioned
in section .: epistemic, priority, and as shown in (), also ability.
() After global warming takes hold, cockroaches will be more able to survive
than people.
These observations are not new. As noted above, the idea that possibility and neces-
sity are gradable notions has been part and parcel of formal semantic theories of
modality since at least Kratzer (). Over the years a debate has developed over
how to formally analyse gradability in the modal domain, and in recent years two
main approaches have been developed. One approach has aimed to marry Kratzer-
style modal semantics with the independently evolving scale-based theory that was
introduced in the previous section. The other has argued for the rejection of the
traditional quantificational approach in favor of an analysis of modality based on the
mathematical formulation of probability.
. Villalta (, , ) makes one of the first attempts to articulate a fully
scalar account of modals building on Kratzer’s semantics. Focusing on Spanish,
she proposes that subjunctive-selecting predicates, including many of interest
here (such as important, necessary, and probable) are similar to gradable adjectives
in taking a degree argument. The degrees they reference are defined as sets of
propositions that form equivalence classes according to the Kratzerian ordering of
‘better possibility’, which is used to compare propositions according to how desir-
able, necessary, or likely they are. Problems have been pointed out for Kratzer’s
original definition of ‘better possibility’ (see Lassiter  for a summary), and
in response, later work in this tradition explores the interaction of priorities
and likelihood (Katz et al., ; von Fintel, ; Charlow, ), alternative
definitions of comparative possibility (Kratzer, ), and other additions to the
Kratzerian analysis (Cariani et al., ).
. Swanson () and Yalcin (, ) use the mathematical notion of prob-
ability to give a scalar semantics for explicit expression of probability using
such terms as probable and likely. This general approach is applied to priority
modality types by Goble (); Finlay (, ) and Lassiter (), with
Finlay combining probability with a given ‘goal’, while Goble and Lassiter use a
quantitative definition of expected utility.
There is an interesting debate between Lassiter (, ) on the one hand, and
Klecha () on the other, concerning how to explain semantic relations among
possible, likely, and certain. Lassiter interprets all three on the totally closed scale of
probability, with possible taking a minimum standard, likely a relative standard, and

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

extreme and non-extreme deontic modals 

certain a maximum standard. Klecha points out, however, that this analysis conflicts
with the principles relating an adjective’s standard to its scale structure. In particular,
Lassiter’s proposal for likely runs counter to the argument that adjectives with closed
scales do not take relative standards. Klecha () analyzes gradable modality in a way
which borrows from both of the above approaches. Focusing on epistemics, he denies
that possible is a scalar adjective (see also Herburger and Rubinstein, ), and argues
that likely and certain are interpreted on different scales, despite the entailment relation
between them. Because Klecha does not analyse the adjectives based on a single scale,
the entailment from certain to likely is not automatic, and must be stipulated.
The debate between Lassiter and Klecha leaves us somewhat uncertain about the
scales and standards for epistemic adjectives, as neither approach seems to give a
full explanation of their modification patterns. In the next section, we will argue that
another parameter of classification can help to elucidate the properties of gradable
modals.

. Extreme modals

The scalar approach to modal semantics has both increased our understanding of the
properties of modal expressions and inspired new ideas for their semantic analysis.
However, there are differences between classes of gradable adjectives not captured by
the idealized perspective reviewed so far. One such difference is exemplified by the
pair pretty/gorgeous, both members of which presumably are related to a basic scale of
beauty. In this section, we argue that weak/strong modals such as important/crucial
and should/must (holding constant their judgment type) differ in the same way as
pretty/gorgeous. Like gorgeous, must and crucial are extreme expressions (Bolinger,
; Paradis, ; Rett, ; Morzycki, ), describing arguments which have
the property denoted by their non-extreme counterparts (pretty, should, important)
to a degree higher than would normally be conveyed by the non-extreme words.
Extreme adjectives have a number of properties which show that they have scale-based
meanings which are intimately related to the semantics of non-extreme expressions,
but they are also distinct both semantically and pragmatically from ordinary gradable
adjectives.
Section .. reviews the relatively small literature on extreme adjectives, outlines
some of their important properties, and summarizes one prominent analysis of them.
Then in section .., we use these properties to show that strong necessity modals
should be classified as extreme expressions, while weak necessity modals are non-
extreme. These points set the stage for our own effort to provide a semantic analysis
for deontic and other priority necessity modals in Section .

.. extreme adjectives

In an often-cited early discussion, Paradis () states that extreme adjectives involve
scales in their meaning (diagnosed by the fact that they come in antonym pairs), but

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 paul portner and aynat rubinstein

that they only occur with a subset of modifiers associated with gradability. Specifically,
they are used with what she calls ‘totality modifiers’ (absolutely, totally, utterly), but not
with approximators (almost). According to Paradis, some but not all people accept
them with the comparative and superlative.
() a. The salsa is absolutely terrible.
b. *The salsa is almost terrible.
c. ?The salsa is more terrible than the guacamole.
To explain this pattern Paradis says that extreme adjectives ‘represent the ultimate
point of a scale,’ but can be coerced into meanings which apply to elements not at the
maximum point of the relevant scale.
Paradis’ intuitions about the semantics of extreme adjectives are appealing, but it
does not seem correct to say simply that they describe their arguments as being at the
endpoint of their scale. For example, (a) does not mean that the salsa is of the highest
possible (or even the highest actually realized) degree of badness. There certainly are
more terrible things, and quite possibly more terrible condiments on the table. Later
work on extreme adjectives has therefore pursued the intuition that extreme adjectives
utilize subscales related to the more normal scales of their non-extreme counterparts.
Rett () assumes that extreme adjectives utilize totally closed scales which are
subscales of their non-extreme counterparts. Terrible would utilize a scale which
begins at some threshold t and reaches to the top of the scale. The idea that extreme
adjectives’ scales are totally closed is motivated, it seems, by two factors: first, that
there must be a lower bound defined by t, and secondly that there must be an upper
bound to explain the appearance of maximality modifiers such as totally. However,
against this vision is the fact that modifiers which normally go with lower bounded
scales, like barely, are not acceptable (*barely terrible). Moreover, as we have seen,
maximality operators are not interpreted as expected if they diagnose closed scales for
extreme adjectives. As we noted above, (a) does not mean that the salsa is maximally
terrible, as it would if absolutely placed the adjective’s argument at the top of the scale.
Rather, it means something like that it is without question in the range of terrible
condiments.11
Given these points about barely and absolutely, it seems that we do not have clear
evidence concerning the scale type for extreme adjectives. Morzycki () presents
an analysis of extreme adjectives which, though he does not emphasize the fact, allows
for extreme adjectives to be either bounded or unbounded at both top and bottom. His
approach is that scalar adjectives are interpreted with respect to a contextually salient
set of degrees DC on their scales; non-extreme adjectives relate to values within DC ,

11 Not all maximality operators have precisely the same range of meanings. As noted by Stephen Finlay
(p.c.), Janie is totally gorgeous has a reading ‘Janie is gorgeous in every respect.’ This shows that totally
gorgeous can grade an object on a scale other than the scale of gorgeous. Many adjectives are like this, and
we note that totally pretty has a similar reading.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

extreme and non-extreme deontic modals 

while extreme adjectives relate to values above all of the values in DC .12 The standards
follow suit: the standard for a non-extreme adjective is in DC , while that for an extreme
adjective is above DC . In this analysis, the extreme portion of the scale will be bounded
at the top if S is. This fact may allow us to understand what’s going on with (a).
One of Morzycki’s major contributions was a discussion of a range of properties
which differentiate extreme from non-extreme adjectives and which also can serve as
diagnostics for whether a term is an extreme expression.13 We will look at five such
properties in detail:
. Extreme adjectives occur with extreme degree expressions, or extreme modifiers,
such as flat-out, positively and downright, while non-extreme adjectives (even
those with maximum standards) are less natural.14
() a. ??flat-out good
b. flat-out excellent
c. ??flat-out straight
Though it’s not common to find a maximum standard adjective with an extreme
counterpart, in such cases, only the extreme adjective is natural with these
modifiers:
() a. ??downright full
b. downright brimming
. The near inverse of the previous property is that extreme adjectives strongly resist
modification with very.15 They contrast in this respect with relative adjectives
such as good and absolute ones such as straight.
() a. very good
b. ??very excellent
c. very straight
() a. very full
b. ??very brimming

12 ‘Salience’ could be interpreted in different ways, and Morzycki glosses it in this context as ‘degrees
. . . that we regard, for the purposes of the discussion, as reasonable candidates for values we might want to
consider’ (Morzycki, , pp. –). The issue of how to properly characterize the pragmatics of ‘salient’
vs ‘non-salient’ degrees is difficult and closely connected to the analysis of exclamatives (see for example
Zanuttini and Portner, ; Miró, ; Rett, ).
13 Morzycki makes a distinction between lexical extreme adjectives and contextual extreme adjectives.
The lexical extreme adjectives behave as extreme in any context in which they are used; we are focusing on
this class. Contextual extreme adjectives can behave as extreme with contextual support.
14 Morzycki points out that Cruse () can be seen as having first identified this property. Note that
other items with intuitively similar meanings might have different properties. Words such as categorically
and definitely are not extreme modifiers.
15 Similar adverbs such as really and so show different properties and do not serve as a test for the
extreme/non-extreme contrast.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 paul portner and aynat rubinstein

Note that Kennedy and McNally (, p. ) argue that coercion is involved
when very occurs with a minimum or maximum standard adjective; the key point
here, though, is that these nevertheless differ from extreme adjectives.
. Extreme adjectives are often degraded in the comparative (as in (c), repeated
below as (a)), but they do occur and improve with even (b).
() a. ?The salsa is more terrible than the guacamole.
b. The salsa is even more terrible than the guacamole.
. Though they resist the comparative, extreme adjectives fully accept the equative
construction.
() The salsa is as terrible as the guacamole.
. Extreme adjectives are unacceptable when placed in comparatives with their non-
extreme counterparts.
() a. *The salsa is more terrible than the guacamole is bad.
b. *The salsa is worse than the guacamole is terrible.
() a. *The bucket is more brimming than the cup is full.
b. *The cup is fuller than the bucket is brimming.
As Morzycki notes, the data in ()–() contrast with other examples where two
adjectives on related scales are compared:
() The field is longer than it is wide.
Morzycki () lists several other properties which distinguish extreme and non-
extreme adjectives, but does not provide a formal analysis of them. Briefly, these
properties are: (i) Extreme adjectives can readily be intensified by prosodic means
(Cruse, ); (ii) in discourse, an extreme adjective is perceived as disagreeing (not
agreeing) with an understatement; and (iii) extreme adjectives are readily used in
hyperbole. We will not discuss these properties further.
Recall that within Morzycki’s analysis, non-extreme adjectives associate their argu-
ment with a contextually salient degree, while extreme adjectives associate their
argument with a degree above the contextually salient portion of the scale. He assigns
big and gigantic the following meanings, where C by default refers to the contextually
salient set of degrees (= DC above):
() a. [[ bigC ]] = λxλd . d ∈ C ∧ x is d-big
b. [[ giganticC ]] = λxλd . d > max(C) ∧ x is d-big
Morzycki states that this difference in meaning goes along with a difference in
contextual standard: the standard for big must be within C while that for gigantic must
be above C. This proposal allows Morzycki to provide explanations for the patterns
involving extreme and non-extreme adjectives observed above.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

extreme and non-extreme deontic modals 

. Extreme modifiers make reference to a set of degrees C+ which extends the salient
set C into higher degrees, and they presuppose that the contextual standard is
within the extreme portion of this extended scale (i.e., C+ − C).16
() [[ downrightC ]] = λaλx : std(a) ∈ (C+ − C) . ∃d[d ≥ std(a) ∧ aC+ (x)(d)]
Because the contextual standard for an extreme adjective is above C, it can satisfy
the presupposition of the extreme modifier, but because the standard for a non-
extreme adjective is within C, it cannot. For a non-extreme adjective to satisfy
the presupposition of an extreme modifier, it must be treated as extreme in the
context. In other words, it must be given a standard above C.
. Morzycki (, p. ) analyzes very+ADJ as asserting that the individual it takes
as argument is at the top of the contextually salient set of degrees of the scale of
ADJ:
() [[ very C ]] = λaλx . ∃d[d ≥ std(a) ∧ small(max(C) − d) ∧ aC (x)(d)]
The infelicity of ??very gigantic is thus explained on this analysis through the
contradictory requirements it places on the measure of the object described;
specifically, it asserts that the argument is big to a degree which is both below
max(C) (via very) and above max(C) (via gigantic).
. Morzycki’s explanation of why extreme adjectives often resist the comparative is
complex, but the basic idea is that it is not natural to compare degrees which are
not in the salient range. Ideally, we would have a theory of salience of degrees
to provide a more specific reason why this would be so,17 but intuitively it
makes sense that if the salsa is terrible, it is so overwhelmingly bad that it might
be difficult or pointless to decide whether it is better or worse than the (also
terrible) guacamole. After all, if it’s terrible, you know all you need to know:
you’re not going to eat it. But this difficulty or pointlessness can be overcome
pragmatically, as in the context of a Mexican cooking class where an instructor’s
detailed feedback on the student’s failed salsa and guacamole would be relevant.
Moreover, it seems that even (specifically the presupposition it triggers) supports
the kind of context in which extreme degrees can be compared; see Morzycki
(, §.) for further discussion.
. In contrast to comparatives, Morzycki argues that the semantics for equatives
does not involve comparison. For example, () states that there is a degree of
badness beyond the salient set C which is the degree of the salsa, and which is the

16 Notice that Morzycki has a

C+ in the final conjunct. This implies that downright gigantic describes
its argument as big to a degree beyond even C+ . He also has aC+ rather than aC in the restriction on
the standard, but he states that an adjective’s standard does not depend on the set of degrees with which
it is interpreted (i.e., std(aC ) = std(aC+ )). See Morzycki (, p. ). Our analysis of extreme modals
diverges from his by not using degrees above C+ .
17 Our analysis will actually be framed in terms of a more explicit explanation of what ‘salience’ amounts
to for modal adjectives.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 paul portner and aynat rubinstein

degree of the guacamole. Seen this way, the equative only asserts the existence
of a single degree, and this is why it does not show the pragmatic incongruity
sometimes seen with comparatives.
Morzycki’s explanation for the comparative and equative facts may not be
entirely convincing, for two reasons: it’s not clear why comparing extreme degrees
in the comparative should be less felicitous than referring to them in the equative,
and the formulation of the equative semantics without comparison is non-
standard, and so might be seen as ad hoc. Our intuitions about the comparatives
and equatives are slightly different from his. It seems to us that, in a simple context,
any distinctions above C along the relevant dimension are collapsed, and so there
is only a single extreme degree. In such contexts, a sentence comparing two
extreme degrees would be logically false, unless they are equated. Context could,
however, articulate the portion of the scale above C, making multiple extreme
degrees available, and in such a context, a comparative could be informative.
Further work will be needed to determine the best analysis of the behavior of
extreme adjectives in comparatives and equatives.
. In Morzycki’s analysis, comparisons involving an extreme and a non-extreme
adjective, like (), have trivial truth conditions. (a) states that salsa’s degree
of badness is beyond C, and moreover beyond the guacamole’s degree of badness,
which is within C. This must be the case. (b) states that the salsa’s degree of
badness, which is within C, is above the guacamole’s degree of badness, which is
beyond C. This cannot be the case.

.. strong necessity modals are extreme expressions

Strong necessity modals, but not weak ones, have the properties associated with
extreme adjectives:
. Strong necessity modals readily take extreme modifiers:18
() a. Susan positively/flat-out/downright must/*should call her mother.
b. It is positively/flat-out/downright crucial/*important for Mary to call
her mother.
c. It is positively/flat-out/downright certain/*likely that Mary will call her
mother.
. Strong necessity modals are less natural with very:19

18 In the corpora we have examined, extreme modifiers are extremely rare with the modal elements
on which we have focused our discussion. In the BNC, there is one occurrence of positively must and
no other combinations of interest. In COCA (Davies, ), the raw counts and percentages of flat-
out/positively/downright with the modal elements are: must= (.%), should= (.%), certain= (.%),
likely, crucial, and important=. Overall, the quantitative trends conform to our hypothesis, but the numbers
are too small to count as strong evidence by themselves.
19 Corpus patterns in the BNC support this difference between adjective classes. Percentage occurrence
with very and association measures are as follows: crucial=.% (PMI=.), important=.% (.),

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

extreme and non-extreme deontic modals 

() a. Susan very much *must/should call her mother.

b. It is very *crucial/important for Mary to call her mother.
c. It is very *certain/likely that Mary will call her mother.
. Strong necessity modals resist the comparative, and in the comparative, they are
better with even:20
() a. Susan ?must/should call her mother more than she ?must/should call
her father.
b. It is more ?crucial/important for Mary to call her mother than her father.
c. It is more ?certain/likely that Mary will call her mother than her father.
() a. Susan must call her mother even more than she must call her father.21
b. It is even more crucial for Mary to call her mother than her father.
c. It is even more certain that Mary will call her mother than her father.
In evaluating the example with must, it is important to keep in mind that
the relevant reading is one where degrees of obligation are being compared;
the example has a grammatical reading where frequencies of calling are being
compared.
Notice that more crucial and more certain are not entirely unacceptable in (),
but this was the case with the non-modal extreme adjectives such as more terrible
in () as well. When they are usable, they seem synonymous with the versions
containing even in ().
. Both strong and weak necessity modals are acceptable in the equative:
() a. Susan must call her mother just as much as she must call her father.
Susan should call her mother just as much as she should call her father.
b. It is as crucial/important for Mary to call her mother as it is for her to
call her father.
c. It is as certain/likely that Mary will call her mother as it is that she will
call her father.
. Strong and weak necessity modals cannot be placed in the comparative with one
another:

certain=.% (-.), likely=.% (.); non-modal adjectives for comparison: gorgeous=.% (.),
pretty=.% (.).
We used very much in place of very alone with the auxiliaries (a). The intuition data in this case confirm
our predictions, but the combinations are very rare. Neither occurred in the BNC, while the raw counts
in COCA were very much must= and very much should= (<.%). The single example is interesting,
though, and fits our analysis: You should know—actually, for complete, suspenseful enjoyment of the film, you
very much should not know, but the word is out, so we’re obliged to tell you—that Heavenly Creatures is based
on a notorious murder case. (‘A Heavenly Trip Toward Hell’, Richard Corliss Time, Nov , ).
20 Corpus data related to the modals in () was provided in footnote  above.
21 This sentence is somewhat degraded in comparison to the examples with crucial, certain, but
considerably better than the corresponding (a) without even.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 paul portner and aynat rubinstein

() a. *Susan must call her mother more than she should call her father.
*Susan should call her mother more than she must call her father.
b. *It is more crucial for Mary to call her mother than it is important for
her to call her father.
*It is more important for Mary to call her mother than it is crucial for
her to call her father.
c. *It is more certain that Mary will call her mother than it is likely that she
will call her father.
*It is more likely that Mary will call her mother than it is certain that she
will call her father.
We have shown that the strong necessity modals pattern with extreme adjectives, and
conclude that they should be included within this class. The weak necessity modals, in
contrast, should be treated as non-extreme expressions.

. Summary and plan

The observation that strong necessity modals are like extreme adjectives is important
because it points the way towards a better analysis of the strong/weak necessity
distinction and, more broadly, the gradability of modal expressions. Before we turn
to a formal analysis of the strong and weak priority modals must/should and cru-
cial/important, we think it’s useful to think about the intuitive basis for assigning
a scale structure to deontic expressions divided into salient/non-extreme vs. non-
salient/extreme portions. What would the degrees be, and how would they be ordered?
And what would be the difference between those ‘within the salient range’ and those
‘beyond the salient range’?
As to the first question, we propose that degrees of necessity correspond to the
weight of the requirements which favor a particular proposition. For example, ‘you
don’t murder’ has a higher degree of necessity than ‘you do the dishes’, because the
requirements favoring it are weightier. And as to the second question, the idea is that
the salient degrees are those assigned to propositions which are worth discussing, and
they are worth discussing because they are based on requirements the interlocutors
might or might not follow. Non-salient degrees, in contrast, are based on settled
or unquestioned requirements. For example: On an elementary school assignment
students have a hard requirement that the student write his name on the paper. With no
name, the assignment will not be graded at all, an outcome these good students would
never contemplate, and so the degree of necessity of ‘I write my name’ is ‘beyond the
salient range.’ Assuming it is graded, though, points will be assigned according to a
rubric with elements like ‘Illustrate your report with a hand-drawn, colored picture:
 points.’ If the student fails to color the picture, he might receive only four points
based on this criterion. For the students, the alternatives ‘I draw but do not color
the picture’ and ‘I draw and color the picture’ are worth discussing, and hence the
degrees of necessity of the alternatives are salient in the context. In this context, writing

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

extreme and non-extreme deontic modals 

your name has an extreme degree of necessity, while coloring your picture has a non-
extreme degree.
In the next section, we work towards a formal analysis of necessity modals which
incorporates the intuition that the scale of necessity is divided into a salient, or non-
extreme, part and a non-salient, or extreme, part. This approach not only provides
an explanation for the empirical differences between strong and weak modals enu-
merated in this section. It also promises a pragmatic grounding for the difference. On
our approach, must and should, and crucial and important, are all necessity modals;
they differ in that the strong ones paint an alternative as necessary and reject further
discussion, while the weak ones portray it as being necessary but still up for discussion.
In terms of issues of scales and standards, our approach can be seen as pursuing
a synthesis of the earlier ideas of Lassiter and Klecha.22 In considering how to rep-
resent the relationships between semantically related modal adjectives, these authors
represent two reasonable strategies. Lassiter follows the strategy of putting the pair
likely and certain on the same scale, and thus makes the differences between them
fairly minimal (literally, a matter of degree). However, as we have seen, the differences
between extreme and non-extreme adjectives amount to more than that. Klecha
follows a different strategy, where likely and certain are associated with different scales,
stipulating a relation between those scales. However, because the relation between the
scales is not motivated by any broader picture of adjective semantics, this leaves an
explanatory gap. Our approach puts the extreme and non-extreme pairs on subscales
of the same basic scale. Hence, as in Lassiter’s analysis, the members of the pairs differ
in degree. However, they are on pragmatically different parts of this scale, and so there
is a sense in which Klecha is right as well in saying they use different scales. Crucially,
the difference between the extreme and non-extreme parts of modal scales is based on
a motivated pragmatic distinction and is parallel to what is observed with non-modal
adjectives.

 Analysis
The intuition behind theories of extreme adjectives is that they characterize an
argument as going beyond the normal or expected range of its scale. As we have
discussed, this intuition is apt for extreme modals; for example, just as downright huge
things are big to a degree beyond the expected, the positively crucial actions are so
important that we do not expect to discuss them. We aim to capture this intuition by
defining a general scale-type labeled the scale of necessity. This scale will be used as the
basis for the analysis of both weak and strong necessity modals. It may be compared
to a general scale of size, which would provide the core of the analysis of both big
and huge.

22 It should be kept in mind, though, that their debate focused on epistemic modals, and so we must
extrapolate a bit to attribute them contrasting positions concerning deontics.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 paul portner and aynat rubinstein

Our formulation builds on existing work on strong and weak necessity within a
premise semantics framework. Specifically, von Fintel and Iatridou () proposed
an analysis of must and should that involves at least two ordering sources, rather
than the single one in the standard version of Kratzer’s framework. With a strong
necessity modal, only the primary ordering source is relevant, and it determines the
modal’s domain of quantification in the way familiar from Kratzer’s work. With a weak
necessity modal, the secondary ordering source operates on the set of worlds which
are relevant according to the modal base and primary ordering source, to further
restrict the domain of quantification. Specifically, the semantics first identifies the set
of worlds compatible with the modal base m(w) which are best-ranked according
to the primary ordering source, o (w), and then picks out the best-ranked worlds
among those according to the secondary ordering source o (w), as in (). We assume,
following von Fintel and Iatridou (), that there is always a set of ‘best’ worlds
among those ordered by an ordering source.23
() a. u ≤O v iff {p ∈ O : v ∈ O} ⊆ {p ∈ O : u ∈ O}
b. Best(M, O) = {v ∈ M : ¬∃u[u <O v]}
() Best(M, O , O ) = Best(Best(M, O ), O )
In this way, they account for the fact that strong necessity modals logically entail their
weak counterparts, and provide some degree of intuition about what the difference
between them amounts to.
Although their ideas advance our understanding, von Fintel and Iatridou do not
provide much explanation of the crucial difference between primary and secondary
ordering sources (see their footnote ). Building on their work, Rubinstein ()
proposes that this difference is to be understood in terms of pragmatics, where the
difference between the two ordering sources tracks the distinction between what is
treated as non-negotiable versus what is treated as negotiable. The primary ordering
source, in these terms, contains priorities that the speaker presents as ones she is
not willing to give up, while the secondary ordering source contains priorities that
might give way in the face of other factors.24 We can illustrate Rubinstein’s analysis by
considering the examples at the beginning of the paper:
() a. Our uninsured citizens should get insurance.
b. Our uninsured citizens must get insurance.
Sentence (a) is appropriate in a context in which the speaker is contributing to a
negotiation about health-care reform, while (b) is appropriate when the speaker

23 See Katz et al. () for an alternative way of combining multiple ordering sources. The definitions
cite the premise sets given by the modal base and ordering source(s) at the world of evaluation, e.g., M =
m(w).
24 Stephen Finlay (p.c.) makes the interesting suggestion that the priorities in o might be secondary

not because they are seen as unsettled, but rather because it is at issue whether they are achievable. This
suggestion might allow us to make a connection between this analysis and the information sensitivity of
deontic modals (Kolodny and MacFarlane, ; Charlow, ; Cariani et al., ; Cariani, ).

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

extreme and non-extreme deontic modals 

wants to rule out further debate. Combining these ideas with the semantics from von
Fintel and Iatridou (), the prediction is that must(φ) says that φ follows from
non-negotiable requirements, while should(φ) says that φ follows taking into account
requirements about which the speaker allows negotiation.
Our analysis connects the semantic and pragmatic differences between primary and
secondary ordering sources to the conceptualization of scale structure developed in
work on gradable and extreme adjectives. Specifically, we propose that the extreme
portion of the scale (the non-salient portion) is defined only in terms of the priorities
in the primary ordering source. This makes sense of the intuition that in the extreme
portion of the scale, distinctions between degrees are not salient. They are not salient
because any such distinctions would require treating one non-negotiable requirement
as more important than another. The non-extreme (salient) portion of the scale,
in contrast, is defined by taking into account both the primary and the secondary
ordering sources, and given Rubinstein’s understanding of the secondary ordering
source, this captures the intuition that distinctions within this portion of the scale are
salient. They are salient because they are based on priorities which the speaker treats
as important, but not set in stone.
Formally speaking, the non-extreme part of the scale is derived by considering
subsets of the secondary ordering source o (w). In what follows f is a contextually
provided function from sets of premises o to sets of non-null subsets of o, where the
members of f (o) are totally ordered by ⊆:
() Sc,w is defined only if f (o (w)) is defined, for every world w. When defined,
Sc,w =
Dw , ≤: Dw = {{p : Best(m(w), o (w), o ) ⊆ p} : o ∈ f (o (w))}
≤ = {
d , d  : d , d ∈ D ∧ d ⊇ d }

According to (), a degree in D is a set of propositions, specifically the set which is

best according to the modal base, primary ordering source, and a particular subset
of the secondary ordering source. Degrees are ordered by the superset relation, and
because f (o (w)) is totally ordered by assumption, the set of degrees is also totally
ordered.
We construct an extension of the scale Sc,w to include extreme degrees in a similar
way by considering subsets of the primary (non-negotiable) ordering source o (w).
The complete scale which includes the extreme extension is S+ c,w :

() S+
c,w is only defined if f (o (w)) and Sc,w are defined, for every world w. When
defined,
S+ + + +  
c,w =
Dw , ≤ : Dw = Dw ∪ {{p : Best(m(w), o ) ⊆ p} : o ∈ f (o (w))}
+ +
≤ = {
d , d  : d , d ∈ D ∧ d ⊇ d }

Notice that S+
c,w is defined using the two-place version of Best, since the secondary
ordering source is not relevant. The extreme part of the scale S+ + +
c,w is
(Dw − Dw ), ≤ ∩
+
(Dw − Dw ) .


i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 paul portner and aynat rubinstein

(a) defines a measure function μ which will allow us to give a scalar semantics
for necessity modals. The lexical entries in (b–c) capture the two levels of strength
observed with necessity modals: weak ones, specifically should and important, are
measures of propositions of non-extreme degree, whereas strong ones, namely must
and crucial, are restricted to higher values.
() a. Measure function: For any proposition p, world w, and necessity scale ,
μ (p, w) is the highest degree d in D+
w such that p ∈ d.
b. Weak necessity modal: [[ N ]] = [λpλdλw . d ∈ Dw ∧ d = μ (p, w)]
c

c. Strong necessity modal: [[ N + ]] c = [λpλdλw . d > max(Dw ) ∧ d =

μ (p, w)]
We can see how the sketched system works with a concrete example. Suppose
that the speaker is supportive of three goals: eating ramen (R), seeing Casablanca
(C), and meeting up with a friend who’s in town for the weekend (F). f (o (w))
reflects how strongly held these goals are, and can be construed as a series of sets
of goals which result from successively dropping the weaker ones. Assume that
f (o (w)) = {{F}, {C, F}, {R, C, F}}, reflecting that F is the most important and R
the least, and also assuming (for the time being) that the three goals are compatible.
These goals require going to different parts of town: F requires going to Georgetown;
C requires going to Silver Spring; and R requires going to Adams Morgan. In this
context, ‘We go to Georgetown’ has the maximal non-extreme degree on the scale of
necessity. ‘We go to Silver Spring’ has an intermediate degree, and ‘We go to Adams
Morgan’ has the lowest degree. It is more important that we go to Silver Spring than
Adams Morgan is therefore true.
We assume that unmodified necessity modals combine with their argument via the
covert degree specifier pos in () (Kennedy, ). In parallel to how pos interacts
with ordinary relative adjectives, when it combines with a necessity modal, it will
require the prejacent to exceed a contextually determined standard. Plausibly, the
contextual standard is set at the degree of ‘We go to Georgetown’, and in such a context
It is important that we go to Georgetown is true but It is important that we go to Silver
Spring/Adams Morgan are false.
This analysis generates some good predictions: Since extreme degrees are con-
structed as subsets of non-extreme degrees in ()–(), our analysis derives the
entailment relations between strong and weak necessity modals (). It derives the
meaning proposed for should and must by von Fintel and Iatridou () and Rubin-
stein () as special cases: the contextually specified degree for should is derived from
a non-null subset of o , while that specified for must is derived using no premises from
o . In addition, Morzycki’s () explanations of the properties of extreme adjectives
and modifiers directly carry over to the present account, in particular the restrictions
on very and extreme modifiers and the infelicity of comparing extreme and non-
extreme adjectives.
The way of constructing the scales in ()–() has a serious logical flaw, however.
It does not allow that incompatible propositions can have the same degree of necessity.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

extreme and non-extreme deontic modals 

Change the scenario so that eating ramen and seeing Casablanca are incompatible yet
equally strong goals. As a result, going to Adams Morgan (to have ramen) and going
to Silver Spring (to see Casablanca) are equally important, but we cannot do both.
In this scenario, f (o (w)) is the set of premise sets {{F}, {R, C, F}}, from which the
premise set {R, C, F} generates two disjoint maxima in the ordering, worlds where we
eat ramen (in Adams Morgan) and worlds where we see Casablanca (in Silver Spring).
(In both sets of worlds, we also go to Georgetown to see the friend.) Neither destination
is a necessity with respect to {R, C, F}, so neither has any degree of necessity. This is
incorrect.
We would like to be able to use both of the maxima defined by {R, C, F} to associate
propositions with the same degree of necessity. In other words, because eating ramen
entails going to Adams Morgan, and seeing Casablanca entails going to Silver Spring,
{R, C, F} should determine the same degree for going to Adams Morgan and going
to Silver Spring. One approach would be to have f select not just subsets of the
ordering source, but sets of consistent subsets of the ordering source, for example
f (o (w)) = {{{F}}, {{R, F}, {C, F}}}. Here, the second member encodes the fact that
{R, F} and {C, F} are equally important collections of goals. For each set of sets of
premises (e.g., {{R, F}, {C, F}}), we would then collect the necessities with respect to
each member ({R, F} and {C, F}) to define a degree of necessity. Going to Adams
Morgan and going to Silver Spring would thus have the same degree of necessity (and
this degree would be lower on the scale than the degree of going to Georgetown).
As an alternative to the above, we get a similar effect by allowing inconsistent
premise sets to be selected by f , and defining degrees by looking at all of the maxima in
the world-ordering defined by each premise set. In particular, the following definition
might work, where GP(A, B, C)(p) indicates that p is a ‘good possibility’ (in the sense
of Kratzer ) with respect to A, B, and C:
() S c,w is only defined if f (o (w)) is defined, for every world w. When defined,
Sc,w =
Dw , ≤: Dw = {{p : GP(m(w), o (w), o )(p)} : o ∈ f (o (w))}
≤ = {
d , d  : d , d ∈ D ∧ d ⊇ d }
Assuming that f (o ) = {{F}, {R, C, F}} (encoding that R and C are equally important),
this definition assigns the highest degree of necessity to going to Georgetown, and
equal but lower degrees to going to Adams Morgan and Silver Spring. This definition
has the surprising property that the scale of necessity is defined not in terms of
standard necessity (‘entailed by the union of the maxima’), but rather in terms of the
weaker notion of good possibility (‘entailed by at least one maximum’).
The revised definition allows us to account for examples with incompatible goals
like the following.
() a. It is as important to preserve the wetlands as it is to build the new housing
(which would drain the wetlands).
b. We should preserve the wetlands just as much as we should build the new
housing.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 paul portner and aynat rubinstein

As Morzycki () pointed out, equatives are also natural with extreme expressions,
and there seems to be little problem in interpreting cases where they describe incom-
patible propositions as having the same degree:
() a. It is as crucial to preserve the wetlands as it is to build the new housing.
b. We must preserve the wetlands just as much as we must build the new
housing.
These sentences are clearly acceptable, and describe situations which are true dilem-
mas. They could be analysed using the same strategy as ().
Although extreme and non-extreme modals are both perfectly acceptable with
equatives, Cruse (); Paradis (); Bolinger (); Morzycki () note that
extreme expressions are less natural with comparatives:
() a. ?It is more crucial to preserve the wetlands than it is to build the new
shopping center.
b. ?We must preserve the wetlands more than we must build the new shop-
ping center.
As we’ve mentioned, we interpret this fact in a slightly different way than does
Morzycki. We propose that the extreme part of the scale has only a single degree
(though it can be coerced into multiple degrees). We can model this idea by assuming
that f (o (w)) as a default is a singleton; this stipulation makes sense given the
assumption that o (w) contains only non-negotiable propositions, and so in many
contexts there would be no useful way for f to distinguish among them. However,
under coercion we can make distinctions among these propositions, in which case f
can select multiple subsets of o (w), giving rise to multiple extreme degrees.
Our analysis handles positive forms as well. Both (a) and (b) are true in the
scenario envisioned.
() a. It is important to preserve the wetlands.
b. It is crucial to preserve the wetlands.
Our account predicts that (b) can be true even if a proposition inconsistent with
preserving the wetlands also has the extreme degree. This is clearly possible, for
example when it’s uttered after (). More typically, (b) would imply that we
preserve the wetlands in all accessible worlds, and so build the new housing in none.
This result follows if, in the normal case, o (w) is consistent. Note that when f (o (w)) is
a consistent singleton, our semantics is equivalent to the standard Kratzerian analysis
of must.
We have seen how our analysis accounts for some of the subtle differences between
strong and weak deontic modals in a way which parallels the treatment of other
extreme/non-extreme expressions. Specifically, we have mentioned the distinction
between very and extreme modifiers such as positively and discussed contrasts in the
acceptability and interpretation of equatives, comparatives, and the positive form. We

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

extreme and non-extreme deontic modals 

have not returned to the final three properties listed in section ., and doing so must
await another occasion.

 Discussion
We have given strong evidence that strong and weak necessity modals with priority
readings should be classified as extreme and non-extreme expressions, respectively.
Moreover, we have argued that this observation, when combined with other ideas
from the scalar approach to modals, points to a better understanding of the semantic
and pragmatic differences between the two types of modal necessity. At an empirical
level, our analysis accounts for some of the subtle differences between strong and
weak deontic modals in a way that parallels the treatment of other extreme/non-
extreme expressions. Specifically, we have mentioned the distinction between very and
extreme modifiers such as positively and discussed contrasts in the acceptability and
interpretation of equatives, comparatives, and the positive form.
We have focused specifically on deontic and other priority readings of necessity
operators. In future work, we hope to extend these ideas to other judgment types, in
particular epistemics. Once an analysis of epistemics is in hand, it will be important to
compare the kind of approach we have advocated here with the probabilistic semantics
for gradable epistemic modals developed by Yalcin () and Lassiter (). We also
hope to extend our ideas to non-priority root modals such as can, able, and adjectives
formed with -able. These are not necessity operators, but are nonetheless gradable. The
distinct logical properties of these two groups of non-priority modals will allow us to
further refine the formal analysis of gradable modality.

Acknowledgments
We thank Graham Katz and Elena Herburger for discussions which have shaped our under-
standing of the topic and Jessica Rett, Nate Charlow, Matthew Chrisman, and Stephen Finlay
for comments on previous versions. Any errors are of course our responsibility, not theirs.
We gratefully acknowledge the support of National Science Foundation grant BCS- to
both authors, and support from the Mandel Scholion Interdisciplinary Research Center in the
Humanities and Jewish Studies to the second author.

References
Bartsch, R. and Vennemann, T. () Semantic structures: A study in the relation between syntax
and semantics. Frankfurt: Athenäum.
Beck, S. () Comparison constructions. In von Heusinger, K., Maienborn, C., and Portner, P.
(eds.) Semantics: An International Handbook of Natural Language Meaning. Volume .
pp. –. Berlin: de Gruyter.
Bhatt, R. and Takahashi, S. () Reduced and unreduced phrasal comparatives. Natural
Language and Linguistic Theory. . pp. –.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 paul portner and aynat rubinstein

Bierwisch, M. () The semantics of gradation. In Bierwisch, M. and Lang, E. (eds.) Dimen-
sional Adjectives. pp. –. Berlin: Springer.
Bolinger, D. L. () Adjective comparison: a semantic scale. Journal of English Linguistics. ..
pp. –.
Bolinger, D. () Degree words. The Hague: Mouton.
Burnett, H. () The puzzle(s) of absolute adjectives: On vagueness, comparison and the
pragmatic origin of scale structure. In Paperno, D. (eds.) UCLA Working Papers in Linguistics:
Special Issue on Semantics and Mathematical Linguistics. Los Angeles: University of California.
Burnett, H. () A delineation solution to the puzzles of absolute adjectives. Linguistics and
Philosophy. .. pp. –.
Cariani, F. () Deontic modals and probabilities: One theory to rule them all? In Charlow, N.
and Chrisman, M. (eds.) Deontic Modality. Oxford: Oxford University Press.
Cariani, F., Kaufmann, M., and Kaufmann, S. () Deliberative modality under epistemic
uncertainty. Linguistics and Philosophy. . pp. –.
Charlow, N. () What we know and what to do. Synthese. . pp. –.
Cresswell, M. J. () The semantics of degree. In Partee, B. H. (ed.) Montague grammar.
pp. –. New York: Academic Press.
Cruse, D. A. () Lexical semantics. Cambridge: Cambridge University Press.
Davies, M. () BYU-BNC (based on the British National Corpus from Oxford University
Press). [Online] Available from: http://corpus.byu.edu/bnc/. [Accessed  Nov. .]
Davies, M. () The Corpus of Contemporary American English:  million words, –
present. [Online] Available from: http://corpus.byu.edu/coca/. [Accessed  Nov. .]
Davies, M. () Corpus of Global Web-Based English: . billion words from speakers in 
countries. [Online] Available from: http://corpus.byu.edu/bnc/. [Accessed  Nov. .]
Demonte, V. () Adjectives. In von Heusinger, K., Maienborn, C., and Portner, P. (eds.)
Semantics: An International Handbook of Natural Language Meaning. Volume . pp. –.
Berlin: de Gruyter.
de Swart, H. () Aspect shift and coercion. Natural Language and Linguistic Theory. .
pp. –.
Doetjes, J. S. () Quantifiers and Selection: On the distribution of quantifying expressions in
French, Dutch and English. Leiden: Holland Academic Graphics.
Finlay, S. () Oughts and ends. Philosophical Studies. . pp. –.
Finlay, S. () What ought probably means, and why you can’t detach it. Synthese. .
pp. –.
Fintel, K. von () The best we can (expect to) get? Challenges to the classical semantics for
deontic modals. Manuscript. MIT Paper presented at the Central APA,  February . MIT.
Fintel, K. von and Iatridou, S. () How to say ought in foreign: The composition of weak
necessity modals. In Guéron, J. and Lecarme, J. (eds.) Time and Modality. pp. –. Berlin:
Springer.
Goble, L. () Utilitarian deontic logic. Philosophical Studies. .. pp. –.
Halpern, J. () Reasoning about Uncertainty. Cambridge, Mass.: MIT Press.
Heim, I. () Notes on comparatives and related matters. Manuscript. MIT.
Herburger, E. and Rubinstein, A. () Is ‘more possible’ more possible in German? In Snider,
T., D’Antonio, S., and Weigand, M. (eds.) Proceedings of SALT . New York University, 
May to  June. pp. –.
Hintikka, J. () Modality and quantification. Theoria. . pp. –.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

extreme and non-extreme deontic modals 

Katz, G., Portner, P., and Rubinstein, A. () Ordering combination for modal semantics. In
Chereches, A. (ed.) The Proceedings of SALT , University of Chicago, Chicago, IL, –
May. pp. –. CLC Publications.
Kennedy, C. () Projecting the Adjective: The Syntax and Semantics of Gradability and
Comparison. Ph.D. thesis. University of California Santa Cruz.
Kennedy, C. () Vagueness and grammar: the semantics of relative and absolute gradable
adjectives. Linguistics and Philosophy. . pp. –.
Kennedy, C. and Levin, B. () Measure of change: The adjectival core of degree achieve-
ments. In McNally, L. and Kennedy, C. (eds.) Adjectives and Adverbs: Syntax, Semantics and
Discourse. pp. –. Oxford: Oxford University Press.
Kennedy, C. and McNally, L. () Scale structure, degree modification, and the semantics of
gradable predicates. Language. .. pp. –.
Kenny, A. () Human abilities and dynamic modalities. In Manninen, J. and Tuomela, R.
(eds.) Essays on Explanation and Understanding: Studies in the Foundations of Humanities
and Social Sciences. pp. –. Reidel, Dordrecht.
Klecha, P. () Positive and conditional semantics for gradable modals. In Guevara, A. A.,
Chernilovskaya, A., and Nouwen, R. (eds.) Proceedings of Sinn and Bedeutung . Volume ,
Utrecht, – September . pp. –. MITWPL.
Klecha, P. () Bridging the Divide: Scalarity and Modality. Ph.D. thesis. University of Chicago.
Kolodny, N. and MacFarlane, J. () Ifs and oughts. The Journal of Philosophy. ..
pp. –.
Kratzer, A. () The notional category of modality. In Eikmeyer, H.-J. and Rieser, H. (eds.)
Words, Worlds, and Contexts. pp. –. Berlin: de Gruyter.
Kratzer, A. () Modality. In von Stechow, A. and Wunderlich, D. (eds.) Semantik/Semantics:
An International Handbook of Contemporary Research. pp. –. Berlin: de Gruyter.
Kratzer, A. () Modals and Conditionals. Oxford: Oxford University Press.
Lassiter, D. () Gradable epistemic modals, probability, and scale structure. In Li, N. and Lutz,
D. (eds.) Proceedings of SALT , University of British Columbia and Simon Waser University,
Vancouver, BC,  April to  May. pp. –. CLC Publications.
Lassiter, D. () Measurement and Modality: The Scalar Basis of Modal Semantics. Ph.D. thesis.
New York University.
Lechner, W. () Ellipsis in Comparatives. Berlin: Mouton de Gruyter.
Miró, E. C. () Wh-exclamatives in Catalan. Ph.D. thesis. University of Barcelona.
Moens, M. and Steedman, M. () Temporal ontology and temporal reference. Computational
Linguistics. . pp. –.
Morzycki, M. () Degree modification of extreme adjectives. In Bochnak, R., Nicola, N.,
Klecha, P., Urban, J., Lemieux, A., and Weaver, C. (eds.) Proceedings of the Chicago Linguistic
Society . – April .
Morzycki, M. () Adjectival extremeness: degree modification and contextually restricted
scales. Natural Language and Linguistic Theory. .. pp. –.
Paradis, C. () Adjectives and boundedness. Cognitive Linguistics. .. pp. –.
Portner, P. () Modality. Oxford: Oxford University Press.
Portner, P. and Rubinstein, A. () Mood and contextual commitment. In Chereches, A. (ed.)
The Proceedings of SALT . University of Chicago, Chicago, IL, – May. pp. –. CLC.
Publications.
Rett, J. () Degree Modification in Natural Language. Ph.D. thesis. Rutgers University.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 paul portner and aynat rubinstein

Rotstein, C. and Winter, Y. () Total adjectives vs. partial adjectives: Scale structure and
higher-order modifiers. Natural Language Semantics. .. pp. –.
Rubinstein, A. () Roots of Modality. Ph.D. thesis. Amherst: University of Massachusetts.
Sapir, E. () Grading. a study in semantics. Philosophy of Science. .. pp. –.
Sassoon, G. W. () Vagueness, Gradability and Typicality: A Comprehensive Semantic Analy-
sis. Ph.D. thesis. Tel Aviv University.
Sassoon, G. W. () Vagueness, gradability and typicality: The interpretation of adjectives and
nouns. Leiden: Brill.
Stechow, A. von () Comparing semantic theories of comparison. Journal of Semantics. .
pp. –.
Swanson, E. () Interactions with context. Ph.D. thesis. MIT.
Unger, P. () Ignorance. Oxford: Clarendon Press.
Villalta, E. () Spanish subjunctive clauses require ordered alternatives. In Jackson, B. and
Matthews, T. (eds.) The Proceedings of SALT , Cornell University, – June. pp. –. CLC
Publications.
Villalta, E. () Context dependence in the interpretation of questions and subjunctives. Ph.D.
thesis. Tübingen.
Villalta, E. () Mood and gradability: an investigation of the subjunctive mood in Spanish.
Linguistics and Philosophy. .. pp. –.
Yalcin, S. () Epistemic modals. Mind. .. pp. –.
Yalcin, S. () Probability operators. Philosophy Compass. .. pp. –.
Zanuttini, R. and Portner, P. () Exclamative clauses: At the syntax-semantics interface.
Language. .. pp. –.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i


Rationalization and the Ross
Paradox
Benj Hellie

 Introduction
I must finish this chapter—and here I am, working away. To type this sentence, as I am
doing, I must push the hyphen key three times—lo and behold, I just did it. Conversely,
although I surely should donate a lot more to causes fighting the prison-industrial
complex, and doubtless oughta do much more on behalf of ending fossil-fuel use, I
will not do either (hey, nobody’s perfect!).
The link between ‘must’ and intentional action appears tight. Whatever one thinks
one must do (whether important or trivial), one does. More cautiously—one might be
stymied, after all—one acts with the intention of doing it.
Indeed, among the English modal auxiliary verbs I understand, the connection
appears uniquely tight. What one thinks one can, could, may, might, should, or would
do (whether important or trivial), one often enough does not even try to do (Sloman,
, p. ); and much of what one will do, one won’t do intentionally.1 With the end
of understanding relations among agency, language, and modality, must stands out as
central, the remaining modals retreat to the periphery.2
Other linguistic phenomena connected to intentional action are imperative sen-
tences and the speech act of command. If Rance accepts Brent’s command ‘Pay the
rent!’ , Rance forms an intention to pay the rent, thinks he must pay the rent, and
endorses the sentence ‘I must pay the rent.’ . As for Brent, he issues the command
endorsing ‘Pay the rent!’, in the belief that Rance must pay the rent, endorsing the
sentence ‘Rance must pay the rent.’ , and in order to put on display an intention for

1 What one shall do is a matter I fear I cannot make much sense of: ‘shall’ is not in good standing in my
idiolect, and corpus uses are freighted with legalese and distorted by prescriptivism.
2 If ‘ethics’ is understood as the theory of value, ethics may be appropriately aligned with should/ought.
But if ‘ethics’ is understood as the theory of practical reason, ethics is appropriately aligned with must.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 benj hellie

Rance to pay the rent.3 Running in the opposite direction, if Rance intends to pay the
rent, he arguably does so thinking he must pay the rent and endorsing ‘I must pay the
rent.’4 —and, with some plausibility, also endorsing the self-directed imperative ‘Pay
the rent!’.5
These various phenomena—individual and general-will intention, imperative sen-
tences, the issuance and acceptance of commands, and the associated obligative
readings of ‘must’—I group together under the rubric of obligation.6 This chapter maps
out the territory of obligation. At its center is the psychological state of intention. An
intention for Rance to pay the rent has a content R of a kind I call a procedure—
a structured, nonpropositional representation of a plan of action.7 The imperative
sentence ‘You pay the rent!’, with Rance as the addressee, shares this procedural
content R.8 Brent’s command addressed to Rance using this sentence puts on display
an intention—harbored by the ‘general will’ of a community including Rance and
Brent, voiced by Brent—with the content R. Accepting the command, Rance forms
an intention with content R.9
My core thesis is that a modal sentence given an obligative reading is a modalized
imperative (Castañeda, ; Hellie, ).10 Read obligatively (and as addressed to

3 Castañeda (, p. ) brings Kant and Hare to his side in maintaining that ‘imperatives are at the
center of the connection of ‘ought’[]-assertions with both action and rationality’. To my mind, this settles
on the wrong modal.
4 Contrast the view that the obligative meaning of must-sentences is ineliminably social (Ninan, ,
§.; Portner, , p. ).
5 Contrast the view that the obligative meaning of imperative sentences is ineliminably social (Lewis,
a; Kaufmann, ).
6 The deontic logic tradition studies an operator . That operator is given the intuitive glosses ‘ought’
and ‘it is obligatory that’. On my terminology, ‘obligation’ is expressed with must rather than ought. If
ordinary ‘obligation’ conforms to my usage,  is either underspecified or semantically defective.
7 Contrast the view that intentions are a kind of belief (Setiya, , p. ).
8 A selection of opposing views of imperative contents: properties (Portner, , p. ); propositions
(Kaufmann, ); satisfaction–violation pairs (Vranas, ); ‘actions’ understood as ‘relations over worlds’
(Barker, ).
9 For a nuance, see note .
Charlow () salutarily links imperatival content to practical psychological states: for α an imperative,
[[α]] is a condition on ‘plans,’ objects with at least the structure of a set of action-kinds (); plans are (near
enough, perhaps) contents of a practical psychological state of ‘planning to do’ (Charlow, , passim). The
communicative meaning of an act of command using α is then (roughly) to constrain the content of the
audience’s ‘planning to do’-state so that it satisfies [[α]].
Charlow (, IV) advances this as improving upon a ‘Standard Account’ affirming the Display Thesis, on
which the content of σ is the content p of a mental state conventionally put on display by speech acts with
σ , and the orthodox Propositional-Attitude Thesis that all mental states are propositional attitudes. To my
mind, Charlow makes a compelling case that the Standard View founders on practical language phenomena.
On the present approach, the Display Thesis serves as a fixed point, while linguistic and psychological
content float in unison. I fear that without the Display Thesis, both linguistic and psychological theory
would be underconstrained. And (as a matter of personal taste) I would find questions about linguistic
meaning far less interesting if they did not shed light on psychology. I also wonder about the stability of
Charlow’s case against the Display Thesis: if ‘plans’ are psychological contents but are not propositional, the
Propositional-Attitude Thesis is already abandoned; but that makes room to preserve the Display Thesis.
10 In this slogan, a modal sentence is a sentence with a modal as its highest operator, while a modalized
blah is a modal sentence in which the modal taking widest scope operates on a sentence of type blah—for

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

rationalization and the ross paradox 

Rance), ‘You must pay the rent.’ operates syntactically on the imperative ‘You pay the
rent!’ and semantically on its content R.11 As is standard (Kaplan, ), I analyze
modals as operators on intensions, and assign intensions the job of representing entail-
ment; but I analyze entailment as the preservation not of truth, but of endorsement (a
mental attitude of ‘implicit acceptance’ toward a sentence: Hare, , p. ; contrast
Castañeda, ).12 In consequence, modals are not ‘metaphysical’ but rather ‘per-
spectival’ operators. Gathering these theses, the predicted meaning of ‘You must pay
the rent.’ is to express one’s perspective as endorsing ‘You pay the rent!’—namely, to
express one’s intention with content R.
A logical peculiarity unifies the phenomena of obligation, and contrasts them with
belief and its affiliated language of declarative sentences and epistemic/metaphysical
modality: the literature on obligative language labels this contrast the Ross Paradox
(Ross, ; compare Williams, ).13
For belief, conjunction both introduces and eliminates classically, and disjunction
introduces classically. I believe that I both posted this letter and drank up your wine
just if I believe that I posted this letter and I believe that I drank up your wine. And if
I believe that I posted this letter, I believe that I either posted this letter or drank up
your wine. Similarly for declarative sentences. ‘You posted this letter and drank up my
wine.’ is equivalent to the pair ‘You posted this letter.’ and ‘You drank up my wine.’ (the
conjunction entails each conjunct, the conjuncts together entail the conjunction). And
‘You posted this letter.’ entails ‘Either you posted this letter or you drank up my wine.’ .
Similarly for epistemic/metaphysical uses of must. ‘You must have posted this letter
and drunk up my wine.’ is equivalent to the pair ‘You must have posted this letter.’ and
‘You must have drunk up my wine.’ And ‘You must have posted this letter.’ entails ‘You
must have posted this letter or drunk up my wine.’ .

example, a modalized declarative and a modalized imperative are both modal sentences. I presume that
every modal sentence is declarative, so that a modalized imperative is a declarative modal sentence.
According to Geach (, p. ), ‘[in] a language that entirely lacked proper imperatives[,] we could
just use the plain future tense to give orders. [For example], “The patient will go down to the operating
theater.” ’. But if all directive modal sentences are modalized imperatives, this is wrong: the operand of ‘will’
is the indispensable imperative ‘The patient go down to the operating theater!’.
11 Contrast the ‘orthodox’ view that sentential-scope modals operate semantically on propositions
(Lewis, , , a, , §.; Kratzer, , ).
12 The conception of entailment as endorsement-preservation has been revived in more recent literature:
see H. Kamp, ; Veltman, ; Gillies, ; Yalcin, , inter alia.
This literature appears to presuppose endorsement-preservation to be incompatible with the intensional-
operator conception of modals—to my mind, incorrectly (Yli-Vakkuri, ; Hellie, ).
13 According to G. Kamp (, p. ), ‘Ross’s paradox [is] not only a reason to question the adequacy
of this or that deontic-logical rule or axiom, but provide[s] sufficient reason to drop the deontic-logical
approach altogether’ (quoted as translated in Hansen (, p.  n. )). If so, casualties include Geach
(); Chellas (, p. ); Lewis (b); Parsons (). For a more extensive survey of the logical and
linguistical impact, respectively, see Hansen () and Starr ().
The subsequent literature on the Ross Paradox is extensive. A selection of notable recentish work:
Lascarides and Asher (); Aloni (); Vranas (); Willer (); Kaufmann (); Portner (,
§.); Starr (, §.); Charlow (, §..); Cariani ().

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 benj hellie

In contrast, for intention, conjunction both introduces and eliminates classically,

but disjunction does not introduce classically.14 I intend to both post this letter and
drink up your wine just if I both intend to post this letter and drink up your wine.
But if I intend to post this letter, I may well yet not intend to either post this letter
or drink up your wine. Similiarly for imperative sentences. Proceeding on a loose
and intuitive understanding of imperative entailment, ‘Post this letter and drink up
my wine!’ appears to entail both ‘Post this letter!’ and ‘Drink up my wine!’, and to be
entailed by the pair of them: imperative conjunction both introduces and eliminates
like declarative conjunction. But ‘Post this letter!’ appears not to entail ‘Post this letter
or drink up my wine!’. Similarly for obligative must: ‘You must post this letter and drink
up my wine.’ appears to entail both ‘You must post this letter.’ and ‘You must drink up
my wine.’, and to be entailed by the pair of them. But ‘You must post this letter.’ does
not entail ‘You must post this letter or drink up my wine.’.15
This chapter is divided roughly / between psychology (sections –) and
language (sections –). The psychological part advances a technical and conceptual
explanation of the psychological Ross Paradox. On the technical side (sections –), I
propose that while belief-contents are propositional (‘fundamentally’ so, in a sense
to be developed in section .), the procedural contents of intentions determine a
granulated proposition—in effect, a pair of a proposition and a partition. While the
partition allows procedure-conjunction to introduce and eliminate classically, it sets
up a barrier to classical introduction for procedure-disjunction.16
On the conceptual side, I trace this contrast to the opposing polarities of belief
and intention in direction of rationalization (sections –): for belief, rationalization
points ‘inward’; for intention, rationalization points ‘outward’. A well-known, related
hypothesis similarly concerns the ‘normative order’ of psychology: a direction of fit
(Austin, ; Anscombe, , section ; Searle, ) runs from ‘world-to-mind’
for belief but from ‘mind-to-world’ for intention. But in contrast (section ), where
fit connects the world to the mind, rationalization is psychology-internal, keyed to
structural constraints on intelligibility (Heal, ; Hellie, ).

14 Contrast the view that the Ross Paradox is essentially linguistic (‘the idea here is to introduce
something like disjunction at the level of speech acts’: Charlow (, p. ), original emphasis).
15 I am inclined to agree with Charlow (, p. ) that the Ross Paradox for oughta is not quite as
compelling: ‘I oughta post this letter, so I oughta either post this letter or drink up all your wine.’ is better
than the must/bare imperative versions; the same goes for should. This suggests that oughta and should
cannot operate on procedures, but only on propositions.
Schroeder () and Chrisman () agree that oughta operates sometimes on propositions; Schroeder
but not Chrisman thinks oughta operates also on ‘actions’. Charlow’s observation supports Chrisman on the
question about oughta; though on the general question whether deontic modals can operate on things other
than declarative sentence semantic values, I side with Schroeder; though on the nuanced question whether
they operate on imperative semantic values, Chrisman (p. ) expresses sympathy: compare Chrisman
(, chapter ).
16 But procedure-disjunction does eliminate classically. Procedure-negation is similarly classical in
elimination but not in introduction.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

rationalization and the ross paradox 

Rationalization in belief (section .; Hellie, ) starts in evidence about local
physiological very recent history and flows toward belief about the more abstract and
remote, along channels of expectation. The origination in evidence makes information
stored in belief, such as that Rance posted a certain letter, transferrable: Fred preserves
information for his future reference, or transmits it to Brent. But if Fred wants the
information later or Brent wants it now, neither typically cares about why Fred wanted
initially to get it: their interest is limited to the raw propositional information. In that
sense, belief is ‘fundamentally’ propositional.
By contrast, rationalization in intention (section .) originates in abstract long-
term aims and/or open-ended projects and flows toward intention for the local
physiological very near future, along channels of instrumentality: to walk to work,
I walk halfway then the rest of the way; to walk halfway, I walk a quarter of the way
then the rest of the way; and so on (Anscombe, ; Thompson, ). Because the
rationalization of intention traces back to aims for which it is instrumental, intention
has a constrained ‘mandate.’ An intention to post a certain letter in service of paying
the rent is rationalized only by the end of paying the rent; the intention to post the letter
has no independent power to rationalize anything orthogonal to the end of paying the
rent, such as drinking up all Brent’s wine. That excludes assigning significance in the
content of intention to raw propositional information.
The closing % or so of the chapter brings obligative language into contact with
this psychological story. Assigning granulated proposition-determining procedures as
their contents, an abstract form of Ross-Paradoxicality percolates up to imperative
sentences (section ).17 The alignment between linguistic and psychological content
subserves a unification of intention with natural-language imperative endorsement
and command (section ): a ‘statics’ for imperative sentences grounding their endorse-
ment in intentions, together with an analysis of sentential intensions (and the affiliated
notions of validity and entailment) as reflecting patterns of endorsement (rather than
of truth),18 predicts natural-language Ross-Paradoxicality; a ‘dynamics’ for command
models the distribution of intentions using imperative sentences in their canonical
speech act. Obligative readings of must-sentences are brought into the fold by a
combination of means: a familiar statics for declarative sentences and dynamics for
their assertion; a grammatical analysis of obligative must-sentences as modalized
imperatives; and an analysis of modals as intensional operators (section ).19 This
package predicts that ‘You pay the rent!’ and ‘You must pay the rent.’ have the same

17 Boolean connective semantic values are classical Boolean operations, while composition adjusts:
namely, Booleanizing imperatives, the Boolean operation acts on the propositions while the partitions are
multiplied.
18 I therefore agree with Charlow (this volume, chapter ) about the desirability of a ‘reorientation
of semantic theorizing away from the aim of delivering a sentence’s truth-condition, toward the aim of
modeling the state of competently accepting a sentence’.
19 Cariani () analyses φ as true just if (i) of the best salient options partitioning modal space, all
entail [[φ]], (ii) no option entailing [[φ]] is ‘undesirable’, and (iii) the options do not cross-cut [[φ]]. Cariani
aims to solve, inter alia, the Ross Paradox and the ‘Professor Procrastinate’ puzzle (it is not true that he

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 benj hellie

endorsement-condition and are therefore equivalent; and that an assertion using the
latter can have the same communicative meaning as a command using the former.
The phenomenon of permission appears only in concluding remarks:20 I consider
why the dual of obligative must is in a sense ‘elusive’ in natural language; and I propose
a solution to Lewis’s (b) ‘problem about permission’.

 Concern and Indifference in Belief

A contrast between concern with a certain issue and indifference to it will be central
to my case against propositional content for intention. So I want to be perfectly clear
about what I mean by that contrast and why it matters to questions about content. I set
up for that in this section by moving at a deliberate pace through potentially familiar
terrain regarding belief.
On the familiar system of representation, when Fred is in a certain psychological
state, that fact can be factored into content and attitude (Lewis, a). For example,
if Fred believes that goats eat cans and that lambs eat ivy but does not believe that
ibexes eat amphorae, Fred bears the attitude belief toward the propositional contents
that goats eat cans and that lambs eat ivy but not toward the propositional content that
ibexes eat amphorae; if Fred wonders whether horses eat hay but not whether ibexes eat
amphorae, Fred bears the attitude wondering toward the content of whether horses eat
hay but not that of whether ibexes eat amphorae (whatever the contents of wondering
may be).
The identification of propositions with sets of possible worlds (aka subsets of modal
space) is familiar (Stalnaker, ; Lewis, ), if not completely uncontroversial. The
proposition that goats eat cans is identified with the set of just those worlds in which
goats eat cans; the proposition that lambs eat ivy with the set of just those worlds in
which lambs eat ivy. A proposition is true or false depending whether it contains the

oughta accept an assignment, because he presumably won’t carry it out, but he oughta accept the assignment
and carry it out).
Some worries. First, perhaps there is no Ross Paradox for oughta: see note . Second, Cariani complains
his opponent ‘must resort [] to completely different resistance strategies for’ the Ross Paradox and the
Professor Procrastinate puzzle. He also ‘refrain[s] from drawing any conclusions about the imperative
version of the [Ross Paradox]’ (p. , n. ). But there is no imperative version of the Professor Procrastinate
puzzle, so a unified treatment is not warranted. Third, if it is salient that I play with the dog all afternoon
and not ‘undesirable’ that I play with the dog but do not post this letter, then (post this letter) entails
(post this letter ∨ play with the dog all afternoon). Fourth, if it is salient that I drink up all your wine,
then (post this letter) entails ((post this letter ∧ drink up all my wine) ∨ (post this letter ∧ don’t drink
up all my wine)).
20 On a contrasting approach originated by Åqvist () (compare Aloni, ; Veltman, ; and
Charlow,  and , §..), an imperative has both permission- and obligation-content. I am sym-
pathetic to the spirit behind this account, and I conjecture that its apparatus affords the same predictive
resources as mine. But permission is a ‘positive’ phenomenon only socially, and not psychologically:
postulating a distinctive variety of content for permission is theoretically odd, somewhat akin to doing
so for epistemic possibility.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

rationalization and the ross paradox 

actual world: horses do eat hay, so the set of worlds where they do includes actuality,
so the proposition that they do is true; goats do not eat cans, so the set of worlds where
they do excludes actuality, so the proposition that they do is false.
The content of Fred’s total belief state can be identified with a certain set of
worlds, Fred’s doxastic possibilities (Hintikka, ; Stalnaker, ; Lewis, ): the
set containing w just if Fred, so to speak, takes w seriously as a candidate for actuality.
In believing that goats eat cans, Fred has ruled out of candidacy for actuality all worlds
in which goats do not eat cans; in his uncertainty whether horses eat hay, Fred takes
seriously at least some worlds in which they do and some in which they do not.
More generally, Fred believes that p just if his doxastic possibilities are a subset of
the p-worlds, disbelieves that p just if his doxastic possibilities are a subset of the not-
p-worlds, and is otherwise uncertain.
Belief is said to ‘aim at the truth’ (compare Frege, ). If belief aims at the whole
truth—aims to get doxastic possibilities down to just the actual world—belief will
never hit this target, because we never uncover the whole truth. By contrast, missing
the target is easy: in ruling out all worlds in which goats do not eat cans, Fred has ruled
out the actual world, and (unless he changes his mind) he will never get it back in his
sights. As far as the whole truth is concerned, the best we can hope for is to have not
yet missed it.
Aiming at the whole truth might seem pointless: we can never succeed. And it
might seem a fool’s errand: we can be reasonably sure of failing right out of the gate.
Why bother? Fortunately, individual truths are much bigger targets. If Fred aimed at
the truth that lambs eat ivy, he hit it: his doxastic possibilities are a subset of that
proposition, so (unless he changes his mind) Fred will never lose his grip on that truth.
Does belief aim at the whole truth, or just at truths? —Let us not yet legislate, but
instead distinguish protean belief, which aims at the whole truth, from granular belief,
which aims at particular truths.
When aiming at the truth, in which direction do we point? —If Fred knew to aim at
the truth that lambs eat ivy, he would already have it in hand as a truth, and would not
need to aim at it. That at which we aim need not be the truth down to the last detail,
but only the truth about a certain issue: of the worlds where lambs eat ivy and the set
of worlds where they do not, which of those sets includes the actual world. Aiming
at an individual truth is setting up a question to be answered; letting the arrow fly is
answering it, by settling belief into one of its answers.
Setting up a question to be answered is wondering about it.21 If one wonders, one
wonders whether goats eat cans, or who shot JR, or where Waldo went. Papering over
certain complexities, whatever one wonders sets up a range of answers on any of
which one might settle. A range of answers can be pictured as a bunch of lines of
subdivision cutting through modal space: what the wonderer aims to settle is which
of the subdivided lots contains the actual world. Or, more formally, as partition of

21 I retell an orthodox story: classic texts are Belnap () and Groenendijk and Stokhof (, ).

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 benj hellie

modal space: a set of propositions—the cells of the partition—which, pairwise, do not

overlap one another and which, collectively, subsume the entirety of modal space. For a
‘polar’ (yes/no) question such as whether goats eat cans, the partition is the pair set of a
proposition p (like that goats eat cans) representing the affirmative answer and another
proposition p, its complement, including exactly all other possible worlds, representing
the negative answer. When p is a proposition, I notate the question whether p with
?p := {p, p}.
If Fred wonders whether horses eat hay and whether chickens bathe in dust, he
wonders on which side of the horses do/don’t eat hay-line actuality lies, and also on
which side of the chickens do/don’t bathe in dust-line actuality lies. Assuming Fred
thinks the answers are independent, these two questions subdivide modal space into
four lots. More formally, when π and π  are partitions, their product, π × π  , is the set
containing the intersection of C and C , whenever C is a cell of π and C is a cell of π  .
The notion of a subpartition will be useful: intuitively, a subpartition draws every line
drawn by its superpartition, and maybe draws more. More formally, ρ is a subpartition
of π (ρ  π ) just when for every C ∈ ρ, there is some C ∈ π for which C ⊆ C .
It is straightforward to see that the product of two partitions is always a subpartition
of each.
The totality of the questions Fred has asked establishes a partition that is the product
of the partitions giving the content of any question Fred has asked. The totality of
Fred’s granular belief can be represented as a pair of the propositional content of Fred’s
protean belief and the partitional content of the totality of Fred’s questions. Say that a
certain proposition p is discriminable within the partition π just if p is the union set
of some subset of π —p does not distinguish any pair of worlds occupying the same
cell of (and therefore undistinguished by) π . Assuming Fred cannot answer a question
he has not asked, the content of his protean belief will never distinguish worlds that
are both within a single subdivided lot, and will therefore be discriminable within the
partition-content of his questions.
The logics of protean and granular belief-content differ: protean content is classical,
granular content Ross-Paradoxical. The story for protean belief is familiar: Boolean
operations on propositions are Boolean set-theoretic operations; entailment is subset.
So p∧q = p∩q (intersection), p∨q = p∪q (union), and ¬p = p (complementation).
And, allowing  to stand for the content-entailment relation (as contrasted with the
sentential-entailment relation , to be discussed later) q , . . . , qn  p just if q ∩ · · · ∩
qn ⊆ p. It is straightforwardly seen that, on this analysis, the propositional connectives
obey their classical introduction and elimination rules.22
The story for granular belief lacks a comparably orthodox treatment, but the
following seems natural. Let some premiss-partitions entail a conclusion-partition
(ψ , . . . , ψn  π ) just if the product of the premisses is a subpartition of the

22 Namely, ∧-intro: p, q  p ∧ q; ∧-elim: p ∧ q  p, q; ∨-intro: p  p ∨ q; ∨-elim: if p  r and q  r then

p ∨ q  r; ¬-intro: if p, q  ∅, then p  ¬q; ¬-elim: p, ¬p  ∅; ¬¬elim: ¬¬p  p.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

rationalization and the ross paradox 

conclusion: ψ × · · · × ψn  π (Groenendijk and Stokhof, ). Let a ‘granulated

proposition’ be a pair π , p such that p is discriminable within π . Some notation
will be convenient: where X, Y is a granulated proposition, let π X, Y = X (the
partition-component) and pX, Y = Y (the proposition-component). Then a Boolean
operation O on granulated propositions P and Q may be stipulated to yield that
granulated proposition with, as its partition-component, the product of the partition-
components of P and Q, and with, as its proposition-component, the action of O on
the proposition-components of P and Q: P ∧ Q = π(P) × π(Q), p(P) ∧ p(Q) ,
P ∨ Q = π(P) × π(Q), p(P) ∨ p(Q) , and ¬P = π(P), ¬p(P) .
Granulated-proposition entailment (Q , . . . , Qn  P) can then be identified with
both (i) entailment by the partition-components of the premisses of the partition-
component of the conclusion (π(Q ), . . . , π(Qn )  π(P)) and (ii) entailment by the
proposition-components of the premisses of the proposition-component of the con-
clusion (p(Q ), . . . , p(Qn )  p(P)). Conjunction introduces and eliminates classically,
and disjunction and negation eliminate classically. But disjunction does not introduce
classically: in order that π , p  π × ψ, p ∪ q , it must be that π  π × ψ;
but that requires π = π × ψ and therefore π  ψ. For example, let g be the
proposition that goats eat cans and h be the proposition that ibexes eat amphorae.
Then the disjunction of their ‘polar granulations’ G = ?g, g and H = ?h, h is
G∨H = ?g ×?h, g ∪h = {g ∩h, g ∩h, g ∩h, g ∩h}, g ∪h . The partition component of
G∨H is a subpartition of π(G) = {g, g}, while the proposition component is a superset
of p(G) = g: accordingly, neither G nor G ∨ H entails the other. Arguing similarly,
nor does negation introduce classically: Let P, Q  ρ, ∅ then π(P) × π(Q)  ρ
and p(P) ∩ p(Q) = ∅; while it follows from the latter that p(P)  ¬p(Q), all this is
compatible with π(P)  π(Q).
The metaphor of aiming, hitting, and missing can be discharged using talk of
norms and standards, of success and failure, correctness and incorrectness, and
related notions (Wedgwood, ). The standard for belief is the actual world, in
the following sense. The actual world fixes which propositions are true and which
false: false propositions are incorrect, true propositions not incorrect. If one’s doxastic
possibilities are incorrect, one has failed at protean belief; if not, one has not (yet) failed
at protean belief.
Granular belief may yet be successful. Fred has asked whether goats eat cans and
answered it incorrectly. But he has also asked whether little lambs eat ivy—and
answered correctly. The partition content of Fred’s questions discriminates each mem-
ber of the set of the goats eat cans-worlds and the set of the lambs eat ivy-worlds from
its complement. The content of Fred’s protean belief is within the incorrect cell of the
former, but within the non-incorrect cell of the latter. That makes Fred’s granular belief
whether goats eat cans a failure, but his granular belief whether lambs eat ivy a success.
More generally, suppose Fred’s questions determine the partition π , with the actual
world in cell C @ ; and let Fred’s protean beliefs have the content b, for b discriminable
within π . Then consider some proposition p that is also discriminable within π . If b

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 benj hellie

is a subset of p, we may say that Fred has the granular belief that p. That proposition p
is not incorrect just if C@ is a subset of p. That is so just if Fred’s granular belief that p
is a success; otherwise it is a failure. If b is not a subset of p, then Fred does not believe
p: perhaps b is a subset of the complement of p, so Fred disbelieves p. Otherwise, b
leaves it open whether p, and Fred is uncertain whether p; moreover, when one is
uncertain about a proposition discriminable within the questions one has asked, say
one is indecisive about that proposition. The success/failure conditions of disbelief are
obviously the opposite of those of belief.
What of those propositions p which are not discriminable within π ? For example,
I supposed earlier that Fred had not asked whether ibexes eat amphorae. Perhaps that
proposition is independent of the totality of questions Fred has asked: any cell in π
contains worlds where ibexes eat amphorae and cells where they do not. Then Fred
can have no opinion on the matter: Fred does not believe that ibexes eat amphorae,
nor does he disbelieve it.
In that case, Fred has no opinion on the matter because nothing he cares to find out
could settle it. But other cases of propositions not discriminable within the totality of
Fred’s questions are not that way. Consider the proposition that either goats eat cans or
ibexes eat amphorae. Fred believes the stronger proposition that goats eat cans. Does
he thereby believe the weaker proposition? It is not clear whether intuition pronounces
on the matter; perhaps theoretical motivations may be found for saying one, perhaps
for saying the other; perhaps for giving one answer for some cases, the other for
others. Fortunately, we have apparatus enough to allow us to say either. Fred protean-
believes it because all doxastic worlds are members of the disjunctive proposition.
As for granular belief, we could group the threefold division of belief, disbelief, and
indecision under a category of concern, and set that supercategory against an exclusive
category of indifference: and we could say that Fred has concern whether p just if p is
discriminable within the content of his questions, and is otherwise indifferent. If we
decided to use the apparatus in that way, we would rule Fred as indifferent regarding
the disjunctive proposition.
How does the normative status of indecision (recall, uncertainty regarding a ques-
tion one has asked) compare to that of indifference? Having asked a question but
not answered it, one’s situation is clearly neither (yet) a success nor a failure by the
standards of belief. Arguably, though, indecision has a negative normative status along
a different dimension: questions are ‘to-be-answered’; having asked one, we are in a
situation that is in a sense ‘unsettled’ until we settle it. Indifference, by contrast, is
none of the above: neither successful nor failed, neither settled nor unsettled.

 Concern and Indifference in Intention

Belief can perhaps be sensibly spoken of either as protean or as granular; intention, by
contrast, is always granular, never protean.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

rationalization and the ross paradox 

So I claim. Intention is far less well-understood than belief, so I do not hope to ride
the winds of orthodoxy, but will have to argue the point. Unfortunately, establishing a
stable foundation requires drilling down far to hit bedrock: familiar formally-driven
stories lack phenomenological plausibility; familiar phenomenologically-driven sto-
ries lack formal rigor. So rather than with tools or doctrine, I begin by collecting data—
in particular, about the bearing of various worldly circumstances on the success or
failure of intention. I will then run these through the apparatus of the previous section
to generate my thesis that the content of intention is always granular.
For data-gathering, I exploit Lewis’s (b) device of a ‘questionnaire’—to my
mind, and I hope you will agree, it shows we have some ‘very firm and definite
opinions’ about the success and failure of intention.
First question. On Monday, you are strolling along on the way to work, preoccupied with
thoughts about deontic modals. Your attention is briefly drawn to a stone skittering along the
sidewalk, which you appear to have just kicked. You pay it little mind and continue on your
way. Consider the process-type kicking a stone. Does your having just undergone a process of
this type constitute your success or failure at anything?
Answer. Of course not. Whether you kick any stones is a matter of complete indifference to you.
Had you not kicked any stones at all on the way to work, that would not have constituted your
success or failure at anything either.
Second question. On Tuesday, it is as on Monday, except that you have set out with a mindfulness
project requiring not kicking any stones—but, again preoccupied, you kick a stone. Does your
having just undergone a process of the type kicking a stone constitute your success or failure at
anything this time?
Answer. Yes. You have failed at your intention to avoid kicking any stones.
Third question. On Wednesday, vexed at yesterday’s failure, you have set yourself the intention
of giving a good hard kick to some stone. Spotting one, you kick it. Does your having just
undergone a process of the type kicking a stone constitute your success or failure at anything
this time?
Answer. Yes. You have succeeded at your intention to kick a stone.
Last question. On Thursday, when you arrive at work, you set yourself the intention of either
reading the new article by N or the new article by M that day. You manage to steal enough time
to read the new article by N over the course of the day. Does your having undergone a process
of the type reading the new article by N constitute your success or failure at anything?
Answer. Yes. You have succeeded at your intention to read either the new article by N or the
new article by M. You would have equally well succeeded had you instead read the new article
by M. And you would have succeeded at that intention had you been able to read both.

If some of this does not seem obvious, please reconsider; if it all does, splendid! —Let
us use it as evidence for what follows.
It is not the answers to the questionnaire by themselves that establish the content of
intention to be never protean. For that matter, they do not even establish the content
of intention to be ever granular. The obvious proposal for protean content is this: the

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 benj hellie

content of intention includes, on Monday, both worlds where one kicks stones and
worlds where one does not; on Tuesday, only worlds where one does not kick stones;
on Wednesday, only worlds where one does kick stones; and on Thursday, only worlds
where one reads either the article by N or the article by M (or both).
But a case for granular content emerges in working through the following exercise:
Exercise. On Friday, you are strolling along on the way to work, self-possessed, confident in
your way of being. You have recognized that true mindfulness consists in equanimity about the
trivial. Regarding the vexatious kicking of stones, you have set yourself the intention to either
kick (one or more) stones or not kick (any) stones. As it happens, you kick a stone. Does your
having just undergone a process of the type kicking a stone constitute your success or failure at
anything this time?
Answer. Yes. You have succeeded at your intention to either kick stones or not kick stones. But
you would have equally well succeeded had you not kicked any stones.
After all, recall that on Thursday, one has resolved to do at least one of two things: read the
article by N; read the article by M. So long as one does at least one of them, one will succeed,
and will not fail. One will fail only if one does neither.
To generalize, whenever one intends to act in at least one of two ways, one fails only by doing
neither, succeeds by doing either (or both). It should not matter whether the thing in question
is ‘positive’ or ‘negative’. If one resolves either to read the article by N or drink five or more cups
of coffee, one fails only if one neither reads the article nor drinks five or more cups of coffee,
otherwise succeeds. And if one resolves either to read the article by N or not drink five or more
cups of coffee, one fails only if one does not read the article and does drink five or more cups of
coffee, otherwise succeeds.
Now, when an intention is to satisfy at least one of two properties, one of them the negation of the
other, it should still be that one fails only by doing neither, succeeds by doing either (or both).
Why should the content make a difference? Normative psychology does not ordinarily involve
that kind of exception, so best not to postulate it now, either. Now of course, when one intends
to act at least by kicking a stone or not kicking a stone, one cannot do neither and cannot do
both. Because the only remaining alternatives are do the one—kick a stone—or do the other—do
not kick a stone—one must do exactly one of those. Either way, one succeeds.
So sometimes one’s intentions are guaranteed to succeed. I am inclined to see this consequence
as an occasion for rejoicing, rather than resistance.

Intention, like belief, can involve a distinction between concern with an issue—the
prospect of success or failure, depending on how the world meshes with the content
of one’s attitude—and indifference, in which the answer to the question cannot imbue
the relevant aspect of one’s mental state with either success or failure.
Indeed, distinguishing concern and indifference is at least as warranted for intention
as for belief. Contrast Monday and Friday. On Monday, whether one kicks a stone is a
matter of no significance for the success or failure of any intention: one is indifferent.
But on Friday, there is an intention such that, if one kicks a stone, it will succeed, and
if one does not kick a stone, it will succeed: one’s attitude is equanimity.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

rationalization and the ross paradox 

The distinction between indifference and equanimity, like the distinction in belief
between indifference and indecision, cannot be represented with protean content. On
the earlier attempt at a protean characterization, the Monday content contained kick a
stone-worlds and don’t kick a stone-worlds; the Tuesday content contained only don’t
kick a stone-worlds; the Wednesday content contained only kick a stone-worlds. The
Friday content should not distinguish kick a stone and don’t kick a stone-worlds; and
we surely do not want it to do so by excluding all worlds; so the Friday content is like
the Monday content in containing both kinds of world.
Saying what the content is does not yet settle the bearing of this on the
success/failure-status of one’s intention. Presumably false content makes for failure.
Perhaps true content makes for success (as with granular belief). Then one’s intention
succeeds no matter what on Friday, desirably—but then it also succeeds no matter
what on Monday, undesirably. Or perhaps true content makes for mere non-failure
(as with protean belief). Then one’s intention neither fails nor succeeds no matter what
on Monday, desirably—but then it also neither fails nor succeeds no matter what on
Friday, undesirably.
So protean intention cannot make enough distinctions. By contrast, this is no
problem for granular intention: Monday does not partition kick a stone from don’t
kick a stone-worlds, but Tuesday, Wednesday, and Friday do; Tuesday, in addition,
designates the don’t kick a stone-answer, by involving a proposition that is a subset of
the don’t kick a stone-cell; Wednesday designates the kick a stone-answer, by involving a
proposition that is a subset of the kick a stone-cell; and Friday designates both answers,
by involving a proposition overlapping both cells.23

 Direction of Rationalization
Belief is arguably sensibly spoken of both as protean and as granular; intention is at
least sensibly spoken of as granular. Can intention be sensibly spoken of as protean?
No: this section and the next trace the contrast between intention and belief to their
opposing polarity in what I call direction of rationalization.
I will sidle up on this notion by reflecting on the challenges faced by the venerable if
obscure notion of direction of fit. In the tradition, belief and intention (often: ‘desire’)
are said to display oppositely polarized directions of fit with regard to the world: in
some sense duly capturing their fundamental normative contrast, belief is to fit the
world, whereas the world is to fit intention—where ‘A is to fit B’ means B somehow
exerts normative pressure on A to fit B. Alluring, to be sure. But what on earth could
it mean?

23 I explicate designation in terms of what makes for success, in contrast with in terms of what is permitted.
In that sense, my view represents disjunction as ‘alternative-presenting’ in contrast with ‘choice-offering’:
Rescher and Robison (, p. ); Åqvist ().

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 benj hellie

For ‘fit,’ I imagine truth of content is just enough. Substituting, beliefs are often true
thanks to the world’s exerting normative pressure on them, while intentions are often
true thanks to their exerting normative pressure on the world. This is not so plausible:
how is the world ‘normatively pressured’ to make intentions true (Humberstone, ;
Frost, )? If an errant bus undermines my intention to walk to work, that may
be problematic in many ways—but I doubt one of them will be the world’s regret at
violating its half of the bargain struck in my forming my intention.
A popular fix has been to exchange normative pressure for causal pressure (Smith,
; Velleman, ; Setiya, ). Beliefs are often true thanks to the world’s exerting
causal pressure on them, while intentions are often true thanks to their exerting causal
pressure on the world. Of course the world impinges on our sensory organs, and in this,
we acquire evidence, which then ramifies in further, more abstract beliefs; when these
do not go too far from evidence, they are often true. And of course, we implement our
more abstract intentions through intentional motions of our bodies, which exert gentle
and local causal pressure on the world; when we planned well and matters beyond
our locality do not impede us, our intentions often come true. But the body is where
causality ends, for belief, and starts, for intention. Characterizing the post/pre-bodily,
psychological part of the story as just more causation does not square comfortably
with ordinary understanding of psychology (is ‘phenomenologically off-key’). But the
point of discussing direction of fit is to shed light on psychology: so this fix changes
the subject unhelpfully.
This suggests that the remedy is to constrain our ambition: we need to limit the
world’s exposure to normativity. We can continue to distinguish belief and intention
in direction of normative pressure if the norm of truth is exchanged for a more
psychology-internal, structural normative constraint based in a notion of rationaliza-
tion (Hellie, , ). If psychology and the world are isolated, this will overshoot,
kicking the world out of the story. Fortunately, psychology and the world overlap at
the body: for belief, it is where causality ends, but also where rationalization begins;
for intention, it is where causality begins, but also where rationalization ends. I will
add detail to this sketch after introducing ‘rationalization’.
The root structural psychology-internal norm is intelligibility.24 For an example of
its absence, suppose Fred asserts ‘it is false that either goats eat cans or horses eat hay’
and then asserts ‘goats eat cans’. How is Fred thinking of the world? Without further
information, I am in the dark.
This case does not distinguish the internal norm of intelligibilty from the external
norm of truth: Fred appears to have contradicted himself, so his statements at face
value cannot possibly be true. But intelligibility and truth can be distinguished by
means of Moore’s Paradox. Suppose we ask Fred whether goats eat cans, and he

24 Compare ‘If we say “it does not make sense for this man to say he did this for no particular reason” we
are not “excluding a form of words from the language”; we are saying “we cannot understand such a man” ’
(Anscombe, , §).

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

rationalization and the ross paradox 

emphatically replies ‘yes, goats do indeed eat cans’. One of us did not catch his answer,
and follows up: ‘sorry, was it your opinion that goats eat cans?’ Fred answers ‘Now
you are asking me my opinion—I thought this was supposed to be about goats. No,
I am not of that opinion’. Here too it is hard to make sense of what is going on with
Fred. But this is not because it is impossible for goats to eat cans though Fred lacks the
opinion that they do. Instead, it is because there is no intelligible perspective which
pictures goats as eating cans without lucid recognition under at least some mode of
presentation that it does so.
Why think intelligibility is a ‘norm’? —Not because intelligibility is guaranteed
to be beneficial (contrast Parfit, , §; Harman, , § .) or is connected to
anything else of value or to any other sort of value. Rather, Fred’s intelligibility is a
boundary condition on our using psychological categories to describe or explain him.
No intelligibility, no location in the ‘space of reasons’ at all: intelligibility is a ‘necessary
precondition on the possibility of normativity’. If that is not enough for counting as a
norm, but only as a ‘norm*’, I will concede the point—but in the interest of saving ink
and trees, I leave the ‘*’ implicit.
Intelligibility provides a basis for defining a variety of ‘normative pressure’ exerted
by mental states on one another: this rationalization. To illustrate, Sam thinks her pet
rabbit, Rupert, has escaped. Why? Sam replies ‘if Rupert is in the house, I always know
which room he is in; and I don’t know which room Rupert is in’—modus tollens. The
first explanans is the state supporting the conditional: a pattern of expectation that is
notably ‘stable’ in Sam’s mental life. The second is the rejection of the consequent, a
temporary condition of uncertainty in Rupert’s whereabouts. Having accepted these
explanantia, we would find Sam unintelligible if she did not think Rupert had escaped.
Unintelligibility is a ‘boundary condition’ on psychological characterization; knowing
all this, we therefore have no choice but to accept the explanandum (that Sam
thinks Rupert has escaped). So what we learned explains the belief; the bankshot
off of intelligibility is what makes the explanation a rationalization. In general, then,
some states rationalize others when the former explain the latter by way of its being
unintelligible to have the former without the latter. Because rationalization is an
explanatory relation, it inherits all the directedness of explanation.
I propose that the opposite polarity of belief and intention pertains to the direction
of rationalization. For belief, rationalization flows, in a sense, ‘inward’; for intention,
it flows ‘outward’. But inward or outward toward what and from what?
To see the answer, suppose Brent is off to Lower Slobbovia. He thinks Sam knows
how one will eat in Lower Slobbovia, and wants to know how he will eat, so he asks
her. Sam replies ‘you will eat terribly in Lower Slobbovia’. So: should Brent believe he
will eat terribly in Lower Slobbovia? And should he intend to eat terribly in Lower
Slobbovia?
Very plausibly, learning Sam’s answer rationalizes Brent believing he will eat terribly
in Lower Slobbovia. Brent thinks Sam knows how one will eat in Lower Slobbovia, and
has learned that Sam believes he will eat terribly in Lower Slobbovia. I am inclined to

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 benj hellie

say that if A thinks B knows the answer to a certain question, and A thinks B’s answer
to the question is P, then A thinks P is true—on pain of unintelligibility (%‘Rance
knows whether goats eat cans and thinks they don’t—but I don’t know whether goats
eat cans’). So it is unintelligible for Brent not to believe he will eat terribly in Lower
Slobbovia.
But Sam’s answer does not rationalize Brent’s intending to eat terribly in Lower
Slobbovia. Brent might very well, knowing this, strive with great energy—albeit
pessimistically—to eat well in Lower Slobbovia. Perhaps intention in some sense
requires belief. But only for the completely acquiescent/megalomanical fatalist is belief
always accompanied by intention!
To explain the contrast, I will quickly rehearse a model of ‘rational architecture’ I
have defended elsewhere (Hellie, , § .; , §§ –). Belief begins in evidence:
evidence must be true, must be believed, consists of facts broadly about ‘perception’,
and is given under the ‘mode of presentation’ of perception (Lewis, b). So, for
example, if I see a dagger before me, my evidence is some true proposition about my
‘visual condition’—perhaps that I see a dagger before me, perhaps that my retina is
stimulated daggerishly, perhaps that my brain is agitated daggerishly. Evidence must
be true and must be believed, but there remains plenty of room for falsehood in belief.
The perceptual ‘mode of presentation’ typically must be ‘recoded’ into something less
evanescent and more ‘conceptualized’ (Hellie, , § .–); even then, my evidence
concerns just my perceived surroundings up to the present, so my full rich set of beliefs
requires extensive ‘amplification’ by my pattern of expectations.25
(If I lucidly dream of a dagger before me, my evidence is again some true
proposition about my ‘visual condition’—perhaps that I am imagining a dagger, or
that my visual cortex is agitated daggerishly; either way, it is certainly possible to
have mistaken ‘background beliefs’ about my condition and therefore ‘recode’ the
perceptual mode of presentation in a way that conflicts with my actual evidence: in
that case, I wind up with inconsistent beliefs, and am not fully intelligible: Hellie, ,
§ ; Hellie, , § ..)
Because evidence must be believed, evidence pushes belief around. Our expecta-
tions extract a great deal more from evidence than the content of evidence strictly
entails; still, this leveraging move is not without risk, in that evidence might always
surprise us and wind up contradicting what we expected: when that happens,
we lapse from full intelligibility until consistency is restored (making ‘belief willy
nilly’ generally to be avoided). In this sense, rationalization in belief flows ‘from
evidence’.
That evidence can only have true content is good fortune, given that it must be
believed. But this also gives evidence a foot on both sides of the fence: evidence is

25 A congenial theory of the mechanics of expectation exploits the sort of construction behind ‘presum-
ably’ in Veltman () and exploited by Yalcin (this volume, chapter ) in the analysis of should; I conjecture
that ‘ranking functions’ (Spohn, ; Huber, ) afford an interchangeable apparatus.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

rationalization and the ross paradox 

‘in the world’, and also ‘in the mind’. In that sense, when rationalization in belief flows
‘from evidence’, it flows ‘inward’; and that is the good sense to be made of the claim
that belief ‘fits the world’.
One’s evidence at a given time pertains to the stretch of history ending at the present
moment. But one’s intentions pertain to the stretch of history beginning at the present
moment. So intention cannot radiate from evidence: evidence constrains intention at
best in partnership with expectation. So rationalization for intention does not (for that
reason anyway) flow ‘inward’.
Indeed, it flows ‘outward’. The paradigmatic rationalization in intention is from
a longer-term, more ambitious ‘governing intention’ to a structure of shorter-term,
less ambitious ‘implementing intentions’ undertaken as instrumental to executing the
governing intention (Anscombe, ; Thompson, ). Suppose I set myself to walk
to the office. I have to implement this intention somehow: I can’t just blink my eyes
and it happens. In particular, I have to walk the first half of the way, and then walk the
second half of the way. But to do all this intentionally, I need intentions to do both of
those things. And they too need to be implemented: to do the first, I have to walk the
first quarter of the way, then the second quarter of the way. And so on. At some point I
am implementing my intention to walk to the corner by walking halfway to the corner,
implementing that by taking three steps, implementing that by taking one step, . . .
implementing that by exerting all sorts of micro-forces throughout my body, . . . .
More generally, long-term ambitious intentions rationalize shorter-term less ambi-
tious intentions, and so on—eventually arriving at motor coordination in the here
and now. Now, motor coordination in the here and now is at least somewhat like
evidence: each pertains to the here and now, and to matters impinging pretty closely
on physiology. If they differ, it is in that motor coordination is about moving, while
evidence is about sensing. And maybe moving and sensing, at the level of physiology,
are not really dissociable: perhaps all there is is the ‘sensorimotor’. So either way,
rationalization of belief flows ‘inward’ from evidence about the sensory (or perhaps
the sensorimotor), while rationalization of intention flows ‘outward’ toward intention
for the motor (or perhaps the sensorimotor).
So, in sum, belief and intention are oppositely polarized in their direction of
rationalization: rationalization in belief flows inward from the physiological here and
now qua evidence; rationalization in intention flows outward toward the physiological
here and now qua near-momentary intentional action.

 Polarity and Content

. Inward-directedness and protean content
The job assigned to granular belief was giving a point to inquiry: without questions
to answer, success could only amount to omniscience, which we will never reach, and
failure would be absolute and permanent with the smallest initial mistake.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 benj hellie

Why is that not a compelling reason to jettison protean belief entirely? Because
belief is inward-directed, it has a kind of ‘objectivity’: my beliefs are not just for me, nor
yours just for you; rather, mine can be for you and yours for me. The granulation Fred
imposes on his beliefs, conversely, lacks ‘objectivity’: it is an idiosyncratic matter of
how Fred packages his information for himself, indispensable to Fred but potentially
of no interest to anyone else. Fancy words? This ‘objectivity’ is easy enough to
operationalize: the most salient symptom is that accepting testimony is rationalizable.
I therefore argue as follows. First: because belief is inward-directed, accepting
testimony is rationalizable. Second: because accepting testimony is rationalizable,
belief has protean content.
I do not expect the first premiss to be controversial. If Brent thinks Fred’s beliefs
about animal husbandry have lost their moorings in evidence, it would be unintelli-
gible for Brent to accept anything Fred says on the matter. Conversely: even if Rance
thinks Sam’s beliefs about the culinary amplify her evidence considerably, that is not
in itself a reason not to accept what she says. After all, it may be that Rance’s own
beliefs amplify evidence in more or less the same pattern as Sam’s. If so, Sam’s beliefs
about the culinary can be expected to be roughly the same as Rance’s would be if he
had her evidence. So if this is Rance’s view, Rance will expect Sam’s testimony to just
put on display reasonable expectations on the basis of some (in absentia) evidence. If
so, when Sam asserts that p, and Rance is concerned whether p, for Rance to remain
indecisive whether p would be either just as hard to make sense of as ignoring evidence
on a matter of concern or just as hard to make sense of as skepticism about his own
reasonable expectations. So without inward-directedness, accepting testimony would
not be rationalizable, while with inward-directedness, what is not rationalizable is
indecisiveness regarding at least favorable cases of testimony over matters of concern.
So in that sense, it is because of the inward-directedness of belief that accepting
testimony is rationalizable.
Now to the second premiss. The idea here is that sometimes when accommodating
your issues would be a barrier to my acquiring information from you, I can still do it;
but then I can gather your information without accommodating your issues; and if so,
I think of your information as protean belief.
Suppose each of Rance and Brent was indifferent whether Amber would be at the
party when he arrived, and indifferent whether Bruno would be there. But while
individually, neither Amber nor Bruno was a matter of concern to either Rance or
Brent, Amber and Bruno collectively were a matter of concern to each—though in
different ways. Rance was worried that both Amber and Bruno might be there, while
Brent was worried that Amber might be there without Bruno. The issue of concern to
each, then, was a polar question: for Rance, ?(a ∩ b); for Brent, ?(a ∩ b). Upon arrival,
each was happy to answer his question in the negative—Rance believing a ∩ b, Brent
a ∩ b—and gave it no further thought.
Some time later, Sam, who does not know Bruno, is wondering whether Amber was
at the party: of concern to Sam is ?a, but not ?b. Knowing Rance and Brent were there,

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

rationalization and the ross paradox 

she asks each whether Amber was there. Rance replies ‘I don’t know, but Amber wasn’t
there with Bruno’; Brent replies ‘I don’t know, but Amber wasn’t there without Bruno’.
Sam concludes that Amber was not at the party—believing a and thereby settling her
question in the negative.
Suppose now for reductio that belief is essentially granular. In that case, the content
of Rance’s posterior belief entails neither ?a, W nor ?(a ∩ b), W , the content of
Brent’s posterior belief entails neither ?a, W nor ?(a ∩ b), W , and the content of
Sam’s posterior belief entails neither ?(a ∩ b), W nor ?(a ∩ b), W .
Grant me that if an assertion communicates a content X, for the speaker’s posterior
belief-content S and the audience’s posterior belief-content A, both S  X and A 
X. So whatever was communicated to Sam by either Rance or Brent entails neither
?a, W , nor ?(a∩b), W , nor ?(a∩b), W . But the only other granulated proposition
even conceivably in the ballpark for what was communicated is the trivial granulated
proposition  := {W}, W —which could not have settled anything for Sam, as the
communications evidently did.
So, I conclude, belief is not essentially granular. Rance’s assertion communicated his
protean belief that a ∩ b, and Brent’s communicated a ∩ b, with the consequence that
Sam’s posterior protean belief entails a—and her question ?a is settled in the negative.
. Outward-directedness and granular content
The direction of rationalization of intention is oppositely polarized, in the sense
that what rationalizes a given intention is (ordinarily,26 and in significant part) the
‘governing’ intention it implements.
For this reason, intentions are not in any reasonable sense transferrable. Suppose
Rachel intends to hang a picture Joan thinks is ugly. Joan need not intend to hang that
picture herself; or support Rachel in doing so; or even stay out of the way: Joan might
quite reasonably try, perhaps, to foil Rachel. The intention itself does not transfer. Each
of Joan and Rachel might be confident their side will prevail, with Rachel believing she
will hang the picture and Joan believing Rachel will not hang the picture. Any belief
associated with the intention of either need not transfer either. The reason for ascribing
protean content to belief does not apply to intention.
Conversely, the business of intention—implementing its governing intention—
requires its content to be granular. A given intention has a limited mandate: it is to
carry out its role in the plan for implementing the governing intention, and then
expire.
For example, Sam intends to do her job today, and to do that, she has to get to the
office. So she sets out after breakfast with an intention to walk to work at the office.

26 The hedge is to accommodate the rationalization of ‘highest’ intention: it has to top out at some stage,
so at that stage intention is either self-rationalizing or arational or rationalized by a governing non-intention.
The status of the ‘ultimate sources of motivation’ is orthogonal to the planning relations that are the focus
of my argument.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 benj hellie

She plans, and therefore intends, to do this by first walking halfway and then the rest
of the way; she plans to walk halfway by first walking the first quarter of the way and
then walking the second quarter of the way; and so forth. The intention to walk the
second quarter of the way is rationalized by the intention to walk halfway. Its mandate
is just to do its part in making Sam walk halfway, and then expire. While exercising that
mandate requires the creation of other intentions it in turn rationalizes, the intentions
it creates are rationalized only through the contribution they make to attaining the
goal of walking the second quarter of the way.
So in particular, the intention to walk the second quarter of the way could rationalize
an intention to walk the third eighth of the way. But it could not rationalize a
disjunctive intention to either walk the third eighth of the way or empty the bank
account and go on a drinking spree after work. That disjunctive intention succeeds
unless one neither walks the third eighth of the way nor empties the bank account and
goes on a drinking spree after work. In particular, it succeeds if one empties the bank
account and goes on a drinking spree after work. An intention rationalizes whatever
might be appropriate for realizing its success-condition; so the disjunctive intention
rationalizes, if appropriate, emptying the bank account and going on a drinking spree.
So if the intention to walk the second quarter of the way rationalizes the disjunctive
intention, it can (derivatively) rationalize emptying the bank account and going on a
drinking spree (just so long as it also rationalizes walking the third eighth of the way).
But emptying the bank account and going on a drinking spree makes no contribution
whatever to walking to work; so because the intention to walk to work is what governs
the intention to walk the second quarter of the way, the rationalization bestowed on
the latter is restricted to its contribution to the aim of walking to work. If it ‘goes rogue’,
so to speak, it does so without any rationalization, and therefore without any power to
rationalize the wacky splinter plan.
So a disjunctive intention is not in general rationalized by its disjunct-intentions.
But a conjunctive intention does seem to rationalize its conjunct-intentions: if I intend
to have breakfast and then go back to sleep, I intend to have breakfast and I intend to
go back to sleep (after eating breakfast).
Disjunction-introduction, understood as a rule of rationalization for the contents
of intentions, seems to be invalid; but conjunction-elimination is valid. That means
that rationalization by the contents of intention sometimes traces decreasing strength
of protean content, but not always. If not, the contents of intentions cannot be protean.
Whatever these contents may be—I call them procedures—it is plausible that
they at least determine a partition over something. As above, entailment over
granular propositions validates classical conjunction-introduction and -elimination,
and disjunction- and negation-elimination, but neither disjunction- nor negation-
introduction. Intentions, we have seen, validate conjunction-elimination but not
disjunction-introduction: and they seem to conform to the remaining patterns as well.
Conjunct-intentions do seem to rationalize their conjunctive intention: if I intend to
work and I intend to whistle, I intend to whistle and work (conjunction-introduction).

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

rationalization and the ross paradox 

If I intend to have breakfast, I cannot intelligibly intend to fail to have breakfast

(negation-elimination). Something like disjunction-elimination seems valid: to get
downtown, you gotta ride a bus; to get uptown, you gotta ride a bus; so to get uptown
or downtown, you gotta ride a bus. But negation-introduction seems invalid. The case
against disjunction-introduction works equally well: the intention to go to work is
incompatible with the intention to neither go to work nor to have a drinking spree;
but that is so without the intention to go to work requiring an intention to either go
to work or have a drinking spree.
A range of further phenomena militates for ascribing still further varieties of
structure to the procedural contents of intentions, which provide affordances for
further intricacies of natural language.27 I set these aside for treatment elsewhere.

 Imperative Semantics
The conclusion of the foregoing discussion is that belief and intention contrast in
that belief is optionally (and perhaps fundamentally) protean, while intention is
mandatorily granular. This mandates Ross-Paradoxical behavior for the content of
intention while permitting classical behavior for the (perhaps fundamental) content
of belief.
I now transition from the psychology to the language of obligation. In these
concluding sections, I will describe a series of increasingly complex languages.
Collectively, these (i) show how agents conforming to my psychological hypotheses
might perform speech acts to put (individual) beliefs and (general-will) intentions on
display, in order to collect individual beliefs and parcel out collective intentions; and
(ii) illustrate meanings within such a practice for expressions designed to resemble
ordinary-English Boolean connectives and modal operators. These languages con-
form, in various nuanced ways, to the phenomena of obligation with which I opened
the chapter: the connection between the language and psychology of obligation;
certain commonalities and distinctions between command uses of imperatives and
obligative uses of must, including the Ross Paradox; the elusiveness of their ‘duals’,
permission and permissive can/could.
The initial language L contains simple imperative sentences, with procedure-type
semantic values, and classical Boolean connectives: with its natural entailment rela-
tion, the  of section , L is Ross-Paradoxical. Next, L introduces subject-predicate
structure to imperatives, allowing connectives to take scope: narrow-scope disjunction
‘softens’ Ross-Paradoxicality, as in English.

27 Three examples. Delayed triggering: wait five minutes, then stir; remove on July ; leave at sundown;
smile when they walk in; fold in the whipped cream, then put it in the oven. Conditionalization: if it is A, do
α; if it is B, do β; if you can’t tell, do γ . Delegation: the contrast between Rance’s intention to pay the rent,
which is essentially ‘de se’ or subjectless, and our intention for Rance to pay the rent, which adds a subject.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 benj hellie

In section , L wraps a ‘language game’ around L : command uses of imperatives

distribute an initially centralized stock of general-will intentions to appropriate par-
ticipants in a conversation. With context now playing a significant role, L modernizes
entailment as the ‘diagonal’ notion of preservation of endorsement in all contexts.
In section , L extends this language game, with declarative sentences used in
assertion to centralize a distributed stock of beliefs. The final stage language L adds
modal operators: these are affiliated with the theory of entailment, and accordingly
behave either epistemically, when operating on declaratives, or obligatively, when
operating on imperatives.
Stipulations for L
In the elementary language L , as well as in its extensions to follow, Boolean connec-
tives are semantically ‘classical’ in that they have Boolean set-theoretic operators as
their semantic values: as usual, ‘the semantic value of ξ ’ is abbreviated [[ξ ]]; ‘the seman-
tic value of ξ in context c’, as [[ξ ]]c .28 The non-classical Ross-Paradoxical behavior is
generated by assigning procedures as imperative semantic values, in combination with
a composition rule directing the action of a connective to the proposition-components
of its operanda and composing their partition-components by multiplication.

syntax
Lexicon: (i) a stock of elemental imperative sentences A, B, etc.; (ii) Boolean
connectives ∧, ∨, ¬.
Composition: if α , . . . , αn are imperative sentences and C is an n-place Boolean
connective, C(α , . . . , αn ) is an imperative sentence.

semantics
Ontology: Let PR be the set of all procedures; and for present purposes, let a
procedure be a granulated proposition: recall, a partition paired with a proposition

it discriminates (so PR := {π , p : π a partition of W ∧ (∃S ⊆ π )(p = S)});
for  = X, Y , let π() = X and p() = Y.
Lexicon: (i) When α is an elemental imperative, for some p ⊆ W, [[α]] = {p, p}, p ;
(ii) [[∧]] = ∩, [[∨]] = ∪, [[¬]] = λx x.29
Composition: [[C(α , . . . , αn )]] = π([[α ]]) × · · · × π([[αn ]]), [[C]](p([[α ]]), . . . ,
p([[αn ]])) .

28 The treatment of compositionality falls short of full rigor, but without I hope undermining my central
aims.
29 I am not actually so confident that negation can take wide scope over imperatives: ‘Fred don’t open
the door!’/‘Fred open the door and Brent take out the trash!’/?‘Don’t (Fred open the door and Brent take
out the trash)!’.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

rationalization and the ross paradox 

entailment
If β , . . . , βn , α are sentences of L , then β , . . . , βn  α just if [[β ]], . . . , [[βn ]]  [[α]].

discussion
It is easy to see that classical conjunction-introduction and -elimination, and negation-
and disjunction-elimination are valid, but negation- and disjunction-introduction are
invalid: L is, in effect, a projection into the syntactic of the procedure-contents of
intention, with an entailment relation isomorphizing content-entailment. (Example:
let [[A]] = {q, q}, q and [[B]] = {r, r}, r ; then [[A ∨ B]] = {q, q} × {r, r}, q ∪ r ;
the partition component of [[A ∨ B]] is then {q ∩ r, q ∩ r, q ∩ r, q ∩ r}  {q, q}, while
the proposition component is q ∪ r ⊇ q, so whenever q and r are orthogonal, neither
A  A ∨ B nor A ∨ B  A.)30
Stipulations for L
The imperatival language L extends L with subject–predicate structure and permits
Boolean connectives to take either wide (sentential) or narrow (predicate) scope.
This produces a ‘softer’ Ross-Paradoxical effect arising when disjunction takes narrow
scope.
My intent with this is to model the datum that imperative disjunction-introduction
can be coerced into sounding less objectionable. It takes some advance preparation.
First reflect on the excellence of the conjunction-elimination argument. Then high-
light that if something is a certain way, it is thereby any less specific way—so in partic-
ular, if someone undergoes a process of a certain type, they thereby undergo a process
of any more inclusive type. Then think of a disjunctive action-characterization as just
less specific than either disjunct action-characterization. And finally, coerce the dis-
junctiveness in the conclusion into the disjunctiveness of its action-characterization,
as follows: ‘Fred post this letter!—So: Fred (either-post-this-letter-or-take-out-the-
trash)!’. That sounds better than before, if not yet strikingly valid.31
The predicates of L pertain to action-kinds. Actions are sustained processes; an
action-kind groups together a bunch of processes under a certain respect of similarity.
In the modal spirit, a property of individuals is a set of ‘centered worlds’ aka world–
time–individual triples: red is the set containing w, t, j just if, in w, at t, j is red. A
similarity-class of processes could then be thought to determine a set of ‘extended
centered worlds’ aka world–interval–individual triples: running around the block is
the set containing w, (t, t ), j just if, in w, at t, j commences a run around the block
which concludes at t . But for the purposes of the speech act theory to be introduced

30 Starr’s otherwise congenial (if intricate) approach (Starr, , § .) unfortunately predicts that it is
always preferable to carry out both disjuncts (when compatible) to just carrying out one (Charlow, ,
pp. –).
31 Starr (, pp. –) maintains that disjunction takes scope either over or under the imperatival mood
operator !, and takes pains to secure the equivalence of the updates associated with both.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 benj hellie

for language L , it is only the onset of the action that will matter: the impact of our
imperative language will paper over matters of practical reasoning and the structuring
of actions for which this matters; the issue of interest will be setting each other to action
through issuing commands, for which it is only the onset that matters.32

syntax
Lexicon: (i) a stock of elemental action-predicates F, G, etc.; (ii) a stock of names a,
b, etc.; (iii) Boolean connectives ∧, ∨, ¬.
Composition: (i) if  , . . . , n are action-predicates and C is an n-place Boolean
connective, C(  , . . . , n ) is an action-predicate; (ii) if ν is a name and is an
action-predicate, ν is an elemental imperative sentence; (iii) if α , . . . , αn are
imperative sentences and C is an n-place Boolean connective, C(α , . . . , αn ) is an
imperative sentence.33

semantics
Ontology: Let W, T, and J be (respectively) the sets of worlds, times, and individuals.
Let A be the set of all action-kinds, where an action-kind determines a set of world–
time–individual triples, with which I will identify it: where K is an action-kind, I will
notate {w : w, t∗ , j∗ ∈ K} as K(t ∗ , j∗ ).
Let a context c determine a moment of time tc .
Lexicon: (i) For an elemental action-predicate , [[ ]]c ∈ A; (ii) for a name ν, [[ν]]c ∈
J; (iii) [[∧]]c = ∩, [[∨]]c = ∪, [[¬]]c = λx x.
Composition: (i) [[C(  , . . . , n )]]c = [[C]]c ([[  ]]c , . . . , [[ n ]]c ); (ii) [[ν ]]c =
?([[ ]]c ([[ν]]c , tc )), [[ ]]c ([[ν]]c , tc ) ; (iii) [[C(α , . . . , αn )]]c = π([[α ]]c ) × · · · ×
π([[αn ]]c ), [[C]]c (p([[α ]]c ), . . . , p([[αn ]]c )) .

entailment
If β , . . . , βn , α are sentences of L , then β , . . . , βn  α just if, for all c, [[β ]]c , . . . ,
[[βn ]]c  [[α]]c .

32 This actually raises an annoying side-issue: as the pragmatics for L is set up, commands occur in a

temporal sequence, and require immediate commencement. But they also put on display stable intentions
that are not in a position to be acted upon until speech acts occur. In natural language, that would be resolved
by tacit or explicit temporal adverbialization; or perhaps by looseness about what counts as the onset of
action (compare: ‘is there much to choose between “She is making tea” and [] “She is going to make tea”?
Obviously not’: Anscombe, , §).
33 For further details of the syntactic and semantic composition of imperatives, see Hellie (). In
eschewing an imperatival operator, I side with Portner (, , ); and I oppose Hare (, § .);
Kenny (); Stenius (); Chellas (); Lewis (b); Kaufmann (); Charlow (, ); and
Starr ().

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

rationalization and the ross paradox 

discussion
The scope of negation does not matter: [[¬(ν )]]c = [[ν(¬ )]]c . After all, [[¬(ν )]]c =
?([[ ]]c ([[ν]]c , tc )), [[¬]]c ([[ ]]c ([[ν]]c , tc )) = ?([[ ]]c ([[ν]]c , tc )), [[ ]]c ([[ν]]c , tc ) =
?([[¬]]c [[ ]]c ([[ν]]c , tc )), [[¬]]c [[ ]]c ([[ν]]c , tc ) = [[ν(¬ )]]c .
But the scope of disjunction does matter. Let p be the proposition that Fred opens
the door, q that he takes out the trash. The content of ‘Fred take out the trash!’, of
course, is ?q, q . And the content of ‘Fred (open-the-door-or-take-out-the-trash)!’
is ?(p ∪ q), p ∪ q , while that of ‘(Fred open the door) or (Fred take out the trash)!’ is
?p×?q, p ∪ q .
The sentential-scope disjunction raises more issues than the disjunct: it offers
Fred the option of success by opening the door in addition to by taking out the
trash. Consequently, the partition component is stronger than that of the disjunct. By
contrast, the predicate-scope disjunction offers Fred only one option, that of taking
the disjunctive action: the resulting partition is neither stronger nor weaker than that
of the simple imperative.
Perhaps we reason about the relative strengths of natural-language imperatives α
and β by comparing the relative strengths of their partitions and of their propositions.
If so, we might perceive the predicate-scope Ross-inference as ‘more valid’ than the
sentential-scope: with predicate scope, the premiss proposition is stronger than the
conclusion, the partitions orthogonal; with sentential scope, the premiss proposition
is stronger, the conclusion partition is stronger. With sentential scope, respects of
strength are balanced between premiss and conclusion; with predicate scope, by either
measure, whenever their strength is comparable, the premiss is stronger than the
conclusion.

 Command Pragmatics
Next, the language L assembles a command-‘language game’ around L -syntax/
semantics. Conversations in L are centered on the intentions held in a ‘general will’
of the assembled conversationalists—which I understand as to be a group mental state
in a fully robust sense (Rovane, ; Hellie, ), in many salient respects just like
individual intention.34 (Why? If canonical episodes of command require some sort of
‘group project’, it would be peculiar for non-Catholics to issue commands to the Pope
(compare Ninan, , § .); involving a collective mental state permits reusing struc-
tures for individual practical rationality. A more Lewis (b)/Portner ()-like
approach, on which commands merely constrain without rationalizing, must also find
some alternative explanation of the imperative Ross Paradox.) Conversation begins in
common ignorance of how the general will intentions obligate the conversationalists,
and moves toward total common recognition of the obligations of each; the acts of

34 For differences, see fn. .

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 benj hellie

command in L serve to distribute this common recognition (I set to the side the
epistemological question of how anyone knows what the general will intends; note also
that L stripped of its speech act theory can represent individual practical calculation).
In mechanical terms, L extends L with a statics and a dynamics: the statics
postulates contexts to record the general will intentions, the intentions of each con-
versationalist, and the extent of common knowledge of the distribution of general will
intentions;35 the dynamics sets the start- and end-states of the language game and the
effect of speech-acts on the state of play.36
In its postulation of psychologically-laden contexts, the statics of L permits a defi-
nition of endorsement: implicit acceptance of a sentence by a mental state (individual
or group), represented by positive evaluation at a context of the content of the sentence,
interpreted at that context.37 Armed with endorsement, entailment can be defined as
endorsement-preservation: any context endorsing all premisses endorses the conclu-
sion.38 Despite this, the entailment-relation of L does not differ extensionally from
that of L : that happens only in L , with the introduction of intensional operators to
shift the context at which a sentence is evaluated for endorsement.
The objective is to illustrate these points: how the contents of commands and of
intentions can be identified; how the meaning of a command can be obligative, in
the sense of binding its recipient to the course of action commanded; and, more
programmatically, what is required of the relations among content, mental states,
speech act meaning, and entailment.
Stipulations for L
statics
The central device of statics is the set C of all contexts. The approach uses a particular
context c ∈ C to represent a mental state relevant to language use. Those include
the mental states of individual language-users, of course, but also collective mental
states of groups of individuals gathered in conversation. A context representing an
individual I call solo; a context representing a conversational group I call social. (I
allow that contexts may represent not only ‘sincerely’ held mental states, but also
other states ‘treated as’ held—as for hypothesis, fiction, contingency planning, and
so forth: Stalnaker, , p. . I henceforth pass over this qualification without
acknowledgment.)

35 In this mentalistic conception of context, I side with Stalnaker (, ), Lewis (a), and Veltman
(, p. ), and oppose Kaplan () and Lewis (a); compare Hellie () and, to a considerable
extent, Yalcin ().
36 My story here is a spin on the familiar Stalnakean pragmatics (Stalnaker, , ).
37 In this ‘diagonal’ conception of entailment, I side with Kaplan () and against Stalnaker (),
Veltman (), and Yalcin ().
38 In thinking of entailment as endorsement-preservation, I side with Hare (), Veltman (), and
Yalcin (), and against Stalnaker () and Kaplan ().

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

rationalization and the ross paradox 

A mental state needs to know whose it is, and when—matters of particular signifi-
cance for agency, because if I am uncertain whether I am BH in Toronto as I write this
or Julius Caesar just before his crossing the Rubicon, I will have no idea whether to
type or command my army (and I certainly have no idea how to split the difference!).39
Accordingly, I continue to let c determine its time tc . I also require c to determine a
nonempty set of individual conversationalists Vc ⊆ J: c is solo just if Vc is a singleton
set, otherwise social. For solo c, Vc is the singleton of the agent whose state of mind is
represented by c; for social c, Vc represents the members assembled in conversation as
represented by c.
In order to link imperative meaning to intentions, a context must bear a representa-
tion of the content of someone’s or some group’s intention. Accordingly, let c determine
a procedure c ∈ PR: for solo c with Vc = {j}, c is the procedural content of the
intentions of j at tc . For social c, c represents the obligations imposed on one another
by the ‘general will’ of the assembled members of the conversation (Castañeda, ,
pp. –).
I next insist that, for social c, for any j ∈ Vc , there is a solo context cj such that
Vcj = {j}. After all, j can only converse in an imperative language-game if it can place
itself under obligations (by having intentions, and thereby being representable by a
device cj representing j as itself and the content of its intentions as c j ); I impose the
requirement that cj be solo to exclude the bizarre prospect that a conversation is a
conversationalist in another conversation. And next, I require that tc = tc j : at any
time, it is the same time for everyone.
The purpose of assembling in L -conversation is to parcel out the general will. If
that is what we are doing, we need a sense for how much of it we have parceled out.
To this end, I require a social context c to determine a to-do list function tdc assigning
a procedure to each member of the conversation (tdc : Vc → PR):40 tdc represents
the common recognition among the conversationalists of how much of the general
will has been parceled out, and to whom. In order for tdc to parcel out the general
will, I require that it assign no procedure to anyone unless the general will intends that
procedure to be carried out by them (for j ∈ Vc , c  tdc (j)). And in order for a to-do
list to reflect the parceling out of the general will—for it to be commonly recognized
who is under an obligation to do what—it needs to be that it accurately reflects what
the intentions of the assembled conversationalists actually are (c j  tdc (j)).

entailment
Sentential-entailment in L is grounded in its statics: the relation at the definitional
basis L -entailment, (content)-endorsement, holds between a context c (a creature of

39 In requiring this ‘immunity to error through misidentification’ I side with Stalnaker () and against
Kaplan () and Lewis (a).
40 I borrow the ‘to-do list’ terminology from Portner (, , ) (my conception of these objects
adjusts Portner’s to meet with my apparatus and doctrine, as needed, and without comment). Portner
acknowledges Han () and Potts () as precursors; and of course (Charlow, , § .–) the to-do
list is a generalization of Lewis’s single-subject ‘sphere of permissability’ (Lewis, b).

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 benj hellie

the statics) and a content , in virtue of the content-entailment of  by a parameter

of c (namely c ). In formal terms: c   := c  . Because contexts represent
mental states, this makes endorsement psychological; because endorsement is funda-
mental to entailment, it in turn makes entailment fundamentally psychological, and
excises the linguistic (including speech-acts) from the basis of entailment.41
Sentential entailment, of course, is a relation between sentences. L -entailment
insists on the fundamental role of the psychological by bringing language in as
a product of content-endorsement and semantic-valuation. In formal terms: c c-
endorses α (c c α) := c  [[α]]c .
This relation, c-endorsement, is a context-relative binary relation between a context
and a sentence. Contexts, therefore, do two jobs when c c-endorses α: there is the
job of c, which I call interpretation; and there is the job of c , which I call evaluation.
The interpreting context c is in charge of saturating context-dependent parameters
in the semantic-valuation clause for α (for example, c turns the sentence ‘You pay the
rent!’ to the procedure committing the addressee of c to set about paying the rent at the
moment of c). The evaluating context is in charge of assessing whether that interpreted
semantic value is content-endorsed (for example, whether it is intended by the general
will that, at the moment of c, the addressee of c set about paying the rent).
This dual role of context sets up a ‘two-dimensional intension’ for a sentence (the
horizontal is the evaluating axis, the vertical the interpreting axis, the ‘extensions’
are endorsement- rather than truth-values): I follow the two-dimensionalist tradition
(Kaplan, ; Chalmers, ) in affiliating rational power with the diagonal. The
diagonalization of c-endorsement, endorsement (simpliciter), is c-endorsement by c:
c  α just if c c α.
Finally, (sentential-)entailment is endorsement-preservation: universal endorsement
of the conclusion conditional on endorsement of all premisses. In formal terms:
β , . . . , βn  α just if whenever c  β , . . . , c  βn , then c  α. Collecting together:
just if whenever the general will-intention is sufficient to endorse the premisses, it is
sufficient to endorse the conclusion.

dynamics
The central device of dynamics is a set of events E, the set of all speech acts. I have
troubled to tease apart the speech acts from the contexts—the former are linguistic,
the latter mental—in order to emphasize the purification of sentential-entailment of
any linguistic componentry. In this respect, the approach contrasts with ‘dynamic
semantics’.
The dynamics pictures a temporally-ordered sequence of speech acts—a
conversation, conv = e , . . . , eω —with each ei ∈ E. A speech act e determines a
moment of time t(e) at which it occurs: the temporal sequencing of conv means that
t(ei ) < t(ei+ ).

41 In this latter respect I side with Yalcin () and against ‘dynamists’ (Veltman, ; Starr, ).

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

rationalization and the ross paradox 

A speech act determines an individual s(e), its speaker, by whom it is performed.

In L , a speech act also determines a distinct addressee, a(e) = s(e), at whom it is
directed. A speech act of L can occur only inside of circumstances in which, at its
time, its speaker and addressee are assembled for conversation. That requires a social
context in which speaker and addressee are conversants. Accordingly, e determines a
social context c(e): the time of the context and speech act are the same (t(e) = tc(e) );
and the addressee and speaker are conversationalists in c(e) ({s(e), a(e)} ⊆ Vc(e) ).42 I
assume that conversationalists do not come and go: V(conv) := V(cei ) = V(cei+ ).
In a speech act, a sentence is spoken: accordingly, e determines a sentence σ (e).
The purpose of the speech act is to adjust its context c(e)—the prior context—into
a posterior context cc(e). In L , perhaps sanding down certain intricacies of natural
language, the adjustment occurs in line with the conventional force of σ (e) and with
the content of σ (e) interpreted against the prior context. (I assume that it is not the
role of speech acts to adjust the general will: c(e) = cc(e) .)
The aim of an L -speech-act is to adjust the to-do list in a way that furthers
the ‘aim’ of the conversation. The aim of the conversation is characterized by its
commencement- and conclusion-states, as reflected in its initial prior and final poste-
rior contexts. Conversation commences absent any common recognition of the obliga-
tions placed by the general will on any conversationalist: the to-do list for each member
according to the prior context of the initial speech act e assigns the trivial procedure
 (more formally, for each j ∈ V(conv), tdc(e ) (j) = ). Conversation concludes in
a state in which there is nothing in the general will remaining to be parceled out to

some appropriate conversationalist (more formally, j∈V(conv) tdcc(eω ) (j) = cc(eω ) ).
In that sense, the aim of the conversation as a whole is to spread the whole general will
over the to-do list; an individual speech act contributes to that aim by moving a bit of
the general will to the to-do list. That makes these speech acts commands.43
In its mechanics, a command is subject to the constraints of relevance and legitimacy.
A command is relevant just if its content pertains to its addressee (so for any e, if
σ (e) = ν , [[ν]]c(e) = a(e): compare Portner, , p. ); a command is legitimate
just if endorsed by the general will (so for any e, c(e)  σ (e): compare Castañeda, ;
contrast Portner, , § .). The effect of e is then to determine a posterior context
cc(e) adjusting the prior context c(e) just by adding the content of σ (e) to the to-do
list of the addressee (tdcc(e) (a(e)) = tdc(e) ∧ [[σ (e)]]c(e) : compare Portner, , p.
—but Portner (, p. ) then goes on (to my mind, incongruously) to explain
this force as an effect of ‘pragmatic reasoning’).44

42 Speech acts therefore require contexts, and not the other way around. In this sense, contexts are
essentially psychological, and when associated with speech acts, are only accidentally so: compare Chalmers
() against ‘contextual two-dimensionalism’; contrast Stalnaker (, passim); Kaplan (); and Lewis
(a). This wiggle-room introduces the prospect that meanings of personal pronouns are grounded in
dynamic strategies for resolution of anaphora rather than static rules of interpretation.
43 Imperatives can be used for a laundry list of speech acts other than commands (Kaufmann, ;
compare Portner, , , Starr, , § ., Charlow, , § .). Because of space, I set this aside.
44 See fn.  for a complication.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 benj hellie

If contexts evolve between speech acts, they do so in the direction of strength:

c(ei+ )  cc(ei )—the posterior context of the leading speech act is entailed by the prior
context of the trailing speech act. What is entailment between contexts? The objective
is to ensure that what speech acts do does not get undone: if my command or assertion
is accepted, behind-the-scenes manoeuvring won’t later ‘de-accept’ it (Charlow, ,
p. ). So in particular, whenever xc is a parameter of context adjustable by speech
acts, c  c only if xc  xc  .
This requirement is weak in two ways. First, it says nothing about parameters that
cannot be adjusted by speech acts. Second, it does not require identity between speech-
act parameters of leading posterior and trailing prior contexts. The latter weakness
leaves room for the important phenomenon of accommodation (Lewis, c): roughly,
the natural-language practice of ‘working outside of the rules’ to permit speech acts to
succeed. For example, ‘Pick up Sam’s sister at the airport!’ presupposes that Sam has a
sister (compare Stalnaker, , III, IV): a command using that sentence fails unless
that presupposition is common belief in the prior context of its command. But if the
prior-context common beliefs need not just be the leading posterior-context common
beliefs, the conversationalists may strengthen the latter with the presupposition so as
to ready the former for updating with the command.45
A final point: truth is nowhere to be seen in L .46 Imperative contents cannot
be truth-conditions: they are too finely-grained for that. So truth is nowhere in
the semantics. The entailment-relation on L is, as I have argued, fundamentally
and ineliminably psychological. While entailment is preservation of something-or-
other, it is not truth that is preserved, but the psychological relation of endorsement.
So truth is nowhere in the statics. Finally, the aim and practice of command is of
distributing the general will by putting it on display. There is nothing like a ‘truth-
norm’ for command—nor any recasting of a truth-norm in other terms, such as
a ‘goodness-norm’ (the good may suffuse should and oughta, but imperatives and
intentions are rather less permeable: contrast Castañeda, ). So truth is nowhere
in the dynamics.47

45 Another use for this looseness is negotiation of the ‘general will’ (compare Lewis, c, example 
on the collective negotiation of a plan): the constraint that c(e) = cc(e) requires merely that there is no
speech act essentially for adjusting the general will. That is compatible with the prospect of adjusting the
general will ‘between speech acts’.
46 Do I thereby abandon the ‘utility of the truth-conditional paradigm’ (Charlow, , p. )—
namely, the ‘standard’ notion of entailment as incompatibility of the truth-condition of the negation of the
conclusion with the truth-conditions of the premisses, and the ‘familiar’ conception of compositionality as
requiring the contribution of the truth-condition of an embedded sentence to the meaning of its embedding
sentence? I do not feel discomfitingly anxious. What is actually the ‘standard’ (Kaplan, ) notion of
entailment is truth of the conclusion at the index of the context whenever the premisses are true at the
index of the context; as noted above, endorsement does as well in this role as truth. And needless to say,
objects other than truth-conditions can embed.
47 Might truth still be lurking at the bottom, in the contents of intentions, by (as suggested in the early
sections) grounding success of intentions in truth of their contents? But with intelligibility in hand as a
norm, I suspect truth can be stripped even of that job, and content can be given a purely phenomenological
reduction. Further discussion must await another forum.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

rationalization and the ross paradox 

 Obligative Modals
Extending L with modals requires not just new semantic clauses for modals, but an
extension of the pragmatics. Modal sentences are declarative; accordingly, sticking
with the endorsement-preservation conception of entailment requires a declarative
statics; moreover, the use of obligative ‘must’-claims to command is among the
explananda, requiring a declarative dynamics. The language L extends L with a
declarative pragmatics; L , at last, adds modals.
The move to L is easy, because the general form of a declarative pragmatics
is relatively clear: each asserts propositions they believe to enrich common belief.
Endorsement is belief; the ‘essential effect’ of assertion intersects the proposition
asserted with the ‘context set’. While the rigors of segregating statics and dynamics
call for a few subtle adjustments, the pragmatics is in the main orthodox.
The semantics for modals presented in L is less so. The founding datum of this
chapter, recall, is that one thinks one must do something just if one acts with the
intention of doing it. Or, expressed in the jargon, a solo context (sentence-)endorses
‘I must pay the rent.’ just if it (content-)endorses the procedure pay the rent.48
The orthodoxy says that modals are intensional operators, comparing their
operanda to aggregations of those things at which intensions are evaluated, what-
ever they may be. It says also that modals quantify over worlds, comparing their
operanda to aggregations of worlds. The orthodoxy can say both things because it
thinks intensions are evaluated at worlds. But I think entailment is endorsement-
preservation, intensions are evaluated at representations of mental states—what I have
been calling contexts—so I have to choose. (Revanchists attempting to avoid this choice
by retreating from endorsement-preservation must find some alternative approach to
the imperative Ross Paradox.)
One option is to break the contact between modals and intensions but preserve
contact between modals and worlds:49 on this approach, ‘I must pay the rent.’ is
sensitive to whether, in all worlds that have been somehow selected by the context,
I pay the rent. Selected how? Perhaps they are those where I do as I intend. Suppose I
intend to pay the rent; then in all worlds where I do as I intend, I pay the rent; then in
all worlds selected by the context, I pay the rent. And that is what it takes for ‘I must
pay the rent.’ to be ‘good’ (true, or acceptable, or what have you). Unfortunately, in
all worlds where I pay the rent, I either pay the rent or drink up all your wine. So if
I intend to pay the rent, ‘I must either pay the rent or drink up all your wine.’ is also
‘good’. So if I intend to pay the rent, I think I must either pay the rent or drink up all
your wine. But I don’t! The Ross Paradox works for obligative modals, too.

48 See fn. .

49 Compare Yalcin (): intensions are evaluated at information states: ‘φc is accepted in information
state s iff for all worlds w in s, [[φ]]c,s,w is true’; ‘ i φ just in case for every context c and body of information
s, if every member of c is accepted in s, then φ is accepted in s’ (p. ). But modals quantify over worlds:

‘[[φ]]c,s,w is true iff ∀w ∈ s : [[φ]]c,s,w is true’ (p. ).

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 benj hellie

The other option holds out more promise of a treatment of the Ross Paradox for
must: break the contact between modals and worlds but preserve contact between
modals and intensions. On this option, modals are intensional operators, comparing
their operanda to aggregations of contexts, looking for patterns of endorsement. What
makes a reading of ‘must’ obligative is its use to compare its procedural operand to
the intentions of the context at which it is interpreted; with the consequence that
a solo context endorses obligative ‘I must pay the rent.’ just if it endorses pay the
rent and ‘I must pay the rent or go drinking.’ just if it endorses pay the rent or go
drinking.

Stipulations for L
syntax
Lexicon: A stock of elemental declarative sentences P, Q, and so forth.
Composition: If φ , . . . , φn are declarative sentences and C is an n-place Boolean
connective, C(φ , . . . , φn ) is a declarative sentence.

semantics
Lexicon: For an elemental declarative sentence φ, [[φ]]c ⊆ W.
Composition: [[C(φ , . . . , φn )]]c = [[C]]c ([[φ ]]c , . . . , [[φn ]]c ).

statics
Let a context c ∈ C determine an information state, a nonempty set ic ⊆ W (compare
Stalnaker, , pp. –; also (inter alia) Veltman, ; Yalcin, ). For solo c,
ic represents the content of the beliefs of the subject c represents; for social c, ic
represents the common beliefs of the participants to the conversation c represents (the
strongest proposition each believes each believes . . . each believes). The nonemptiness
requirement on ic stipulates that no (nondefective) belief state can be inconsistent. A
union-set requirement on social contexts stipulates that nothing can count as common

belief unless it is severally believed: if c is social, then j∈Vc icj ⊆ ic .

entailment
To bring in declarative entailment, it suffices to add a ‘module’ for content-
endorsement of declarative content: namely, contextual endorsement of a proposition
is believing it (c  p := ic  p).

dynamics
As with L , a conversation in L is a temporally ordered sequence of speech acts
conv = e , . . . , eω . Conversation progresses from complete common ignorance to
complete common belief in that which is not controversial (more formally: ic(e ) = W

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

rationalization and the ross paradox 

 
and c(e ) = ; icc(eω ) = j∈V(conv) icj and cc(eω ) = j∈V(conv) tdcc(eω ) (j)).
By contrast with command, which aims at parceling out content which is already
common to us, assertion aims at gathering up content which had been scattered among
us. This opposite polarization of assertion and command recapitulates the opposite
polarization of belief and intention: command is polarized from the social to the solo,
assertion is polarized from the solo to the social.
As with L , an L -speech act e ∈ E determines a sentence σ (e). If σ (e) is an
imperative sentence α, e is a command; if σ (e) is a declarative sentence φ, e is an
assertion. But the opposing polarities of command and assertion make for opposing
polarities in the appropriateness-conditions of the speech-act types. Commanding α
is appropriate when the content of α is intended by the general will, where this is
represented as endorsement by the social context. But when an assertion contains only
information already included in common belief, it is redundant, and will typically be
undesirable: asserting φ is generally inappropriate when the social context endorses
φ, appropriate when the speaker’s context endorses φ.
Accommodating this requires an extension of the apparatus. In addition to its prior
and posterior contexts c(e) and cc(e), e determines also an imputed context mc(e): the
imputed context of a speech act is, in effect, the context ‘authorizing’ the speech act,
or ‘in the voice of which’ the speech act is advanced. When e is a command, mc(e) is
typically c(e), the social context of the conversation; when e is an assertion, mc(e) is
typically cs(e) , the solo context of the speaker. Either way, it is required that the imputed
context does indeed endorse the sentence advanced in the speech act (mc(e)  σ (e)).
Finally, the update rule for assertion is the familiar intersective ‘essential effect’:
icc(e) = ic(e) ∩ [[φ]]mc(e) .

Stipulations for L
statics
If modals are intensional operators, they ‘quantify’ over whatever intensions are
evaluated at. If entailment is endorsement-preservation, intensions are evaluated at
contexts. So modals quantify over contexts. As on the orthodoxy, the quantification
is typically restricted. I represent the restriction with a set of contexts Cc determined
by c. As on the orthodoxy, restriction represents a limitation in which entities are ‘taken
seriously’ in c; those not meeting the restriction are ‘ignored’.
The orthodoxy typically imposes a requirement of realism (Kratzer, , p. ):
actuality is never ignored, always taken seriously. I impose a similar restriction: a
context never ignores itself, always takes itself seriously. Formally: c ∈ Cc .

syntax
Lexicon: Modal operators , .

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 benj hellie

Composition: If σ is a sentence and M is a modal operator, Mσ is a declarative

sentence.

semantics
Lexicon: [[]]c = λF {w : (∀c ∈ Cc )(c  F(c))}, [[]]c =
λF {w : (∀c ∈ Cc )(c  F(c))}.
These clauses represent a modal as an intensional operator, comparing the Cc -
restricted set of contexts to the content of its operand: for , whether all members of
Cc endorse that content; for , whether not all members of Cc antiendorse the content. I
return to ‘antiendorsement’ in the wrap-up discussion. What it is for a modal sentence
to be ‘good’ is for it to be endorsed, as interpreted, however evaluated: as a declarative
sentence, that is for it to be noncontingent—trivial if true, absurd if false.50

Composition: [[Mσ ]]c = [[M]]c (λc [[σ ]]c ).
. Obligative necessity
Every context takes itself seriously: c ∈ Cc . But under reflexive restriction, a context
takes only itself seriously: Cc = {c}. Let  and be unofficial symbols for  and 
under reflexive restriction: then [[]]c = λF {w : (∀c ∈ Cc = {c})(c  F(c))} =
λF {w : c  F(c)}; similarly, [[ ]]c = λF {w : c  F(c)}.
Accordingly, [[σ ]]c = {w : c  [[σ ]]c } = {w : c  σ } = W just if c  σ ,
otherwise ∅: from which it follows that c  σ just if c  σ , from which it follows
that σ  σ and σ  σ : any sentence is equivalent to its reflexively-restricted
necessitation (compare Yalcin, , p. ).
And accordingly, [[ σ ]]c = {w : c  [[σ ]]c }. Now, as it happens, for imperative
σ = α, [[α]]c is undefined, and with it [[ α]]c : I return to this point in the concluding
remarks. But for declarative σ = φ, [[ φ]]c = {w : c  ¬φ}. It follows that c  ¬ φ
just if c  ¬φ, from which it follows that ¬φ  ¬ φ and ¬ φ  ¬φ: the negation
of any declarative sentence is equivalent to the negation of its reflexively-restricted
possibilization (compare Yalcin, , pp. , ).
Because modal sentences are declaratives, they conform to disjunction-
introduction.51 In particular, α  α ∨ β. But because σ and σ are equivalent,
the Ross-Paradoxical nonentailment α  α ∨β results in a comparable nonentailment
α  (α ∨ β). (Compare: I must post this letter, so either I must post this letter or
I must drink up all your wine—sounds fine to me.) Of course φ  (φ ∨ ψ).
An assertion e of α can have a force resembling that of a command e of α.
How so? Let the imputed context mc(e) = c(e), the social context of the speech act

50 Because the completely ignorant belief state must still endorse a ‘good’ modal sentence, the content of
that sentence can be no more specific than the trivial proposition.
51 Not so disjunction-elimination. Read epistemically,  gives this counterexample: φ  φ, ¬φ 
¬φ,  φ ∨ ¬φ, but  φ ∨ ¬φ.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

rationalization and the ross paradox 

(while assertion is ordinarily imputed to the speaker, that is not a hard-and-fast rule:
in general, epistemic uses of ‘they might have already left’ to probe for compatibility
with common belief are imputed to the social context). Then, by stipulation, mc(e) =
c(e)  α. That is so just if c(e)  α. That is so just if c(e)  [[α]]c .
Now, by stipulation, c(e )  σ (e ) = α as well. But command has the conventional
force of introducing [[α]]c to the to-do list of its addressee, while assertion has no such
conventional force. To bridge the gap, observe with Stalnaker (, pp. –) that
common belief includes commonly salient information, in addition to the content
so far asserted. Common extraction of the content of a speech act requires common
recognition of the act’s occurrence as an act with a certain conventional meaning. So
typically this common recognition regarding a speech act will be incorporated into its
prior context (recall that a trailing prior context need only be at least as strong as its
leading posterior context).
So if the assertion e of α has been accepted, it is commonly recognized that the
social context has accepted a speech act requiring  c(e)  [[α]]c and in that sense
commonly recognized that the general will endorses α. But that has the same effect
as putting this endorsement on display, the conventional force of command. The
objective of the conversation conv is to parcel out the content of (conv) among the
conversationalists by putting the general will on display; they should not care whether
that is done by a speech act with that conventional force or via this more scenic route.

. Obligative possibility?

I will close by reflecting on the dual of obligative necessity, namely obligative possi-
bility. I consider an explanation for the difficulty in perceiving obligative readings of
natural language strong possibility modals. And I propose a resolution of David Lewis’s
‘problem about permission’ (Lewis, b).

on its remoteness
Suppose we let go this paperweight here. Fred says it must fall, Brent says it can/could
fail to fall. It is easy to access a reading on which they disagree, hard to access readings
on which they do not; it would be unnatural for Brent to rejoin %‘There are a lot of
things this object can/could fail to do that it must do—so what?’. Suppose we toss this
chemical into water. Fred says it can/could explode, Brent says it must fail to explode.
Again, it is easy to access a reading on which they disagree, hard to access readings on
which they do not; it would again be unnatural for Brent to rejoin %‘There are a lot
of things this chemical can/could do that it must fail to do—so what?’. Giving them
‘natural capacity’ readings, it is not easy to avoid coordinating must and can/could so
that they are dual to one another.
What to do about the rent? Brent says Rance must pay the rent, Fred says Rance
can/could fail to pay the rent. It is easy to access readings on which they do not
disagree; indeed, Brent quite sensibly rejoins ‘There are a lot of things Rance can/could

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 benj hellie

fail to do that he must do—so what?’. What to do about Brent’s wine? Brent says Rance
must not drink it all up, Fred says Rance can/could drink it all up. Again, it is easy
to access readings on which they do not disagree; again, Brent sensibly rejoins ‘There
are a lot of things Rance can/could do that he must not do—so what?’. Giving must its
obligative reading, it is easy to avoid coordinating can/could so that they are dual to
one another.52
Granting my core thesis, the obligative readings of can/could Fred hopes to access
would be reflexively restricted possibility-modalized imperatives: Brent says β, Fred
says ¬β; Brent says ¬α, Fred says α. In L , [[ ¬β]]c = {w : c  [[¬β]]c };
[[ α]]c = {w : c  [[α]]c }. What is the problem?
The problem is that, for a granulated proposition P, it is not straightforward to
define its complement P. Now, for a (protean) proposition p, its complement p is
the unique proposition for which the ⊆-maximally weak ‘trivial’ proposition  is the
strongest proposition weaker than both p and p and the ⊆-maximally strong ‘absurd’
proposition ⊥ is the weakest proposition stronger than both p and p.
Is there a comparable definition of partition-complementation? In the set of par-
titions of W, the -maximally weak partition  is the one-cell partition {W}, while
the -maximally strong partition  is the partition into singletons {{w} : w ∈ W}.
But a partition does not generally have a unique complement—namely, given π ,
there is sometimes more than one π  for which  is the strongest partition weaker
than both and  is the weakest partition stronger than both. Example: over the set
{A, B, C, A, B, C} let π be the ‘column’ partition {{A, A}, {B, B}, {C, C}}; let
ρ be the ‘row’ partition {{A, B, C}, {A, B, C}}; and let ρ  be the ‘skew’ partition
{{A, C}, {B, C}, {A, B}}. Then each of ρ and ρ  meets the conditions required of
a complement for π : for both of the pairs (π and ρ) and (π and ρ  ),  is the strongest
partition weaker than both and  the weakest partition stronger than both.
If partitions cannot be complemented, what sense can be made of the complement
of a granulated proposition? We manage to determine [[¬α]]c 53 by complementing
the proposition-component of [[α]]c . So perhaps while [[ α]]c = {w : c  [[α]]c }
is incomprehensible as {w : c  π([[α]]c ), p([[α]]c ) }, we can make sense of it as
{w : c  π([[α]]c ), p([[α]]c ) }. Still, the latter competes with the ‘natural capacity’
reading of ‘Rance can drink up Brent’s wine.’ as the modalized declarative (Rance
drinks up Brent’s wine). Because the proposition Rance drinks up Brent’s wine can be
complemented without digging under the hood, we ceteris paribus prefer the natural

52 Our parents would irritate us by answering ‘Can I be excused from the table?’ with ‘You can if you
may.’ . Our parents were correct that the issue was compatibility with a gradated standard of good manners.
We did not embrace those standards as our own, but instead felt them to have the force of natural bonds—
they imprisoned us at the table. Our irritation was at being not only imprisoned, but compelled to ritually
legitimize the imprisoning standards as a condition of freedom.
53 If we do: see note .

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

rationalization and the ross paradox 

capacity reading to the obligative reading. That is why it is easy to avoid coordinating
the obligative readings of must and can/could.

on lewis’s problem about permission

Lewis’s ‘problem about permission’ (Lewis, b) is this. Allow (against the view of
this chapter) that obligation and permission are dual necessity and possibility modals
over propositions: p is obligatory just when, for a ‘sphere of permissibility’ s, s ⊆ p; p is
permitted just when s  p. Suppose s obligates p, then the obligation is lifted, resulting
in s . What is s ? If s = s ∪ p, then everything compatible with the once forbidden
is now permitted—so retracting an obligation to carry rocks all day would issue a
permission to drink up all my wine. Canvasing various alternatives, Lewis concluded
none worked.
But if obligation is a mandate to settle the question ?p in favor of p, permission—the
retraction of obligation—is just the retraction of the mandate to settle the question.
The obligation had surely made the not-p worlds ‘prohibited’. But retracting the
obligation does not make the p worlds ‘permitted’—it just makes the not-p worlds
‘unprohibited’. The retraction just makes it no longer a matter of significance whether
p. And then the fact that in a possibility, not-p, does not make that possibility in itself
either prohibited or required, nor permitted. Instead, that fact is restored to its initial
evaluative neutrality—for a time a matter of concern, it subsides once again back into
indifference.

Acknowledgments
My great thanks to the editors, both for their sternness regarding matters of content and
their indulgence regarding matters of process, but most importantly for their wise, savvy, and
sympathetic guidance. Thanks also to Jessica Wilson, and to Luke Roelofs, Adam Murray,
Jonathan Payton, Sergio Tenenbaum, Andrew Sepielli, Bryan Reece, Peter Ludlow, Charlie
Cooper-Simpson, Chris Meyns, Maarten Steenhagen, and Steven Coyne.

References
Aloni, Maria () Free choice, modals, and imperatives. Natural Language Semantics. .
pp. –.
Anscombe, Gertrude Elizabeth Margaret () Intention. Cambridge, MA: Harvard University
Press.
Åqvist, Lennart () Choice-offering and alternative-presenting disjunctive commands. Anal-
ysis. . pp. –.
Austin, J. L. () How to talk: Some simple ways. Proceedings of the Aristotelian Society. .
pp. –.
Barker, Chris () Imperatives denote actions. In A. A. Guevara, A. Chernilovskaya, and
R. Nouwen (eds.) Proceedings of Sinn und Bedeutung . Volume I. pp. –.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 benj hellie

Beaney, Michael (ed.) () The Frege Reader. London: Blackwell.

Belnap, Nuel D. () The Logic of Questions and Answers. New Haven: Yale University Press.
Cariani, Fabrizio () ‘Ought’ and resolution semantics. Noûs. . pp. –.
Castañeda, Hector-Neri () Imperative reasonings. Philosophy and Phenomenological
Research. . pp. –.
Castañeda, Hector-Neri () Thinking and Doing: The Philosophical Foundations of Institu-
tions. Dordrecht: Reidel.
Chalmers, David J. () The foundations of two-dimensional semantics. In Manuel
García-Carpintero and Josep Macià (eds.) Two-Dimensional Semantics: Foundations and
Applications. Oxford: Oxford University Press.
Charlow, Nate () Imperative statics and dynamics. Manuscript. University of Michigan.
Charlow, Nate () Practical Language: Its Meaning and Use. Ph.D. thesis. University of
Michigan.
Charlow, Nate () Logic and semantics for imperatives. Journal of Philosophical Logic. .
pp. –.
Chellas, Brian F. () Imperatives. Theoria. . pp. –.
Chrisman, Matthew () ‘Ought’ and control. Australasian Journal of Philosophy. .
pp. –.
Chrisman, Matthew () The Meaning of ‘Ought’: Beyond Expressivism and Descriptivism in
Metaethics. Oxford: Oxford University Press.
Frege, Gottlob () Über Sinn und Bedeutung. Zeitschrift für Philosophie und philosophische
Kritik. . pp. –. Translated (), as ‘On sense and reference’, by Max Black; Reprinted
in Beaney, .
Frost, Kim () On the very idea of direction of fit. Philosophical Review. . pp. –.
Geach, Peter () Imperative and deontic logic. Analysis. . pp. –.
Gillies, Anthony S. () A new solution to Moore’s paradox. Philosophical Studies. .
pp. –.
Groenendijk, Jeroen and Stokhof, Martin () Semantic analysis of wh-complements. Linguis-
tics and Philosophy. . pp. –.
Groenendijk, Jeroen, and Stokhof, Martin () Questions. In Johan van Benthem and Alice
ter Meulen (eds.) Handbook of Logic and Language. Amsterdam: Elsevier.
Han, C. H. () The Structure and Interpretation of Imperatives: Mood and Force in Universal
Grammar. Ph.D. thesis. University of Pennsylvania.
Hansen, Jörg. () The paradoxes of deontic logic: Alive and kicking. Theoria. . pp. –.
Hare, R. M. () The Language of Morals. Oxford: Clarendon Press.
Harman, Gilbert () Rationality. In E. E. Smith and D. N. Osherson (eds.) Invitation to
Cognitive Science. Cambridge, MA: MIT Press. Reprinted in Harman, .
Harman, Gilbert () Reasoning, Meaning, and Mind. Oxford: Oxford University Press.
Heal, Jane () Mind, Reason, and Imagination. Cambridge: Cambridge University Press.
Hellie, Benj () There it is. Philosophical Issues . pp. –.
Hellie, Benj () Love in the time of cholera. In Berit Brogaard (ed.) Does Perception Have
Content? Oxford: Oxford University Press.
Hellie, Benj () Obligation and aspect. Inquiry.
Hintikka, Jaakko () Knowledge and Belief. Ithaca, NY: Cornell University Press.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

rationalization and the ross paradox 

Huber, Franz () The logic of theory assessment. Journal of Philosophical Logic. .
pp. –.
Humberstone, I. L. () Direction of fit. Mind. . pp. –.
Kamp, Georg () Logik und Deontik. Paderborn: Mentis.
Kamp, Hans () Free choice permission. Proceedings of the Aristotelian Society. .
pp. –.
Kaplan, David () Demonstratives. In Joseph Almog, John Perry, and Howard Wettstein
(eds.) Themes From Kaplan. () Oxford: Oxford University Press.
Kaufmann, Magdalena () Interpreting Imperatives. Berlin: Springer Science & Business
Media.
Kenny, A. J. () Practical inference. Analysis. . pp. –.
Kratzer, Angelika () What ‘must’ and ‘can’ must and can mean. Linguistics and Philosophy.
. pp. –.
Kratzer, Angelika () The notional category of modality. In H. J. Eikmeyer and H. Reiser
(eds.) Words, Worlds, and Contexts. Berlin and New York: de Gruyter.
Kratzer, Angelika () Modality. In Arnim von Stechow and Dieter Wunderlich (eds.) Hand-
buch Semantik/Handbook Semantics. Berlin: de Gruyter.
Lascarides, Alex and Asher, Nicholas () Imperatives in dialogue. Pragmatics and Beyond
New Series. . pp. –.
Lewis, David () General semantics. Synthese. . pp. –.
Lewis, David () Counterfactuals. Cambridge, MA: Harvard University Press.
Lewis, David () Adverbs of quantification. In Edward L. Keenan (ed.) Formal Semantics of
Natural Language. Cambridge: Cambridge University Press. Reprinted in Lewis, .
Lewis, David (a) Attitudes de dicto and de se. The Philosophical Review. . pp. –.
Reprinted in Lewis, .
Lewis, David (b) A problem about permission. In Essays in Honour of Jaakko Hintikka on the
Occasion of His Fiftieth Birthday on January , . pp. –. Dordrecht: Reidel. Reprinted
in Lewis, .
Lewis, David (c) Scorekeeping in a language game. Journal of Philosophical Logic. .
pp. –.
Lewis, David (a) Index, context, and content. In Stig Kanger and Sven Öhman (eds.)
Philosophy and Grammar. Dordrecht: Reidel. Reprinted in Lewis, .
Lewis, David (b) A subjectivist’s guide to objective chance. In Richard C. Jeffrey (ed.) Studies
in Inductive Logic and Probability. Volume II. Berkeley: University of California Press.
Lewis, David () Philosophical Papers. Volume I. Oxford: Oxford University Press.
Lewis, David () On the Plurality of Worlds. London: Blackwell.
Lewis, David () Papers in Philosophical Logic. Cambridge: Cambridge University Press.
Lewis, David (a) Papers in Metaphysics and Epistemology. Cambridge: Cambridge Univer-
sity Press.
Lewis, David (b) Why conditionalize? In Lewis, a. Handout from .
Lewis, David () Papers in Ethics and Social Philosophy. Cambridge: Cambridge University
Press.
Ninan, Dilip () Two puzzles about deontic necessity. MIT Working Papers in Linguistics. .
pp. –.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 benj hellie

Parfit, Derek A. () Reasons and Persons. Oxford: Oxford University Press.
Parsons, Josh () Command and consequence. Philosophical Studies. . pp. –.
Portner, Paul () The semantics of imperatives within a theory of clause types. In R. B. Young
(ed.) Proceedings of SALT . pp. –.
Portner, Paul () Imperatives and modals. Natural Language Semantics. . pp. –.
Portner, Paul () Permission and choice. In G. Grewendorf and T. Zimmermann (eds.)
Discourse and Grammar. Berlin: Mouton de Gruyter.
Potts, Chris () Keeping world and will apart: A discourse-based semantics for imperatives.
Handout: Talk at NYU,  October.
Rescher, Nicholas and Robison, John () Can one infer commands from commands?
Analysis. . pp. –.
Ross, Alf () Imperatives and logic. Philosophy of Science. . pp. –.
Rovane, Carol () Group agency and individualism. Erkenntnis. . pp. –.
Schroeder, Mark () Ought, agents, and actions. The Philosophical Review. . pp. –.
Searle, John R. () The intentionality of intention and action. Inquiry. . pp. –.
Setiya, Kieran () Reasons Without Rationalism. Princeton: Princeton University Press.
Sloman, Aaron () ‘Ought’ and ‘better’. Mind. . pp. –.
Smith, Michael () The Humean theory of motivation. Mind. . pp. –.
Spohn, Wolfgang () Ordinal conditional functions: A dynamic theory of epistemic states.
In William L. Harper and Bryan Skyrms (eds.) Causation in Decision, Belief Change, and
Statistics. Volume II. Dordrecht: Kluwer.
Stalnaker, Robert C. () Pragmatics. Synthese. . Reprinted in Stalnaker, .
Stalnaker, Robert C. () Indicative conditionals. Philosophia. . Reprinted in Stalnaker, .
Stalnaker, Robert C. () Assertion. Syntax and Semantics. . pp. –. Reprinted in
Stalnaker, .
Stalnaker, Robert C. () Inquiry. Cambridge, MA: MIT Press.
Stalnaker, Robert C. () On the representation of context. Journal of Logic, Language, and
Information. . Reprinted in Stalnaker, .
Stalnaker, Robert C. () Context and Content. Oxford: Oxford University Press.
Stalnaker, Robert C. () Our Knowledge of the Internal World. Oxford: Oxford University
Press.
Starr, William () A preference semantics for imperatives. Semantics and Pragmatics.
[Online] Available from: williamstarr.net/research/ a-preference_semantics-for-
imperatives.pdf. [Accessed  Nov .]
Stenius, Erik () Mood and language-game. Synthese. . pp. –.
Thompson, Michael () Naive action theory. In Life and Action. Cambridge, MA: Harvard
University Press.
Velleman, David () The Possibility of Practical Reason. Oxford: Oxford University Press.
Veltman, Frank () Defaults in update semantics. Journal of Philosophical Logic. .
pp. –.
Veltman, Frank () Or else, what? Imperatives on the borderline of semantics and
pragmatics. Lego Seminar Presentation, Institute for Logic, Language, and Computation.
 November. [Online] Available from: https://staff.fnwi.uva.nl/u.endriss/teaching/lolaco/
/slides/veltman.pdf [Accessed  Nov .]
Vranas, Peter () New foundations for imperative logic I: Logical connectives, consistency,
and quantifiers. Noûs. . pp. –.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

rationalization and the ross paradox 

Wedgwood, Ralph () The aim of belief. Philosophical Perspectives. . pp. –.
Willer, Malte () Modality in Flux. Ph.D. thesis. University of Texas at Austin.
Williams, Bernard () Imperative inference. Analysis. . pp. –.
Yalcin, Seth () Epistemic modals. Mind. . pp. –.
Yli-Vakkuri, Juhani () Comments on Hellie, ‘Actuality and triviality in mindset semantics’ .
Personal communication,  May.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i


Dynamic Foundations
for Deontic Logic
Malte Willer

 The Plot
Several authors—most notably Horty (, , , , ) but also Asher and
Bonevac (, ), Belzer (), Bonevac (), McCarty (), Nute (), and
Ryu and Lee ()—have observed that the tools and techniques developed within the
field of nonmonotonic logic provide a fruitful framework for the theoretical study of
deontic discourse and reasoning.1 The classical motivation for this approach is the idea
that certain obligations are only prima facie obligations and may thus be overridden
by other considerations (see Ross, ). Here is a familiar case. Jones has promised
to meet Mary for lunch and so ought to meet her for lunch. But it may very well be
that Jones finds himself in a situation in which he would need to miss lunch to save
a drowning child, and then he is permitted, arguably even ought, to miss lunch and
instead save the drowning child. Such situations, while perfectly ordinary, pose trouble
for the standard monotonic approach to deontic reasoning.2
To see what the issue is, assume that the logical consequence relation is monotonic
in the following sense:
Monotonicity: If φ , . . . , φn  ψ, then φ , . . . , φn , φn+  ψ
The initial observation is that ()–() are consistent:
() If Jones has promised to meet Mary for lunch, he ought to meet Mary for lunch.
() If a drowning child needs Jones’s help, then he ought to help the drowning
child.
1 Some classical frameworks in the nonmonotonic tradition: model-preference theories (beginning
with McCarthy’s () theory of circumscription); fixed-point theories (see, e.g., McDermott and Doyle,
 and Reiter, ); logics for argument-based defeasible reasoning (see Nute, , Pollock, , and
Touretzky,  for seminal discussions).
2 Classical deontic logic goes back to the foundational work in von Wright (, ). See Åqvist ()
and McNamara () for overviews.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

dynamic foundations for deontic logic 

() Jones has promised to meet Mary for lunch.

() A drowning child needs Jones’s help.
Here () and () are instances of the general principles that one ought to keep one’s
promises and that one ought to help children in need, respectively. Clearly, ()–()
entail () and ():
() Jones ought to meet Mary for lunch.
() Jones ought to help the drowning child.
This is as it should be, yet monotonicity requires that if ()–() entail () and (), so
does the set of premises consisting of ()–() and ():
() Jones would need to miss lunch with Mary to save the drowning child.
And that leaves no room for the intuition that () relieves Jones from the obligation
to meet Mary for lunch in the sense that breaking his promise is perfectly permissible
(perhaps even required) if this is what it takes to save the drowning child. One may, of
course, say that we should have never subscribed to () in the first place since, after
all, () might come out true. But then it would be just as mistaken to rely on ()
in concluding that Jones ought to save the drowning child—this obligation may be
overridden in certain cases as well—and in general it would be hard to see how prima
facie obligations could play any interesting role in deontic reasoning.
Those sympathetic to Harman’s () strict distinction between logic and reasoned
change in view may insist that the issue under consideration has nothing to do with
principles of semantic entailment and is better left to principles about how to revise
one’s beliefs in discourse and reasoning. Here I will mostly focus on what form a
semantic explanation of the nonmonotonicity of deontic discourse and reasoning
could take, leaving a discussion of whether it should receive such an explanation to
the concluding section. But that the overridability of prima facie obligations is of some
importance for semantic theorizing should be uncontroversial since it underwrites the
consistency of () and (), which (by everyone’s agreement) any satisfying semantics
for ifs and oughts must explain:
() If Jones has promised to meet Mary for lunch but needs to break his promise
to save a drowning child, he may/ought to miss lunch with Mary.
Accounting for this fact, it turns out, motivates a departure from the classical analysis
of modals and conditionals—one that relies on a proper model of how defeasible
norms interact with what is taken for granted in discourse and reasoning to determine
what is deontically ideal. Once this is done it is only a small step from the observation
that conditionals fail to be monotonic in the antecedent (resist antecedent strengthen-
ing) to a nonmonotonic semantic consequence relation.3

3 Charlow (a) argues that deontic conditionals such as “If you want to go to Harlem, you ought to
take the A-train” exhibit a resistance to antecedent strengthening that is grounded in the goal sensitivity of

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 malte willer

Previous attempts to account for the nonmonotonicity of deontic discourse and

reasoning draw their principal inspiration from the striking parallel between prima
facie obligations and defeasible generalizations: the former may be overridden while
the latter may be defeated.4 For instance, the assumption that Tweety is a bird together
with the general principle that birds fly licenses the inference that Tweety flies, but
this inference is only tentative and to be retracted in case we also assume that, say,
Tweety is a penguin. The parallel, of course, is that penguins are an exception to
the general rule that birds fly just as situations in which one would need to break a
promise to save a drowning child are an exception to the general rule that one ought
to keep one’s promises. Reasoning with obligations, so the hypothesis then goes, is
nonmonotonic precisely because certain obligations are like defeasible generalizations
in that they allow for exceptions, and this is good news since we may now rely on the
well-established nonmonotonic analyses of how to reason with rules that allow for
exceptions to get the facts about deontic reasoning straight. But the parallel between
obligations that may be overridden and defeasible generalizations highlights only one
particular source of the nonmonotonic nature of deontic reasoning: the phenomenon,
or so I shall argue here, is much more general and thus the lessons from default logic,
important as they are, provide a too narrow conceptual foundation for exploring the
phenomenon of nonmonotonicity in deontic reasoning.
One way to see the point starts with the obvious fact that obligations may not only
allow for exceptions but also be outright violated. What is more, our misdeeds are like
exceptions in that they also have the potential to transform certain duties into actions
that are wrong. Here is a case that is inspired by Chisholm’s () classical discussion
of contrary-to-duty obligations: obligations telling us what to do in case we neglect
our duties so that we make the best of the bad situations to which our misdeeds have
led. Consider the classical inference rule of deontic detachment (Greenspan, ):
Deontic Detachment: (If φ)(Ought ψ), Ought φ  Ought ψ
Deontic detachment licenses the inference of () from () and ():
() Jones ought to go to the aid of his neighbors.
() If Jones goes to the aid of his neighbors, then he ought to tell them he is
coming.
() Jones ought to tell his neighbors that he is coming.
But suppose that:
() Jones does not go to the aid of his neighbors.

their consequents and cannot be captured by the classical analysis of modals and conditionals. I set such
conditionals aside but everything I am about to say here is sympathetic to Charlow’s discussion.
4 An exception is the proposal by van der Torre and Tan (), which resembles mine because of its
dynamic spirit but differs substantially in motivation, execution, and scope.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

dynamic foundations for deontic logic 

Then it may very well be that Jones ought not tell his neighbors that he is coming:
leaving them hanging is already bad enough, but creating false expectations in addition
to that would make things even worse.5 Once again it looks as if strengthening a set
of premises with an extra bit of information blocks an inference that the original set
licenses.
Earlier I said that there is a striking parallel between prima facie obligations and
defeasible generalizations: the former may be overridden while the latter may be
defeated. What we have seen just now is that there is also a striking parallel between
exceptional circumstances and violations of obligations: both may transform a duty
such as keeping a promise or telling one’s neighbors that one is coming into a wrong
action. The obvious thing to say then is that reasoning with violable obligations is
just as susceptible to monotonicity failures as is reasoning with obligations that allow
for exceptions. But while the lessons from default reasoning may very well help us
understand why reasoning with prima facie obligations exhibits monotonicity failures,
they do not explain why reasoning with violable obligations should exhibit the same
phenomenon: it is, after all, a truism that violations of obligations are not exceptional
circumstances in which the obligation is no longer binding (Prakken and Sergot, ,
). In particular, a situation in which Jones does not go to the aid of his neighbors
is not an exceptional circumstance that relieves him from the duty articulated by
()—his obligation to go—and it cannot be an exception to the conditional obligation
articulated by () either, since the latter does not even pertain to situations in which
Jones does not go. Reasoning with violable obligations has a nonmonotonic flavor
alright, but insights from the logic of reasoning with exceptions do not immediately
illuminate why this should be so.
Reasoning with obligations under epistemic uncertainty arguably has a nonmono-
tonic flavor as well, though the case is slightly more involved. Consider the miners
paradox from Kolodny and MacFarlane (). Ten miners are trapped either in shaft
A or in shaft B, but we do not know which one. Water threatens to flood the shafts. We
have enough sandbags to only block one shaft but not both. If one shaft is blocked, all
of the water will go into the other shaft, killing every miner inside. If we block neither
shaft, both will be partially flooded, killing one miner.

Action if miners in A if miners in B

Block A All saved All drowned

Block B All drowned All saved
Block neither shaft One drowned One drowned

5 The intended interpretation of deontic ought here is the deliberative one in the sense of Thomason
(): the one that figures prominently in advice and takes the facts as given. The claim that deontic ought
thus interpreted is subject to monotonicity failures—the advice we give to Jones is not preserved under
information strengthening—is compatible with the intuition that it also has interpretations less sensitive to
what is taken for granted.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 malte willer

In this scenario, () seems to be the right thing to say since all we know is that () is
true. However, we are also willing to accept () and ():
() We ought to block neither shaft.
() Either the miners are in shaft A or they are in shaft B.
() If the miners are in shaft A, we ought to block shaft A.
() If the miners are in shaft B, we ought to block shaft B.
()–() seem consistent. But let  be the set of premises consisting of ()–() and
observe that () intuitively entails both () and ():
() It is not the case that we ought to block shaft A.
() It is not the case that we ought to block shaft B.
() and () just negate the consequents of the conditionals in () and (), respec-
tively, and so given monotonicity and modus ponens,  ∪ {inA}  ⊥ and hence
  ¬inA by reductio. For parallel reasons,   ¬inB. Hence  is inconsistent after
all since it entails the negation of ().
The paradoxical argument just given appeals to a monotonic conception of logical
consequence—specifically, it assumes that if  entails (), then so do  ∪ {inA} and
 ∪ {inB}—and so one might hope for a nonmonotonic escape route. But reasoning
with obligations under epistemic uncertainty is not another instance of reasoning with
obligations that allow for exceptions. For suppose that the miners’ being in shaft A is
just a case in which the obligation to block neither shaft is no longer in force: then so
would be a case in which the miners are in shaft B, and since the miners are in one
of the two shafts we know that we are facing an exceptional situation and thus the
obligation to block neither shaft should not hold in the first place.
One reaction to the issues I have just outlined is to say that nonmonotonicity in
deontic reasoning is indeed limited to cases in which prima facie obligations are in
play: Chisholm’s scenario and the miners paradox highlight other problems with the
classical conception of deontic logic.6 But another reaction—the one I will pursue
here—is to say that nonmonotonicity in deontic reasoning is a pervasive phenomenon
that does not reduce to reasoning with rules allowing for exceptions.7 This reaction

6 Rejecting deontic detachment blocks Chisholm’s paradox but the prominent proposals by Lewis ()
and Kratzer (, ) preserve the validity of this rule (though they deny modus ponens for deontic
conditionals, which is sometimes labeled “factual detachment”). Appealing to tense is another option but
does not help with “timeless” Chisholm scenarios (see Carmo and Jones,  and McNamara,  and
references therein). Rejecting modus ponens is Kolodny and MacFarlane’s preferred solution to their miners
paradox, but Willer () argues that this is an overreaction and also says a bit about Bledin’s ()
and Yalcin’s () suggestion to respond to the paradox by rejecting reductio (see also § for additional
remarks). Silk () critically discusses Dowell’s () and von Fintel’s () suggestion that the deontic
modal ought receives different interpretations in () and ()/(), respectively.
7 See also the discussion by van der Torre and Tan (), who argue that the violability of obligations
should be understood as just such a kind of defeasibility, though they focus on a quite narrow conception
of defeasibility as failure of antecedent strengthening for deontic conditionals.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

dynamic foundations for deontic logic 

is appealing for reasons other than theoretical simplicity. The key conceptual driver
behind nonmonotonic logic is the idea that the validity of an inference is not only
sensitive to the presence of information—as in classical logic—but also to the absence
of information, and it does not take much to apply this idea to the cases just dis-
cussed. Prima facie obligations are binding unless there are exceptional circumstances,
violable obligations may entail other duties unless they are violated, and obligations
under epistemic uncertainty may no longer hold in case the underlying uncertainty
is resolved by additional information. In brief, while the cases discussed are at some
level of description quite disparate, there is more than a merely initial appeal to the idea
that they also have something important in common that a nonmonotonic analysis of
deontic logic is ideally suited to capture.
What is needed, then, is a comprehensive perspective on deontic discourse and
reasoning that captures the variety of ways in which deontic inference is sensitive
to the absence of information. Such a perspective, I maintain, arises in a dynamic
semantic analysis of ifs and oughts that, while being non-classical, incorporates ideas
from the familiar linguistics literature by analysing deontic modals and conditionals
using possible worlds and in light of a context determining what is deontically ideal.8
The resulting framework is flexible enough to allow for the interpretation of deontic
ought in light of contextually provided norms that may be overridable, violable, or
information-sensitive. While the potential overridability, violability, and information-
sensitivity of norms are distinct features that are captured by different aspects of the
formal apparatus, the semantic framework also highlights a common denominator:
they all induce deontic commitments that carry in their wake distinct epistemic
commitments and since the latter are defeasible, so are the former. More precisely, in
the framework I propose the previously observed sensitivity of deontic inference to the
absence of information uniformly manifests in entailment relations between deontic
ought and epistemic might. Once we have explained why epistemic might allows for
monotonicity failures—a fact that finds a natural explanation in dynamic semantics—
we also know why deontic ought does so as well.
My strategy is straightforward. § briefly outlines a basic dynamic semantics for
epistemic modals and conditionals, highlights its nonmonotonic outlook on logical
consequence, and briefly explains how this outlook may serve as a comprehensive
foundation for a nonmonotonic perspective on deontic discourse and reasoning. §
offers a simple dynamic semantics for deontic ought that makes sense of reasoning
with prima facie obligations and with obligations under epistemic uncertainty by
highlighting its sensitivity to epistemic possibilities. § extends the simple proposal
to avoid a few shortcomings when it comes to the analysis of contrary-to-duty

8 See Lewis () for an overview of early studies of deontic logic within possible world frameworks
(including his own proposal in Lewis, ). In the linguistics literature, the work by Kratzer (see Kratzer,
 and references therein) is seminal.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 malte willer

obligations.9 § concludes the discussion by briefly comparing the story told here with
alternative views that treat nonmonotonic effects as a topic for belief revision theory
rather than semantics.

 Basics
A basic dynamic story about conditionals and epistemic modals has been told before
(see Veltman,  and Gillies,  for seminal discussion), and so I will just briefly
highlight its key ideas and explain why they matter for current purposes. Semantic
values are stated in terms of context change potentials (CCPs): functions that take
an information state as input and return another state—the result of updating the
input state with the sentence under consideration—as output.10 Thinking of semantic
values in such a way owes some inspiration to the classical story about assertion
from Stalnaker (), who observes that context-content interaction is not a one-
way road: context affects what proposition an assertion expresses, but an assertion in
turn affects the context by adding the proposition expressed to the common ground.
In Stalnaker’s picture the latter effect is a pragmatic afterthought to classical truth-
conditional semantics, but here the facts about context-content interaction are integral
to how we model semantic values.
Thinking of semantic values dynamically as CCPs motivates an equally dynamic
conception of logical consequence as guaranteed preservation of rational commit-
ment. On this view, an argument is valid just in case its conclusion is accepted by
any carrier of information σ under the supposition of its premises:
φ , . . . , φn  ψ just in case for all σ : σ [φ ] . . . [φn ]  ψ
Here σ [φ] is the result of strengthening σ with the information carried by φ, and to
say that σ  φ is to say that σ is rationally committed to φ. The question whether
logical consequence is nonmonotonic now hinges on the question whether additional
premises in discourse and reasoning are guaranteed to preserve existing rational com-
mitments and, what is more, considerations about epistemic might strongly suggest
that information agglomeration fails to be commitment-preserving in such a way: the

9 Nothing I am about to say here by itself enforces the validity of deontic detachment (which, recall,
figures prominently in Chisholm’s paradox about contrary-to-duty obligations) but the upcoming story can
be modified so that it does. Motivating the required twists and turns in detail and explaining why they do
the trick goes beyond the scope of this paper—in brief, what is needed are some plausible restrictions on the
notion of a deontic context (§.) and some minimal modifications of the notion of retraction (§)—and so I
will focus here on the (no less important) task of telling a story about Chisholm’s case that is compatible with
the assumption that deontic detachment is valid and explains the data from a nonmonotonic perspective.
See Willer () for a detailed discussion of deontic detachment that sits happily with everything said in
this paper but does not mention many of the issues that matter here.
10 Classical dynamic frameworks include Discourse Representation Theory (see Kamp et al.,  for an
up-to-date discussion), Dynamic Predicate Logic (Groenendijk and Stokhof, ), File Change Semantics
(Heim, ), and Update Semantics (Veltman, , ).

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

dynamic foundations for deontic logic 

belief that, say, Mary might be in Chicago needs to go once one learns that Mary is not
in Chicago (see Levi, , Fuhrmann, , and Rott,  for seminal discussion).
More generally:
σ  Might φ −→
 σ [¬φ]  Might φ
A state may be committed to Might φ but fail to be thus committed if strengthened
with ¬φ. And that already leads to monotonicity failures given the following
minimal assumptions:
Acceptance: For all σ : σ  φ just in case σ [φ] = σ
Epistemic Success: For all σ : σ [Might φ]  Might φ

The first principle says that φ is accepted in σ just in case σ already carries the
information carried by φ. The second principle requires that the result of updating
some carrier of information with an epistemic modal claim accepts that claim. It
follows immediately that Might φ  Might φ for all φ, and we also know that for
some σ and φ, σ  Might φ yet σ [¬φ]  Might φ and thus σ [Might φ][¬φ] 
Might φ. Accordingly, Might φ  Might φ yet Might φ, ¬φ  Might φ and thus
the proposed notion of logical consequence is nonmonotonic.
To make all of this more precise, define the initial target of our semantic analysis as
a simple propositional language extended with the conditional connective if (⇒) and
epistemic might (e ):
Definition  (Basic Language). L is the smallest set that contains a set of atomic
sentences A = {p, q, r, . . .} and is closed under negation (¬), conjunction (∧),
epistemic might (e ), and the conditional connective if (⇒). Inclusive and exclusive
˙ the material conditional (⊃), the biconditional (≡), and epistemic
disjunction (∨, ∨),
must (e ) are defined in the usual way. L is defined as the propositional fragment
of L.
Semantic values take a state as input and return one as output. Such carriers are
sets of indices consisting of a possible world coupled with a frame (more on frames
momentarily):
Definition  (Information States). σ is an information state iff σ ⊆ I, where I is the set
of all indices. wi is the possible world parameter of some index i, where w is a possible
world iff w : A → {, } and W is the set of all possible worlds.  is the set of all
information states. σ is the initial information state and identical with I.
Information states are sets of indices, and to avoid the need for redefinitions at a later
stage we will not simply identify an index with a plain possible world, though this
would be good enough if we were interested only in the semantics for L.
Recursively define the update rules for elements of L as follows:
Definition  (Basic Update Rules). Associate with each φ ∈ L a context change rule
[φ] :  →  as follows:

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 malte willer

() σ [p] = {i ∈ σ : wi (p) = }

() σ [¬φ] = σ\σ [φ]
() σ [φ ∧ ψ] = σ [φ][ψ]
() σ [φ ⇒ ψ] = {i ∈ σ : σ [φ][ψ] = σ [φ]}
() σ [e φ] = {i ∈ σ : σ [φ] = ∅}
According to (), updating σ with an atomic sentence eliminates all indices with a
world-parameter at which that sentence is false. Updating σ with ¬φ comes down
to subtracting from σ the result of updating σ with φ (cf. ()). The clause in ()
requires that updating with a conjunction amounts to an update with the first conjunct
followed by an update with the second conjunct. The clause in () takes Ramsey’s ()
test procedure for conditionals as a guide to the semantics of if. Ramsey suggested
that a conditional is to be accepted just in case its consequent is (hypothetically)
accepted under the supposition of its antecedent, and here we say that a conditional
tests whether the input state σ has a certain informational structure: once we have
updated σ with the antecedent, updating with the consequent idles (does not induce
any additional change), which is just to test whether the consequent is accepted under
the supposition of the antecedent. If this is so, the test is passed and the conditional
is accepted; otherwise, the conditional is rejected. Finally, the meaning of epistemic
might is given in terms of acceptance conditions as well: it tests whether its prejacent
is compatible with the input state.
With each element of L we may associate a proposition in the familiar way:
Definition  (Propositions). The function [[·]] : L → W assigns to each φ ∈ L a
proposition understood as a set of possible worlds:
() [[p]] = {w ∈ W : w(p) = }
() [[¬φ]] = W \ [[φ]]
() [[φ ∧ ψ]] = [[φ]] ∩ [[ψ]]
Propositions will not play their classical role as carriers of meaning but be useful at a
later stage.
The following definition just restates the dynamic conception of logical conse-
quence that was mentioned earlier but also says what it takes for a commitment to
be defeasible.
Definition  (Acceptance, Defeasibility, Entailment). Consider any σ ⊆  and
φ  , . . . φn , ψ ∈ L:
. σ accepts φ, σ  φ, iff σ [φ] = σ
. σ defeasibly accepts φ iff σ  φ and for some σ  ⊆ σ : σ   φ
. ψ defeats a commitment to φ in σ iff σ  φ but σ [ψ]  φ
. φ , . . . φn entails ψ, φ , . . . φn  ψ, iff for all σ ∈  : σ [φ ] . . . [φn ]  ψ
A state σ accepts φ just in case updating σ with φ idles. This commitment is defeasible
just in case there is some strengthening of σ that no longer accepts φ, and ψ is a

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

dynamic foundations for deontic logic 

defeater just in case updating σ with ψ results in such a state. An argument is valid
just in case any information state that is updated with its premises is committed to its
conclusion.
It will be helpful to highlight three facts about the outlined framework, the first one
being that modus ponens is valid:
Fact  φ ⇒ ψ, φ  ψ
Take any σ and consider σ [φ ⇒ ψ][φ]: if σ [φ ⇒ ψ] = ∅, the conclusion is accepted
for trivial reasons. Otherwise, σ [φ ⇒ ψ] = σ and σ [φ]  ψ, which just proves the
point.
Secondly, observe that the dynamic logical consequence relation is nonmonotonic:
Fact  For all contingent φ ∈ L : e φ  e φ and e φ, ¬φ  e φ.
The reason is familiar: might-commitments are not guaranteed to be preserved as new
information is acquired, and in particular a commitment to e p fails to be preserved
by the information that p is not the case.
The third observation of relevance is that the defeasibility of epistemic might may
enforce the defeasibility of other rational commitments:
Fact  Consider any σ ∈  such that σ  φ and suppose that φ  ψ: if χ defeats
a commitment to ψ in σ , then χ defeats a commitment to φ in σ .
This is just to say that whenever a commitment to φ brings in its wake a commitment
to ψ, then the commitment to φ is defeasible in case the commitment to ψ is. And that
is important since there is every reason to think that commitments to deontic ought
frequently bring in their wake a commitment to epistemic might: the defeasibility of
deontic commitments can be illuminated in terms of the by now established dynamic
defeasibility to certain epistemic commitments. Let us go through the details.

 Simple Oughts
This section presents a simple dynamic framework that makes sense of reasoning
with prima facie obligations but is also flexible enough to make sense of some
other nonmonotonic effects in deontic reasoning. Doing so requires that we offer a
dynamic analysis of the deontic modal ought (d ) as well as of a defeasible conditional
connective (>) that will be used in stating prima facie obligations. The extended target
language is defined as follows:
Definition  (Extended Language). Define a language L so that L ⊆ L and
whenever φ, ψ ∈ L , then d φ ∈ L and also φ > ψ ∈ L . L+ is the smallest
set that contains all elements of L and is closed under negation (¬), conjunction (∧),
epistemic might (e ), and the conditional connective if (⇒). Other connectives are
again defined as usual.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 malte willer

Earlier I stated the semantics of epistemic modals and conditionals in terms of

acceptance conditions, and it makes good sense to do the same when it comes to
deontic ought. The proposal is that deontic ought is accepted just in case its prejacent
is accepted in the state got by focusing on those indices that accord best with the rules
articulating what is deontically ideal, where those rules are determined by context and
may very well be defeasible. The idea will be slightly refined at a later stage, but it is
good enough to illustrate why the setup pursued here has some real promise.
Definition  (Deontic Ought (Basic Version)). Extend the update rules for L with the
following entry:
() σ [d φ] = {i ∈ σ : σd  φ}
Here σd is the set of indices compatible with σ that are deontically ideal in light
of some deontic context d, and σ  φ holds whenever σ expects φ to be true: φ
is true in every possible world that is compatible with σ and accords best with the
defeasible generalizations accepted by that state. I will first say more about defeasible
generalizations (§.) and then explain what a deontic context does (§.).

. Defeasible Rules

Defeasible generalizations articulate what to expect in case certain conditions are
satisfied, and since we do not care too much about sentential structure we can make
this more precise by saying that such conditionals select a set of default worlds from
the possibilities depicted by the antecedent. Taking some inspiration from Veltman
() the semantic analysis is provided using frames:
Definition  (Frames). A frame π maps each scenario s ⊆ W to a subset of s. [[φ]] is a
default in π(s) iff π(s) ⊆ [[φ]]. w is a normal world in π(s), w ∈ N π(s), iff w ∈ s and
for all s ⊆ s: if w ∈ s , then w ∈ π(s ). π is coherent iff for every nonempty s ⊆ W,
there is a normal world in π(s). is the set of all such coherent π ’s.
Frames allow us to say, for instance, that it normally rains but there is an exception:
if there is an easterly wind, it normally does not rain. If this is all we want to say, we
have π(W) = [[rain]], π([[east]]) = [[east]] \ [[rain]], and π(s) = s for all other
scenarios. Consider the following distribution of truth-values across possible worlds:
rain east
w T T
w T F
w F T
w F F

Then π(W) = {w , w } and π([[east]]) = {w }. Since w violates the expectations that
come with an easterly wind—w ∈ [[east]] yet w ∈ / π([[east]])—it does not count as
a normal world. In contrast, w does not violate any of our expectations—notice that
w ∈/ [[east]] and so even though it rains at w , no expectation is violated—and so

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

dynamic foundations for deontic logic 

N π(W) = {w }: normally, it rains and there is no easterly wind. But assuming that
there is an easterly wind, we expect no rain since N π([[east]]) = {w }.
We may now make the notion of an index precise and determine the update rule for
the defeasible conditional connective:
Definition  (Indices). i is an index iff i ∈ W × , that is, an index is a pair consisting
of a possible world w and a coherent frame π . πi is the frame parameter of i.
Definition  (Defeasible Conditional Connectives). Extend the previous update
rules for L and deontic ought with the following entry:
() σ [φ > ψ] = {i ∈ σ : πi ([[φ]]) ⊆ [[ψ]]}
An information state keeps track of the hard information modeled by a set of possible
worlds but also of what is to be expected, which is represented by a set of coherent
frames. A formula of the form φ > ψ eliminates from an information carrier all
those indices whose frame parameters fail to treat [[ψ]] as a [[φ]]-default.
Veltman () offers a model for how to reason with defeasible generalizations in
dynamic semantics, but for our purposes there is no need to repeat his story here.
I am first and foremost interested in the role of prima facie obligations for deontic
discourse and reasoning, and for this I need to say precisely what is expected by a
carrier of information. As a preparation, observe that we can associate with each such
carrier a unique scenario and frame as follows:
Definition  (Depicted Scenarios and Frames). Consider arbitrary σ ∈ :
. The scenario depicted by σ is defined as sσ = {wi : i ∈ σ }.
. The frame depicted by σ is defined so that [[φ]] is a default in πσ (s) iff for all i ∈ σ :
[[φ]] is a default in πi (s).
The scenario depicted by σ is simply the set of possible worlds compatible with the
hard information carried by σ . And σ treats [[ψ]] as a [[φ]]-default just in case σ accepts
the defeasible generalization articulated by φ > ψ.
Suppose then that one accepts a set of default rules. The simple intuition is that
a default rule should play a role in forming expectations just in case the rule is
triggered—one’s information entails the premise of the default rule—and at the same
time undefeated—one’s information does not depict an exceptional scenario in which
the rule is no longer binding. The only wrinkle we have to add to the story is that
default rules may sometimes lead to conflicting expectations, and so we must define
what it takes for a set of default rules to be unconflicted in expectation forming.11 The

11 My story about how to form expectations on the basis of accepted defaults owes inspiration to
Veltman’s () and Horty’s () accounts. But the difference between the upcoming proposal and theirs
are substantial, not least because it offers a possible-world analysis of the defeasible conditional connective
(thus differing from Horty’s) and ties the applicability of a default rule to a triggering condition (thus
differing from Veltman’s).

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 malte willer

key notions involved in this characterization of how expectations are formed on the
basis of accepted default rules can be made more precise as follows:
Definition  (Triggers, Defeats, Conflicts). Consider arbitrary π ∈ , σ ∈ , and
say that a world w complies with a [[φ]]-default [[ψ]] iff w ∈
/ [[φ ∧ ¬ψ]]:
. A [[φ]]-default [[ψ]] is triggered in σ iff σ  φ.
. A [[φ]]-default [[ψ]] is undefeated in σ iff for every s such that sσ ⊆ s: there is some
w ∈ N π(s) that complies with the [[φ]]-default [[ψ]].
. A set of defaults is unconflicted in σ iff for every s such that sσ ⊆ s: there is some
w ∈ N π(s) that complies with each of its members.
To say that a world complies with a default rule is just to say that it does not violate
the expectation articulated by that rule. For a default rule to be triggered in σ , its
condition must be entailed by the scenario depicted by σ . But the rule is defeated in
σ if another, more specific rule is triggered that conflicts with it, and a set of default
rules is conflicted in σ just in case no normal world in the scenario depicted complies
with each of its members. Note here that whenever a single default rule is defeated in
σ , no set of rules containing it can be unconflicted in σ .
To give a very simple example of the notion of defeat, go back to the case in which
we believe that it normally rains but if there is an easterly wind, the weather is normally
dry: [[Rain]] is W-default but [[¬Rain]] is an [[East]]-default. Then intuitively a
scenario in which the wind comes from the east is an exception to the rule that it
normally rains, and this is just what the framework predicts. For suppose σ accepts that
the wind comes from the east: then sσ ⊆ [[East]] and of course every normal [[East]]-
world is one at which the weather is dry. So no normal [[East]]-world complies with
the W-default [[Rain]], which is just to say that the default rule is defeated in σ .
To illustrate the possibility of conflicting expectations, and to go through an
example that will be of relevance in the upcoming discussion, start with the initial
information carrier σ and consider the state σ = σ [Promise > Meet][Need >
Help][Promise ∧ Need], with the atomic sentences being translated as follows:
promise: Jones has promised to meet Mary for lunch.
meet: Jones will meet Mary for lunch.
need: The drowning child needs Jones’s help.
help: Jones will help the drowning child.
Then clearly σ accepts that Jones will, all things being equal, meet Mary for lunch
if he promised to do so and that he will, all things being equal, help a child if
it is in need of his help. Not surprisingly, both default rules are triggered and
undefeated in the scenario we consider—that Jones has promised to meet Mary
for lunch and that there is a drowning child in need of Jones’s help. The worlds
compliant with both rules are just those at which Jones keeps his promise and saves
the drowning child.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

dynamic foundations for deontic logic 

Things become more interesting if we assume that Jones cannot do both, which I will
do here by considering τ = σ [meet ∨ ˙ help].12 Both default rules remain triggered and
undefeated—the scenario depicted by τ is not an exception to any of those rules—but
they conflict with each other: no world in the scenario under consideration complies
with both expectations. For at any such world, Jones does not meet Mary for lunch
even though he promised to do so—thus violating the expectation that he will keep
his promise—or fails to help the child even though it is drowning—thus violating the
expectation that he helps the child in need. It makes sense to expect that Jones will at
least keep his promise or save the drowning child (rather than doing neither) but there
is, without additional assumptions about Jones’s character anyway, no reason to favor
one default rule over the other in thinking about what he is going to do in the scenario
under consideration.
Now suppose we want to say that in Jones’s case, the default rules are not of
equal strength: Jones will not let the child in need drown just to keep his promise
to meet Mary for lunch. One way to go is to appeal to a priority relation between
default rules (see Horty,  for recent discussion) but here I exploit the already
existing feature that more specific rules trump less specific ones in case of a conflict.
For consider χ = τ [((promise ∧ need) ∧ (meet ∨ ˙ help)) > help], that is, the
result of strengthening τ with the additional default rule that Jones will help the child
in need if faced with the unhappy choice between doing so and keeping his promise.
This update would have no interesting effect if it were epistemically possible for Jones
to both keep his appointment and help the drowning child, but it makes a big difference
in the case under consideration.
The crucial observation about χ is that it differs from τ in treating a scenario in
which Jones must choose as an exception to the rule that, all things being equal, he
will meet Mary for lunch if he has promised to do so. To see this, simply observe that
a world now counts as normal in πχ (([[promise ∧ need) ∧ (meet ∨ ˙ help)]]) only if
John helps the drowning child—and thus breaks his promise to meet Mary for lunch—
at that world. No such world can thus comply with the [[promise]]-default [[meet]],
which just means that the default rule is defeated in the scenario under consideration.
Since the other default rules remain triggered and undefeated, we expect that Jones
will help the drowning child.
Much more could be said about how we form expectations in everyday discourse
and reasoning, but the emerging picture is good enough for our purposes. What
matters here is that the role of a default rule in expectation forming is sensitive to
the presence or absence of global features of an information state that is not preserved
under the process of strengthening. Accordingly, expectations may be defeated as new

12 This is simplifying a bit: it is desirable to distinguish the claim that John makes a choice between
keeping his promise and saving the child—which is what “meet ∨ ˙ help” literally says—from the stronger
one that Jones must make this choice. Here and throughout it will do no harm to avoid additional
complications and simply assume that Jones makes the choice only if he cannot do both.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 malte willer

information comes into view, and the following definition makes it easier to see why
this is so.
Definition  (Expectations). Consider any σ ⊆  depicting a frame π :
. i ∈ σo iff i ∈ σ and wi complies with a maximal set of π -defaults such that (i) the
set is unconflicted in σ and (ii) each member is triggered in σ .
. i ∈ σon+ iff i ∈ σon and wi complies with a maximal set of π -defaults such that
(i) the set is unconflicted in σon and (ii) each member is triggered in σon .

. The set of optimal indices in σ is defined as σo = n≥ σon .
. σ expects φ, σ  φ, iff σo  φ.
To say that φ is expected in σ is to say that φ is accepted by the state got by focusing on
the optimal indices in σ . The optimal indices are defined in a stepwise fashion: first,
determine σo , that is, the indices whose world-parameter complies with any maximal
unconflicted set of triggered default rules in σ . Then determine σo , that is, the indices
in σo that come with a world-parameter complying with any maximal unconflicted
set of triggered default rules in σo , and so on. The set of optimal indices—defined as

n≥ σo —is just the state of information that the deliberating agent will arrive at after
n

carrying out the reasoning process indefinitely.

Whenever the applicability of a default rule is defeasible, so are expectations formed
on the basis of that rule. And the special case that interests us is the one we considered
earlier: given certain expectation patterns, the [[promise]]-default [[meet]] is defeated
in some carrier of information σ such that σ  need unless for some i ∈ σ , wi ∈
[[meet ∧ help]]. Specifically:
Fact  Take any σ so that πσ = πχ : if σ  meet, then σ  e (¬need ∨ (meet ∧
help))
Expectations that derive from defeasible generalizations, in short, may be sensitive to
the presence or absence of certain epistemic possibilities, and this is just how things
should be. This is the moral we need to make sense of the nonmonotonicity of thought
and talk about prima facie obligations. The purpose of the next section is to explain
why.

. Deontic Ought

Earlier I said that deontic ought is accepted just in case its prejacent is optimal in light
of some deontic context. It is a familiar idea from the linguistics literature that a deontic
context determines what is deontically ideal by fixing an ordering source. Here I will
adopt this idea but not think of ordering sources classically as sets of propositions but
dynamically as sets of context-change potentials.
Definition  (Deontic Contexts). A deontic context d determines for each σ ∈ 
a deontically ideal state σd by providing an ordering source o ⊆   , that is, a set
of CCPs. od is the ordering source provided by d. Given some σ ∈ , σ = {i ∈

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

dynamic foundations for deontic logic 

σ : ∃i ∈ σ. wi = wi } and i <σd i iff i, i ∈ σ and (i) for all [φ] ∈ od : if i ∈ σ [φ],
then i ∈ σ [φ] and (ii) for some [φ] ∈ od : i ∈ σ [φ] and i ∈ / σ [φ]. i is minimal in σ
given d iff ¬∃i : i <σd i. σd is then just the set of indices that are minimal in σ given d.13
Whenever the ordering source exclusively consists of elements of L , we may just as
well have propositions fix what is deontically ideal. The current framework is more
flexible, however, and in particular we will put the possibility of stating ordering
sources using epistemic modals and defeasible conditionals to good use. Let me
explain.
Start with Jones, who has promised to meet Mary for lunch but also stumbles upon a
drowning child. Suppose that the deontic context fixes the following ordering source:
˙
od = {[promise > meet], [need > help], [((promise ∧ need) ∧ (meet∨
help)) > help]}
Here the first member of od articulates the rule that Jones’s promise to meet Mary for
lunch creates the prima facie obligation to meet her for lunch, while the second says
that a child’s need for help creates the prima facie obligation to help that child. The
third rule effectively resolves a potential conflict between the first and second prima
facie obligation in favor of helping the child in need.
Take any σ ∈  and let π = πσd : then π([[promise]]) = [[promise ∧ meet]],
π([[need]]) = [[need ∧ help]], π([[(promise ∧ need) ∧ (meet ∨ ˙ help)]]) =
[[(promise ∧ need) ∧ (¬meet ∧ help)]], and π(s) = s for all other scenarios. The
good news is that we already know what it takes for it to be the case that σd  meet
from the earlier discussion in §.: of course, it is required that σ  promise,
but for the [[promise]]-default [[meet]] to apply in σd , it must also be the case that
σ  need ∧ ¬(meet ∧ help)—otherwise, the [[promise]]-default [[meet]] is defeated.
And that is just to say that a commitment to () needs to go in light of the additional
information that (i) there is a drowning child in need of Jones’s help and (ii) Jones
cannot help that child and keep his lunch appointment. It follows immediately that ()
and () entail () but no longer do so if strengthened by the information carried by
(). Here is a summary of the predictions:
Fact  Consider d as fixed for Jones’s scenario:
. promise ⇒ d meet, promise  d meet
. promise ⇒ d meet  e (¬need ∨ (meet ∧ help))
. ˙ help)  promise ⇒ d meet
promise ⇒ d meet, promise, need ∧ (meet ∨
. ˙ help)  d meet
promise ⇒ d meet, promise, need ∧ (meet ∨
The first observation follows from the general validity of modus ponens while the
second is a direct consequence of the applicability conditions for defeasible general-

13 A disclaimer with a familiar ring: analysing deontic ought as a quantifier over a set of possible worlds
that are minimal in σ given d is technically convenient but requires that there is always such a set. The
classical analogue is the Limit Assumption (Lewis, ).

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 malte willer

izations together with the semantics for ifs and oughts. It follows from the defeasibility
of epistemic might and its connection with deontic ought that the information carried
by () defeats a commitment to (), and this yields the final observation together with
the semantics of the Ramsey conditional. It is easy to verify that Jones’s obligation to
help the drowning child, while induced by a norm that is in principle defeasible, is not
affected by ().14
I am about to conclude that the framework developed so far offers a straightforward
treatment of prima facie obligations: such obligations may be overridden, and this
is not surprising since they stem from deontic rules that are articulated using the
defeasible conditional connective. But one may wonder whether the story told here
really captures Ross’s () idea that prima facie obligations, even if overridden,
continue to count in favor of some action rather than lose their status as a reason
for action altogether. The answer, I think, depends on what it means to accept the
antecedent of a defeasible norm (that is, to be in a state that triggers the default norm).
We are free to suggest that doing so provides a reason for action, albeit one that may be
outweighed by other reasons. Since the state of accepting the antecedent of a defeasible
norm is guaranteed to be preserved even if new information is acquired, it follows
that the overridability of prima facie obligations is compatible with their persistence
as moral reasons for action. But this is not the only available interpretation: perhaps
it is better to say that moral valence is entirely a matter of what norms are triggered
and undefeated in the scenario under consideration, thus allowing for the possibility
that a feature counting in favor some action given one scenario may not be a reason at
all given another, or perhaps even count against choosing that action. If so, accepting
the antecedent of a defeasible norm by itself does not give one a moral reason for
acting in one way or another. Both views have some intuitive appeal, and both can be
accommodated by the story told here.15

14 One might complain that the proposal made here hardly tells us everything we want to know about
Jones’s case since it simply hardwires our intuitions about what he is normally required to do and about what
counts as an exceptional case into an ordering source. However, it is arguably not the task of a framework
for deontic reasoning to explain our normative intuitions but rather to correctly predict which inferences,
given those intuitions, are valid and invalid, and here I add the additional twist that these predictions derive
from plausible assumptions about the semantics of modals, conditionals, and defeasible generalizations.
Integrating techniques from frameworks such as Horty’s () allows us to derive the ordering source for
Jones’s scenario from its first two rules and the assumption that the requirement to save a child in need takes
priority over the requirement to keep a lunch appointment, but clearly one would then still have to rely on
normative intuitions that are not derived from more basic ethical principles.
15 The first interpretation is in line with Horty’s () conception of reasons as triggered defaults, though
he develops his framework further so that it leaves room for valence switching. The second interpretation
is in line with a contextualist perspective on reasons, which has been developed in detail by, for instance,
Lance and Little in a series of papers, starting with Lance and Little () (which they originally labeled
“particularist”). What all these views have in common is that they appeal to general, albeit defeasible,
principles in the derivation of obligations, thus being at odds with moral particularism. See Dancy ()
for a recent articulation of the case for particularism that also criticizes the appeal to defeasible principles as
guides to reasoning about obligations. Alas, discussing the merits of these criticisms goes beyond the scope
of this chapter.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

dynamic foundations for deontic logic 

All of that preserves the advocated connection between prima facie obligations and
defeasible generalizations, and it translates the classical idea that deontic reasoning
is sensitive to the absence of information into the claim that deontic commitments
are sensitive to epistemic commitments: since the latter may be defeated by addi-
tional information—in particular, commitments involving epistemic might are not
necessarily preserved by information agglomeration—it is not surprising that deontic
commitments are no less susceptible to preservation failure. The even better news
is that the attested sensitivity of deontic ought to epistemic might highlights other
potential nonmonotonic effects in deontic discourse and reasoning that would remain
hidden if we merely focused on the parallel between prima facie obligations and
defeasible generalizations. Let me explain.
Taking a nonmonotonic perspective on discourse and reasoning, I have said,
promises an attractive response to the problem surrounding the miners paradox. But
whatever the underlying mechanism is, it cannot be that we are dealing with a prima
facie obligation to block neither shaft that is overridden whenever the miners are in
shaft A (shaft B). To see this, suppose we think of the deontic context in the miners
scenario as follows:
od = {[(inA ∨ inB) > ¬(blA ∨ blB)], [inA > blA], [inB > blB]}
Any frame that treats [[¬(blA∨blB)]] as an [[inA∨inB]]-default while treating [[blA]]
as an [[inA]]-default and [[blB]] as an [[inB]]-default is incoherent. For let s = [[inA ∨
inB]] and consider any w ∈ s : then w ∈ N π(s) only if w ∈ [[¬(blA ∨ blB)]]. But
clearly either w ∈ [[inA]] and in that case w ∈ / π([[inA]]) or w ∈ [[inB]] and in that
case w ∈ / π([[inB]]), which is just to say that N π(s) = ∅.
The last observation does not show that there is something wrong with our story
about defeasible generalizations and prima facie obligations, but just that a nonmono-
tonic escape route from the miners paradox cannot rely on what this particular story
has to say. The good news is that the dynamic framework developed so far does not
tie nonmonotonic effects in deontic discourse and reasoning to the defeasibility of
certain generalizations but rather to the defeasibility of epistemic might. The following
deontic context does justice to our intuitions about the miners scenario while allowing
for nonmonotonic effects to play a role in reasoning about that scenario:
od = {[blA ≡ e inA], [blB ≡ e inB], [¬(blA ∨ blB) ≡ (e inA ∧ e inB)]}
Let σ be the information we have about the miners’ whereabouts. Then σ  e inA ∧
e inB and accordingly σ  ¬e inA and σ  ¬e inB. It follows that the minimal
indices in σ are those at which we block neither shaft. But consider σ  = σ [inA]: then
σ   e inA and accordingly, the minimal indices in σ  are those at which we block
shaft A. For parallel reasons, σ [inB]  d blB. Here is a summary of the output:
Fact  Consider d as fixed for the miners paradox and let σ be the information
we have about the miners’ whereabouts:

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 malte willer

. σ  d ¬(blA ∨ blB)
. σ  inA ⇒ d blA
. σ  inB ⇒ d blB
. σ  inA ∨ inB
The framework thus accounts for our intuitions about the miners scenario and resolves
its air of paradox. For while σ  d ¬(blA ∨ blB), we also know that σ [inA] 
d ¬(blA ∨ blB) and so even though σ [inA]  d blA, nonetheless σ [inA]  ⊥
and so σ  ¬inA. For similar reasons, σ  ¬inB. What makes all this possible is that
adding the information that the miners are in shaft A (in shaft B) defeats one’s rational
commitment to the claim that we ought to block neither shaft, and the fact underlying
this observation is the following:
Fact  Consider d as fixed for the miners paradox: then for all σ ∈  such that
σ  e blA∧e blB it holds that if σ  d ¬(blA∨blB), then σ  e inA∧e inB
A commitment to the claim that we ought to block neither shaft (even though we might
block either of them) is dependent on one’s ignorance about the miners’ whereabouts.
Once this ignorance is removed the deontic commitment has to go as well.16
We may now highlight what Jones’s case has in common with the miners scenario
and where they differ. Both scenarios require that an action that is deontically optimal
with respect to some state σ fails to be deontically optimal with respect to a contracted
state even though that action is still choosable. The possibility of meeting Mary
for lunch, albeit deontically optimal in case Jones might do so and help the child
in need, fails to be deontically optimal once we assume that he needs to choose
between keeping his appointment and helping the child—even though he still might
do the former instead of the latter. Likewise, the possibility of blocking neither shaft,
albeit deontically optimal in case we do not know where the miners are, fails to be
deontically optimal once we assume that the miners are in shaft A—even though we
still might block neither shaft. So in both cases deontic ought exhibits a sensitivity to
epistemic might, but for different reasons. In Jones’s scenario, the underlying fact is that
information strengthening may override certain obligations since certain rules may
be defeated as additional information becomes available. This is not so in the miners
scenario since, for instance, the rule to block none of the shafts just in case the miners
might be in either of them is not defeasible in the first place. Instead, what is going on

16 Carr () and Cariani () argue that the semantics of deontic modality is sensitive to probabilistic
considerations, and it would be possible to expand the story developed here so that ordering sources can
be stated using probability operators. This would not change the fact that the framework developed here
is “quietist” in Charlow’s () sense since it does not explicitly encode decision-theoretic considerations
in the semantics: rather, such considerations play an implicit role in selecting the relevant ordering source
for the case under consideration. Charlow argues that so much quietism is not a good thing, but for our
purposes there is no need to resolve this issue here: whether or not decision-theoretic considerations are
to be explicitly represented in the semantics, the point remains that our intuitions about what ought to be
done in the miners scenario are sensitive to the presence or absence of certain epistemic possibilities in a
way that the framework developed here is well equipped to capture.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

dynamic foundations for deontic logic 

here is that contraction may result in the rule requiring an action—blocking at least
one of the shafts—that it forbids in light of a weaker carrier of information (and simi-
larly for the other rules in the ordering source). In other words, the rules that are at play
in the miners scenario are absolute in the sense that they apply within any scenario,
but they may require different actions since they are stated using epistemic modals.
Horty () considers various obstacles to integrating the full insights from default
logic in a Kratzerian possible-worlds analysis of deontic modals. I have not demon-
strated that all these obstacles can be overcome—for instance, I have said nothing
about higher-order default norms in deontic reasoning—but the progress made here
should create some confidence that such a project is anything but futile. Let me briefly
highlight one particular issue that Horty discusses in some detail. Suppose that there is
a general prohibition against eating with your fingers ( > ¬F) with one exception: if
you are served cold asparagus, eating with your fingers is required (A > F).17 Horty’s
preferred framework for conditional oughts predicts that in light of these norms, “You
ought not eat with your fingers” and “If you are served cold asparagus, you ought to eat
with your fingers” are accepted in an out-of-the-blue context. It also predicts that the
norms under consideration do not license what he labels the “asparagus inference” to
the conclusion that you ought not be served cold asparagus. And this is a good result
since rules with exceptions should not render their exceptions as deontically sub-ideal
by design.
Horty also correctly observes that Kratzer-style possible worlds analyses of deontic
conditionals license the unfortunate asparagus inference. The inference pattern, how-
ever, is grounded in a “stability” feature of deontic contexts that is characteristic of
classical frameworks but not of the one presented here: that if an index i is optimal in
σd and included in some τ ⊆ σ , then i is guaranteed to be optimal in τd as well.18
It is a well-worn story that so much stability is problematic when it comes to the
miners paradox (see Cariani et al., , Charlow, b, and Kolodny and MacFarlane,
)—the crucial observations here are, first, that the feature is no less problematic
when it comes to prima facie obligations, and, secondly, that it is avoided by the story
about defaults told here. For consider od = {[ > ¬F], [A > F]} and let σ be the
initial information state: then clearly σd  ¬F yet σd  ¬A even though σ [A]d  F,
and so σ  d ¬F and σ  A ⇒ d F yet σ  d ¬A, as desired.19 It follows that

17 Following standard protocol, unconditional default rules are understood as defaults for the trivial
scenario W;  is just any tautology.
18 To see the connection, suppose that σ   ¬φ and thus that for some optimal i in σ , w ∈ [[φ]].
d d i
Clearly i ∈ σ [φ] and thus, assuming that d is stable, optimal in σ [φ]d . And if σ  d ¬ψ as well, then
wi ∈ [[¬ψ]] and hence σ  φ ⇒ d ψ. Accordingly, if σ  d ¬ψ and σ  φ ⇒ d ψ, then σ  d ¬φ
under the assumption that d is stable.
19 The underlying fact here is that only the W-default [[¬F]] but not the [[A]]-default [[F]] is triggered in
σd . Accordingly, optimality in σd merely requires compliance with the W-default [[¬F]] and so an index
may be optimal even if its world-parameter is an A-world. When it comes to the miners paradox, suppose
again that σ is the information we have about the miners paradox: then all the indices in σd are optimal
since no default norms are involved and for some i ∈ σd , wi ∈ [[(¬blA ∧ ¬blB) ∧ inA]]. Such an index
is an element of σ [inA] but not included in σ [inA]d , which shows that the deontic context for the miners
scenario is just as unstable as the one for the asparagus case.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 malte willer

a possible-worlds analysis of conditionals and deontic modals—if developed with a

dynamic spin—is not committed to licensing the unattractive asparagus inference.
A dynamic analysis of deontic ought offers an attractive foundation for a non-
monotonic perspective on discourse and reasoning. It readily handles cases in which
prima facie obligations are overridden—cases that naturally call for a nonmonotonic
analysis—but also offers a nonmonotonic perspective on thought and talk about ought
under epistemic uncertainty, and all of this while keeping track of what these cases
have in common and of where they differ. I take this result to be motivation enough to
look at how the dynamic framework can be refined so that it captures the connection
between nonmonotonicity, epistemic possibility, and the violability of obligations.

 Loose Ends
The framework developed so far has something useful to say about contrary-to-duty
obligations, though the story does not turn out to be entirely satisfying. To see the
positive aspect of the story, fix the ordering source for Chisholm’s scenario as follows:
od = {[go], [go ⊃ tell], [¬go ⊃ ¬tell]}
Since we are concerned here only with absolute rules, considerations about frames are
irrelevant for current purposes and so may simply consider the following distribution
of truth-values across possible worlds:
go tell
w T T
w T F
w F T
w F F

If sσ = {w , w , w , w }, then for any i ∈ σd , wi = w and accordingly σ  d tell.

But consider τ = σ [¬go]: then sτ = {w , w } and so for any i ∈ τd , wi = w . Hence
σ [¬go]  d tell and in fact, σ [¬go]  d ¬tell, which just captures our intuition
that Jones ought not tell his neighbors that he is coming under the assumption that he
does not go. The general observations are the following:
Fact  Consider d as fixed for Chisholm’s scenario:
. d go, go ⇒ d tell  d tell
. d go, go ⇒ d tell, ¬go  d tell
Once again we can predict that thought and talk about deontic ought is nonmonotonic:
in the present case, an additional bit of information that licenses the derivation of
a contrary-to-duty obligation—Jones’s obligation not to tell his neighbors that he is
coming—defeats one’s commitment to a conflicting obligation—Jones’s obligation to
tell his neighbors that he is coming. The validity of deontic detachment and thus of
the move from () and () to () is compatible with the intuition that () and () no

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

dynamic foundations for deontic logic 

longer license the inference of () under the additional assumption that Jones does
not go.
All of this is good news, but there remains the question why reasoning with violable
obligations is susceptible to monotonicity failures, and it is not clear that the story told
here gives quite the right answer. For notice that the current framework predicts that
d φ  e φ for all φ ∈ L, and so the reason for why ¬go defeats d tell is that
¬go defeats d go. But of course we want to say that Jones ought to go regardless of
whether or not he goes, and so the treatment of the case under consideration is not
fully satisfying. Let me outline how the basic framework may be expanded so that we
can do better (see Willer,  for additional discussion).
I rely on Frank’s () proposal that deontic ought requires that the input context
be “non-trivial” in the sense that its prejacent as well as its negation must be open
possibilities. This requirement was satisfied in all the cases we considered earlier but
it naturally becomes an issue when we look at cases in which obligations are violated.
Frank suggests that nontriviality violations result in the retraction of information to
arrive at an appropriate state. How to retract information from a state is a very complex
issue, and here I choose a very simple approach and assume that context associates
with each σ ∈  a system of spheres S(σ ) that is ordered by ⊆ and centered on σ . The
intuitive role of S(σ ) is to capture which commitments stand and fall together, and it is
then possible to define a downdating operation on carriers of information as follows:
Definition  (Downdating). Consider arbitrary σ ∈  and φ ∈ L. S(σ ) ◦ φ = {σ  ∈
S(σ ) : σ   φ}. The result of downdating σ with φ, σ ↓ φ, is the minimal element of
S(σ ) ◦ φ in case S(σ ) ◦ φ = Ø, and σ otherwise.
Downdating σ with φ removes any commitment to φ by weakening σ to its minimal
revision that is no longer committed to φ. Notice that downdating idles whenever the
input state already fails to be committed to φ.
The modified proposal then is that deontic ought is a universal quantifier over the
set of possible worlds that are deontically optimal in light of a carrier of information
that leaves room for the prejacent as well as its negation to be a possibility. Precisely:
Definition  (Deontic Ought with Retraction). Extend the update rules for L with
the following entry:
() σ [d φ] = {i ∈ σ : (σ ↓ φ ↓ ¬φ)d  φ}
Downdating thus guarantees that we consider an appropriate carrier of information
even if nontriviality is violated by the input state (as long as the prejacent is contingent).
It is easy to see that the refined proposal avoids the unfortunate result that any
update with ¬φ defeats a commitment to d φ. For consider again the distri-
bution of truth-values across possible worlds from the beginning of this section, let
sσ = {w , w , w , w } and consider σ [¬go]: then downdating with the prejacent of
d tell or d ¬tell idles and so σ [¬go]  d ¬tell. But downdating with the

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 malte willer

prejacent of d go or d ¬go re-introduces some indices at which Jones does go to

the help of his neighbors since σ [¬go]  ¬go. In particular, observe that on the
most natural conception of the fallback relation, σ [¬go] ↓ go = σ , the intuitive
idea being that if Jones were to go, he might and might not tell his neighbors he is
coming. But if that is right, then for any i ∈ (σ ↓ go ↓ ¬go)d , wi = w . Given
minimal assumptions about the fallback relation that figures in downdating, we can
thus predict that, assuming that Jones does not go, he ought not tell his neighbors that
he is coming, but he (still) ought to go.
The revised story is not entirely complete since it is currently incompatible with
the validity of deontic detachment: as before σ [¬go]  d ¬tell, but now we also
have σ [¬go]  d go and σ [¬go]  go ⇒ d tell. So if this were the final word,
there would be no nonmonotonic story to be told here in the first place; however, it is
a familiar assumption that conditionals in the indicative mood (such as the one that
figures in Chisholm’s case) presuppose that their antecedent is compatible with the
input context:
Definition  (Presupposition). For all σ ∈ : σ [φ ⇒ ψ] is defined iff σ [φ] = ∅.
Here presupposition is modeled as a definedness condition on updating.20 In these
lights, the notion of logical consequence should be refined as follows (see Starr, 
and also Beaver,  and von Fintel, ):
Definition  (Logical Consequence with Presupposition). φ , . . . φn  ψ iff for all
σ ∈  : if σ [φ ] . . . [φn ][ψ] is defined, then σ [φ ] . . . [φn ]  ψ.
Logical consequence remains understood as preservation of rational commitment, but
we now set aside those input states for which updating with the premises and then with
the conclusion is undefined. This is all we need for the final proposal.
Observe that while the commitment to the claim that Jones ought to go is not sen-
sitive to the epistemic possibility that Jones in fact goes, the conditional obligation is:
Fact  Consider d as fixed for Chisholm’s scenario: then for all σ ∈  it holds that
if σ  go ⇒ d tell, then σ  e go.
The hypothesis then is that the phenomenon of nonmonotonicity in Chisholm’s sce-
nario can once again be explained in terms of the sensitivity of deontic commitments
to the existence of epistemic possibilities: the conditional obligation licensing the
inference to the conclusion that Jones ought to tell his neighbors that he is coming
presupposes that Jones might in fact go to the help of his neighbors.
Earlier I said that just as reasoning with prima facie obligations is sensitive to
the absence of information, so is reasoning with violable obligations. The informal

20 See Heim () and, among others, von Fintel (), Gillies (, ), and Starr () for the
presuppositional analysis of conditionals. The basic idea goes back to Stalnaker ().

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

dynamic foundations for deontic logic 

proposal, remember, was that such obligations may entail other obligations unless they
are violated, and I can now make this more precise. Say that an inference presupposes
φ just in case any state for which updating with its premises and its conclusion is
defined is committed to φ: then an inference licensed by deontic detachment—the
rule allowing the step from () and () to ()—presupposes that the satisfaction of
the unconditional obligation is compatible with the input state. It follows that even
if deontic detachment is valid, the inferences it licenses may be defeated through
information agglomeration.
On the view I have proposed here, then, an obligation [[φ]] is violated but binding
in σ just in case σ  ¬φ ∧ d φ. Such cases differ sharply from those in which an
obligation [[φ]] is defeated in σ , which are such that σ  ¬d φ yet for some σ  so
that σ ⊆ σ  , σ   d φ. And intuitively, whenever [[φ]] is a prima facie obligation
conditional on [[ψ]] but defeated in σ , this is because σ depicts an exceptional
circumstance just in case σ  ψ ∧ ¬d φ.
So in our Chisholm-style scenario, Jones violates his obligation to go but his obli-
gation to tell his neighbors is defeated and thus no longer binding. And once again we
have identified a nonmonotonic effect in deontic discourse and reasoning that cannot
be explained by appealing to the fact that certain obligations allow for exceptions. To
see this, suppose an alternative deontic context d providing the following ordering
source consisting of defeasible rules:
od = {[ > go], [go > tell], [¬go > ¬tell]}
Then clearly any σd treats [[go]] as a W-default, [[tell]] as a [[go]]-default, and
[[¬tell]] as a [[¬go]]-default. The simple observation then is that σ [¬go]  d go,
which is just to say that σ [¬go] does not depict an exceptional circumstance in which
the default obligation to go fails to be binding—this just captures the intuition that we
should not collapse violations and exceptions. And while the prima facie obligation
[[tell]] conditional on [[go]] is defeated in σ [¬go], this is not so because σ [¬go]
depicts an exceptional circumstance either since σ [¬go]  go—this just captures the
intuition that a prima facie obligation conditional on [[go]] fails to pertain to situations
in which Jones does not go, and so a scenario in which Jones does not go hardly
qualifies as an exceptional circumstance. That certain obligations allow for exceptions
does not explain why reasoning with violable obligations has a nonmonotonic flavor.
The fact that the framework developed here has no trouble accounting for the
connection between defeasibility and violability demonstrates once again that it
captures a rich variety of nonmonotonic effects in discourse and reasoning without
reducing them to exceptional circumstances. While providing a promising account
of violable obligations requires adding some non-trivial complications to the basic
dynamic account told in §, its success at fulfilling a range of key desiderata for a
nonmonotonic story about reasoning with violable obligations suggests that the story
is on the right track.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 malte willer

 Conclusion
Deontic discourse and reasoning, so a familiar story goes, is nonmonotonic, but the
variety of ways in which it is sensitive to the absence of information has traditionally
been underappreciated. The traditional focus on prima facie obligations overlooks
that information agglomeration may trigger nonmonotonic effects without creating
contexts in which a prima facie obligation no longer applies, and I have highlighted
two reasons for why this is so: the fact that obligations may be violated and the fact
that obligations may be sensitive to epistemic uncertainty. The analysis developed
here differs from what has happened before in the literature in that it offers a
comprehensive perspective on the nonmonotonic nature of deontic discourse and
reasoning, and it does so by translating the sensitivity of deontic inferences to the
absence of information into the sensitivity of deontic ought to epistemic might. Since
information agglomeration may defeat commitments to epistemic might, it may also
defeat commitments to deontic ought: the nonmonotonicity of deontic thought and
talk can be illuminated in terms of the familiar nonmonotonicity of epistemic thought
and talk that finds a natural articulation in dynamic semantics.
I have remained silent on another popular motivation for nonmonotonic
approaches to deontic logic: the possibility of genuine moral dilemmas. This is
deliberate, for even if deontic logic must leave room for such dilemmas there is some
reason to think that their existence is compatible with a monotonic outlook (see,
e.g., van Fraassen, ; Cariani, ; and von Fintel, ). But let me briefly come
back to the question of whether attempting to arrive at a suitably nonmonotonic
logical consequence relation puts an explanatory burden on semantic theorizing that
is better reserved for some other component of a complete story about meaning,
communication, and reasoning.
The question connects with the issue whether dynamic semantics has any advan-
tages over a truth-conditional alternative coupled with an adequate pragmatic story
about conversational dynamics (see Rothschild and Yalcin (forthcoming) for discus-
sion). To get the issue into better view, suppose we recursively define truth-conditions
for our target language relative to an index i and information state σ that deliver the
following result for all φ of our target language: if i ∈ σ , then [[φ]]i,σ is true just in case
i ∈ σ [φ]. We may then define two distinct notions of logical consequence:
. Neoclassical Consequence: φ , . . . , φn  ψ iff for all i and σ such that i ∈ σ :
if [[φ ]]i,σ is true and . . . and [[φn ]]i,σ is true, then [[ψ]]i,σ is true.
. Informational Consequence: φ , . . . , φn  ψ iff e ψ is a neoclassical
consequence of e φ , . . . , e φn .
These semantic entailment relations have been extensively discussed in the context
of the miners paradox: neoclassical consequence avoids the problem by denying
modus ponens (see Kolodny and MacFarlane, ) while informational consequence
does so by allowing for reductio failures (see Bledin,  and Yalcin, ). But

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

dynamic foundations for deontic logic 

what about Ross- and Chisholm-style scenarios, which stand at odds with the fact
that neoclassical and informational consequence are monotonic by design? The most
promising strategy is to appeal to the already familiar story about conversational
dynamics from Stalnaker (): additional information in discourse and reasoning
affects, as a matter of pragmatics, the informational parameter in light of which
subsequent utterances are evaluated. More precisely, say that σ + φ = σ ∩ [[φ]]σ .
Then we may define a pragmatic inference relation (inspired by Stalnaker, ) that
treats the inference of ψ from φ , . . . φn as reasonable (rather than valid) just in case
for all i and σ : if i ∈ σ + φ + . . . φn , then i ∈ [[ψ]]σ +φ +...φn . Given the striking
similarity between reasonable inference and dynamic logical consequence, it should
not be surprising that the former is just as nonmonotonic as the latter is: epistemic and
deontic commitments are, as we have seen, defeasible and hence a sentence may be true
in light of some carrier of information but fail to be true with respect to some stronger
state. We may thus leave it to pragmatics to explain why additional information in
discourse and reasoning can defeat existing deontic commitments, and of course this
shift of explanatory burden would live happily with the suggestion that belief revision
theorists, rather than those concerned with semantic entailment, should worry about
nonmonotonicity. Why let semantics do all the work?
Let me make two points in response. First, many key aspects of the previous
discussion remain unaffected by the decision on how to divide the labor between
semantics and pragmatics (or between logic and belief revision theory). Specifically,
the point remains that the nonmonotonicity of deontic discourse and reasoning stands
in need of an explanation that goes beyond the one provided by theories that take
their inspiration from default logic, and that such an explanation can be provided by
a possible worlds semantics for conditionals and deontic modals that takes dynamic
effects in discourse and reasoning seriously. Furthermore, I contend that any such
explanation needs a model of how norms are sensitive to the flow of information along
the lines I have provided in this paper. These aspects of my story, as well as the key
ideas behind their formal elaboration, remain in place regardless of the semantics-
pragmatics distinction.
Playing a bit more offense, the choice of a monotonic consequence relation has
repercussions that go beyond the need to put substantial explanatory burden onto
a pragmatic story about information dynamics. Most strikingly, any monotonic con-
ception of logical consequence will inevitably be at odds with the semantic fact that
conditionals are nonmonotonic in the antecedent (resist antecedent strengthening),
and in particular we have cases in which   ψ yet   φ ⇒ ψ for some
φ, ψ ∈ L+ and  ⊆ L+ .21 But of course on a monotonic conception of logical
consequence, , φ  ψ whenever   ψ, which is just to say that any such
conception must flat out reject the semantic equivalent of the deduction theorem:

21 To choose an equivalent formulation that carries strengthening failures on its sleeves:    ⇒ ψ

yet   ( ∧ φ) ⇒ ψ. One example would be the premises of the miners paradox.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 malte willer

if , φ  ψ then   φ ⇒ ψ. As Yalcin (, footnote ) observes, this is a

problematic feature of neoclassical consequence—it disconnects conditionals from
consequence in unexpected ways—but it is in fact shared by all monotonic entailment
relations, including informational consequence. And of course this is just what we
expect if we ban dynamic effects from the semantic entailment relation but also
follow the standard protocol of adopting a Ramsey-inspired semantics for conditionals
that evaluates such constructions by evaluating their consequents in light of the
result of updating some carrier of information with the antecedent. Only a dynamic
entailment relation along the lines I have suggested here preserves the intuitive match
between the semantic evaluation procedure for conditionals and the one for logical
arguments. And since this conception of logical consequence essentially relies on
the assumption that semantic values are relations between information states rather
than indices of evaluation, we also have a point here at which a dynamic perspective
on semantic theorizing differs non-trivially from a truth-conditional alternative that
pushes dynamic effects into the pragmatics.22
There is, to be clear, some legitimacy to a monotonic perspective on discourse
and reasoning. Specifically, there is every reason to think that logical consequence
should be insensitive to the absence of information if validity amounts to guaranteed
preservation of truth at a point of evaluation, and in fact dynamic logical consequence
turns out to be monotonic in case we restrict attention to the propositional fragments
of our target languages. The point here is that validity as guaranteed preservation of
truth is just a special instance of the more general conception of logical consequence as
guaranteed preservation of rational commitment, and this connects with the semantic
fact that truth-conditional meaning is just a special instance of meaning understood
as context change potential.23 Just like the classical conception of meaning as truth-
conditional meaning, monotonicity has some place in our best theory about discourse
and reasoning—the dynamic perspective on meaning and communication, so the
moral of the story, makes sense of the scope as well as the limit of the classical view.
The comprehensiveness of the framework developed here does not reduce to the
fact that it has something nonmonotonic to say about thought and talk pertaining to
obligations that may be overridden, violated, or are sensitive to epistemic uncertainty.
It showcases what these cases have in common and where they differ, and in addition
assigns the proper place to important ideas from the classical literature. The fact
that we have every reason to take seriously a nonmonotonic perspective on deontic

22 For sure, one may rely on CCPs in semantic theorizing and just arrive at a framework that makes the
same predictions about consistency and entailment as its truth-conditional alternative (see von Fintel and
Gillies, ). Such proposals arguably fail to be interestingly dynamic, but clearly the one I have developed
here is not one of them.
23 The underlying observation from van Benthem () is that updating with an element of L is always

mediated by a classical proposition. This is no longer the case once we consider modal formulas and state
their semantics in terms of acceptance conditions that are sensitive to global features of the input state.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

dynamic foundations for deontic logic 

discourse and reasoning is well-established. I submit that we have just as much reason
to take seriously a dynamic perspective on deontic discourse and reasoning.

Acknowledgments
Key ideas of this paper were presented at a workshop on nonstandard modals at the University of
Leeds (organized by Paolo Santorio). Special thanks to the audience for insightful comments and
discussion, as well as to Fabrizio Cariani, Nate Charlow, Matthew Chrisman, Frank Veltman,
and an anonymous peer commentator for detailed feedback that proved to be invaluable
in preparing the final manuscript. Support from the Franke Institute for the Humanities is
gratefully acknowledged.

References
Åqvist, Lennart () Deontic Logic. In Gabbay, Dov M. and Guenthner, Franz (eds.) Hand-
book of Philosophical Logic. nd Edition. Volume . pp. –. Dordrecht: Kluwer.
Asher, Nicholas and Bonevac, Daniel () Prima Facie Obligation. Studia Logica: An Interna-
tional Journal for Symbolic Logic. . . pp. –.
Asher, Nicholas and Bonevac, Daniel () Common Sense Obligation. In Nute, Donald E.
(ed.) Defeasible Deontic Logic. pp. –. Dordrecht: Kluwer.
Beaver, David I. () Presupposition and Assertion in Dynamic Semantics. Stanford, CA: CSLI
Publications.
Belzer, Marvin () Reasoning with Defeasible Principles. Synthese. . . pp. –.
Bledin, Justin () Logic Informed. Mind. . . pp. –.
Bonevac, Daniel () Against Conditional Obligation. Noûs. . . pp. –.
Cariani, Fabrizio () ‘Ought’ and Resolution Semantics. Noûs. . . pp. –.
Cariani, Fabrizio () Deontic Modals and Probabilities: One Theory to Rule Them All? In
Charlow, Nate and Chrisman, Matthew (eds.) Deontic Modality. Oxford: Oxford University
Press.
Cariani, Fabrizio, Kaufmann, Magdalena, and Kaufmann, Stefan () Deliberative Modality
under Epistemic Uncertainty. Linguistics and Philosophy. . . pp. –.
Carmo, José and Jones, Andrew J. I. () Deontic Logic and Contrary-to-Duties. In Gabbay,
Dov M. and Guenthner, Franz (eds.) Handbook of Philosophical Logic. nd Edition. Volume .
pp. –. Dordrecht: Kluwer.
Carr, Jennifer () Deontic Modals without Decision Theory. In Chemla, E., Homer, V., and
Winterstein, G. (eds.) Proceedings of Sinn und Bedeutung . École normale supérieure, Paris,
– September. pp. –.
Charlow, Nate (a) Conditional Preferences and Practical Conditionals. Linguistics and
Philosophy. . . pp. –.
Charlow, Nate (b) What We Know and What To Do. Synthese. .. pp. –.
Charlow, Nate () Decision Theory: Yes! Truth Conditions: No! In Charlow, Nate and
Chrisman, Matthew (eds.) Deontic Modality. Oxford: Oxford University Press.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 malte willer

Chisholm, Roderick M. () Contrary-to-Duty Imperatives and Deontic Logic. Analysis. . .
pp. –.
Dancy, Jonathan () Ethics Without Principles. New York, NY: Oxford University Press.
Dowell, Janice L. () Contextualist Solutions to Three Puzzles about Practical Conditionals.
In Shafer-Landau, Russ (ed.) Oxford Studies in Metaethics. Volume . pp. –. New York,
NY: Oxford University Press.
Fintel, Kai von () The Presupposition of Subjunctive Conditionals. In Sauerland , Uli and
Percus, Orin (eds.) The Interpretive Tract (MIT Working Papers in Linguistics ). pp. –.
Cambridge, MA: MITWPL.
Fintel, Kai von () NPI Licensing, Strawson Entailment, and Context Dependency. Journal
of Semantics. .. pp. –.
Fintel, Kai von () The Best We Can (Expect to) Get? Challenges to the Classic Semantics
for Deontic Modals. Manuscript, MIT. [Online] Available from: http://mit.edu/fintel/fintel-
-apa-ought.pdf. [Accessed  Nov .]
Fintel, Kai von and Anthony S. Gillies () An Opinionated Guide to Epistemic Modality.
In Gendler, Tamar and Hawthorne, John (eds.) Oxford Studies in Epistemology. . pp. –.
New York, NY: Oxford University Press.
Frank, Anette () Context Dependence in Modal Constructions. Ph.D. thesis. University of
Stuttgart.
Fuhrmann, André () Reflective Modalities and Theory Change. Synthese. . . pp. –.
Gillies, Anthony S. () Epistemic Conditionals and Conditional Epistemics. Noûs. . .
pp. –.
Gillies, Anthony S. () On Truth-Conditions for If (but Not Quite Only If ). Philosophical
Review. . . pp. –.
Gillies, Anthony S. () Iffiness. Semantics and Pragmatics. . . pp. –.
Greenspan, Patricia S. () Conditional Oughts and Hypothetical Imperatives. The Journal of
Philosophy. . . pp. –.
Groenendijk, Jeroen and Stokhof, Martin () Dynamic Predicate Logic. Linguistics and
Philosophy. . . pp. –.
Harman, Gilbert () Change in View. Cambridge, MA: MIT Press.
Heim, Irene () The Semantics of Definite and Indefinite Noun Phrases. Ph.D. thesis. Univer-
sity of Massachusetts, Amherst.
Horty, John F. () Moral Dilemmas and Nonmonotonic Logic. Journal of Philosophical Logic.
. . pp. –.
Horty, John F. () Nonmonotonic Foundations for Deontic Logic. In Nute, Donald E. (ed.)
Defeasible Deontic Logic. pp. –. Dordrecht: Kluwer.
Horty, John F. () Reasoning with Moral Conflicts. Noûs. . . pp. –.
Horty, John F. () Reasons as Defaults. Philosophers’ Imprint. . . pp. –. [Online] Available
from: http://dx.doi.org//spo.... [Accessed  Nov .]
Horty, John F. () Reasons as Defaults. New York, NY: Oxford University Press.
Horty, John F. () Deontic Modals: Why Abandon the Classical Semantics? Pacific Philosoph-
ical Quarterly. . . pp. –.
Kamp, Hans, van Genabith, Josef, and Reyle, Uwe () Discourse Representation Theory. In
Gabbay, Dov M. and Guenthner, Franz (eds.) Handbook of Philosophical Logic. nd edition.
Volume . pp. –. Dordrecht: Springer.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

dynamic foundations for deontic logic 

Kolodny, Niko and MacFarlane, John () Ifs and Oughts. Journal of Philosophy. . .
pp. –.
Kratzer, Angelika () Modality. In von Stechow, Armin and Wunderlich, Dieter (eds.)
Semantics: An International Handbook of Contemporary Research. pp. –. Berlin: de
Gruyter.
Kratzer, Angelika () Modals and Conditionals. New York, NY: Oxford University Press.
Lance, Mark N. and Little, Margaret O. () Defeasibility and the Normative Grasp of
Context. Erkenntnis. . –. pp. –.
Levi, Isaac () Iteration of Conditionals and the Ramsey Test. Synthese. . . pp. –.
Lewis, David () Counterfactuals. Cambridge, MA: Harvard University Press.
Lewis, David () Semantic Analyses for Dyadic Deontic Logic. In Stenlund, Sören (ed.)
Logical Theory and Semantic Analysis: Essays Dedicated to Stig Kanger on His Fiftieth Birthday.
pp. –. Dordrecht: Reidel.
McCarthy, John () Circumscription—A Form of Non-Monotonic Reasoning. Artificial
Intelligence. . –. pp. –.
McCarty, J. and Thorne, L. () Defeasible Deontic Reasoning. Fundamenta Informaticae. .
–. pp. –.
McDermott, Drew and Doyle, Jon () Non-Monotonic Logic I. Artificial Intelligence. . –.
pp. –.
McNamara, Paul () Deontic Logic. In Gabbay, Dov M. and Woods, John H. (eds.) The
Handbook of the History of Logic, vol. : Logic and Modalities in the Twentieth Century.
pp. –. Amsterdam: Elsevier.
Nute, Donald E. () Defeasible Reasoning and Decision Support Systems. Decision Support
Systems. . . pp. –.
Nute, Donald E. () Apparent Obligation. In Nute, Donald E. (ed.) Defeasible Deontic Logic.
pp. –. Dordrecht: Kluwer.
Pollock, John L. () Defeasible Reasoning. Cognitive Science. . . pp. –.
Prakken, Henry and Sergot, Marek () Contrary-to-duty Obligations. Studia Logica. . .
pp. –.
Prakken, Henry and Sergot, Marek () Dyadic Deontic Logic and Contrary-to-Duty Obli-
gations. In Nute, Donald E. (ed.) Defeasible Deontic Logic. pp. –. Dordrecht: Kluwer.
Ramsey, Frank P. () The Foundations of Mathematics. London: Routledge and Kegan Paul.
Reiter, Raymond () A Logic for Default Reasoning. Artificial Intelligence. . –. pp. –.
Ross, William D. () The Right and the Good. Oxford: Oxford University Press.
Rothschild, Daniel and Yalcin, Seth (forthcoming) On the Dynamics of Conversation. Noûs.
Rott, Hans () Conditionals and Theory Change: Revisions, Expansions, and Additions.
Synthese. . . pp. –.
Ryu, Young U. and Lee, Ronald M. () Deontic Logic Viewed as Defeasible Reasoning. In
Nute, Donald E. (ed.) Defeasible Deontic Logic. pp. –. Dordrecht: Kluwer.
Silk, Alex () Truth Conditions and the Meanings of Ethical Terms. In Shafer-Landau, Russ
(ed.) Oxford Studies in Metaethics. Volume . pp. –. New York, NY: Oxford University
Press.
Stalnaker, Robert C. () Indicative Conditionals. Philosophia. . . pp. –.
Stalnaker, Robert C. () Assertion. In Cole, Peter (ed.) Syntax and Semantics . pp. –.
New York, NY: New York Academic Press.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 malte willer

Starr, William B. () A Uniform Theory of Conditionals. Journal of Philosophical Logic. ..
pp. –.
Thomason, Richmond H. () Deontic Logic as Founded on Tense Logic. In Hilpinen, Risto
(ed.) New Studies in Deontic Logic. pp. –. Dordrecht: Reidel.
Touretzky, David S. () The Mathematics of Inheritance Systems. Los Altos, CA: Morgan
Kaufmann.
van Benthem, Johan () Essays in Logical Semantics. Dordrecht: Reidel.
van der Torre, Leendert W. N. and Tan, Yao-Hua () The Many Faces of Defeasibility. In
Nute, Donald E. (ed.) Defeasible Deontic Logic. pp. –. Dordrecht: Kluwer.
van der Torre, Leendert W. N. and Tan, Yao-Hua () An Update Semantics for Deontic
Reasoning. In Proceedings of DEON. Bologna, January . pp. –. Amsterdam: IOS
Press.
van Fraassen, Bas () Values and the Heart’s Command. Journal of Philosophy. . . pp. –.
Veltman, Frank () Logics for Conditionals. Ph.D. thesis. University of Amsterdam.
Veltman, Frank () Defaults in Update Semantics. Journal of Philosophical Logic. . .
pp. –.
Willer, Malte () A Remark on Iffy Oughts. Journal of Philosophy. . . pp. –.
Willer, Malte () Dynamic Thoughts on Ifs and Oughts. Philosophers’ Imprint. . .
pp.: –. [Online] Available from: http://dx.doi.org//spo.... [Accessed 
Nov .]
Wright, von George H. () Deontic Logic. Mind. . . pp. –.
Wright, von George H. () A Note on Deontic Logic and Derived Obligation. Mind. . .
pp. –.
Yalcin, Seth () A Counterexample to Modus Tollens. Journal of Philosophical Logic. . .
pp. –.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i


Dynamic Expressivism about
Deontic Modality
William B. Starr

As a first approximation, expressivism about deontic discourse holds that an utterance

of Zoyd must share expresses certain motivational attitudes towards sharing, but with-
out referring to and describing, or representing, those attitudes—that would amount
to a form of subjectivism that makes disagreement difficult to capture (Moore, ).1
(I can hardly disagree with your descriptions of your attitudes.) Some philosophers
have sought to connect deontic discourse to motivational attitudes such as desire
because they think it is the only way of explaining how deontic language and thought
motivate us to do things: passive descriptions of things out there in the world lack
the requisite motivational gravitas (Mackie, ). So, the aim of expressivism is
to ground deontic discourse in motivational psychology without making its subject
matter motivational psychological states.2 From this perspective, expressivism is only
distinctive and interesting to the degree that motivational and representational psy-
chological states are distinguishable. I believe this presents a dilemma for expressivists.
If representational thought and talk isn’t substantially different from its motivational
counterparts, then expressivism becomes elusive and uninteresting. If representational
thought and talk is substantially different from its motivational counterparts, the
existing tools of truth-conditional semantics tailored for representational purposes
should not work for motivational thought and talk. After all, if motivational attitudes
are quite different one would expect their kinematics, hence their expression and hence
their semantics to also differ. It seems to follow that if expressivism is an interesting
and distinctive position, it faces the Frege-Geach problem. It is not clear, as Frege
() stressed, what logical compounds of non-representational sentences mean and,

1 Note that even on this approximation, expressivism should not be equated with a non-propositional
semantics. A non-propositional semantics, e.g. Hamblin’s () for interrogatives, needn’t involve motiva-
tional attitudes. Further, one could offer a propositional expressivist semantics as long as one had a story
about how those contents are connected to motivational attitudes rather than representational ones. For
discussion see Chrisman (), Schroeder (), Ridge (), and Charlow ().
2 For a more detailed version of this narrative see Schroeder (a, Chapters –).

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 william b. starr

as Geach () stressed, it is not clear how non-representational sentences could bear
relations of consequence and consistency.
One goal of this chapter is to press this dilemma against recent expressivists that use
the tools of truth-conditional semantics to model expressive meaning and communi-
cation (Gibbard, ; Dreier, ; Yalcin, , ; Silk, ) (§§–). I will argue
that these accounts either rely on representational concepts that are inappropriate
for expressive motivational discourse, or they end up blurring the distinctions that
make expressivism a distinctive and interesting thesis. In doing so, I will propose
a new way of decomposing the negation problem (§) and argue that expressivism
cannot succeed without a distinctive theory of expressive communication (§). My
other goal is more positive. I will assume for the purposes of the argument set
out that representational language inherits its distinctive semantic properties from
the function of representational attitudes: to track how the world is. Further, I will
assume that motivational attitudes have a very different function, namely to motivate
choice and action, and that motivational language inherits rather different semantic
properties from this motivational function. This is not so much because I am confident
in these assumptions, but because they provide the most interesting and challenging
setting to defend expressivism. Against this backdrop, I will use dynamic semantics, a
relatively recent semantic framework discussed in computer science and linguistics,
to offer a model of expressive and representational language where both kinds of
discourse are seamlessly integrated, and where uniform explanations of composition-
ality and logical relations are forthcoming (§§–). Dynamic semantics provides an
algebra of processes rather than contents. I will show how this embodies a more
general conception of meaning and logic that yields motivational and representational
meaning as special cases.3
Here at the outset, I should clarify that I will not be offering a positive linguistic
or philosophical argument in favor of expressivism. Yet, linguists and empirically-
minded philosophers of language may find something of interest here. The empirical
investigation of ‘expressive meaning’ is a burgeoning field, whose object of study has
remained somewhat difficult to classify and integrate in compositional systems.4 The
perspective on expressivism, communication and compositional semantics offered
here should be useful for these purposes. The semantics developed here also makes
interestingly different predictions than more orthodox approaches (e.g. Kratzer,
), thereby connecting to a broader debate over non-truth-conditional semantics

3 This emphasis on processes (acceptance, rejection) rather than content is a theme in Gibbard (),
but is not embraced in Gibbard’s formalism. Dynamic semantics provides the tools for doing so. Alwood
() pursues a similar dynamic approach and situates it in more detail with work in metaethics. I aim
to provide a more thorough technical implementation here and connect such an approach to the issue of
communication in expressivism. One virtue of this implementation is not having to distinguish, as Alwood
() must, between two kinds of negation. I will offer one semantics for negation which models both
propositional and expressive negation.
4 E.g. evidentiality (Murray, ; Faller, ), attention (Bittner, ), narrative structure (Hobbs,
), slurs (Anderson and Lepore, ) and other phenomena (Potts, ; McCready, ).

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

dynamic expressivism about deontic modality 

(Veltman, , p. ; Portner, , p. )5 and expressivist pragmatics (Yalcin, ,
) for modality. Throughout, I will be assuming that expressivism is a semantic
thesis about the linguistic meaning of sentences which challenges the idea that all
communication proceeds by referring to things and describing them. Along the way
(§), I will present difficulties for the idea that expressivism is merely a pragmatic
thesis (Yalcin, , ). Unfortunately, I will not have space to discuss important
recent work that pitches expressivism as a metasemantic thesis (Chrisman, ;
Ridge, ) about how certain expressions get their traditional meanings, or what
it means to ascribe those traditional meanings to those expressions (Carballo, ).
It is interesting to note, however, that the linguistic meanings ascribed in my dynamic
approach are abstract objects that model the processes which Lewis () and Grice
() took to endow sentences with their contents. However, in taking these processes
themselves to be the meanings the dynamic approach arrives at additional semantic
resources for solving the Frege-Geach problem.6 Only further, more general, work
on metasemantics will reveal the comparative virtues of semantic and metasemantic
approaches.

 The Virtues of Dynamic Expressivism: a Sketch

Subsequent sections present dynamic expressivism and its alternatives in full technical
detail. This section offers a less formal sketch of the positions and issues with the hope
that this will bring order to the wealth of detail that is to follow. I will start with a simple
model of the role representational and motivational states play in rational agency. Then
I will use this framework to articulate deontic expressivism and show how it leads to
the Frege-Geach problem. This will allow me to say what dynamic expressivism is and
how it is better suited to solve this problem than accounts which use the tools of truth-
conditional semantics.
According to a familiar, if limited, model of rational agency an agent’s beliefs
are modeled with a probability space and their values with a utility function. The
probability space measures how likely they take certain propositions to be and the
utility function measures how valuable they take certain states of affairs to be. Agents
make rational choices by selecting those actions that lead to outcomes with the highest
‘expected utility’: the value of an outcome weighted by how likely the agent takes it to
be. It is crucial to note that beliefs and values play quite distinct theoretical roles in
this model: agents’ decisions are primarily driven by what they value, and tempered
only by what they believe. An agent may strive for a very unlikely outcome if it is far
more valuable than the other options. Further, if they had no values they would have

5 In an earlier version of this work I adopted a semantics much closer to Veltman (, p. ) and
Portner (, p. ). The new semantics is now a bit of a hybrid between that semantics, one from dynamic
logic (van Benthem and Liu, ) and Starr’s () dynamic semantics for imperatives.
6 This approach has much in common with Millikan (, ).

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 william b. starr

no basis for choice, but if they had no beliefs they could still (naively) choose actions.
These different theoretical roles justify rather different assumptions about probabilities
and utilities. If an agent has credence . that Zoyd shared, then that agent is assumed
to have credence . that Zoyd didn’t share. By contrast, if Zoyd’s sharing has utility 
for an agent, it is not assumed that Zoyd’s sharing gets utility t −  where t is the total
amount of utility actions can have for the agent. The assumption for probabilities is
reasonable since they capture how the agent represents the world: in the world either
Zoyd shared or he didn’t. This assumption is not plausible for utilities because they
don’t represent the world.
Against this backdrop, an expressivist might say that utterances of Zoyd must share
express utility functions. Perhaps functions where Zoyd’s sharing is assigned a high
utility relative to some standard. In simple cases, this might serve to bring one agent’s
values in line with another’s. By contrast, the subjectivist would say that deontic
utterances of Zoyd must share express propositions about utility functions. In simple
cases, these utterances would serve primarily to align agents beliefs (i.e. probability
spaces). Though this might result in agents’ aligning their values, that’s not the primary
function of deontic discourse according to the subjectivist. A third position, call it
simple naturalism, maintains with subjectivism that deontic utterances express beliefs
about the world. But it rejects that these beliefs are about utility-relative facts: they
are simple beliefs about the utility-independent world. Seeing how the Frege-Geach
problem arises in this setting will allow me to say what dynamic expressivism is and
how it differs from other accounts.
Frege (, ) famously proposed the standard explanation of how represen-
tational content behaves in a compositional way. But, less famously, Frege () also
contended that those explanations do not work for non-representational meaning. The
argument was simple: the meaning of negation is exhausted by its role of turning a true
content into a false one and vice versa. This explains why any sentence and its negation
are inconsistent, i.e. cannot both be true. Here’s the problem for expressivists: if (a)
does not represent the world, then how does negating it do anything meaningful, let
alone inconsistent with (a)?
() a. Zoyd must share.
b. It’s not the case that Zoyd must share.
This is the Frege-Geach Problem, whose title reflects its application by Geach ()
to non-representational theories of moral language.7 To solve this problem, it seems
that expressivists must say what non-representational attitude (a) expresses, how the
attitude expressed by (b) is generated by negating the expression of an attitude by (a)
and what it is for two non-representational attitudes to be inconsistent. Parallel issues
arise with other connectives and explaining entailment patterns.

7 See also Schroeder (d) and references therein.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

dynamic expressivism about deontic modality 

One can see the challenge more precisely from within the simple model of rational
agency above. There is a relationship between a rational agent’s credence in Zoyd
shared and Zoyd didn’t share: no credence should be assigned to the possibility of
both being true. This, as mentioned above, follows from background assumptions
about truth, representation and the functional role of beliefs—more on that in §.
It is less obvious that such a relationship exists between how much a rational agent
values Zoyd sharing and how much they value Zoyd not sharing. One can certainly
imagine a scenario where both outcomes have a high utility for the agent. Further, the
relationship between how an agent values Zoyd sharing and how they value Zoyd not
sharing speaks to the consistency of (a) and (b), rather than (a) and (b).8
() a. Zoyd must share.
b. Zoyd must not share.
Although (b) intuitively entails (b), the opposite is not true. Note that it is only
consistent to follow up (b) with but Zoyd may share. This leads to an important
observation: there are really two ‘negation problems’ for the expressivist. Suppose one
can successfully analyze (a) and (b) as expressing ‘motivationally conflicted’ values:
being in favor of sharing and being in favor of not sharing. This would be a solution to
the internal negation problem. Since one can endorse (b) without endorsing (b), an
analysis of (b) is still needed. This leaves open the possibility that a different account
of the conflict between (a) and (b) will be needed. Call this the external negation
problem. I believe the two problems require different solutions. While this point will
be argued for more precisely later, the basic idea can be articulated with the resources
currently on hand. It will then serve as a basis for contrasting dynamic expressivism
with others (Gibbard, ; Dreier, ; Yalcin, , ; Silk, ).
Even if (a) and (b) express motivationally conflicted states of mind, it is hard
to see how the same is true of (a) and (b). Initially, this is because it is hard to see
how a sentence with external negation expresses a particular motivational attitude at
all. If one thinks of Zoyd must share as serving to instill pro-sharing utilities in the
hearer, then It is not true that Zoyd must share rejects instilling those utilities. But
what particular utilities does this rejection promote? I believe that focusing on this
question gets to the heart of the problem for expressivists. Previous attempts to say
what attitude is expressed have failed to render (a) and (b) inconsistent in a way that
is clearly non-representational.
Gibbard (, pp. –) offers the following expressivist model:
() a. ‘You must share’ expresses: Agreeing with sharing
b. ‘You must not share’ expresses: Disagreeing with sharing
c. ‘It is not the case that you must share’ expresses: Disagreeing with agreeing
with sharing

8 Unwin () highlights the importance of this distinction in the context of expressivism.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 william b. starr

This model does not say anything more about what disagreeing and agreeing are which
explains why agreeing with X and disagreeing with X are in rational tension. In the
setting above, this amounts to saying that You must share and It is not the case that you
must share express conflicting utility functions without saying what utility functions
these are and how exactly those functions or their acceptance conflict. Some have
worried whether such a semantic theory is adequately explanatory (e.g. Schroeder,
a, d). § will press that concern in another way by arguing that expressivists
need to offer a positive account of how expressive communication works and how
it differs from representational communication. It is hard to see how this could work
without characterizing representational and expressive communication as importantly
different processes.
Building on Dreier (, ), Silk () develops what might be called prefer-
ence expressivism, which formulates a deontic semantics in terms of preference order-
ings. In Starr () I pursue roughly the same strategy to explain the incompatibility
of contrary imperatives such as Dance! and Don’t dance!, without assuming that imper-
atives are representational. This aims to fill the gaps in Gibbard’s approach. The basic
idea is that deontic sentences express preferences. Much like utilities, the function of
preferences is to motivate choice.9 Crucially, this means that preferences which prevent
a choice will be dysfunctional. The idea is then that two deontic sentences can be
inconsistent if they express preferences which, taken together, are dysfunctional in this
way. A theory along these lines offers an interesting expressivist solution of the internal
negation problem. But I do not think it can solve the external negation problem.
A (strict) preference for you sharing could be modeled as an ordering which ranks
any world w where you share over every world w where you don’t: w  w . Now,
consider a semantics on which:
() a. ‘You must share’ expresses: w  w
b. ‘You must not share’ expresses: w  w
For now, set aside how the meaning for (b) is determined from the meaning of (a)
and a semantics for negation. Here is how one might explain the inconsistency of
these two sentences on such a semantics.10 The first step is to determine what it is to
take these two preference orderings ‘together’. Unioning them into a bigger preference
ordering is the most natural idea: w  w , w  w .11 It is a platitude about
choice that you shouldn’t choose A if you strictly prefer B to A and you can choose B

9 The earlier utility expressivist proposal was that Must(A) expresses utility functions which assign a
utility to A-worlds that is high according to some standard. By contrast, preference expressivism says: the
utility of A-worlds is greater than the utility of ¬A-worlds.
10 Neither Dreier (, ) nor Silk () explicitly formulates the decision-theoretic constraints on
preference and choice that they take to constitute preferential coherence. Following Starr (), I assume in
the main text above a non-dominance principle about choice and tie this to the need for acyclic preferences.
I do not know whether this is what Dreier (, ) and Silk () had in mind.
11 Note that  is the set {w , w } and  is the set {w , w }. The fact that union is the operation
     
appropriate to combining preferences is discussed by van Benthem and Liu (). This is fairly clear from
thinking about how to combine the orderings from Must(A) and Must(B). Intersection predicts that A ∧ B-
worlds, ¬A ∧ B-worlds and A ∧ ¬B-worlds are best, while union predicts that only A ∧ B-worlds are best.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

dynamic expressivism about deontic modality 

instead. This means that any cyclic strict preference ordering is irrational. Something is
preferred to every alternative, so no alternative can be rationally chosen. This is exactly
the situation in  . Since it fails to motivate a choice, it is motivationally defective. This
seems like exactly the kind of explanation of inconsistency that expressivists seek. But,
I claim, it does not work for the external negation problem.
Following standard accounts of modality, Silk () proposes that Must(A)
requires all the most preferred worlds to be A-worlds, while ¬Must(A) requires some
of the most preferred worlds to be ¬A-worlds. The issue is that taking together—
unioning—two preference orderings that meet these respective descriptions does not
always yield a dysfunctional one. To illustrate this, consider a model in which we
are concerned only with A and B worlds. Using a capital letter to indicate truth
and a lowercase letter for falsity, Must(A) can express an ordering such as  while
¬Must(A) can express an ordering such as  : wAB and wAb are the most preferred
worlds according to  , and waB is among the most preferred worlds according to  .
As the graphs illustrate, there is nothing dysfunctional about taking these preferences
together. Together, they will motivate the agent to choose wAB or wAb . These combined
preferences are not motivationally dysfunctional.

AB Ab AB Ab AB Ab

aB ab aB ab aB ab

1 2 1∪ 2
Must(A) ¬Must(A)

From the dynamic perspective, the preference expressivist went wrong in two
assumptions: (i) accepting ¬Must(A) is the same kind of process as accepting
Must(A) and (ii) consistency is a relationship between contents rather than processes.
On the dynamic account developed in §, the meaning of an expressive sentence is
characterized in terms of how accepting it changes a preference ordering. Must(A)
introduces a strict preference for all live A-worlds over ¬A-worlds, and tests that
doing so makes A a practical necessity: all the choices it motivates result in A-worlds.
(The success of this test will depend on which preferences have been previously
introduced.) ¬Must(A), on the other hand, does not introduce a preference at all,
so its inconsistency with Must(A) is not an inconsistency of preference. ¬Must(A)
changes a preference ordering by removing any preferences that would be added by
Must(A). Must(A) and ¬Must(A) are dynamically inconsistent: since one sentence
adds what the other removes there is no ordering which contains the effects both have
on a preference ordering. But that is crucially different from saying that they lead to
orderings which, combined in the standard way, are rationally defective. Must(A) and

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 william b. starr

¬Must(A) are inconsistent in the same way that certain actions are. Consider a robot
that rolls around a bar to each table. It fills an empty shot glass with vodka and it
vacuums up the contents of non-empty shot glasses. Then it cruises on. The action of
pouring a shot and the action of vacuuming one are incompatible in the sense that no
state of a given glass that directly results from the robot’s actions contains the effects of
both actions. But this is not because the effects themselves are incompatible: the empty
space in the shot glass and the shot of vodka are clearly compatible in that one can fill
the other. As will be discussed, a non-dynamic account does not make room for the
idea that φ and ¬φ have radically different communicative effects. On a non-dynamic
account, negation takes one content (e.g. preference-ordering or proposition) and
yields another of the same kind. Their communicative effects are modeled by a single
pragmatic process which combines contents of that kind (the previously accepted
contents incremented with the new content).
The dynamic approach to the external negation problem combines elegantly with
the preference expressivists’ approach to the internal negation problem. In §, I define
dynamic preferential inconsistency as follows: φ and ψ are dynamically preferentially
inconsistent just in case there is no single practically rational preference ordering that
contains the effects on preferences that both would have. In the case of Must(A)
and Must(¬A), there exists a preference ordering that contains preferences both
introduce, but it is irrational because those preferences are cyclic. Must(A) and
¬Must(A) are different: there is no preference ordering which contains the effects on
preferences that each sentence produces. If an ordering has the preferences Must(A)
adds, then updating with ¬Must(A) would remove them. If an ordering lacks the
preferences ¬Must(A) would remove, then updating with Must(A) would add them.
Formalizing the dynamic account requires a logic of actions rather than contents,
one which speaks to questions of how connectives combine actions and how actions
can bear logical relations to each other. Fortunately, logics of this kind—called dynamic
logics (Harel et al., )—have been extensively developed in computer science and
have inspired new semantic approaches to natural language (Heim, ; Kamp, ;
Groenendijk et al., ). The basic idea of this approach is to formally describe how
accepting each type of sentence changes an agent’s state of mind. Surprisingly, one can
reformulate classical logic in this way and then add non-classical operators to produce
operations that do not amount to operations on content (§). Hence, this is not an
alternative approach to semantics, but a more general one. By importing the dynamic
semantics for disjunction from Starr () below, I will also be able to navigate a chal-
lenge presented by Schroeder () for previous dynamic analyses and analyses such
as Yalcin’s (, ): they incorrectly predict that believing Must(A) ∨ B entails
either believing Must(A) or believing B.12 As I will emphasize, parallel observations
hold for what disjunctions communicate. This connection between an expressivist

12 Schroeder () makes a more general observation: this prediction holds even if both disjuncts are
expressive. The semantics in § also blocks this result.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

dynamic expressivism about deontic modality 

account of communication and solving the negation and disjunction problems is

surprising. As I’ve argued above, solving the negation problem requires Must(A) and
¬Must(A) to do very different communicative jobs. This is unlike A and ¬A, so
it seems to follow that the viability of expressivism relies crucially on the idea that
expressive communication works differently to representational communication. In
the following section, I argue that the need for a distinctive expressivist model of
communication follows from more general features of the expressivist program. This
will also allow me to consider what is distinctive about expressive communication.

 Expressivism and Communication

Philosophers who maintain that some bit of discourse does not work by describing
distinctions in the world often rally under the label expressivism, while those who
disagree rally under descriptivism. Defining expressivism could easily occupy a chapter
of its own, so I will not authoritatively impose such a definition on the discussion here.
Instead, I wish to consider three characteristic expressivist theses:13
Expressivist Theses
. Communication: ‘To express a state of mind is not to say that one is in it’ (Gibbard,
, p. ).
. Explanation: ‘The semantic properties of sentences are to be explained, funda-
mentally, in terms of properties of the attitudes conventionally expressed by
utterances of those sentences’ (Silk, , §).
. Non-representation: The states of mind expressed by sentences are non-
representational, and, more specifically, motivational.
As suggested above, the point of the explanatory thesis is to replace distinctions ‘in the
world’ with psychological distinctions, all while keeping the subject matter of deontic
discourse non-psychological. Viewed in this light, a surprising fact becomes apparent:
the explanatory expressivist thesis assumes the one about communication. Suppose
the communication thesis did not hold, and consider whether the explanatory one
could. Without the communication thesis, expressing one’s preference for sharing is
to describe oneself as preferring sharing. Then one’s attitude of preference is involved
in the same way as books are involved in the sentence Books are made from paper.
The semantic fact that You must share and You must not share are inconsistent is no
more explained by the properties of your preferences than the parallel inconsistency
between Books are made from paper and Books are not made from paper is explained by
the properties of books. In both cases, one needs things around—books, preferences—
and distinctions between them. So vindicating the explanation thesis is at best a partial

13 The quotation from Silk (, §) nicely captures the operative expressivist thesis at play in much
recent work including Gibbard (, p. ), Schroeder (c, p. ) and Dreier (, , p. ).

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 william b. starr

vindication of expressivism. Contemporary expressivists have done little to meet this

other important objective.14
As summarized in §, the account of Gibbard () declines to positively char-
acterize the content of ¬Must(A). But without such a characterization, it is not
possible to positively characterize how an agent’s state of mind changes when someone
communicates ¬Must(A) to them. What are the prospects for supplementing analyses
such as those found in Gibbard () and Silk () with an expressivist pragmatics
that clearly distinguishes expressing attitudes from expressing representational con-
tents about those attitudes? Yalcin (, §) and Silk () propose similar deontic
semantics and Yalcin (, §) sketches an expressivist pragmatics for his account.
On that account, expressive communication works in a way that is formally parallel
to representational communication. § will present a challenge for this account.
As suggested in §, to solve the negation problem one must reject the idea that
expressive sentences convey content in the same way as representational sentences.
This is because motivational contents combine and conflict in different ways than their
representational counterparts. But that suggests that a model which fits the expression
of motivational attitudes into the representational model is not the model expressivists
need. It is more naturally interpreted as a model of how one conveys information about
motivational attitudes. This is no accident, I claim. On a robust representational inter-
pretation of truth-conditional semantics, there is an essential reliance on reference.
All contents are therefore abstractions of referential relations. This means that one
cannot employ a robust interpretation of truth-conditional semantics, and use truth-
conditional semantics to either solve the Frege-Geach problem or give an expressivist
theory of communication.
Although the expressivist communication thesis is my focus, a word about the
explanatory thesis is in order.15 It is hard to say whether that thesis applies to the view
offered here since this view does not proceed by assigning sentences to the attitudes,
or states of mind, they express. Further, the semantic properties of all sentences could
be said to be explained in terms of properties of the mental states they update. For
example, in one sense of inconsistency, the inconsistency of two descriptive sentences is
explained by the fact that an update with both leads to a state of mind whose function is
to represent the world, but fails to do so.16 The inconsistency of two expressive deontic
sentences is ultimately explained by the fact that an update with both requires a state
of preference whose function is to motivate agents to choose one option over another,
but fails to do so. One can recover something in the vicinity of the explanatory thesis

14 Schroeder (b) also highlights this gap and explores several expressivist accounts of ‘expression’
which pair a sentence with the attitude it expresses. The expressivist accounts I’ll consider, like Yalcin (),
do not make use of such a pairing, and focus on the interpersonal relation of communication rather than
the individualistic notion of expression. So while Schroeder (b) makes important and related points,
its criticism is complementary to that presented here.
15 See also Silk () and Charlow () who discuss the explanatory thesis.
16 I will say more about what I take ‘explaining inconsistency’ to be in §.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

dynamic expressivism about deontic modality 

by replacing talk of sentences expressing mental states with updating mental states,
and focusing on explanations which appeal to the non-representational dimensions
of those states (e.g. attention, preference).
Dynamic Expressivist Explanatory Thesis Semantic properties of expressive
sentences are explained, fundamentally, in terms of the way they update non-
representational dimensions of language users’ mental states.
I’ve said a bit above about why expressivism is a thesis worthy of attention in
metaethics, but I think it is also important to say why it is worthy of attention in the
study of natural language meaning.
Expressivism may seem like an esoteric philosophical doctrine, but its communi-
cation thesis is familiar from commonsense observations. Suppose we’re looking at
the same field of wild flowers. I’ve noticed a bee fly on a nearby wild flower and I’m
attending to it. I point at it and say, that’s what a bee fly looks like. I was attending
to the bee fly and you are now attending to the bee fly. I’ve communicated my state
of attention to you. But I did not do so by referring to my state of attention and
stating a fact about it, i.e. I did not say something like My state of attention is thus-
and-so. The way I communicated my state of attention was very different than the way
I communicated the object of my attention, to which I referred with a gesture. Further,
one can imagine two agents’ having identical visual representations of a scene but their
attention to be centered on different objects. Since such a difference is likely to impact
actions where they coordinate, it seems that successful communication will involve
agents attuning their state of attention, despite the fact that they are not referring
to those states and talking about them. Indeed, given that agents seem capable of
communicating certain features of their mental states without even being aware of
them, expressivism in these domains seems inevitable. Attention, of course, is not the
only expressive dimension where this difference seems to matter. When I tell you
that bee flies lay their eggs in beetle nests and that after hatching their larvae feed
on their hosts’ eggs, you might express disgust at this parasitism with a grimace. You’ve
thereby communicated your disgust to me.17 But you did not do so by referring to your
state of disgust and stating a fact about it. You did not say anything like My state of
disgust is thus-and-so. Nothing in your facial expression seems to have brought your
affective state into the conversation in the way my pointing brought the bee fly into
the conversation, nor does anything in your facial expression describe your affective
state. These two examples illustrate the expressivist communication thesis at work in
a more general setting. In these cases it seems plausible enough. But neither example
involved essentially deontic language, or language at all: gestures were the expressive

17 Add to the story that I intended to do so if you are inclined to the Gricean orthodoxy that requires
an effect to be intended for it to be communicated. Note that the cost of such orthodoxy is positing a
higher-order mental state which involves the speaker attending to their state of attention. Expressivism
is naturally allied with the various attempts to do without the Gricean intellectualization of communication
and convention (e.g. Burge, ).

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 william b. starr

signals. Thus, an expressivist must say how this kind of communication operates in the
linguistic realm in general, and the deontic realm in particular. Note, however, that it is
not problematic for the expressivist to admit that deontic discourse does not appear to
be psychological. On an expressivist view, psychological states are behind the scene but
not on stage. In fact this is likely another advantage to be celebrated by the expressivist.
The proper characterization of the psychological states relevant to the interpretation
of modals requires theoretical sophistication beyond our everyday grasp of our own
minds, e.g. orderings or probability spaces.
I will now proceed by presenting an entirely standard truth-conditional semantics
for propositional logic and highlighting how its approach to logic, compositionality
and communication depends on reference (§). I then critically examine attempts to
use an extension of this framework to solve the negation problem and to provide an
expressivist account of communication in §. That discussion will propel us towards a
dynamics semantics. I will first present a dynamic semantics for propositional logic to
illustrate the basic ways in which it departs from classical semantics (§). I will then,
in §, provide a dynamic semantics for deontic discourse and explain how dynamic
expressivism tackles the difficulties for other approaches.

 Logical Semantics, Classically

Logical semantics has provided a model of linguistic meaning that is useful in three
ways. First, it accounts for how the meanings of complex sentences are built up from
the meanings of their parts. Secondly, it provides an account of why some sentences
follow from others and why some are inconsistent with others. Thirdly, it provides an
account of the basic information communicated by any use of the sentence. All three
of these applications rely on exclusively referential semantic concepts. That is, they rely
on the idea that linguistic expressions function by referring to and making distinctions
between things in the world. To be concrete but keep the formalism simple, I will
illustrate this with an intensional (possible-worlds) semantics for a propositional
logic.
In intensional logical semantics the meaning of a sentence is usually modeled as
a set of possible worlds, namely the ways the world could be if the sentence is true.
Call such a set a proposition.18 The aim of a compositional semantics is then to assign
each sentence φ to a proposition, [[φ]]. But before turning to those details, consider the
more intuitive and pertinent account of communication built on this semantics. An
idealized but useful model of how representational language is used to communicate is
familiar from Stalnaker (, ): it provides information by excluding possibilities.
One agent has some information they want to convey to another. This information can
be thought of as ruling out some possibilities w , . . . , wn . The first agent can then utter

18 This would be more clear in a first-order system where the ‘referent’ of an atomic sentence is
determined by the referents of its terms and predicate. But that would introduce unnecessary complexities.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

dynamic expressivism about deontic modality 

a sentence φ that excludes those possibilities, i.e. w , . . . , wn ∈

/ [[φ]], and by combining
this information with theirs i, the second agent can rule out some possibilities, i.e. i ∩
[[φ]]. The important part is that on this view, gaining information amounts to pointing
to some ways the world could be and saying that the world isn’t those ways. Or, to use
the exact phrasing from earlier: it refers to a distinction, say p or ¬p, and describes
the world as being on one side of that distinction. An expressivist cannot endorse this
account. After all, they hold that deontic discourse communicates those states of mind
without referring to them.
The compositional semantics employs a space W of possible worlds. Though a
simplification, one can equate a possible world with an atomic valuation, i.e. a function
assigning exactly one truth value (, ) to every atomic sentence—thinking of a truth-
table containing a column for every atomic sentence, each row specifies a possible
world.19 Atomics then denote the set of worlds where they are true, and the Boolean
connectives are treated as complementation, intersection and union, respectively.
Definition  (Compositional Semantics).
() [[p]] = {w ∈ W | w(p) = } () [[¬φ]] = W − [[φ]]
() [[φ ∧ ψ]] = [[φ]] ∩ [[ψ]] () [[φ ∨ ψ]] = [[φ]] ∪ [[ψ]]
What is the intuitive picture of meaning behind this formalism? An atomic sentence
p points to a region of logical space by referring to a particular distinction between
worlds, w(p) =  or w(p) = , and describing that world as being a certain way:
w(p) = . Connectives make it possible to point to regions of logical space that lone
atomic sentences cannot reach. In this framework, the only way for the language to
express distinctions between psychological states is for it to refer to them. The basic
notion of meaning at work is thoroughly referential, whether one is talking about a
sentence pointing at a region of logical space or whether one is talking about the
distinctions between worlds that make that possible. The compositional semantics
for the connectives is not really separable from this representational function. If one
introduced operators that did something non-referential, Definition  would provide
no clue what Boolean combinations would mean—that was Frege’s () point.
This makes it clear how non-trivial it is to add an operator with a non-referential
function to this framework. This compounds the challenge for an expressivist account
of communication. If distinctions between mental states are tracked compositionally
in the fashion of Definition , then they must be treated exactly like the distinctions
that are being referred to. This makes it very hard to see how one can take a content [[φ]]
and separate those distinctions that are being referred to and those that are not. It is this
tension that I will develop into a more detailed criticism of pragmatic expressivism.
Before developing that criticism, I will turn to logical relations.

19 This allows me to compress the distinction between worlds and models, or between ways the world
could be and ways the atomic expressions could refer.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 william b. starr

The standard definitions of logical relations also rely essentially on reference, though
that is less than clear on a first pass.20
Definition  (Consequence).
φ , . . . , φn  ψ ⇔ ([[φ ]] ∩ · · · ∩ [[φn ]]) ⊆ [[ψ]]
Definition  (Consistency).
φ , . . . , φn are consistent ⇔ ([[φ ]] ∩ · · · ∩ [[φn ]]) = Ø
On the referential picture of meaning sentences point to regions of logical space
by pointing at some distinction between possible worlds and describing the points
in the space as being on a particular side of that distinction. The consequence
relation requires φ , . . . , φn to collectively point to a subregion of what [[ψ]] points
to. This means that the distinctions made by φ , . . . , φn guarantee that the distinction
made by [[ψ]] has already been made. Similarly, consequence requires that φ , . . . , φn
collectively point to a non-empty region. This means that the distinctions they describe
can be satisfied by at least one world.
As before, there is a difficulty for an expressivist who would like to assign a meaning
[[φ]] which expresses psychological states without referring to them. Because both
definitions are entirely moored in the referential framework, it is simply unclear what
they would mean when extended to a variety of non-referential meaning. Consider
consistency. If one isn’t trying to refer to distinctions in worlds is it actually problematic
if there is not a world which bears all those distinctions simultaneously? This draws
out a feature that will be important below, and is often suppressed in discussions about
what an expressivist must do to explain the inconsistency of expressive sentences. Even
in entirely standard truth-conditional semantics, there is a teleological dimension to
the explanations of logical relations.
One might think that if an expressivist could assign denotations to non-referential
expressive sentences, then Definition  could be simply applied to them. However,
such a minimal answer is arguably unsatisfactory even for non-expressive sentences.
When one seeks an explanation of A and ¬A’s inconsistency, one often wants an expla-
nation of why an agent cannot believe/assert both without changing their mind and an
explanation of why two agents which believe/assert these contraries are disagreeing.
Simply saying that the denotation of the two sentences has an empty overlap does
not fully address this question. What is wrong with simultaneously asserting/believing
sentences whose denotations have an empty overlap? Answering this question requires
saying something about what denotation, assertion and belief are such that empty
overlap results in some kind of rational tension for an agent. For example, one might
maintain that the function of belief and assertion is to represent the world (Stalnaker,
, ), and the denotation of a sentence captures how it represents the world to
be. Then, semantically inconsistent sentences represent the world as being some way

20 Geach () stresses the difficulty of capturing inferences involving expressive sentences.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

dynamic expressivism about deontic modality 

no world can be, so they are dysfunctional things to simultaneously believe or assert.
This explanation combines three resources.
Explaining Inconsistency Representational explanations of inconsistency explain
the defective nature of simultaneously believing/asserting two sentences by:
. Construing the denotation relation as the representation relation
. Construing the function of asserting/believing those sentences as representing
the world
. A definition of semantic consistency as empty overlap, and an explanation of why
that tracks dysfunctional assertions/beliefs.
Agreement and disagreement can then be thought of as consistency between the
beliefs/assertions of two agents. This broader understanding of what it means to
explain inconsistency clarifies both the challenge and prospects for an expressivist.
The prospect is clear: they need an alternative conception of the denotation rela-
tion, a different construal of the function sentences serve, a formal definition of
consistency and an explanation of how that formal definition tracks dysfunctional
assertions/beliefs.21 Now for the challenge. The expressivist can semantically treat
the sensitivity to psychological states exactly as one would treat referential discourse,
but then they will face two challenges. They must have a different account of why
empty overlaps in denotation are bad, since they believe the function of expressive
discourse is non-representational. Indeed, it seems that anything in the neighborhood
of the explanatory expressivist thesis commits them to appealing here in an essential
way to the non-representational attitudes that are supposed to explain the semantic
properties of expressive sentences. Additionally, the expressivist must explain how that
meaning, which is formally modeled with means suited to representational meaning,
is used in a non-representational form of communication. In the next section I will
argue that recent expressivist analyses (Silk, ; Yalcin, , , ) appear to
explain inconsistency, but it is not actually clear that they meet the first challenge.
Further, I will argue that they do not meet the second.

 Preference Expressivism and Pragmatic Expressivism

Silk (, p. ) and Yalcin (, §) offer quite similar semantic analyses of deontic
modals, both of which resemble the account of expressive content sketched by Gibbard

21 Charlow () construes explaining consistency more narrowly: deriving a contradiction in the
meta-language from the assumption that the sentences are jointly true/satisfied. Charlow (, §.) argues
that a psychologist semantics is not entitled to assume in such a derivation that irrational psychological
states do not exist, since they are logically possible. I’m taking a broader view here, which begins with the
fact that content is functionally grounded. Setting aside which approach is better, note that an account like
the above can happily admit the existence of irrational states of mind and still explain inconsistencies. They
are essential to that explanation!

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 william b. starr

().22 Both approaches adopt a truth-conditional semantics, but one with a format
that is different from the above: it assigns a sentence to a truth-value relative to a world
w and, crucially, a practical parameter. For Silk (), the practical parameter  is a
weak-preference (pre)ordering (reflexive, transitive) on possible worlds, i.e. w  w
means w is preferred at least as much as w , and when w  w also holds w and
w are equally preferred. Yalcin (, §) stays closer to Gibbard (), and uses
‘hyperplans’.23 I’ll focus on Silk’s () semantics, since it will be closer to the view
I develop later. Both analyses follow Gibbard () in proposing truth-conditions
that do not place any constraint on the world parameter, only the practical parameter.
For Silk (), this means placing a constraint only on the  parameter: there is
no mention of the world of evaluation w on the right-hand side of the semantic
clauses. Deontic modals constrain  by requiring something of the worlds that are
best according to , i.e. worlds that are at least as preferable as every other world.
Definition  (Preference Expressivism, Silk , ).
Where Best() = {w | ∀w : w  w if w  w}:
. [[Must(φ)]]w,  =  ⇐⇒ ∀w ∈ Best(): [[φ]]w ,  = 
• Must(φ) is true at w,  iff all the worlds best according to  are φ-worlds.

. [[May(φ)]]w,  =  ⇐⇒ ∃w ∈ Best(): [[φ]]w ,  = 

• May(φ) is true at w,  iff some world best according to  is a φ-world.
This semantics predicts that Must(A) and ¬Must(A) cannot both be true. Given the
standard semantics for negation, 24 the truth of both Must(A) and ¬Must(A) requires:
∀w ∈ Best(): [[A]]w ,  =  and ∃w ∈ Best(): [[A]]w ,  = . In words, the truth
of both requires all -best worlds to be A-worlds and some -best world to be a ¬A-
world. There is no set of worlds, let alone one generated from , that satisfies that
requirement. This semantics clearly predicts the inconsistency of contrary expressive
sentences. But does it predict it in a way that the expressivist can endorse? Silk moves
quickly here, contending that it does:
In response to the negation problem, we can say that must and not must are inconsistent
because their conjunction expresses an incoherent set of weak preferences. It is incoherent
to have all of one’s most weakly preferred alternatives be ones where [A] and some of one’s
most weakly preferred alternatives be ones where [¬A]. There is no (non-empty) preorder that

22 I am tempted, but hesitant, to categorize the account proposed by Charlow (, ) along with
the pragmatic expressivists. The view developed there has a more nuanced and developed metasemantics
that make this categorization unclear. How exactly the points made in this section relate to these voices will
have to be left for another day.
23 Unlike Gibbard (), Yalcin (, §) constructs each hyperplan h out of possibilia: h is a function
that maps a set of worlds s to one of its subsets, namely those worlds permissible according to the hyperplan
given that one is in circumstances s.
24 [[¬φ]]
w, =  ⇐⇒ [[φ]]w,  = 

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

dynamic expressivism about deontic modality 

represents such a body of preferences. Preferential incoherence appears to be the right kind of
inconsistency in attitude to explain the inconsistency between normative sentences ‘φ’ and ‘¬φ’.
(Silk, , )

The key idea is that the two sentences express preferences that are incoherent. It is
crucial for the expressivist that this incoherence is non-representational. This is where
I think there is room for skepticism. The kinds of non-representational coherence
constraints placed on preferences are like the one discussed in §: no cyclic orderings.
That is because such orderings exist and would fail to motivate choice. There is no
parallel coherence constraint on preferences in decision theory that says one cannot
have all of one’s most weakly preferred alternatives be ones where A and some of
one’s most weakly preferred alternatives be ones where ¬A. It’s not that there is a
preference ordering which does have these two properties but would fail to motivate
choice. It’s that no preference ordering meets the two descriptions at all. It is no more a
practical constraint on preference orderings than the fact that they cannot both range
over ten alternatives and twelve alternatives. The nonexistence of such a preference
ordering does not follow from any background theory about how preference motivates
choice, but from the fact that they are inconsistent ways of representing preferences.
More generally, if there is a genuine appeal to preferential incoherence and not
representational inconsistency here, then one should be able to construct an ordering
which has these two properties and show how it fails to motivate choice. Since the
descriptions are logically incompatible, I do not see how this is possible.
Another way of illustrating the problem here is to return to the observations of
§. There I motivated the idea that the natural way to combine preferences is to
union them. This means that the motivational consistency of two preference orderings
amounts to whether their union can serve its function of motivating choices. Now,
consider two orderings  and  over one A-world w and one ¬A-world w .
Suppose then that w  w and w   w , while the other ordering is just the
opposite: w  w and w    w . Must(A) then expresses  while ¬Must(A)
expresses  . If the two orderings are combined into one, call it  , a perfectly
rational ordering results where one is indifferent between A and ¬A: w  w
and w  w . Indeed, since  is also expressed by Must(¬A) the same problem
arises for explaining the rational tension produced by committing oneself to both
Must(A) and Must(¬A).25 Yet,  is a perfectly rational state of preference. There are
no background decision-theoretic principles the preference expressivist can appeal to
here to render this ordering incoherent. There is, however, an important lesson one can
take away from this discussion. Solving the external negation problem by appealing to
incoherent preferences may not be possible.

25 To some degree, this latter point is a bug of Silk’s formalization using weak preferences. Weak
preferences are ill-suited to the job here because it is impossible to distinguish between an irrational
symmetric strict preference and indifference when you take weak preference as basic and define strict
preference as an asymmetric preference. For this reason, I use strict preferences in §.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 william b. starr

Let me now turn to the question of how deontic modals with the semantics in
Definition  could be used to communicate preferences without referring to them
and describing them, i.e. representing them. This is where Yalcin (, §) makes a
valuable proposal.26 Beginning with the dynamics of language use, the view is that the
common ground of a conversation consists not just of mutually supposed information
s = {w , . . . , wn , . . .} but also a set of preferences P = { , . . . , n , . . .} (Yalcin, ,
p. ). While descriptive discourse constrains the world and eliminates worlds from s,
deontic language constrains preferences and eliminates preference orderings from P.
In brief: each type of sentence affects the common ground in a different way.27 Since
s is a representational state and P is a motivational state this distinction appears apt
for distinguishing representational from motivational discourse. Yalcin’s (, p. )
sketch suggests a formalization in the following territory:28
Pragmatic Expressivist Discourse Model
. A descriptive sentence φ expresses a constraint on how the world can be, i.e. the
set of left-indices that make it true: φ  = {w | ∃: [[φ]]w, = }
. An expressive deontic sentence φ expresses a constraint on preferences, i.e. the
set of right-indices that make it true: φ = { | ∃w: [[φ]]w, = }
. Common ground: C = sC , PC , where sC = {w , . . . , wn }, PC = { , . . . , m }
• Background assumption: sC is the ways the world could be given what the
agents are taking for granted
• Background assumption: PC is the preferences compatible with the situations
the agents’ are in favor of
. Representational communication: C  φ = sC ∩ φ , PC 
. Expressive communication: C  φ = sC , PC ∩ φ
How does this allow one to distinguish referring to and describing preferences
from expressing preferences? The idea seems to be that referring to and describ-
ing preferences eliminate worlds while expressing preferences eliminates orderings.
The problem is that this interpretation of the model is entirely optional. Since the
dynamics of expressive and representational content are identical, the system could
be reformulated in terms of updating a set of world-ordering pairs, or centered-
worlds, which distinguish ways the world could be from the perspective of a preference
ordering. Then one can easily describe the system as behaving as referring to and
describing preferences, just as worlds centering on individuals are said to refer to and
describe an individual (Egan, ). The only way to clearly distinguish expressive

26 See Rothschild (); Swanson (forthcoming); and Moss () for related work. Ninan ()
presents a view that is ambivalent between semantic and pragmatic expressivism. The examination of
pragmatic expressivism through examining embedded must claims is a strategy Ninan (, §.) already
highlights as an important way to probe these views. See note  in § for more discussion of Ninan ().
27 The terminology of expressing constraints comes from Swanson (; forthcoming).
28 The notation of φ  and  φ is novel here, but iconic: [[φ]] is a set of pairs w,  while φ  is the
set of left members (worlds) and  φ is the set of right members (orderings).

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

dynamic expressivism about deontic modality 

from representational communication is to motivate and adopt a model where the two
kinds of discourse have clearly different dynamics, say one where expressive sentences
do not refer to orderings and place constraints on them. It is hard to see how one can
do this without departing from truth-conditional semantics. As discussed above, this
difference also seems necessary for a solution to the negation problem. It turns out
there is a rather similar issue with disjunction.
Mixed disjunctions provide a hard challenge for the expressivist approach above.
It is hard to see how they can be explained while maintaining that expressive and
representational communication have distinct communicative effects. The challenge is
related to Schroeder’s () concerns about mixed disjunctions in belief ascriptions.
But my focus will be on what mixed disjunctions communicate.29 Suppose you know
that exactly two things make Alex happy, namely Pat being in town or Alex receiving
permission to skateboard, and I ask Why is Alex happy? you might answer:
() Either Pat is in town or Alex may skateboard.
Consider how this sentence would update the common ground on the pragmatic
expressivist model above. The standard semantics for disjunction gives us the follow-
ing truth-conditions and corresponding constraints:
() [[P]]w, =  or [[May(S)]]w, = 
() P ∨ May(S) = {w | ∃: [[P]]w, =  or [[May(S)]]w, = }
() P ∨ May(S) = { | ∃w: [[P]]w, =  or [[May(S)]]w, = }
The problem for the world constraint is that the deontic disjunct will allow every world
to meet this condition. There is a preference ordering that makes May(S) true and if
one world-ordering pair featuring that ordering makes it true, then any other pair
with that ordering makes it true: deontics do not constrain the world parameter. The
same problem arises for the preference constraint, since there is a world that makes
P true and P does not constrain the ordering parameter. This predicts that a mixed
disjunction cannot communicate anything, contrary to our intuitions about ().30
There is an alternative analysis of disjunction that pragmatic expressivists could
adopt to capture this data. Suppose we took common grounds to be sets of the old
common grounds: C = {s , P , . . .}. For any non-disjunctive sentence, proceed as
before, only applying the update to each information state or preference set. But for
disjunctions, don’t use truth conditions, but rather perform a special update process.
Take the result of updating C with the left disjunct and union it with the result of

29 Mixed disjunctions are a focal point in the arguments I’ve offered for dynamic analyses of imperatives
(Starr, , ).
30 Silk (, §) offers a modified semantics designed to solve Schroeder’s () disjunction problem.
But this results in practical sentences that constrain the world parameter and hence update S, and
representational ones that constrain the preference parameter and hence update the preference set. This
solves the problem at the expense of having a way of neatly distinguishing expressive from representational
communication.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 william b. starr

updating with the right disjunct. To illustrate how that would work, consider a simple
common ground where there is one pair: C = {{w , w }, { ,  }}. Suppose w
is a P-world, w is a ¬P-world,  makes May(S) true and  makes May(S)
false. Then updating with () will produce C  = {{w }, { ,  }, {w , w }, { }}.
We may interpret these new common grounds as ‘imprecise common grounds’, on
analogy with imprecise credences. The agents’ states of mind are not precise enough
to distinguish which state they are in.31 This revised analysis can be interpreted in
two ways. Perhaps disjunction does not have a truth-conditional semantics, but only
a distinctive pragmatic update rule that is different from that which applies to every
other kind of sentence. This assumption conflicts with the received methodology of
semantics and pragmatics. No connective has only a pragmatics. It does, however,
suggest a more radical approach: what if all sentence forms were given update rules as
their semantics? This more radical approach is simply dynamic semantics, and will be
pursued in the next section.

 Logical Semantics, Dynamically

On a dynamic semantics, the meaning of a sentence can be modeled as a function [φ]
which specifies how accepting it transforms an agent’s state of mind s into another s .
One specifies the semantics with equations such as s[φ] = s , which say how applying
[φ] to s results in some other state s . Since meaning resides in how sentences change
states, it is said to be dynamic. By contrast, truth-conditional semantics is static: it
specifies a relation between symbols and the world rather than specifying how those
symbols are used. To illustrate this new approach in familiar territory, I will provide
a semantics for propositional logic consisting of equations with the form s[φ] = s .
While many have proposed systems in the vicinity of this, including Kamp (),
Heim (), and Groenendijk and Stokhof (), what will follow is basically the
system presented in Veltman ().32 The intuitive idea is that using φ consists in
moving from state s to state s . With programs, the states were machine states, but
what are states now? The goal here is to use the formal language of propositional logic
to model the use of natural language. I will take s to be the state of information in
the conversation or inquiry: s is the set of worlds compatible with what the agents’
are assuming for the purpose of the exchange. Formally, filling out an equation of the
form s[φ] = s consists in saying for any given set of worlds s, the set of worlds s that

31 This interpretation and solution are a wink to Rothschild () who, essentially, uses lifted states
like this to address a related problem for pure expressive disjunctions. Mixed disjunctions, however, are
not modeled there. This is the only reason Rothschild () can do without a special disjunction pragmatic
update rule. For more on the motivations for imprecise credences and pointers to that literature see Halpern
(, §.).
32 Unlike Veltman (, p. ) but like Groenendijk et al. (, §) I will treat conjunction as
sequential update. Unlike Groenendijk et al. (, §) but like Veltman (, p. ) I will treat disjunction
as forming the union of parallel updates to the initial state.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

dynamic expressivism about deontic modality 

result from interpreting φ. Intuitively, this means saying how processing φ requires
the agents’ to adjust the information they are assuming in their exchange. To save a
few words I will say that s is the result of updating s with φ.
Definition  (Compositional Semantics).
() s[p] = {w ∈ s | w(p) = } () s[¬φ] = s − s[φ]
() s[φ ∧ ψ] = (s[φ])[ψ] () s[φ ∨ ψ] = s[φ] ∪ s[ψ]
Atomic sentences eliminate possibilities incompatible with their truth. For complex
sentences, the equations specify how the interpretation of φ depends on interpreting
its constituents. Negation eliminates the possibilities compatible with its scope. Con-
junctions update with each of their conjuncts in sequence. Disjunctions update with
each of their disjuncts in parallel and form the union of each result. The clause for
atomic sentences directly appeals to truth/reference in specifying the interpretation
procedure for an atomic sentence. But note that unlike in the classical semantics
the connectives here are not given a representational interpretation. Intuitively, they
do not work by increasing the precision with which one can refer to distinctions
between possible worlds. They increase the space of transitions between states the
language can traverse. This different understanding of expressive power is the crucial
difference between dynamic and classical semantics. It allows one to conceive of two
different ways to make the language more expressive: add operators which discern
finer distinctions in the space of possible worlds, and add operators which trace more
transitions through the space of states. The key in understanding the relationships
between the two forms of semantics comes to this: what kinds of transitions amount
to drawing distinctions between possible worlds and what kinds of transitions amount
to something else? One way of answering this in the present context is considering
which operators exploit operating on a whole set of worlds rather than considering
each world in isolation. In systems such as that above where s[φ] ⊆ s, this formal
difference amounts to whether or not the update is distributive.33

Definition  (Distributive). s[φ] is distributive just in case s[φ] = {{w}[φ] | w ∈ s}.
The semantics above is distributive, and so the underlying language does not ‘do
anything’ which does not amount to making distinctions between possible worlds.
This is what I will eventually give up. The possibility of doing so hangs crucially on
the fact that dynamic semantics furnishes logical concepts more general than those
employed in truth-conditional semantics.
A more general conception of meaning is molded by a more general conception of
logic. Instead of truth, dynamic semantics builds its logic on support.
Definition  (Support, Truth in w).
() Support s  φ ⇔ s[φ] = s () Truth in w w  φ ⇔ {w}[φ] = {w}

33 This property is sometimes also called continuity, and has been isolated as a key feature by van Eijck
and Visser (, §.), Muskens et al. (, p. ) and van Benthem (, p. ).

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 william b. starr

Generally, a state s supports φ just in case interpreting φ requires no change in s. Where

s is an information state this means that interpreting φ requires no change in that
information. (Equivalently: [φ] is a fixed point with respect to s.) Support is more
general than truth: truth in a world in a special case of support. φ is true in w just
in case accepting φ requires no change to perfect information about w: {w}.34 This
definition of truth makes plain that non-distributive dynamic meanings are the non-
classical ones: they are the ones where support, rather than truth, is necessary.
Support is the basic logical concept in the sense that consequence and consistency
are defined in terms of it. ψ a consequence of φ just in case updating any state with φ
results in a state that supports ψ.
Definition  (Dynamic Consequence).
φ , . . . , φn  ψ ⇔ ∀s: s[φ ] · · · [φn ]  ψ
The basic idea is that any agent who accepts the premises is committed to accept-
ing the conclusion (or rejecting those premises) because accepting the conclusion
in the resulting state requires no more information than is already possessed; they
are making no additional commitments about what the world is like. Consistency is
also defined in non-propositional terms: there is a non-absurd state of information
which supports each of φ , . . . , φn . That is, there is a single informational perspective
which supports all of the sentences.
Definition  (Dynamic Consistency).
φ , . . . , φn are consistent ⇔ ∃s = ∅: s  φ , . . . , s  φn
Throughout, I’ve emphasized that the dynamic system presented is a generalization of
the classical one. It is worth making this more explicit. On the dynamic semantics, one
can still define propositional content.
Definition  (Propositional Content). [[φ]] = {w | w  φ}
Taken with the dynamic compositional semantics from Definition , this implies the
classical compositional semantics (Veltman, ).
Corollary  (Possible Worlds Semantics).
() [[p]] = {w ∈ W | w(p) = } () [[¬φ]] = W − [[φ]]
() [[φ ∧ ψ]] = [[φ]] ∩ [[ψ]] () [[φ]] ∪ [[ψ]] = [[φ]] ∪ [[ψ]]
Thus, even though propositions do not directly figure in the compositional clauses
and connectives are not propositional operators (they actually combine update pro-
cedures), the dynamic semantics behaves exactly as if these classical ideas were
operative. Propositions do not directly figure in the definitions of consequence or
consistency either. Just as truth is a special case of support, classical consequence and
consistency are special cases of the dynamic notions.

34 This definition is mentioned, but not adopted, by Muskens et al. (, p. ).

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

dynamic expressivism about deontic modality 

Definition  (Classical Consequence).

([[φ ]] ∩ · · · ∩ [[φn ]]) ⊆ [[ψ]] ⇔ ∀w: {w}[φ ] · · · [φn ]  ψ
Definition  (Classical Consistency).
([[φ ]] ∩ · · · ∩ [[φn ]]) = ∅ ⇔ ∃w: {w}  φ , . . . , {w}  φn
In full generality, dynamic consequence and consistency are not equivalent to the
classical notions. But restricted to the language of propositional logic with the dynamic
semantics of Definition , they are. This follows rather directly from the fact that all
updates in the present semantics amount to information gain: eliminating possibilities.
Fact  (Update is Propositional Communication).
Given the semantics of Definition : ∀φ: s[φ] = s ∩ [[φ]]
This fact not only means that the notions of consequence and consistency are equiv-
alent to their classical counterparts. It also means that the account of communication
is equivalent: communication works by providing information, i.e. referring to some
aspect of the world and stating a fact about it. As long as this equivalence holds,
the difference between dynamic and classical semantics is purely conceptual. This
conceptual difference resides in the centrality of support, rather than truth, to dynamic
semantics. This conceptual difference makes room for a non-equivalent account which
preserves the virtues of the classical account, albeit with non-classical definitions
of compositional meaning, consequence, consistency and communication. But the
classical logician cannot charge the dynamic logician with a change of topic. The
classical definitions are a special case of the non-classical ones. The disagreement is
thus not one of subject matter, but of breadth. It is easy to see how this disagreement is
relevant to the debate about expressivism. Frege and Geach took the narrow classical
concepts as given. But what happens when the more general concepts are embraced?

 Dynamic Expressivism about Deontic Modality

The equivalence between the dynamic and classical definitions in § can be broken.
In this section I will provide a dynamic semantics where the updates assigned to
must claims have the distinctive expressivist property: they do not communicate by
providing information, i.e. they do not refer to some part of the world and provide
information about it. Instead, they promote certain preferences and then test that
those preferences make the scope of the modal a practical necessity/possibility. I will
first present the semantics and then turn back to the key issues from the discussion of
expressivism above. The first step will be to model states that reflect not just the agents’
information, but their motivations.

. What are States?

Here is my answer to the question:

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 william b. starr

 
Definition  (States). A state S = {s  , . . . , sn m } is a set of substates si , j .
() Each si is a set of worlds compatible with what the agents are mutually suppos-
ing, and is competing to be the information they are mutually supposing.
() Each preference frame j = j , ∼j  consists of two relations, j and ∼j ,
on W.
• Where w j w just in case w is strictly more preferred than w and j is
competing to guide the agents
• Where w ∼j w just in case w is equally preferred to w and ∼j is competing
to guide the agents
() Each substate si , j  consists of information si and a preference frame j .
() Notation: s := s,  and any set theoretic notation applied to ‘s ’ is
understood as being applied to s, e.g. s ⊆ W means that s ⊆ W.
Of course, it may be helpful to hear why each of these components is here and how
states are being interpreted. Those who wish to skip those niceties, proceed to §.,
but have a peek at Definition  along the way.
I will follow a developed tradition in decision theory and model an agent’s motiva-
tional state as an ordering of worlds, where w  w means that w is strictly preferred to
w , and w ∼ w to mean that w and w are equally preferred.35 While it is common to
just use a weak preference ordering  and define equal and strict preference in terms
of it, this will not work for my purposes. My analysis of inconsistency of Must(φ) and
Must(¬φ) will require modeling the irrational state of strictly preferring φ-worlds to
¬φ-worlds and vice versa. On a weak preference analysis, this cannot be distinguished
from the perfectly rational state of φ and ¬φ-worlds being equally preferred. At the
same time, I cannot do with just strict preference. I will need equal preference to model
a state where ¬Must(φ) has been accepted, but Must(¬φ) has not: since the agents
find φ and ¬φ equally preferable they reject that they must φ, but also reject that
they must ¬φ.36 So I propose to capture the motivational component of states using
a preference frame, which is a pair consisting of these two relations on worlds: , ∼.
To simplify notation, I’ll refer to a frame as a whole with ‘’ but reserve ‘’ or ‘∼’ for
use between worlds.
As has already surfaced, rational preferences have certain properties, e.g. strict ones
are acyclic, equal ones are reflexive. But I’ve already said that I am not limiting this
model to rational states. Some take these properties to be definitional of preferences
so it is important for me to articulate a different view. I do not take preference to

35 Some gloss w ∼ w as indifference but this gives the misleading impression that the agent doesn’t care
about w and w at all.
36 Appropriately capturing the distinction between ¬Must(φ) and Must(¬φ) is a well-documented
difficulty for expressivism (Unwin, ; Schroeder, c; Dreier, ), and so one this analysis must
solve.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

dynamic expressivism about deontic modality 

be reducible to choice behavior, but I will assume that preferences have a function:
to motivate choice. This functional account assumes that rational preferences have
the properties they do because those are the properties they must have to achieve
their function of motivating choice. This idea about the function of preferences
will be important when I discuss consistency, and it should recall the discussion of
consistency at the end of §.
Even though I will not require preferences to be rational, their rationality will
nonetheless be relevant. It is important to be clear about how. Their first relevance
will be in the pragmatics: I will assume that a speech act which brings about irrational
preferences is unstable: the discourse participants will not be satisfied with it. Their
second relevance is in the semantics for deontic modals. For example, Must(φ) will
test whether the alternatives which may be rationally chosen are the φ-alternatives. If
the underlying preferences are not rational, this test is sure to fail since no alternatives
may be rationally chosen from irrational preferences. This means that I must say
more about how alternatives are rationally chosen, and which properties this requires
preferences to have. This is obviously an enormous and delicate issue, but all I will need
below is two plausible claims about rational choice: (i) rational agents can choose any
alternative which is not dominated by—strictly less preferred than—an alternative; (ii)
if w can be rationally chosen and w ∼ w , then w can be chosen.37 The set of rational
alternatives to choose from a space of worlds s can then be defined as follows.
Definition  (Choice Possibilities).
Choice(s, ) := {w ∈ s | w ∈ Nd or ∃w ∈ Nd: w ∼ w }
• Nd(s, ) := {w ∈ s | w ∈ s: w  w}
• Choice possibilities in s are either non-dominated, or equally preferable to some
w ∈ s that is.
On this conception of choice, a cyclic strict preference ordering such as w  w , w 
w has the function of promoting the choice of both w and w , but discounts both. It is
thereby irrational. There are many other properties rational preference must have on
this definition, but asymmetry is the only one that will be crucial below.
I seem to have arrived at a good model of states: a space of worlds and a preference
frame s, . But, for the reasons discussed at the end of §, disjunctions featuring
deontic modals challenge such a model. At least in the static setting this challenge
motivated a higher type for states, namely a set of these pairs. I will also adopt this
higher type. Here’s why. Suppose s,  was the form states take. A standard idea is
that Must(φ) tests that the ‘choice’ worlds in s,  are the φ-worlds.38

37 See Hansson and Grüne-Yanoff (, §.).

38 This is a deontic adaptation of Veltman’s (, p. ) presumably; see also Portner (, p. ),
definition ..

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 william b. starr

Definition  (Simple Test Semantics).


s,  if s, [φ] = Choice(s, ), 
s, [Must(φ)] =
∅,  otherwise
Consider a scenario where Pat is coming to visit Alex, who lives in a neighborhood
where parking is allowed only on one side of the street, depending on the day. Alex
knows that today, parking on the left is required, so he might tell Pat You must park
on the left. The idea, as Veltman (, pp. –) puts it, is for Pat to test whether
the information and preferences in play have the property in question: are the choice
worlds the live worlds where Pat parks on the left? If so, then proceed as before with
s, , otherwise ∅,  spurs Pat to adjust  so that the test is passed. Since the
preferences in this case are new to Pat, he must adjust the preferences he takes to be at
play. Importantly, this adjustment is a pragmatic process on this analysis.
Focusing now on disjunction, one must first generalize the semantics from Defi-
nition . to the new form of states. The simplest idea is to union the information
provided by each disjunct and keep the input ordering:
() s, [φ ∨ ψ] = sφ ∪ sψ , , where s, [φ] = sφ ,  and s, [ψ] =
sψ , .
Consider a similar scenario in which Pat is coming to visit Alex, but Alex cannot
remember which day goes with which side. Alex might then tell Pat:
() You must park on the left side of the street or you must park on the right side
of the street (I can’t remember which).
As before, the preferences are new to Pat, so the disjunction in () leads to the state
∅, . Pat must revise the preferences, but how should we represent the state he enters
when he does so? Switching to a strict ordering L which ranks parking-on-the-left
worlds over parking-on-the-right-worlds, will lead to the state s, L . But this means
that Pat would know what to do: park on the left. The same goes for an ordering
that prefers parking-on-the-right worlds. There is no clear candidate to represent the
uncertainty Alex expresses. What is needed is for the pragmatic process of adjusting
 to happen semantically, when each disjunct is interpreted, and then to somehow
combine the results into one state. The first task is tractable, and will be carried out
in §., Definition ..39 The hard question is how to combine those two augmented
strict orderings L and R into one that captures an undecided state. Forming the
union of the two orderings produces an irrational set of symmetric strict preferences
where parking on the left is strictly preferred to parking on the right and vice versa.
Moving to a frame where parking on the left and on the right are equally preferred
is also problematic. Alex has pretty explicitly expressed that the two options aren’t

39 Veltman’s (, p. ) normally works this way, but I will pursue yet another way.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

dynamic expressivism about deontic modality 

equally good, just that she can’t say which is better. Whatever she is conveying, it
seems to leave both orderings open. A simple move can capture this idea: states are sets
of s , .40 So interpreting Alex’s disjunction results in the state {s , L , s , R }.
The idea is not that agents posses a representation of their preferences which fails to
distinguish these two options. Instead, both preferences are competing for control of
their choices. Given her epistemic state, all Alex can do is induce this competition
among Pat’s preferences.41
. The Semantics
Recall the guiding slogan for the expressivist: Must(φ) promotes certain motivational
attitudes towards φ.42 The difficult part for semantically implementing this slogan has
been ‘promotion’. The essential idea of the semantics I will offer is that this promotion
consists in adding a strict preference for φ-worlds. More precisely, Must(φ) extends
each existing strict preference ordering by ranking each live φ-world over each live
¬φ-world. This preference update is defined below as φ+(). It simply extends 
with a strict preference for each φ-world over each ¬φ-world and leaves ∼ as it was.
Crucially, the semantics also tests, for every substate, that this augmented ordering
makes φ a practical necessity: the choice worlds are the φ-worlds.43 This test relates
the preference update to the information supposed in the interaction.44
Definition  (Dynamic Expressivist Semantics).
⎧
⎨{sφ+() | s ∈ S} if ∀s ∈ S: Choice(sφ+() ) = s
φ
S[Must(φ)] =
⎩{∅φ+() | s ∈ S} otherwise

 
• s
φ := ({s }[φ])

◦ sφ is the set of φ-worlds in s.

40 This allows variation along both axes: {s ,  , s ,  }, {s ,  , s ,  }, and {s ,  , s ,  }
           
are all states. Variation along the informational axis is needed for mixed disjunctions such as A ∨ Must(B).
41 This idea of various states competing for control of an agent’s actions is widespread in artificial
intelligence, in particular in the use of evolutionary algorithms to model adaptive problem solving. See
Franklin (, Chapter ) for a helpful overview. This seems in the spirit of Blackburn’s () being ‘tied
to a tree.’
42 Ninan () is a clear predecessor of this view, but is ambivalent about whether the ordering-
effect is semantic. Making this effect semantic is essential for expressivist purposes: it makes Must(A) and
¬Must(A) inconsistent in a non-representational way.
43 I here define May(φ) as the dual of Must(φ), but a less conservative option is worth noting: May(φ)
creates a new substate in which φ-worlds are preferred to ¬φ-worlds, and tests that φ is consistent
with the choice worlds in each substate. Given the logical definitions to follow this predicts free-choice
permission patterns: May(A) ∨ May(B)  May(A) and May(A ∨ B)  May(A) while ¬May(A ∨ B) 
¬May(A) ∧ ¬May(B).
44 Note that this account captures both performative and descriptive uses of deontic modals within the
same semantics: descriptive ones are those issued in a state that already contains the preference it promotes
and performative ones are those issue in a state that previously lacked that preference. Of course, pragmatic
social conditions concerning authority govern when performative uses are felicitous.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 william b. starr

• φ+() := {w, w  ∈ s × s | w  w or, w ∈ s  

φ & w ∈ s¬φ }, ∼
 
◦ φ+() adds to  a preference for each w ∈ sφ over each w ∈ s¬φ .
• May(φ) := ¬Must(¬φ)

After all, this is a model of declarative modal sentences and interacting with informa-
tion is the function of declaratives. The fact that they have some relation to information
makes them suitable to negate and embed under believe. But they hardly work by
providing information, and they certainly do not eliminate worlds by referring to a
distinction in each world. Their key contribution is the promotion of preferences and
this seems on track for expressivist aims. It is rather crucial for these aims that the
test component of the meaning is not redundant. After all, that would mean that an
update with any deontic modal claim will succeed. To see why the preference update
does not guarantee that the test will be passed, consider the fact that the incoming
preferences may be inhospitable. The test asks whether the alternatives a rational
agent could choose according to the augmented preferences are φ-alternatives. But
if the augmentation leads to an irrational body of preferences, then the choice set
will be empty. (That’s why those preferences are irrational.) To illustrate, consider
a simple scenario where the agents have narrowed in on two worlds, an A-world
w and a ¬A-world w . Further suppose the agents’ only strict preference is for w
over w and the agents only equally prefer each world to itself:  = {w , w } and
∼ = {w , w , w , w }. The agents’ state is then modeled as {{w , w } }. The first
thing that an update with Must(A) will do is create a new preference ordering where
w is also strictly preferred to w : A+( ) = {w , w , w , w }, ∼ . But such a
preference ordering is irrational. Its symmetric preference guarantees that there are
+
no non-dominated worlds, thus: Choice({w , w }A ( ) ) = ∅. That means Must(A) is
testing that there are no A-worlds. Since w is an A-world, this test is failed, resulting
+
in the state ∅A ( ) . The information ∅ in this result state reflects that it is not stable
and the preferences explain why: A+( ) is irrational. The agents then face a choice of
which preference to jettison. I do not have space to elaborate how I see that pragmatic
process unfolding, but suffice it to say that it may involve debate about the particular
preferences, or less conscious deference based on social influences.
As it turns out,  is the ordering that would have resulted from updating a
state with an empty strict ordering and trivial equality ordering {{w , w }∅,∼  } with
Must(¬A). This allows me to sketch how I propose to explain the inconsistency of
Must(¬A) and Must(A): there is no rational preference ordering which contains both
the preferences they promote. If the state is going to contain the preferences Must(A)
promotes, it is going to strictly prefer A-worlds to ¬A-worlds. If the state is going to
contain the preferences Must(¬A) promotes, it is going to strictly prefer ¬A-worlds to
A-worlds. But any state containing both will be irrational in the sense that it contains
dysfunctional preferences. To fill out this explanation, and treat the harder case of
Must(A) and ¬Must(A), a semantics for negation (and atomics) is needed.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

dynamic expressivism about deontic modality 

Apart from negation, the semantics for the rest of the language parallels Definition .
Atomics eliminate worlds where they are false from each substate. Conjunction is
still sequential update and disjunction is again the union of parallel updates to the
initial state. This semantics for disjunction does generate a new behavior, but I will
return to that. Focus now on negation: ¬φ not only removes the information φ would
provide (as before), it removes the preferences that φ would promote.45 This is done
by subtracting any preferences that φ would add to an empty ordering.
Definition  (Dynamic Expressivist Boolean Semantics).
. S[p] = {{w ∈ s | w(p) = } | s ∈ S}
− 
. S[¬φ] = {sφ () − ({s }[φ]) | s ∈ S}
• φ−() :=  − {w, w  ∈ i | {W ∅,= }[φ] = {s m
 , . . . , sn } &   i 


m}, ∼
• φ−() removes from  any pairs that updating with φ would add to an
empty ordering. For non-expressive discourse this will idle. If φ = Must(ψ)
this will extract preferences for ψ-worlds over ¬ψ-worlds.
. S[φ ∧ ψ] = S[φ][ψ]
. S[φ ∨ ψ] = S[φ] ∪ S[φ]
Since descriptive discourse never adds preferences, negations of descriptive sentences
will behave exactly as before. However, when an expressive sentence is negated both the
preferences it would promote and the information it would provide are removed. Con-
sider again the example from above where testing the state {{w , w } } with Must(A)
+
failed, returning ∅A ( ) . What if one updates that same state with ¬Must(A) instead?
The semantics predicts that this returns the initial state {{w , w } } via the following
+
process. It would first arrive at the state ∅A ( ) and then find the informational
difference {w , w } − ∅ and the preferential difference Must(A)−(A+( )). Since
Must(A)−(A+( )) =  ,46 this results in the original state {{w , w } }. The fact that
negation dynamically manipulates preferences in this way guarantees that there will
be no state, not even an irrational one, which contains the preferences both Must(A)
and ¬Must(A) promote. If that state already has the preferences in it that Must(A)
would add, then ¬Must(A) would remove them. Conversely, if the state already lacks
the preferences ¬Must(A) would remove then Must(A) would add them. This is the
gist of my expressivist explanation of why Must(A) and ¬Must(A) are inconsistent.

45 This analysis of negation is inspired by the converse operator of Dynamic Logic (Harel et al., ,
p. ), which serves as a variety of program negation. While there is a growing literature on using dynamic
logic for deontic reasoning, it does not directly speak to the key issues here.
46 Recall that  + − +
 = {w , w }, so A ( ) = {w , w , w , w  }. Then Must(A) (A ( )) =
Must(A)−({w , w , w , w }). To find Must(A)−({w , w , w , w }) one removes any strict preference
that ends up in the result state of {W ∅,= }[Must(A)] = {W {w  ,w  ,...},= }. So w , w  is removed from
{w , w , w , w } leaving {w , w } =  .

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 william b. starr

However, recall that {{w , w } } is a state where the test imposed by Must(¬A)
is also successful. How exactly are ¬Must(A) and Must(¬A) distinguished on this
semantics? This is where states with equally preferred worlds are required.
Consider a state where w ∼ w , and there are no strict preferences. Intuitively,
this state supports ¬Must(A), but not Must(¬A). If one prefers A and ¬A equally,
one definitely isn’t compelled to choose only ¬A-worlds or to choose only A-
worlds. The first fact means that one rejects Must(¬A), while the second means
that one accepts ¬Must(A). To see how the formalism above replicates this intu-
itive reasoning, consider a state {s } where s = {w , w },  = ∅, and ∼ =
{w , w , w , w , w , w , w , w }. Updating {s } with Must(¬A) will augment 
to  = {w , w } (leaving ∼ as it was), and test whether the choice worlds relative to
the augmented ordering are the ¬A-worlds: {w }. The only non-dominated alternative
in  is w , but recall that w ∼ w . Since any choice world is either non-dominated,
or equally preferred to a non-dominated one, the choice worlds are {w , w } and so
the test is failed. In the same way, the test for Must(A) will fail, and result in the
state {∅ }, where  = {w , w }. But this means that updating the original state
with ¬Must(A) will involve taking the informational difference {w , w } − ∅ and
removing the preference w , w . And so an update of {s } with ¬Must(A) returns
{s } while an update of {s } with Must(¬A) returns {∅ }. In other words: {s }
supports ¬Must(A) but does not support Must(¬A).
It is important to note that I do not distinguish ¬Must(A) and Must(¬A) by
appealing to a state where there are two competing substates, one where A is strictly
preferred and one where ¬A is strictly preferred, e.g. {{w , w } , {w , w } }. That
is the kind of state that results from updating with Must(A) ∨ Must(¬A).47 This
is not a state which supports ¬Must(A). Updating with ¬Must(A) after updating
with the disjunction would change the preferences and lead one to rule out the
substate produced by the left disjunct, thereby bringing one to a state where only
the substate produced by the right disjunct remains. In other words, one would
infer Must(¬A). This illustrates why it is so important to distinguish ¬Must(A)
from Must(A) ∨ Must(¬A). Conflating them straightforwardly leads to a collapse of
¬Must(A) into Must(¬A) (Unwin, ; Schroeder, c; Dreier, ). Mention
of logical matters highlights the fact that above I have been relying on the logical
notions of support and consequence roughly in the vein of §. But it is crucial to make
these notions explicit and to note that they arise in the way expressivists require. In
particular it will be important to see how the semantic properties of deontic sentences
are explained by properties of the attitudes they express.

47 This dynamic semantics of disjunction comes from Starr () where it is used to analyze disjunctions
of imperatives, declaratives (and combinations thereof). It works by dynamically generalizing the idea
that disjunctions introduce alternative propositions (Kratzer and Shimoyama, ; Simons, ; Alonso-
Ovalle, ), to the idea that disjunctions introduce alternative updates.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

dynamic expressivism about deontic modality 

. Expressivism Redux

How does dynamic expressivism explain the logical relations between expressive sen-
tences? Can one simply adopt the definitions of support, consistency and consequence
from §, or does the added dynamics of preference make other definitions suitable?
Since support is the central logical concept, that is the place to start. Generalizing
the previous definition of support (§, Definition ) would amount to saying that S
supports φ when S[φ] = S. This is a rather strict form of support which uniformly
requires no change to the information, number of substates or preferences of S. This
notion of support yields an interesting, though radically non-classical, logic. However,
for the purposes here it is somewhat more illuminating and conservative to consider
two parallel notions of support, one specialized to information, one to preferences.48
Definition  (Informational Support). S  φ ⇐⇒ iS = iS[φ]

• iS = {s | ∃: s ∈ S}

Definition  (Preferential Support). S  φ ⇐⇒ PrefS = PrefS[φ]

• PrefS = { | ∃s = ∅: s ∈ S}

Informational support is familiar from §, but preferential support tracks the addi-
tional expressive dimension added by deontic modals. It requires that the preference
frames in S stay the same after updating with φ and that φ does not add any preference
frames.49 This furnishes two notions of consequence and of consistency:
Definition  (Informational Consequence). φ , . . . , φn  ψ ⇐⇒ ∀S: S[φ ] · · ·
[φn ]  ψ
Definition  (Preferential Consequence). φ , . . . , φn  ψ ⇐⇒ ∀S: S[φ ] · · ·
[φn ]  ψ
Definition  (Informational Consistency).
φ , . . . , φn are informationally consistent ⇐⇒ ∃S: iS = ∅ & S  φ , . . . , S  φn
Definition  (Preferential Consistency).
φ , . . . , φn are preferentially consistent ⇐⇒ ∃S: Ch(S) = ∅ & S  φ , . . . , S  φn

• Where Ch(S) = {Choice(s, ) | s ∈ S}

48 There is actually a third to consider, choice-support, which produces a more classical logic for Must:
 
S choice-supports φ just in case {Choice(s, ) | s ∈ S} = {Choice(s, ) | s ∈ S[φ]}. This will
predict that both Must(φ) and Must(ψ) are consequences of Must(φ ∧ ψ). However it also predicts that
Must(φ ∨ ψ) is a consequence of Must(φ). See § for discussion of these principles.
49 It also discounts changes to the preferences that concern only problematic substates whose informa-

tion is contradictory: ∅ . Intuitively, those preferences are to be ignored since they are not constraining
agents’ choices: Choice(∅, ) will always be ∅. This restriction is needed to predict that Must(B) is a
preferential consequence of Must(A) ∨ Must(B) and ¬Must(A).

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 william b. starr

• Recall: if Choice(s, ) = ∅ then  is dysfunctional, i.e. fails to motivate a choice.

– E.g. if  is cyclic over s, Choice(s, ) = ∅
For informational consistency, I require a state that is not representationally dys-
functional to informationally support each of the sentences. By contrast, preferential
consistency requires a state that is not motivationally dysfunctional to preferentially
support each of the sentences. This allows a non-representational explanation of con-
sistency which parallels our representational explanations. For example, A and ¬A are
inconsistent because their function is to provide information to iS , iS has the function
of representing the world, but only a dysfunctional iS has the information each of them
provides. Similarly, Must(A) and Must(¬A) are inconsistent because their function
is to update the preferences in PrefS , PrefS has the function of motivating choices but
only a dysfunctional PrefS contains the preferences each of them promotes. Recall from
above that a state which contains a symmetric strict preference ordering ranking A-
worlds over ¬A-worlds and ¬A-worlds over A-worlds will preferentially support both
Must(A) and Must(¬A). But such a state is motivationally dysfunctional: Ch(S) =
∅. It is important to highlight that Must(A) and ¬Must(A) are inconsistent for
a different reason. It is not that only an irrational state can support both. It’s that
there is no state which supports both because they perform converse updates on the
preferences. If a state has the preferences that Must(A) would add, then ¬Must(A)
would remove them. If a state lacks the preferences ¬Must(A) would remove then
Must(A) would add them. This is an essentially dynamic form of inconsistency: one
formula builds something which the other removes. It is precisely this resource which
pragmatic expressivism lacked, and which made its explanation of the inconsistency
between ¬Must(A) and Must(A) seem suspiciously representational. The beauty
of the dynamic account is that one can add this new element while keeping the
explanations of consistency parallel. In both the motivational and representational
cases, consistency is allied with a function our mental states serve, and inconsistent
sentences are explained by the fact that if they can be embraced by a single state at all,
it is a dysfunctional state.
It was hard to motivate the pragmatic expressivist view that the contents com-
municated by deontics were non-representational. Those contents seemed open to a
representational interpretation since they were reified from a representational theory
of consequence and had identical dynamics. On the dynamic view, the underlying
theory of consequence is not inherently representational. It tracks various fixed points
in the interpretive process—in information, preferences, etc. Even though it has a rep-
resentational edition—when that fixed point concerns information—the conception
of logic is more general. It is precisely this more general approach to meaning and logic
that is essential to capturing the expressive dimension of deontic modals.
On the semantics in Definitions  and , updates involving deontic modals do
not refer to preferences and describe them. Those sentences promote and discount
preferences. This variety of update differs radically from descriptive, informational

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

dynamic expressivism about deontic modality 

updates. Crucially, they differ not just in the contents (preferences vs propositions)
they manipulate, but in how those contents are manipulated.50 The adding and
removing of preferences is quite different from the way descriptive discourse works.
Descriptive discourse provides information by making distinctions between worlds.
As I mentioned in §, this way of providing information can be formally charac-

terized as a distributive and eliminative update: s[φ] = {{w}[φ] | w ∈ s},
and s[φ] ⊆ s. One way of putting the problem for pragmatic expressivism was
that it was indistinguishable from a view where expressive discourse works via a
distributive and eliminative update over more fine-grained worlds such as w, . It
then becomes hard to elucidate the expressivist communication thesis, since it looks as
if expressive communication works by making distinctions between centered worlds,
and one of those distinctions is made by referring to preferences. But on the dynamic
view, the fact that preferences are added, while worlds are removed, guarantees that
preference update and information update cannot be the same process. Indeed, on
the dynamic account there are four kinds of updates: adding preferences (deontic
modals), removing preferences (negation), eliminating worlds (atomics, negation) and
creating substates (disjunction). Language compositionally interleaves these distinct
processes rather than providing a single semantic object which is factored into distinct
contributions.
Ultimately, the dynamic account makes room for the expressivist communica-
tion thesis by breaking from the traditional ‘conduit’ metaphor for communication:
expressing yourself is like putting a package on a conveyor belt between minds.
That metaphor assumes there is a discrete object shared and a single process by
which you unpack the box and sort the contents. On the dynamic approach, a better
metaphor would be two minds connected by a series of vibrating cords which encode
different modalities with different frequency ranges. Language is a virtuosic tool for
weaving these frequencies and modalities together. This departure from the conduit
metaphor is far from a debt for expressivists. It is one urged by researchers not debating
expressivism and spanning the fields of animal communication (Owren et al., ;
Scott-Phillips and Kirby, ) and natural language (Murray and Starr, ).

 A Wider View
I want to conclude this chapter by taking a wider view and being clear about what
has and has not been accomplished. I claim to have offered a dynamic semantics
which vindicates the expressivist communication thesis, and done so better than
competing pragmatic accounts. Dynamic meanings were essential for implementing
the expressive effect of deontic modals, and generalizing our familiar semantics for

50 Varieties of update is an allusion to Murray (), which exploits dynamic semantics to composition-
ally interleave a number of distinct update processes. It is there applied to a number of constructions across
languages which manipulate but differentiate at-issue and not-at-issue content.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 william b. starr

negation so as to interact with that expressive effect. This provided a new expressivist
solution to the negation problem. Must(A) and ¬Must(A) have converse dynamic
effects, which predicts that they are inconsistent because there is no state that can
contain the converse effects of their updates. Must(A) and Must(¬A) are inconsistent
because only irrational preferences support them. It was argued that these explanations
of inconsistency perfectly parallel those for descriptive discourse. There is nevertheless
much work left to do. There are still several constructions which have not been
analysed here and may turn out to be problematic. For example, the interaction
of modals with attitude verbs, modal questions, deontic predicates/adverbs such as
wrong/wrongly, more complex modals such as ought, should, could and non-deontic
uses of must and may all lie outside the scope of the analysis here.51 Further-
more, at some point one would like a positive argument in favor of an expressivist
semantics over the impressive achievements of its descriptive competitors such as
Kratzer (, ). This must wait for another day, but some suggestive remarks
on a few fronts may help chart the territory where such an argument might be
found.
The compositional virtuosity of the dynamic account was essential for analyzing
mixed disjunctions. It explained how communicating Must(A) ∨ B was different
from communicating Must(A) or communicating B. The disjunction involves a
distinctive substate-creating update, where quite different updates are unleashed on
each incoming substate to create two new substates. It is worth highlighting how this
addresses the concern raised by Schroeder () about the interaction of expressive
modals and attitude verbs. That concern is whether BelX (Must(A) ∨ B) will entail
BelX (Must(A)) ∨ BelX (B). BelX (φ) can be analyzed as a descriptive update requiring
of a world w that the state SX modeling X’s state of mind in w both informationally
and preferentially supports φ. But consider a world in which SX is a state with two

substates: SX = {sA , sB }, the first has information s and an ordering which prefers
A-worlds, and the second lacks that particular preference but has the information
carried by B. SX will preferentially and informationally support Must(A) ∨ B. Yet
SX will not preferentially support Must(A), since PrefSX [Must(A)] will lack . Also, SX
will not informationally support B, since iSX [B] will exclude the ¬B worlds in s. Thus
BelX (Must(A) ∨ B) will not informationally entail BelX (Must(A)) ∨ BelX (B). The
success of this analysis would not surprise Schroeder (). He acknowledges that if
belief ascriptions mix information and motivation, and there is a special sort of state
for believing disjunctions, then the entailment can be blocked. What he doubts is a
philosophically plausible interpretation of these constructs. Even the small amount I
have said here suggests a way of resisting this doubt. A substate is an informational
and motivational perspective, and a state containing multiple substates captures the
perspectives competing for control over the agent’s actions. There seems to be no good

51 I think there is some hope for analysing ?Must(A) as S[Must(A)] ∪ S[¬Must(A)].

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

dynamic expressivism about deontic modality 

philosophical or psychological motivation for assuming that beliefs are exclusively

informational or motivational, or that different perspectives can’t compete within
a single agent. Assuming beliefs are ‘purely informational’ seems philosophically
tidy, but psychologically implausible to me. Surely our actual beliefs are influenced
by the motivational perspective we adopt, just as our desires are influenced by the
informational perspective we adopt.52
The dynamic account allows one to articulate two components of a deontic modal’s
meaning: preference promotion and a test about that promotion. This leads one to
wonder whether language contains operators which simply promote preferences. That
is essentially Starr’s () analysis of imperatives:53
+()
Definition  (Imperative Semantics). S[!φ] = {sφ | s ∈ S}
This illuminates one difference between deontic modals and imperatives. Unlike
deontic modals, negation does not scope over imperatives in natural language (Han,
). On the semantics above, negating an imperative would disastrously subtract
each s from itself, returning ∅ for each. Negation thus turns any imperative into
an informational contradiction. This goes some way towards explaining what many
other theories treat as an arbitrary syntactic restriction: negation scoped over an
imperative does not produce a useful meaning, so no human language should bother
to conventionalize those syntactic structures. Explaining this fact by syntactic stip-
ulation looks less satisfying when it becomes clear that languages allow imperatives
to robustly embed under other connectives: !A ∧ !B, A ∧ !B, !A ∨ !B, A ∨ !B (Starr,
). Starr () presents two related arguments that build on this observation.
First, it is argued there that no plausible non-propositional analysis of imperatives can
compositionally account for this behavior without dynamic meanings like those above.
Secondly, it is argued that non-propositional non-dynamic theories such as Portner
() incorrectly predict !(A ∧ B) to be a consequence of (!A∨!B)∨!(A ∧ B), but that
a dynamic theory such as the above does not. This highlights that dynamic meanings
provide not only new compositional resources, but logical ones too. I turn now to this
logical frontier of dynamic thinking.
The deontic realm is a minefield of logical problems, but some recent work sug-
gests that dynamic semantics provides resources for resolving them. Willer (),
for example, shows that a dynamic non-monotonic consequence relation relieves
the pressures placed on modus ponens in deontic settings by Chisholm’s Paradox
(Chisholm, ), the Miner’s Paradox (Kolodny and MacFarlane, ) and the Gentle

52 Charlow () explores these issues in much more depth. He pursues an account which I learned
much from and which bears at least some affinities with the account pursued here.
53 Given this analysis of imperatives, one may draw a comparison between the semantics above and
the analysis in Ninan () where imperatives and deontic must have the same effect on context. On the
analysis above, the semantics of deontic must includes that effect, but adds the crucial element of a test. This
approach also predicts Ninan’s () key data: ¬A ∧ Must(A) will be infelicitous. Further investigation
will be needed to see whether the added test element is empirically advantageous.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 william b. starr

Murder Paradox (Forrester, ). Willer, (this volume, Chapter ) elaborates on the
relationship between this kind of non-monotonic approach and one based on default
logic. The semantics above makes two predictions that may add momentum to this
movement. Suppose a father tells his son that he must take out the trash or do the
dishes. The son proceeds to the trashcan only to find that it has already been taken
out. The son should infer that he must wash the dishes. However, if the son goes to do
the dishes and they too have been taken care of, he may no longer sensibly think he
must wash the dishes. These are facts classical theories do not predict:
() Must(A ∨ B), ¬A  Must(B)
() Must(A ∨ B), ¬A, ¬B  Must(B)
Just because the best worlds are all A ∨ B worlds and we aren’t in an A-world, it doesn’t
follow that all the best worlds are B-worlds. Further, the pattern in () seems to show
exactly the kind of non-monotonicity which classical consequence relations fail to
predict. By contrast, both patterns are predicted by the analysis above.
Now return to the scenario where the son has correctly inferred that he must do the
dishes. Note that he should not infer that he must do the dishes or watch a movie. He
also shouldn’t infer that either he must do the dishes or he must watch a movie:
() Must(B)  Must(B ∨ C)
() Must(B)  Must(B) ∨ Must(C)
This is a modal version of Ross’s Paradox (Ross, ) concerning imperatives. Their
validity in standard approaches to modality has been widely recognized as problem-
atic. Several attempts have been made to leverage pragmatic reasoning and a non-
standard semantics for disjunction to explain them (Zimmermann, ; Simons,
; Geurts, ; Aloni, ). Yet Cariani () and Lassiter () argue that a
strictly semantic approach is justified. While their semantic approach predicts (),
it does not, to my knowledge, predict (). The analysis above predicts both.54 While
none of these prospects constitutes a full empirical argument for the kind of dynamic
analysis pursued here, they illustrate some of the territory where those arguments may
evolve. It would, of course, also be unsurprising if dynamic expressivism proved useful
in the burgeoning literature on non-descriptive meaning.

Acknowledgments
I am extremely thankful to Andrew Alwood, Nate Charlow, Matthew Chrisman, Andy Egan,
Sarah Murray, Paul Portner, Angelika Kratzer, Aynat Rubenstein, Mark Schroeder, Alex Silk,
Justin Snedegar, Malte Willer and audiences at Edinburgh, NYU, Rutgers and USC, some for

54 Like Cariani () and Lassiter () the analysis here predicts that Must(A ∧ B)  Must(A), which
is one way to explain Jackson’s () ‘professor procrastinate’ puzzle.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

dynamic expressivism about deontic modality 

detailed feedback and others for brief, incisive questions. Charlow and Chrisman deserve special
thanks, not only for editing this volume, but also for guiding this project towards properly
engaging with the expressivism literature. I should also acknowledge a debt to Yuna Won. I
have learned at least as much from advising her PhD research on deontic logic as she has learned
from me.

References
Aloni, M. () ‘Free Choice, Modals and Imperatives.’ Natural Language Semantics. . .
pp. –.
Alonso-Ovalle, L. () Disjunction in Alternative Semantics. Ph.D. thesis. UMass Amherst,
Amherst, MA.
Alwood, A. () ‘Nondescriptive Negation for Normative Sentences.’ (To be published in
Philosophical Quarterly.)
Anderson, L. and Lepore, E. () ‘Slurring Words.’ Noûs. . . pp. –.
Beaney, M. (ed.) () The Frege Reader. Malden, MA: Blackwell.
Bittner, M. () ‘Topical Referents for Individuals and Possibilities.’ In Hastings, R., Jackson, B.
and Zvolenszky, Z. (eds.) Proceedings from Semantics and Linguistic Theory . pp. –.
Ithaca, NY. Cornell University.
Blackburn, S. () ‘Attitudes and Contents.’ Ethics. . . pp. –.
Burge, T. () ‘On Knowledge and Convention.’ The Philosophical Review. . pp. –.
Carballo, AP. () ‘Semantic Hermeneutics.’ In Burgess, A. and Sherman, B. (eds.) Metaseman-
tics: New Essays on the Foundations of Meaning. pp. –. New York, NY: Oxford University
Press.
Cariani, F. () ‘ “Ought” and Resolution Semantics.’ Noûs. . . pp. –.
Charlow, N. () ‘Logic and Semantics for Imperatives.’ Journal of Philosophical Logic.
pp. –.
Charlow, N. () ‘The problem with the Frege–Geach problem.’ Philosophical Studies. . .
pp. –.
Charlow, N. () ‘Prospects for an Expressivist Theory of Meaning.’ Philosophers’ Imprint.
. . pp. –. [Online] Available from: http://hdl.handle.net//spo....
[Accessed  Nov .]
Chisholm, R. M. () ‘Contrary-to-Duty Imperatives and Deontic Logic.’ Analysis. . .
pp. –.
Chrisman, M. () ‘On the Meaning of “Ought”. ’ In Shafer-Landau, R. (ed.) Oxford Studies
in Metaethics. Volume . pp. –. Oxford: Oxford University Press.
Dreier, J. () ‘Negation for Expressivists: a collection of problems with a suggestion for their
solution.’ In Shafer-Landau, R. (ed.) Oxford Studies in Metaethics. Volume . pp. –. New
York, NY: Oxford University Press.
Dreier, J. () ‘Relativism (And Expressivism) And The Problem Of Disagreement.’ Philosoph-
ical Perspectives. . . pp. –.
Egan, A. () ‘Epistemic Modals, Relativism and Assertion.’ Philosophical Studies. . .
pp. –.
Faller, M. () Semantics and Pragmatics of Evidentials in Cuzco Quechua. Ph.D. thesis.
Stanford University, Palo Alto, CA.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 william b. starr

Forrester, J. W. () ‘Gentle Murder, or the Adverbial Samaritan.’ Journal of Philosophy. .
pp. –.
Franklin, S. () Artificial Minds. Cambridge, MA: MIT Press.
Frege, G. () Begriffschrift, eine der arithmetischen nachgebildete Formalsprache des reinen
Denkens. Halle: L. Nebert. pp. –. References to M. Beaney’s translation in Beaney ()
pp. –.
Frege, G. () ‘Logische Untersuchungen.’ Beiträge zur Philosophie des deutschen Idealismus.
. pp. –. References to Frege ().
Frege, G. () ‘Compound Thoughts.’ Mind. . . pp. –. Translation of Frege ().
Geach, P. T. () ‘Assertion.’ The Philosophical Review. . . pp. –.
Geurts, B. () ‘Entertaining Alternatives: Disjunctions as Modals.’ Natural Language Seman-
tics. . . pp. –.
Gibbard, A. () ‘An Expressivistic Theory of Normative Discourse.’ Ethics. . . pp. –.
Gibbard, A. () Thinking How to Live. Cambridge, MA: Harvard University Press.
Grice, H. P. () ‘Utterer’s Meaning, Sentence-Meaning, and Word-Meaning.’ Foundations of
Language. . pp. –.
Groenendijk, J. and Stokhof, M. () ‘Dynamic Predicate Logic.’ Linguistics and Philosophy.
. . pp. –.
Groenendijk, J., Stokhof, M. and Veltman, F. () ‘Coreference and Modality.’ In Lappin. S.
(ed.) The Handbook of Contemporary Semantic Theory. pp. –. Oxford: Blackwell.
Halpern, J. Y. () Reasoning about Uncertainty. Cambridge MA: MIT Press.
Hamblin, C. L. () ‘Questions in Montague English.’ Foundations of Language. . . pp. –.
Han, C. H. () ‘Force, negation and imperatives.’ Linguistic Review. . . pp. –.
Hansson, S. O. and Grüne-Yanoff, T. () ‘Preferences.’ In Zalta (ed.), The Stanford Encyclo-
pedia of Philosophy, fall. [Online] Available from: http://plato.stanford.edu/archives/fall/
entries/preferences/. [Accessed  Nov .]
Harel, D., Kozen, D. and Tiuryn, J. () Dynamic Logic. Cambridge, MA: MIT Press.
Heim, I. R. () The Semantics of Definite and Indefinite Noun Phrases. Ph.D. thesis. University
of Massachusetts, Amherst, Massachusetts.
Heim, I. () ‘File Change Semantics and the Familiarity Theory of Definiteness.’ In
Bäuerle, R., Schwarze, C. and von Stechow, A. (eds.) Meaning, Use, and Interpretation of
Language. pp. –. Berlin: Walter de Gruyter.
Hobbs, J. R. () Literature and Cognition. Stanford, CA: Center for the Study of Language
and Information.
Jackson, F. () ‘On the Semantics and Logic of Obligation.’ Mind. . . pp. –.
Kamp, H. () ‘A Theory of Truth and Semantic Representation.’ In Groenendijk, J. A.,
Janssen, T. and Stokhof, M. (eds.) Formal Methods in the Study of Language. pp. –.
Dordrecht: Foris.
Kolodny, N. and MacFarlane, J. () ‘Ifs and Oughts.’ Journal of Philosophy. . . pp. –.
Kratzer, A. () ‘Modality.’ In von Stechow, A. and Wunderlich, D. (eds.) Semantics: An
International Handbook of Contemporary Research. pp. –. Berlin: De Gruyter Mouton.
Kratzer, A. () Modals and Conditionals: New and Revised Perspectives. New York, NY: Oxford
University Press.
Kratzer, A. and Shimoyama, J. () ‘Indeterminate Pronouns: the View from Japanese.’ In
Otsu, Y. (ed.) Proceedings of the Third Tokyo Conference on Psycholinguistics. Tokyo: Hituzi
Syobo.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

dynamic expressivism about deontic modality 

Lassiter, D. () Measurement and Modality: the scalar basis of modal semantics. Ph.D. thesis.
New York University.
Lewis, D. K. () Convention: A Philosophical Study. Cambridge, MA: Harvard University
Press.
Mackie, J. L. () Ethics: Inventing Right and Wrong. New York: Penguin.
McCready, E. () ‘Emotive equilibria.’ Linguistics and Philosophy. . . pp. –.
Millikan, R. G. () Language, Thought and Other Biological Categories. Cambridge, MA: MIT
Press.
Millikan, R. G. () Language: A Biological Model. New York, NY: Oxford University Press.
Moore, G. E. () Ethics. Oxford: Oxford University Press.
Moss, S. () ‘On the Semantics and Pragmatics of Epistemic Vocabulary.’ Semantics and
Pragmatics. . , pp. –. [Online] Available from: http://dx.doi.org/./sp. [Acccessed
 Nov .]
Murray, S. E. () Evidentiality and the Structure of Speech Acts. Ph.D. thesis. Rutgers
University.
Murray, S. E. () ‘Varieties of Update’. Semantics and Pragmatics. . , pp. –. [Online]
Available from: http://dx.doi.org/./sp... [Accessed  Nov .]
Murray, S. E. and Starr, W. B. () ‘The Structure of Communicative Acts.’ Manuscript Cornell
University.
Muskens, R. van Benthem, J. and Visser, A. () ‘Dynamics.’ In van Benthem, J. and ter
Meulen, A. (eds.) Handbook of Logic and Language. st edition. pp. –. Cambridge, MA:
MIT Press.
Ninan, D. () ‘Two Puzzles About Deontic Necessity.’ In Gajewski, J., Hacquard, V. and
Yalcin, S. (eds.) New Work on Modality, MIT Working Papers in Linguistics. Volume .
Cambridge, MA: MIT Press.
Owren, M. J., Rendall, D. and Ryan, M. J. () ‘Redefining Animal Signaling: influence versus
information in communication.’ Biology and Philosophy. . . pp. –.
Portner, P. () Modality. New York, NY: Oxford University Press.
Portner, P. () ‘Permission and Choice.’ In Grewendorf, G. and Zimmermann, T. E. (eds.) Dis-
course and Grammar: From Sentence Types to Lexical Categories. Berlin: Mouton de Gruyter.
Potts, C. () ‘The Expressive Dimension.’ Theoretical Linguistics. . . pp. –.
Ridge, M. () Impassioned Belief. New York, NY: Oxford University Press.
Ross, A. () ‘Imperatives and Logic.’ Theoria. . . pp. –. References to Ross ().
Ross, A. () ‘Imperatives and Logic.’ Philosophy of Science. . . pp. –.
Rothschild, D. () ‘Expressing Credences.’ Proceedings of the Aristotelian Society. . .
pp. –.
Schroeder, M. (a) Being For: Evaluating the Semantic Program of Expressivism. New York,
NY: Oxford University Press.
Schroeder, M. (b) ‘Expression for Expressivists.’ Philosophy and Phenomenological Research
. . pp. –.
Schroeder, M. (c) ‘How Expressivists Can and Should Solve Their Problem with Negation.’
Noûs. . . pp. –.
Schroeder, M. (d) ‘What is the Frege-Geach Problem?’ Philosophy Compass. . .
pp. –.
Schroeder, M. () ‘Two Roles for Propositions: Cause for Divorce?’ Noûs. . . pp. –.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 william b. starr

Schroeder, M. () Attitudes and Epistemics. Expressing Our Attitudes: explanation and
expression in ethics. Volume II. New York, NY: Oxford University Press.
Scott-Phillips, T. and Kirby, S. () ‘Information, Influence and Inference in Language
Evolution.’ In Stegmann. U. (ed.) Animal Communication Theory: information and influence.
pp. –. Cambridge: Cambridge University Press.
Silk, A. () ‘Truth-Conditions and the Meanings of Ethical Terms.’ In Shafer-Landau, R. (ed.)
Oxford Studies in Metaethics. Volume . pp. –. New York, NY: Oxford University Press.
Silk, A. () ‘How to Be an Ethical Expressivist.’ Philosophy and Phenomenological Research.
. . pp. –.
Simons, M. () ‘Dividing Things Up: The Semantics of or and the Modal/or Interaction.’
Natural Language Semantics. . . pp. –.
Stalnaker, R. C. () ‘Assertion.’ In Cole, P. (ed.) Syntax and Semantics : Pragmatics. pp.
–. New York, NY: Academic Press. References to Stalnaker ().
Stalnaker, R. C. () Inquiry. Cambridge, MA: MIT Press.
Stalnaker, R. C. () Context and Content: Essays on Intentionality in Speech and Thought.
Oxford: Oxford: Oxford Univsersity Press.
Starr, W. B. () Conditionals, Meaning and Mood. Ph.D. thesis. Rutgers University.
Starr, W. B. () ‘A Preference Semantics for Imperatives.’ Ms. Cornell University.
Swanson, E. () Interactions with Context. Ph.D. thesis. MIT.
Swanson, E. (forthcoming) ‘The Application of Constraint Semantics to the Language
of Subjective Uncertainty.’ Journal of Philosophical Logic. [Online] Available from:
link.springer.com/article/.%F---. [Accessed  Nov .]
Unwin, N. () ‘Norms and Negation: A Problem for Gibbard’s Logic.’ The Philosophical
Quarterly. . . pp. –.
van Benthem, J. () ‘Semantic Parallels in Natural Language and Computation.’ In
Ebbinghaus, H. D., Fernandez-Prida, J., Garrido, M., Lascar, D. and Artalejo, M. R. (eds.)
Logic Colloquium ’: Proceedings of the Colloquium held in Granada, Spain. Volume  of
Studies in Logic and the Foundations of Mathematics. pp. –. Amsterdam: North-Holland.
van Benthem, J. and Liu, F. () ‘Dynamic Logic of Preference Upgrade.’ Journal of Applied
Non-Classical Logics. . . pp. –.
van Eijck, J. and Visser, A. () ‘Dynamic Semantics.’ In Zalta (ed.) The Stanford
Encyclopedia of Philosophy, winter. [Online]. Available from: http://plato.stanford.edu/
archives/win/entries/dynamic-semantics/. [Accessed  Nov .]
Veltman, F. () ‘Defaults in Update Semantics.’ Journal of Philosophical Logic. . .
pp. –.
Willer, M. () ‘Dynamic Thoughts on Ifs and Oughts.’ Philosophers’ Imprint. . . pp. –.
[Online] Available from: http://hdl.handle.net//spo.... [Accessed  Nov
.]
Yalcin, S. () ‘Epistemic modals.’ Mind. . . pp. –.
Yalcin, S. () ‘Nonfactualism about Epistemic Modality.’ In Egan, A. and Weatherson, B.
(eds.) Epistemic Modality. pp. –. New York, NY: Oxford University Press.
Yalcin, S. () ‘Bayesian Expressivism.’ Proceedings of the Aristotelian Society Supplementary
Volume. . . pp. –.
Zimmermann, T. E. () ‘Free Choice Disjunction and Epistemic Possibility.’ Natural
Language Semantics. . . pp. –.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i


Metanormative Theory and the
Meaning of Deontic Modals
Matthew Chrisman

 Introduction
Metanormative theory is, in part, about the meaning of sentences such as “One ought
always act for reasons for which one could consistently allow everyone else to act as
well,” and “One should proportion degree of belief solely in accordance with evidence.”
There are other normative terms, of course, but ‘ought’ (≈ ‘should’), is clearly one of
the handful of core normative terms, and I will focus on it here.1
Given this focus, here is a way to put an idea that I think many philosophers
find attractive: ought-claims prescribe possible action, thought, and feeling, rather
than describing how things actually stand in reality. Arguably, this idea underwrites
the popularity of the is/ought divide, explains some of the attraction in Moore’s
Open Question Argument, and motivates some of the interest philosophers have had
in noncognitivist, emotivist, prescriptivist, projectivist, and expressivist accounts of
normative thought and discourse.
Here is another popular line of thought that has sometimes seemed to be in tension
with that initial idea: A central part of the best overall account of the meaning of
declarative sentences will explain how their truth conditions can be derived from the
semantic values of their basic components and their logical form. The word ‘ought’
can, of course, figure in nondeclarative sentences, but the ought-sentences primarily
at issue in metanormative theory are grammatically declarative. This implies that
our overall theory of meaning should provide a compositional assignment of truth
conditions to ought-sentences as a part of its treatment of declarative sentences more
generally. This seems necessary for accounting for the fact that it is obviously mean-
ingful to embed ought-sentences in propositional contexts, such as under the truth

1 Returning only very briefly at the end to reflect on the broader issue of how my conclusions here might
bear on similar issues regarding other normative and evaluative vocabulary.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 matthew chrisman

predicate, under epistemic modals, in the antecedent of conditionals, in disjunctions,

or in the context of “S believes/knows/doubts/fears/hopes that . . .,” etc.
One reason this line of thought has seemed in tension with the distinction many
philosophers have wanted to draw between ‘is’ and ‘ought’ is that proponents of the
most prominent family of antirealist metanormative views have sometimes suggested
that normative sentences including ought-sentences are not truth-apt. Or they have
suggested that normative sentences have truth conditions but only in a “deflationary
sense” supposedly incompatible without reliance on truth conditions for substantive
and informative explanations of the meanings of these sentences.
However, it’s important not to underestimate the challenges facing this idea.
Ought-sentences embed just like other declaratives in propositional contexts, and
our best account of the compositional function of these other declarative sentences
is (arguably) one that seeks a perspicuous display of the functional interaction of the
nodes of the correct tree-like representation of the logical form of each sentence, where
these functions are ultimately functions from the semantic values of the parts of the
sentence to the value true—that is to say, a theory which assigns truth conditions
to these sentences. Any theory that would propose otherwise for ought-sentences is
under extremely strong pressure to propose otherwise for all declarative sentences. At
this level of generality, however, the truth-conditionalist approach to the semantics
of declarative sentences is widely (though not universally2 ) assumed to be the best
game in town for articulating the compositional structure of declarative sentences.
This is why many philosophers are strongly inclined to think we should assign ought-
sentences truth conditions, and not as a mere deflationary afterthought but as a
central part of the explanation of how their meaning is a function of the meanings
of their parts.
Does that mean we have to choose: either accept the Herculean task of developing
a new approach to compositional semantics that eschews the successes of truth-
conditionalist approaches even for nonnormative language or treat ought-claims as
describing reality after all (and “locate” the normative properties that would make
sense of that)? Here, I want to argue against thinking that this is an important choice
point at the intersection of metanormative theory and compositional semantics. To
that end, I pursue three main tasks in this chapter. First, I explain a specific version
of the semantic rule for ‘ought’ that I favor. This is a refinement of the standard
treatment of ‘ought’ as a context-sensitive necessity modal. Secondly, I explain two
ways we might interpret the results of truth-conditional semantics: a more familiar
“representationalism” and a less familiar “inferentialism.” I view these as different

2 Nothing I say in this paper will engage substantively with the question of whether a truth-conditionalist
approach to compositional semantics is preferable in general to some dynamic alternative such as one
based on Discourse Representation Theory (Kamp, ), File Change Semantics (Heim, ), or Update
Semantics (Veltman, , ). Here I shall simply assume (along with most metanormative theorists) that
a truth-conditionalist approach is the default position for nonnormative declaratives, and so the interesting
question is what to make of the fact that it is also appropriate for normative declaratives.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

metanormative theory and the meaning of deontic modals 

kinds of metasemantic accounts of what it is in virtue of which declarative sentences

have the contents that they do.3 , 4 I want to explain these views in order to argue that,
with my semantic rule for ‘ought’ in hand, both metasemantic accounts can respect the
apparent prescriptivity of (some) ought-claims and both can accept relatively standard
compositional truth conditions for ought-claims without committing to the idea that
these claims describe how things actually are in reality. What distinguishes them is how
they respect these ideas, not whether they can do so. Finally, I discuss the “ideational-
ist” thesis that sentences are primarily vehicles for expressing our ideas and so get their
truth conditions from their conventional expressive potentials. This may be an attrac-
tive home for some sophisticated contemporary expressivist views in metanormative
theory. I compare them with inferentialism because I think it is not completely clear
how such sophisticated versions of expressivism fit in. Depending on how they are
developed, they might be viewed as compatible with and so perhaps usefully combined
with inferentialism; or they might constitute a third competing metasemantic view.
I hope that comparison of aspects of this somewhat more familiar view to the
inferentialist view I develop earlier in the chapter will help to clarify what is at stake.

 ‘Ought’ and Truth Conditions

As with many other modal auxiliary verbs in English, ought-sentences come in diverse
flavors, e.g.

3 By using language of “in virtue of ” here, I mean to indicate that they are very general grounding
explanations that seek to identify the kinds of nonsemantic facts that undergird the semantic facts postulated
by empirical semantics. Compare Charlow (a, p. ) and Yalcin (, p. ). Some (e.g. Speaks, )
may prefer the term “foundational theory of meaning” or “theory of the foundations of meaning” for what I
am calling metasemantics. Yalcin () surveys arguments for a distinction between content and semantic
value, where the former is (at least) what is said by uses of a sentence and what grounds the intentionality
and causal efficacy of mental states expressed by such uses, and the latter is what is attributed to sentences to
explain various semantic facts, e.g. about productivity of language, entailment relations between sentences,
the un/acceptability or un/interpretability of particular sentences, etc. Given this distinction, there are two
domains about which we might ask second-order questions: we can ask about what it is in virtue of which
speech-acts and the mental states they express have the contents that they do, and we can ask about what
it is in virtue of which sentences have the semantic values that they do. While I am highly sympathetic to
this distinction, observing it in what follows would have unnecessarily complicated an already complicated
discussion. It may be the case that representationalism and inferentialism are best viewed as second-order
accounts of content rather than—or at least in addition to—semantic value. However, I write here under
the simplifying assumption that “metasemantics” is about what I sometimes call “semantic content,” which
is meant to cover both.
4 Perez Carballo () usefully distinguishes a “hermeneutic” question about how to interpret the
formal specifications of truth conditions in truth-conditional semantics from an “explanatory” question
about what it is that explains the (approximate) correctness of the formal specifications of truth conditions
for a sentence. I agree that these are distinct, but I also think the metasemantic views I will discuss here are
all at a high enough level of generality that they can be seen as offering package answers to these questions.
I’m mainly interested in the explanatory question, but sometimes I will ask about how we are to interpret the
results of truth-conditional semantics. By this, I don’t mean just what hermeneutic gloss of the formalism
is appropriate but also what underlying facts about language and language users explain what it is in virtue
of which sentences have the truth conditions that they have.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 matthew chrisman

() We ought to do more to relieve great suffering.

() Bob ought to give up smoking.
() To support a loft-bed in plaster walls, anchor bolts rather than screws ought to
be used.
() Larry, who has had the most unfortunate time of things recently, ought to win
the lottery.
() The storm ought to hit shore before midnight.
In light of this diversity, semanticists typically treat ‘ought’ as a flexible necessity modal
and give its semantics in terms of a universal quantification over possible worlds. For
example, they might use a rule something like:
Modal Rule: [[ought p]]fg = T iff for all w ∈ Wfg , w ∈ p
which says that an ought-sentence is true just in case its prejacent p is true at each
possible world in some contextually determined set W (usually, the set of worlds
consistent with some set of background conditions f and highly ranked by some
ordering function >g ).5
This rule is attractive for the way it generates decent interpretations of all of the
examples ()–() above without positing lexical ambiguity (which is good since the
diversity of flavors of modal verbs is cross-linguistically robust, suggesting that ‘ought’
has univocal meaning). In each case, the difference in flavor is explained as a difference
in the contextual determination of the set of worlds quantified over. As long as context
can determine different values for f and g to input into the modal rule, we can provide
a very systematic interpretation of diverse ought-sentences.
Some metaethicists may be tempted to suggest an old-school reductive analysis of
‘ought’ in terms of obligations, fittingness, or reasons. However, without prejudicing
the separate issue of whether our ought-concept is constitutively connected to these

5 This is a simplified version of the rule predicted by Kratzer () for necessity modals. In addition to
the fact that it suppresses world-relativity of semantic values and elides the distinction between sentences
of the object language and metalinguistic variables, it also prescinds from the interesting issue of how
to account for ‘ought’s apparent weakness compared to ‘must’. Kratzerian proposals for handling this are
explored in Fintel and Iatridou (), Swanson (), and Silk (). Wertheimer (), Finlay (),
and Portner and Rubinstein () make somewhat different proposals, which are nonetheless consistent
with the basic idea that ‘ought’ is a necessity modal and to be modeled in terms of a universal quantification
over a set of worlds. Here I will also suppress the fascinating question of how (if at all) this rule should
be modified to handle the distinction between the objective ‘ought’ of what’s actually best and the so-
called subjective ‘ought’ of what best optimizes ones choices given imperfect information and varying
standards for practical decisions. For discussion, see Kolodny and MacFarlane (), Björnsson and Finlay
(), Dowell (), Cariani (), Charlow (), Carr (), Silk (), Bronfman and Dowell (),
Wedgwood (). Furthermore, there are other metaethically important challenges to this semantics that
will remain off stage here. For instance, the standard account also has difficulty with sentences of the form
“If p, then it ought to be the case that p” (see Zvolenszky, ; Carr, ). It seems to rule out strong
ought-dilemmas by semantic fiat rather than philosophical argument. See Swanson () and Fintel ()
for discussion. I think each of these problems can be addressed within a broadly intensional semantics for
‘ought’, which is all I really need for my argumentative purposes below; but I won’t discuss this further here.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

metanormative theory and the meaning of deontic modals 

other concepts in important ways, we should recognize that the modal rule’s ability to
provide such a systematic treatment of sentences as diverse as ()–() without positing
lexical ambiguity is an impressive feat—arguably one that any semantic analysis of
‘ought’ in terms of obligations, fittingness, reasons, or what’s best will fail to do.6
Nevertheless, I still believe there is a lesson proponents of the modal rule might learn
from the urge amongst metanormative theorists to connect ‘ought’ to notions such as
obligations, fittingness, or reasons. Ought-sentences seem, when used in the sense that
is of most interest in metanormative theory, to call on agents to do something; and
arguably this is a more fine-grained notion than a proposition’s being true at possible
worlds. Here is Geach in a classical statement of the objection:
If the deontic operator applied to whole propositions, then the results of modifying the
active and the equivalent passive form with an ‘ought’ must likewise be equivalent. But such
equivalences often appear not to hold. ‘John beats up Tom’ and ‘Tom is beaten up by John’ are
equivalent; but it looks as though ‘John ought to beat up Tom’ and ‘Tom ought to be beaten up
by John’ are not necessarily equivalent. (, p. )

In light of this good point, rather than abandoning the modal rule altogether,7 I think
we should enhance it by integrating it with a semantics for imperatives. After all, saying
that S ought to φ—in the sense of most interest in metanormative theory—seems
closely related to issuing the imperative, “S φ!”
How might we enhance the modal rule in a way that respects this connection? Here’s
a start: Formally, semantic models typically assign declarative contents (propositions)
truth values relative to points of evaluation, which are conceived as (at least) possible
worlds (which in turn might be thought of in a variety of ways, e.g. as concrete possible
universes, abstract ways reality could be, sets of propositions, or something else). As
far as formal semantics goes, we might also want to assign imperative contents (what
I’ll call “prescriptions”) a kind of semantic value relative to a point of evaluation—
after all, imperatives are also logically-composed sentences of language standing in
semantic relations such as entailment and equivalence to other sentences—but the
kind of semantic value and the conception of the point of evaluation has to be
different (since imperatives do not seem to have truth values).8 One proposal that I
like assigns imperatives the semantic values legitimate/illegitimate relative to possible
norms (which in turn might be thought of in a variety of ways, e.g. as concrete laws of

6 I consider various attempts to reductively analyze ‘ought’ in these terms in chapter  of Chrisman
(); the main stumbling block is plausibly capturing the meaning of epistemic ought-sentences such as
() while retaining a plausible treatment of paradigmatic normative ought-sentences such as () and ().
7 Schroeder () argues that it motivates treating ‘ought’ as ambiguous between raising-verb and
control-verb readings. I discuss this proposal and show why it isn’t linguistically motivated in Chrisman
(a). See also Finlay and Snedegar () for pragmatic explanations of many of the phenomena that
lead Schroeder to posit ambiguity.
8 A point which is argued for in greater detail in Charlow (b, section ), Starr (, section ), and
Chrisman and Hubbs (unpublished, section ).

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 matthew chrisman

possible practices, abstract ways agents might behave, sets of general prescriptions, or
something else).9
It is not important here whether this is the right semantics for imperatives, but if
we posit the imagined prescriptional contents with legitimacy values rather than truth
values, then we can embed this in our semantics for ‘ought’ by adopting something like
the following as a new general rule articulating the truth-conditional contribution of
‘ought’:
Enhanced Modal Rule: [[ought p]]fg = T iff for all <w, n> ∈ Pfg , <w, n> ∈ p
Here we no longer restrict p to propositions that might be expressed by declarative
sentences and modeled as a set of possible worlds; instead we allow that it could also
be the sorts of prescriptions that might be expressed by imperative sentences, and we
model it in general as a set of world-norm pairs (the worlds where the prescription is
legitimate relative to the norm). This rule says that an ought-sentence is true10 just in
case its prejacent is correct relative to each member of some contextually determined
set of world-norm pairs. This means, if we allow that the prejacent p could be either
propositional or prescriptional, we get two species of ought-claims out of a univocal
semantic rule. When the prejacent is propositional, the rule works similarly to the one
above: the ought-sentence is true just in case the prejacent is true relative to each of
some contextually determined set of world-norm pairs, where the norm parameter
is usually idle. By contrast, when the prejacent is practical, the rule evaluates its
legitimacy relative to each of some contextually determined set of world-norm pairs,
where the norm parameter is usually crucial.11

9 Compare Castañeda (, chapters –) for the most detailed way of working out this basic idea about
imperatives (but note that he uses the term “prescription” in a different way to how I am here). Charlow
(b) defends a related idea in an expressivistic framework which treats imperatives and declaratives as
having properties of representations of states of mind as their semantic values, where ordinary declaratives
denote properties of representations of beliefs and imperatives denote properties of representations of
preferences. Hauser (), Segerberg (), Portner (), and Hanks (, chapter ) defend competing
views about the semantic value of imperatives that can still be used to generate sentence-level contents
that are not propositions, which in turn can be thought to be embedded under ‘ought’s. In Chrisman and
Hubbs (unpublished), we show how treating such nonpropositional contents as the semantic values of
imperatives can be motivated by action-theoretic considerations ancillary to the project of compositional
semantics. As indicated above, what is crucial here is not the correctness of this approach to the semantics
of imperatives but rather that it can handle the content embedded under prescriptive-seeming ‘ought’s. A
perhaps more orthodox view is that imperatives have action-related properties as their contents and their
semantic relations to other sentences are derivative of the semantic relations between the propositions that
result from ascribing these properties to individuals. I think this view has trouble explaining the semantic
difference between necessarily coextensive but nonsynonymous imperatives, and I doubt that it gets the
logic of imperatives right. But that doesn’t matter here since I am not relying on the view of prescriptions
above as the right view of the content of imperatives.
10 Ought-sentences are clearly declaratives and so they get truth values and have truth conditions on this
view. Although this treatment of their truth conditions allows that they can embed prejacents which are not
truth-apt, this is not a version of the failed noncognitivist idea that ought-sentences are neither true nor
false.
11 This bears some similarity to the proposal developed in Gibbard (, chapter ) and Gibbard (,
chapters. –) to extend the standard possible-world semantics, adding an extra norm-index relative to

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

metanormative theory and the meaning of deontic modals 

With a rule such as this one, I think we overcome the worry from Geach about
applying truth-conditional semantics to ‘ought.’ The prescriptions John, beat up Tom!
and Tom, be beaten up by John! are different from each other (and from the proposition
that John beats up Tom); they could easily differ in whether they are legitimate relative
to a norm. The enhanced modal rule lets us understand the truth conditions of ought-
sentences in terms of a universal quantification over a contextually determined set of
possibilia.12 Even within a univocal treatment, we can recognize an important seman-
tic difference between obviously propositional ought-sentences, such as () and (),
and obviously prescriptional ought-sentences, such as () and (). It may be less clear
where other ought-sentences fall in respect of this divide, but that is plausibly because
determining this requires more information about the context in which they are used.
In any case, the enhanced modal rule is the rule articulating ‘ought’s contribution
to truth conditions that I will use in the rest of this paper. As I indicated above,
I will not provide any more argument for this semantics of ‘ought’ than is already
in the barebones explanation of the rule, and I fully grant that further refinements
of it are needed for various purposes in semantics. My purpose in explaining it
is mainly to frame the question I want to pursue in the rest of this paper: how
should we interpret this rule? Recall at the outset I suggested that a central debate
in metanormative theory seems to be pulled in two directions based on whether one
takes more seriously the idea that ‘ought’s (at least the ones relevant to metanormative
theory) are distinctively prescriptive, or the idea that all declarative sentences should
be given a truth-conditionalist semantics. With something like the enhanced modal
rule in hand, I think we can see this to be a false choice by exploring two different
metasemantic accounts of what this rule tells us about the facts in virtue of which
ought-sentences have the contents that they do. Both of these accounts are consistent
with the idea that some ought-claims function not as descriptions of what is actually
the case but rather prescriptions for what someone is to do in various circumstances.
The interesting question is how they differ in capturing this idea.

 Truth Conditions as Ways Reality Could Be

The first account stems from what I take to be a pervasive view in contemporary
philosophy. This is that declarative sentences have the contents that they do in virtue

which some sentences could be evaluated. However, Gibbard uses this extension to provide semantic values
for declarative sentences with normative elements, in essence committing to either a kind of noncognitivism
or a kind of norm-relativism about the possibility of true normative declaratives. By contrast, I’m applying
this idea only to nonpropositional contents, like those we might assign to imperatives, which I deny are
truth-apt and for which I think it is much more plausible to assign some kind of semantic value other than
truth.
12 An alternative attempt to do this posits elided “stits” (“sees-to-it-that”) in the prejacent proposition
of ‘ought’s intuitively thought to be practical. As a semantic proposal, I think this is ad hoc and non-
explanatory. But arguing for that is not necessary to the points I want to make here.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 matthew chrisman

of how they represent reality as being. To see how this works, consider first a simple
sentence, such as
() Grass is green.
It is very tempting to think that sentences such as this one are true just in case reality
is the way they represent it as being. That is to say, e.g., that () represents grass as
having the property of being green. Generalizing from this example, it is tempting
to think that pursuit of a truth-conditional semantics boils down to the development
of a recursive method for spelling out the way reality would have to be in order for
each declarative sentence of the language to be true. Using the terminology I prefer,
we can call this a representationalist explanation of truth conditions. Arguably, it is a
core assumption of all traditional realist and error-theoretic views in metaethics. They
would say that
() Stealing is wrong.
is true just in case reality is the way () represents it as being. Then there is room
for debate among these philosophers about the nature of this putative piece of
reality (naturalism vs nonnaturalism) and whether it actually obtains (success-theory
vs. error-theory).
The pervasiveness of this representationalist interpretation of truth-conditional
semantics can help to make some sense of why anyone would have ever denied,
as early noncognitivists did, that normative sentences have truth conditions. For,
on the face of things, it seems obvious that some normative sentences are true and
others are false; and from the point of view of ordinary language, embedding ethical
sentences such as () under the truth predicate and other propositional contexts
such as belief/knowledge attributions and epistemic modals seems beyond reproach.
However, if your view is that normative language at its base is prescriptive rather than
descriptive, and you accept (at least tacitly) the representationalist conception of what
truth-conditional semantics is telling us, then you might be tempted to argue that
normative sentences do not have truth conditions (at least not in any theoretically
interesting sense).
Below I’ll suggest that there’s a way that even those metanormative theorists inspired
by noncognitivism can resist this temptation—as they should, since it is difficult to
see any hope of realizing the aspirations of compositional semantics if we pursue
a bifurcated approach to semantics, using truth conditions as our basic semantic
framework for nonnormative sentences such as () but something else for normative
sentences such as ().13 But before we get there, I want to explain why I think the
metanormative situation with respect to ‘ought’ complicates things in a way that allows

13 This is the main lesson of Hale (), Dreier (), Unwin (). Schroeder () shows how
difficult it is for an expressivist to avoid this even by advancing a psychologistic semantics across the board,
i.e. applied also to paradigmatically descriptive sentences.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

metanormative theory and the meaning of deontic modals 

even the hardcore representationalist to accept, indeed to champion the idea that
normative language (at least insofar as it is “fraught with ought”) is prescriptive rather
than descriptive, undercutting some of the traditional noncognitivist motivation for
denying that normative sentences have truth conditions.
Here’s why things are more complicated. Even if you accept the representationalist
conception of truth conditions in the basic case, you will want to allow there are some
declarative sentences that do something other than describe what is actually the case.
This is because some sentences are formed with intensional operators, e.g.
() Jack and Mary might be in the house.
And intensional operators do not seem to function like normal predicates. Instead,
they are generally thought by semanticists to serve the purpose of shifting some
parameter of the circumstance relative to which we evaluate the content that they
embed. That is, for instance, they tell us to evaluate some embedded proposition not
at the actual world but at some possible world(s). For example, the idea is to think of
() as true just in case the embedded proposition that Jack and Mary are in the house
is true in some world regarded as possible given a particular body of evidence. This is
why its truth conditions are usually given with something like the following:
[[()]]c = T iff [[Jack and Mary are in the house]]c = T in some possible world
consistent with our evidence in this world
As a working hypothesis, this is all relatively familiar from textbook treatments of
intensional operators in semantics.14 It may need to be refined, but the important
question at this stage is not semantic but metasemantic: What is the representationalist
gloss of these truth conditions? One fairly natural thing for the representationalist to
say is that these truth conditions tell us how this sentence represents reality as being,
but the piece of reality it represents is more complicated than anything having to do
merely with what is actually the case. The thought is that () represents how things are
not just in the actual world but also in other possible worlds; this involves representing
some kind of relation between possible worlds. So, if we assume an ontology of
possible worlds,15 we can continue to think of all declarative sentences, even those
formed by intensional operators, as representing ways reality could be. But (here’s the
payoff) we can simultaneously deny that all declarative sentences describe how things
actually are. For on this view some sentences represent a piece of reality that involves

14 Not that it is uncontroversial. Even Kratzer () recognizes that its application to epistemic modals
faces challenges making sense of apparent disagreement, which has led some to follow MacFarlane ()
in complicating the definition of truth for such claims, relativizing to a context of assessment; but see von
Fintel and Gillies (, ) for criticism and a more conservative response.
15 What about linguistic Ersatz-theorists who reject the reality of possible worlds but contend that some
other part of our ontology can play the role of possible worlds in our interpretation of the truth conditions
of modal sentences? Whether they count as representationalists in the sense I’m trying to capture here is
going to depend on how exactly they conceive of the other part of our ontology as playing the role of possible
worlds.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 matthew chrisman

a relation between the actual world and other merely possible worlds. Hence, on the
representationalist interpretation of the standard truth-conditionalist semantics for
them, might-sentences are no mere description of how things actually are.

 Application to ‘Ought’
I hope it’s already clear that this is relevant to the apparent tension between the
idea that ‘ought’s do not describe what is actually the case but prescribe possible
action, thought, and feeling, and the idea that all declarative sentences deserve the
same general semantic treatment, e.g. in terms of compositionally specified truth
conditions. Even if we just use the simple modal rule for ‘ought,’ it can be viewed like
‘might’ as a device for shifting some parameter of the circumstance relative to which
we evaluate the propositions it embeds rather than for describing some feature of
what is actually the case. Then, representationalists will interpret the truth conditions
this rule predicts for ought-sentences as specifying how the sentences represent a
complex modal feature of reality as being—not as merely describing how things
actually are.16 But more to the point, if we use the enhanced modal rule, I think the
representationalist can even capture an important sense in which at least some ‘ought’s
are prescriptive rather than descriptive. For example, we can use this rule to interpret
() Jack and Mary ought to leave.
predicting the following truth conditions:
[[()]]fg = T iff [[Jack and Mary, leave!]] = L in for all < wn >∈ Pfg
This means that () is true just in case its prejacent prescription is legitimate relative
to all of the relevant points of evaluation (in this case world-norm pairs highly
ranked by >g that are consistent with background conditions f ). Then, assuming
an ontology of possible worlds and possible norms, representationalists can interpret
these truth conditions as representing a modally complex way reality could be: a way
the actual world is related to possible worlds and possible norms. More specifically,
representationalists will think the formula above tells us that the prescription for Jack
and Mary to leave is legitimate across a contextually determined range of possible
norms in light of the circumstances common in a contextually determined range of
possible worlds. In this way, () is assigned truth conditions, but since these are truth
conditions involving what is prescribed by norms in various possible worlds, this

16 So it’s possible to reject my suggestion that ‘ought’s can embed prescriptive prejacents and still buy
the rest of what I say here as long as there’s an alternative way to capture the prescriptivity of some ‘ought’s.
Some may be inclined to do this by treating one of the ordering sources evoked by ‘ought’ (according to
the modal rule) as ordering worlds in terms of whether some set of prescriptions is satisfied. I think this is
going to be too coarse-grained to respect Geach’s point from above, but a lot will depend on how we work
out the details of prescriptive ordering sources.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

metanormative theory and the meaning of deontic modals 

sentence can be recognized as prescriptive rather than descriptive of what is actually

the case.
So, if that representationalist story is on the right track, then it’s wrong to think we’re
forced to choose between a truth-conditionalist approach to semantics and the idea
that ought-sentences are often prescriptive rather than descriptive of what is actually
the case. By conceiving of ought-sentences as representing a complex modal piece of
reality, involving possible norms and what they prescribe in various possible worlds,
this apparent dilemma is avoided.
It is avoided, we should note, by divorcing the notions of describing what’s actually
the case and representing reality, allowing that the truth conditions of some sentences
(at least those involving intensional operators) tell us that these sentences have the
contents that they do in virtue of representing more than what is actually the case.
They represent something like merely possible worlds and merely possible norms and
their relations to the actual world. Hence, the representationalist way of avoiding the
dilemma comes with ontological (and associated epistemological) costs.
This is a familiar worry in metanormative theory, but I think the representationalist
interpretation of a workable truth-conditional semantics for ‘might’ or ‘ought’ reveals
how it extends into the metaphysics of modality. By treating intensional operators as
generating representations of reality, the articulation of their truth conditions gestured
at above would force us to understand anyone committed to the truth of sentences such
as () or () as committed to the reality of possible worlds and possible norms (or their
ersatz-replacements). Are the merely possible worlds, e.g. where there are dragons in
the command of a princess, and the merely possible norms, e.g. which require dragons
to blow fire on the princess’s enemies, just as real as the greenness of grass or the
location of people in the house? Once in the representationalist mindset, it may be
difficult to see how the answer could be anything other than “yes”; but taking a step
back, many metanormative theorists are going to worry that this is an ontologically
profligate position affronting to common sense about the difference between what
is real and what is imaginary/virtual/fictional/merely-possible (not to mention the
more theory-laden distinction between what is the case in reality and what reality
might/ought/etc. to be like).17
Perhaps this ontological cost (and any related epistemological costs) should be paid.
I’d certainly be willing to pay provided it was required to get the benefits of system-
aticity and seamlessness provided by truth-conditionalist approaches to compositional
semantics. Below, however, I want to suggest that there are alternative metasemantic
accounts of the results of compositional semantics—ones not committed to the idea
that ought-sentences represent a (complex modal) way reality could be. I think this

17 Are these things perhaps real but not “just as real” as the greenness of grass or the location of people
in the house, because they are abstract? I doubt that the distinction between concrete and abstract elements
of reality helps here, as many of the things we probably want to say are “more real” than possible worlds and
possible norms would, on many ways of drawing this divide, count as abstract.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 matthew chrisman

coheres with the popular stance among many philosophers that possible-worlds talk
in semantics is a convenient fiction or a mere notational device, or that possible worlds
are “pleonastic entities” rather than real parts of our ontology.

 Truth Conditions as Positions in a Space

of Implications
Next, I want to introduce a different metasemantic view inspired by Sellars’s (,
) inferential-role account of semantic and mental content and developed in more
systematic detail in Brandom’s (, ) inferentialist version of pragmatism about
intentional content. I believe this is somewhat less familiar, so I’ll take a bit more
space to introduce it and bring it to bear on the metanormative issues introduced
above. The core idea is to reverse the normal order of explanation that moves from
a view about the semantic content of bits of language to a view about proprieties of
use in a social discursive practice. Instead, these inferentialists seek to explain why
sentences have the contents that they do in terms of what they are vehicles for doing
in a social discursive practice. However, not everything we can do with language is
relevant.18 When it comes to declarative sentences, these are conceived primarily as
vehicles for making assertions, which conception is then understood as undertaking
(acknowledging, or embracing) a particular kind of commitment.
More specifically, looking downstream from a sentence, various further sentences
would typically follow from its truth—e.g. part of the commitment carried by an
assertion of “Grass is green” is to grass’s being colored.19 Then, looking upstream
from a declarative sentence, it would typically follow from the truth of various other
sentences; and one who uses a declarative sentence to make an assertion is open
to a potential challenge to entitle oneself to this commitment by asserting one of
these other sentences—e.g. one who asserts “Stealing is wrong” might, if legitimately

18 This is, I believe, what distinguishes the view from Wittgenstein-inspired “use theories of meaning”
based on skepticism about drawing any principled line between the uses of language that reveal semantic
competence and those that reveal merely pragmatic understanding of the way language can be used to
achieve sundry goals. Brandom has recently characterized his linguistic pragmatism as the combination
of two principles. Methodological pragmatism: “the point of associating meanings, extensions, contents,
or other semantic interpretants with linguistic expressions is to codify proprieties of use,” and semantic
pragmatism: “all there is to effect the association of meanings, contents, extensions, rules or other semantic
interpretants with linguistic expressions is the way those expressions are used by the linguistic practitioners
themselves”(, pp. –). As long as we restrict the focus to the uses of language to make assertions,
which are distinguished in the way they license inference to other assertions and require defense via other
assertions, I think we get close to what most workaday semanticists take as their principal data: ordinary
speaker intuitions about the entailment, inconsistency, equivalence properties of sentences.
19 One may not always acknowledge the inferential consequences of what one asserts. So its being part
of what one is committed to is not the same as believing that it follows. But the more someone fails to
acknowledge what we take to follow from what one asserts, the less we will think she understands what she
is saying.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

metanormative theory and the meaning of deontic modals 

challenged, have to back it up with reasons, for instance arguing that stealing involves
a gross violation of the respect we owe to each other as human beings.20
Importantly, these upstream and downstream inferential connections are not
merely logical entailments but rather in the broader family sometimes called semantic
implications, which are usually defeasible and context-sensitive.21 For this reason,
competence with the meaning of sentences can be thought of, on the inferentialist
picture, as coming in degrees. How much one understands what one is saying depends
on how much one knows one’s way around a space of semantic implication-relations
containing the sentence and the types of things that would defeat these relations.22
Before we can bring this idea to bear on the metanormative issues introduced above,
there are two questions I want to address in order to flesh out the inferentialist idea.
First question: What does endorsing inferentialism mean for truth-conditional
semantics? Brandom sometimes writes as if inferentialism is inconsistent with appeal-
ing to truth conditions in one’s explanation of the semantic composition of declarative
sentences,23 but I doubt inferentialism and truth-conditional semantics are really
in tension, at least not when we understand “truth-conditional semantics” as the
approach in compositional semantics where one attempts to recursively derive some
specification of when a declarative sentence should be assigned the preferred semantic
value true relative to some parameters, thereby displaying the semantic function of the
sentence’s subsentential parts. Because of the pervasiveness of representationalism,
the representationalist gloss on the results of this approach is sometimes labelled
“truth-conditional semantics.” But I think it’s helpful to view representationalism

20 Often one may not actually be challenged to defend what one asserts, and, regarding certain kinds
of topics, one may be default entitled in most or all of the everyday situations where one would make an
assertion; but being the sort of linguistic move whose form makes it challengeable for reasons is essential
for being an assertion. (Brandom , pp. –)
21 We can then seek to recover specifically “logical” entailments by restricting ourselves to consideration
of the semantic implications ensured by the “logical” form and words of the sentence.
22 Sellars and Brandom tend to take a very expansive view of the implication-relations that are relevant
to the determination of meaning. However, any plausible version of their views would embrace the idea
that how central such a relation is to determining the meaning of a particular word/concept is something
that comes in degrees, and the lesson we were supposed to have learned from Quine’s attack of the
analytic/synthetic divide is that it is not possible, in general, to draw a sharp line between the implications
that hold in virtue of meaning alone and the implications that hold in virtue of something else (though if
we hold enough other stuff fixed, we might be able to draw a practically useful line in particular contexts).
23 For example, he writes, “. . . truth is not a concept that has an important explanatory role to play in
philosophy. Appearances to the contrary, are the result of misunderstanding its distinctive expressive role.
The word ‘true’ does indeed let us say things that in many cases we could not say without it. But when we
understand what it lets us say, and how it does that, we will see that the very features that make it expressively
useful make it completely unsuitable to do the sort of theoretical explanatory work for which philosophers
have typically enlisted its aid.” (, pp. –) Later, however, he strikes a slightly more conciliatory tone,
“What about the role of truth in semantic explanation, via a definition of propositional content in terms
of truth conditions? We certainly do use ‘true’ to say what the content of a claim is. . . . But it would be a
mistake to infer from this sort of appeal to truth conditions to express propositional contents that one can
explain what propositional contents are by appeal to the conditions under which sentences are true.” (ibid.,
p. )

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 matthew chrisman

and inferentialism as a metasemantic interpretation of the results of first-order com-

positional semantics (whatever these happen to be). More specifically, I think an
inferentialist should view the specification of the truth conditions of a sentence as an
initial spelling out of what one who asserts the sentence is committed to as a premise
with further downstream inferential consequences, and what one would have to entitle
oneself to by looking to upstream inferential antecedents were one to be legitimately
challenged. In this way, we might say that an assignment of truth conditions to a
sentence is an articulation of the position in a network of implications occupied by
one who uses the sentence to make an assertion in ordinary discursive practice.24
In cases such as () and (), this may appear trivial. Of course, one who asserts
that grass is green has undertaken a commitment describable as a commitment to
the proposition that grass is green; and one who asserts that stealing is wrong has
undertaken a commitment describable as a commitment to the proposition that
stealing is wrong. But when it comes to more complex cases where it is not entirely
clear what truth conditions to assign to a sentence, this way of thinking of things lets
us interpret candidate assignments as articulations of different positions in a network
of implications. For example, if someone asserts
() Visiting relatives can be boring,
is she committed to the proposition that those relatives who visit are boring, or is she
committed to the proposition that going to visit relatives is boring? Disambiguating
this sentence by assigning truth conditions in line with one or the other of these
interpretations has implications for what further commitments downstream from an
assertion of this sentence would carry and what further assertions one would have to
make upstream from this assertion in order to entitle oneself to it if legitimately chal-
lenged. In this way, assigning truth conditions can be viewed as an attempt to identify
what inferential commitments are implicitly endorsed, embraced, or acknowledged by
someone who uses the sentence to make an assertion. Rather than being—in the first
instance—commitments to reality’s being a particular way, they are commitments that
carry downstream inferential consequences and that must, in principle, be defensible
via upstream inferential antecedents.
Of course, representationalists will typically agree that declarative sentences have
upstream and downstream inferential connections to other sentences that are some-
how relevant to understanding semantic content. What distinguishes inferentialism is
that it appeals directly to the spectrum of these connections (rather than to anything
about representing reality) in its metasemantic account of what it is in virtue of which
sentences have the truth conditions that they do.
Second question: If declarative sentences are understood first and foremost as vehicles
for undertaking inferentially articulable commitments, what does this mean for the

24 Compare Williams (, pp. –; ), who also argues that inferentialism is compatible with a
truth-conditionalist approach to compositional semantics.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

metanormative theory and the meaning of deontic modals 

intuitive thought that many declarative sentences describe the way things are in the
world around us? Many philosophers seem to think that inferentialism undermines
the very idea of a sentence’s describing reality. The thought is something like this: if
semantic contents are positions in something we characterize as merely a “game” of
giving and asking for reasons rather than specifications of how reality has to be in
order for the relevant sentence to be true, then surely no sentences should be viewed as
describing reality. Alternatively, inferentialism is sometimes portrayed as taking what
is usually a local denial that some area of discourse is representational (e.g. normative
or modal) and generalizing it to all areas of discourse. Here, the thought is something
like this: If everything we say is somehow about how we should infer, and you think this
“should” is prescriptive rather than descriptive, then surely that means that nothing we
say is a description of how things are in the world around us.25
While granting that some things some inferentialists have written may suggest
that this is the view,26 I think a wholesale rejection of the idea that some declarative
sentences describe the world around us is an extreme version of the view not actually
endorsed by some of its central proponents. For example, Sellars was famously an
inferentialist about content and a realist about science, writing “in the dimension of
describing and explaining the world, science is the measure of all things, of what is
that it is, and of what is not that it is not” (, §). And Brandom characterizes
his inferentialism as supported by a “sophisticated expressivism” about logical, modal,
and normative vocabulary that “. . . is essentially, and not just accidentally, a local
expressivism.” This is because, in the way he understands the expressive role of
these vocabularies, it depends on the existence of other vocabularies that play a
descriptive role: “Autonomous discursive practices must contain vocabularies playing
other expressive roles—for instance, observational vocabulary that reports features of
the non-linguistic bits of the world” (, p. ). In the context of the arguments
here, this is good because any view that implies that ought-sentences aren’t descriptive
because no declarative sentence is descriptive could achieve only a pyrrhic victory.

25 For example, Price imagines what a quasirealist expressivist like Blackburn might say if he adopted a
Brandomian inferentialist conception of assertion. Price suggests the expressivist might say, “I used to think
of my Humean expressivism as a local position, applicable to some vocabularies but not others. . . . However,
I have now come to realise that for no vocabulary at all is it theoretically interesting to say that its function
is to ‘represent’ particular kinds of states of affairs. . . .This. . . does nothing to undermine the interesting
observations that got me started, about the distinctive—and different—functions of moral and modal
vocabulary. On the contrary, it simply implies that they are exemplars of an approach to language we should
be employing everywhere. In other words, what I took to be linguistic islands are simply the most visible
extremities of an entire new continent—a universal program for theorising about language in expressivist
rather than representationalist terms . . .” (, pp. –).
26 Macarthur and Price () appear very sympathetic to this idea. But, in later work, Price (,
chapter ) has developed a more nuanced position, which trades on a distinction between thoughts that
carry information (what he terms i-representations) and a species of these which empirically track the mind-
independent world (what he terms e-representations). See also Blackburn () and Williams () for
useful discussion and repudiation of global expressivism in the sense at issue here.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 matthew chrisman

So, in my view, inferentialism is not committed to global expressivism or the idea

that language use is merely a “game” of giving and asking for reasons, rather than a
central part of how humans conceptualize the world and reason about what to do in
it. To the contrary, I think someone sympathetic with inferentialism might argue that
some or even many sentences have direct inferential consequences for our views about
the way reality is. Knowing how to draw the line between those that are and those that
aren’t is difficult, but maybe “Grass is green.” is such a sentence, in which case, one
who asserts it might be thought to be inferentially committed to the proposition that
grass has the (real, mind-independent, describable) property of being green. That is
to suggest that one of the direct downstream inferential connections of this sentence
is to another sentence deploying explicitly representational notions. Maybe we think
this is the best way to make sense of the idea that the use of such sentences is causally
regulated by tightly corresponding features of the world, such as the color of grass,
and that such use is plausibly treated on a tracking model of the relation between
language use and extralinguistic reality.27 If this were right about this sentence or any
other class of sentences, inferentialism might be seen as overlapping with the core
representationalist idea outlined above but as also offering a deeper explanation of the
relevant phenomena.
Of course, if we maintained that all declarative sentences had such direct infer-
ential consequences for our views about the way reality is, then we would have a
metasemantic view that was supposed to be an alternative to representationalism but
which turned out to be a sophisticated version of it.28 As we’ll see below, the textbook
expressivist move in this context is to insist on a distinction between mental states
with descriptive and directive roles in the psychology of motivation (“belief-like” and
“desire-like” attitudes as they are often called, with different “directions of fit”) and
argue that normative claims express mental states with a desire-like direction of fit
with the world. Whether or not this is consistent with truth-conditional approaches
to compositional semantics, especially as applied to ‘ought,’ is a complicated and

27 Williams (, pp. –) explores a more sophisticated way to cash out this distinction in terms of
various clauses in an explanation of meaning in terms of use. As Williams argues, some such explanations
may be ontologically conservative in that they do not mention referents (properties when it comes to
predicates) of the terms whose use they are appealing to in explaining meaning, but others won’t be. My
suggestion here (following Williams) is that the explanations of the meaning of color vocabulary in terms
of its use might not be ontologically conservative, while the explanation of deontic modal vocabulary is a
plausible candidate for an explanation that can be ontologically conservative.
28 Drawing on Peacocke (; ) and Block (), Wedgwood (, , chapter ) articulates a
theory of conceptual roles as determining the representational function of the words that express concepts
with these roles. On this view, each word refers to something, and what it refers to is determined in part
by the conceptual role of the concept it expresses; and it is the logical composition of these concepts that
determines how each declarative sentence represents reality as being. Because he uses a truth-conditional
semantic rule similar to the modal rule stated above, he ends up defending a conceptual-role interpretation
of this rule that bears important similarities to the one I develop below. However, because he treats all
words as referring to something in reality, his inferentialism makes no room for distinguishing between
representational and nonrepresentational vocabulary. By contrast, I see in the inferentialist metasemantics
resources for motivating a distinction at least when it comes to ‘ought.’

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

metanormative theory and the meaning of deontic modals 

controversial matter that I’ll touch on below. But as I understand inferentialism, at

least of the sort developed by Sellars and Brandom, it offers a different set of resources
for marking out ought-claims from those we typically think of as representing reality.
In short, this is to classify ‘ought’ with other modal terms and then to treat modal terms
along the lines of sentential logical operators. Typically, inferentialists explain why
logical operators have the content they do, not in terms of what in reality they represent
but instead in terms of something they allow us to do: state and endorse/reject claims
about the inferential commitments carried by the use of other sorts of terms. How
exactly does this work for ‘ought’?

 Application to Necessity Modals Including ‘Ought’

There is another idea championed by Sellars and Brandom that I think is not consti-
tutive of an inferentialist metasemantics but a natural extension of it. The idea is that
some terms do not earn their keep from the way they let us represent pieces of reality
but rather by serving as what we might call metaconceptual devices for embracing
commitment to features of the conceptual (inferential) framework in which ordinary
representational terms have determinate application. Sellars (, ) defended
an explicitly metalinguistic and inferentialist version of this claim, arguing that in
addition to descriptive words applying to empirically knowable pieces of reality, our
language contains “categorical” words whose core function is not to describe reality
but to embrace commitment to the inferential relations which must hold between
descriptive words for them to count as descriptive in the first place. In this way,
Sellars viewed philosophically puzzling words as many philosophers view logical
words: not as adding to the descriptive content of the sentences in which they figure
but rather as devices for affirming commitment (at least implicitly) to the inferential
connections between other words (often words which do add to the descriptive content
of the sentences in which they figure). On Sellars’s way of working out this line
of thought, implicit commitment to these kinds of inferential connections is seen
as a crucial element of ordinary competence with other descriptive words; and the
metaconceptual words provide, as Brandom puts it, means for making explicit those
implicit commitments.
There may be other philosophically interesting words that are metaconceptual
rather than representational in this sense.29 However, I think one of the most per-
suasive uses of this idea is for conceptual necessity claims. For example, consider:
() Necessarily, if x is east of y, then y is west of x.

29 Kant (, A/B), for example, identified twelve “pure concepts of the understanding” interpret-
ing them in something like this way. In a similar vein, Sellars () argued that terms such as “substance”
and “quality” are metalinguistic devices that get their content from the way they can be used to endorse the
types of inferential connections between other empirical terms, connections which must be in place, at least
implicitly, for these terms to count as applying to empirically knowable pieces of reality in the first place.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 matthew chrisman

Sellars and Brandom would say that this sentence does not represent a piece of reality
involving the modal relation between locations which are situated east/west of each
other. Rather it is a device for acknowledging the conceptual validity of inferences of
the form
x is east of y → y is west of x
thereby (partially) spelling out and explicitly committing to the inferential connec-
tions between more basic terms such as ‘east’ and ‘west.’30
Does this mean that, according to Sellars and Brandom, conceptual necessity
claims such as () are mere “inference tickets” in a “game” of giving and asking for
reasons, rather than sentences with truth-conditional content? I think it would be
foolish to make that claim.31 Instances of () can clearly be embedded in “It is true
that . . .” and other truth-functional contexts, and they express something that can
be doubted, believed, known, etc. So, it would be very strange to say that they do
not have truth-conditional content. Indeed, something akin to the standard possible-
worlds semantics for alethic modality provides a pretty good articulation of these truth
conditions.
Assuming this is (roughly) right, the crucial metasemantic question is what it is in
virtue of which this truth-conditional specification of the content of some instance of
() is correct (insofar as it is)? As we have already seen, the representationalist answer
is that it is correct in virtue of correctly identifying how that instance of () represents
reality as being. By contrast, an inferentialist following Sellars and Brandom would
say these truth conditions are correct in virtue of correctly modeling the position in
space of implications taken up by someone who asserts that instance of () in normal
discursive practice. One who accepts this sentence is committed to y‘s being to the
west of x, if she is committed to x’s being to the east of y. In most contexts, were it to
come up, in order to entitle oneself to an instance of (), one might need only to affirm
one’s competence with the words ‘east’ and ‘west.’32

30 Compare Thomasson (, ) who argues that this is the appropriate account of metaphysical
necessity claims. Here my suggestion on behalf of Sellars and Brandom is the weaker idea that at least one
species of necessity claim plays this inference-rule endorsing role. If Thomasson is right, however, that’s a
congenial conclusion.
31 Kant and Frege sometimes seem to be making something like this claim, e.g. “. . . the modality of
judgments is a very special function thereof, which has the distinguishing feature that it does not contribute
to the content of the judgment” (Kant, , A/B), and “By saying that a proposition is necessary I
give a hint about the grounds for my judgment. But, since this does not affect the conceptual content of the
judgment, the form of the apodictic judgment has no significance for us” (Frege, , p. ). However it is
probably anachronistic to view them as speaking of the compositional semantic content of modal sentences
rather than something like the empirical or descriptive content of the thoughts expressed by canonical uses
of these sentences.
32 What about conceptual necessity claims whose prejacents are not explicitly conditional in form? As
with all modal claims, we’d still articulate their truth conditions relative to background conditions in the
context of use, and these truth conditions could be viewed as depicting the idea that one who takes those
background conditions to hold is committed to the truth of the explicit prejacent.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

metanormative theory and the meaning of deontic modals 

This is how I see an inferentialist metasemantics, when combined with the further
Sellarsian/Brandomian idea that some terms are metaconceptual devices, as provid-
ing resources for interpreting the truth conditions of some sentences along non-
representationalist lines. So, now we should turn back to the main topic of this paper:
‘ought.’
Above I claimed that the standard treatment of this word in compositional seman-
tics is as a necessity modal rather than an ordinary predicate-forming term. That
already suggests interpreting the truth-conditional contribution of ‘ought’ in terms
of its being a metaconceptual device. The main objection to this basic idea, I think,
is that the standard possible-worlds semantics for necessity modals faces difficulty
in making sense of the distinctive prescriptive uses of ‘ought’ common in normative
discourse. That is why I favor the enhancement of the standard account sketched in §.
This continues to treat ‘ought’s truth-conditional contribution in terms of a universal
quantification over possibilia, but it refines the kinds of prejacent content ‘ought’ is
thought to embed, and it makes correlative enhancements to the conception of the
possibilia at which these prejacents are to be evaluated.
I think this can now be leveraged into a very plausible inferentialist treatment of
‘ought’ as a metaconceptual device, which then opens up space for distinguishing all
ought-claims from descriptive claims. The core idea is to interpret ‘ought’ as getting its
content from its role in embracing commitment to inferential connections between
more basic items, rather than from its representational purport. This is analogous to
the metasemantic story about conceptual necessities provided by Sellars and Brandom.
Then we deploy the enhanced modal rule from before to generate specific explanations
of the types of more basic claims whose inferential connections are embraced via an
ought-claim.
For instance,
() We ought to do more to relieve great suffering.
would be said to have truth conditions something like the following:
[[()]]fg = T iff [[Let’s do more to relieve to relieve great suffering!]] = L in all
<w, n> ∈ Pfg
where this is conceived as a modal operation on a prescription, evaluating its legiti-
macy across a contextually determined set Pfg of world-norm pairs.
However, inferentialists won’t gloss this as an articulation of how () represents
modal reality as being (e.g. as containing various possible worlds and possible norms,
according to which the relevant prescription is legitimate). Rather they could inter-
pret the truth-conditional specification of the content of () as a depiction of a
point in a network of inferentially specifiable commitments one takes on when one
endorses (). Downstream, one is committed to prescriptions to act (in particular
contexts) in ways that comply with the prejacent prescription; and upstream one might
defend () by appealing to more general normative claims that support the norms

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 matthew chrisman

relative to which the prescriptional content is legitimate at the contextually relevant

worlds.
Although inspired by Brandom, it is worth noting that my proposal here is con-
siderably more conservative than his suggestion (in Brandom, , chapter ) that
all normative and evaluative terms should be understood as metaconceptual devices
governing specifically practical inferences. He considers such sentences as:
() You’re obliged, if you’re a bank employee, to wear a necktie to work.
() It is wrong, if it’s done to no purpose, to harm someone.
and claims a sentence like these “. . . makes explicit the endorsement . . . of material
proprieties of practical reasoning” (, p. ). His idea is that, just as conceptual
necessity claims such as () could be seen as devices for embracing inferential
connections between the applications of more basic concepts such as “east” and “west,”
sentences such as () and () should be seen as devices for embracing inferential
connections between more basic concepts such as “bank employee” and “harm to no
purpose” and practical thoughts about what to do, e.g. in the inferences:
I am a bank employee → so I shall wear a necktie to work
Repeating the gossip would harm someone to no purpose → I shall not repeat the
gossip
where these are conceived as “language-exit” transitions because the termini are the
expressions of intentions that—when the time comes, assuming things go normally—
will result in the agent’s doing something other than making further linguistic moves.
Leaving aside what to say about other normative terms, this surely cannot be the
right metasemantic story about why ‘ought’ makes the semantic contribution that it
does. For, as we saw above, not all ought-sentences are about what someone is to do.
For instance,
() Larry, who has had the most unfortunate time of things recently, ought to win
the lottery.
does not seem to prescribe an action to Larry (or anyone else). So, although such sen-
tences may be “normative” or “evaluative” in some broader sense, they are not plausibly
construed as endorsing patterns of narrowly practical inference, i.e. inferences that end
with a language-exit expressible with an “I shall . . . ”.
Nevertheless, I think we can and should make sense of these sentences as non-
descriptive. If we use my enhanced modal rule for ‘ought’ and think of it as taking
propositional rather than prescriptional prejacents for these sentences, then (simpli-
fying slightly) we will get truth conditions for these sentences that say that they are
true just in case some proposition is true at all of some contextually determined set of
possible worlds. When it next comes to a metasemantic interpretation of this result,
we could view the relevant specifications of truth conditions as an initial spelling out

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

metanormative theory and the meaning of deontic modals 

of a position in a space of implications, which determines an inference rule; and these

sentences could then be seen as devices for embracing (acknowledging, endorsing)
this rule. For example, () might be thought of as a device for embracing (and making
explicit one’s commitment to) the defeasible semantic validity of inferences of the
form:
Larry has had the most unfortunate time of things recently → Larry’s winning the
lottery would be most preferable (compared to the salient alternatives).
where the conclusion of this inference isn’t itself a language-exit but a commitment to
a certain constraint on one’s preferences.
That’s admittedly a very rough idea. The purpose of mentioning it here is just
to indicate that full consideration of the types of sentences in which ‘ought’ figures
suggests that it cannot be solely a device for endorsing inferential connections between
propositional commitments and narrowly practical commitments in the way that
Brandom suggests. That does not mean, however, that he’s wrong about its being
a metaconceptual device for endorsing inferential connections more generally, on
the model of conceptual necessity. This is precisely how I think an inferentialist
should interpret the enhanced modal rule for ‘ought’ spelled out above, which always
involves determining the truth conditions of an ought-sentence in terms of a universal
quantification over possibilia. This allows for an evaluation of the legitimacy of an
embedded prescription across a range of norms as a special case but also the truth of
an embedded proposition across a range of worlds as another special case.
Recall the apparent tension between the ideas that (i) ‘ought’s do not describe
what is actually the case but prescribe possible action, thought, and feeling, and
(ii) all declarative sentences deserve the same general semantic treatment, e.g. in
terms of compositionally specified truth conditions. In §, I argued that there’s a
representationalist resolution of this tension, which treats all declarative sentences
as having the truth conditions that they have in virtue of their representing a way
reality could be, but which also allows that ought-sentences don’t describe what is
actually the case because, as their truth conditions suggest, they represent how things
stand with respect to possible worlds and norms. In this section, I have developed
an alternative resolution of this tension. The inferentialist metasemantics says that all
declarative sentences have the truth conditions that they have in virtue of the position
in a space of implications that accepting the sentence commits one to, but it also
allows us to draw a distinction between sentences playing a first-order descriptive
role and those playing a higher-order metaconceptual role. It’s not obvious where to
draw the line between these, but empirical claims we think of as tracking features of
our perceptual environment are a pretty good candidate for the former, and claims
containing logical terms we think of as providing the conceptual scaffolding within
which complex logical interrelations can be specified are a pretty good candidate for
the latter. The truth-conditionalist semantic rule for ‘ought’ encourages us to think of
it along the lines of a necessity modal, and I have suggested that this means we should

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 matthew chrisman

think of it as a metaconceptual device rather than a term for representing pieces of

reality.
This carries an ontological benefit over the representationalist way of resolving the
tension between (i) and (ii): we don’t have to commit ourselves to the reality of possible
worlds and possible norms in order to recognize some ought-claims as true. But, for
this, it incurs the cost of explaining in more detail why language should have terms that
function as metaconceptual devices and how to tell the difference between those terms
and representational terms. Although I am not going to pay it here, it’s a debt I view as
worth paying. (However, I also admit to difficulty in knowing how to think about the
exchange rate between such high-level theoretical costs and benefits.) In any case, my
main claim is that the deciding factor between representationalism and inferentialism
should not be which of ideas (i) and (ii) we are more impressed by but rather what we
think of the relative merits of how these two views respect these ideas.

 Comparison to Expressivism
I want to conclude by discussing a more familiar constellation of ideas that might also
be thought to provide an alternative to representationalist treatments of normative
sentences. This is born out of the noncognitivist thesis that normative claims express
conative attitudes rather representations of reality, the sorts of attitudes that are
supposed to class with desires rather than beliefs in their motivational potentials (or
“directions of fit”) in a Humean psychology of motivation. This noncognitivist thesis
has received considerable attention and refinement over the past half century, resulting
in a number of different positions now defended under the banner of “expressivism.”
Moreover, expressivism has been applied in several domains besides normative dis-
course (modal, epistemic, aesthetic, alethic, etc.) with importantly differing theoretical
constraints. So, I find it far from clear what the core of expressivism is and how it
relates to the inferentialist ideas explored above. As we already saw, some think of
inferentialism as a kind of “global expressivism”—but that’s not the view one finds
in Sellars and Brandom. Moreover, many philosophers conceive of expressivism as
premised on rejecting truth-conditionalist semantics in favor of a “psychologistic”
alternative,33 at least in its application to normative discourse—but that’s not how
I think we should think of inferentialism.
I’m not going to be able to discuss all possible (or even many) versions of expres-
sivism here.34 But I think it will prove useful to discuss one expressivist line of thought
about ‘ought’ that I think comes closest to the inferentialist line of thought developed
above. This will put me in a position to explain how some ways of developing the view
might result in an alternative to inferentialism that has various problems while others

33 This is, for instance, the way expressivism is conceived by critics such as Rosen (, pp. –),
Wedgwood (, p. ), Kalderon (, pp. –), and Schroeder (, p. ).
34 See Chrisman () for my take on some historical and contemporary versions of expressivism.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

metanormative theory and the meaning of deontic modals 

might result in friendly amendments to inferentialism. Hopefully, this will also clarify
the inferentialist view outlined above.
As long as we continue to observe the distinction between semantics and metase-
mantics, I think it should be clear that expressivism needn’t be understood as a view
about compositional semantics.35 We can find a version of expressivism as part of a
larger metasemantic view that has, until now in this paper, remained conspicuously off
stage. This is the view that the truth conditions assigned by compositional semantics
are articulations of what one who asserts one of these sentences ought to think. For
instance, in the most basic case, it’s plausible to say that one who asserts
() Grass is green.
ought to think (judge, affirm) that grass is green. The idea is that someone who asserts
this sentence but does not think this thought has violated the semantic rule associated
with () in virtue of its core communicative role. So, rather than starting with the
view that truth conditions articulate the way reality must be in order for a sentence
to be true, an “ideationalist” metasemantics would encourage us to start with the
view that semantic assignments to sentences tell us what idea one ought to have when
one uses the sentence canonically, in order to conform with the core communicative
norms associated with the sentence in the language.36 This is one plausible way to
flesh out the vague and popular suggestion that different sentences are canonical and
conventional means for expressing different mental states.37 Then, on this ideationalist
view, the truth-conditionalist articulation of the subsentential elements of language
could be viewed as an explanation of how an indefinite number of such expressive
commitments are generated by a finite number of terms and logical forms.
Although ideationalism starts in a different place from representationalism, it’s
wholly compatible with the representationalist view that sentences such as () rep-
resent reality. For an ideationalist can say that the truth conditions of () specify
the way this sentence represents reality as being by virtue of articulating the content
of a particular kind of thought, a representational thought about the way reality is,

35 Compare Suikkanen (), Chrisman (b), Yalcin (), Silk (), Charlow (a, p. ),
and Ridge (, chapter ).
36 Precursors and versions of this idea can be found in Grice (), Blackburn (), Davis ().
Although defending a somewhat similar idea, I think Richard () would say that it’s wrong to think of
these as truth conditions but that compositional semantic values can be interpreted as modeling a space of
commitments to various thoughts.
37 See Schroeder (, chapter ) and Ridge (, §.) for further discussion of the expression
relation appropriate for developing a form of metanormative expressivism. Both end up developing forms
of expressivism with an account of expression very similar to the account I am using here. The key
commonality is in construing the relation as a hypothetical linguistic norm, something like: if one uses
the sentence to make an assertion, one ought to think the relevant thought in order to conform with the
core communicative rules of the language. In Bar-On and Chrisman (), we argue that distinguishing
the way in which users of a sentence express mental states from the way in which sentence-types express
their semantic contents provides for a much simpler and more plausible explanation of the apparently
distinctive connection between moral claims and motivational attitudes than is on offer in standard forms
of metaethical expressivism.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 matthew chrisman

which one who asserts the sentence ought to have in order to conform with the core
communicative rules of the language.
Then, a “metasemantic” form of expressivism begins to come into view when
we observe that, although the ideationalist can agree with the representationalist
about sentences like (), she needn’t agree that all declarative sentences represent
reality. For she can argue that some sentences are conventional means for expressing
nonrepresentational thoughts, the content of which is specified by the truth conditions
the correct semantic theory assigns to the sentence.38 For example, an expressivist
might argue that one who uses a sentence such as:
() Stealing is wrong.
to make an assertion ought to think a particular thought, i.e. that stealing is wrong.
Then, however, she will argue that the functional role of this thought in the psychology
of motivation is directive of action rather than representative of reality.
This move obviously requires giving up on the idea that truth is something like
correct representation, which is why most contemporary expressivists embrace some
sort of minimalism about truth (though I think some kind of pluralism about truth
would also serve this purpose).
But that raises an important question: how could the expression of a directive
thought be something that one could doubt or know to be true, even in a defla-
tionary sense? After all, paradigmatic directive mental states are desires, preferences,
and plans, and these are not the proper objects of doubt and knowledge. In the
present context, I think expressivists might insist that the characterization of directive
thoughts as “desire-like” rather than “belief-like” is simply misleading and distracting.
The important claim is that some thoughts are thoughts about reality, they represent
reality as being a particular way; other thoughts are about what to do, they direct us to
make reality become a particular way. The standard Humean view of the psychology
motivation is that action always results from the cooperation of these two kinds of
thoughts. We should allow expressivists to call both “beliefs” in many core cases; the
important suggestion is that one of them has a directive rather than descriptive func-
tional role in the psychology of human motivation. It is because of the way in which
this role stands in contrast with the descriptive functional role of other thoughts that
this kind of expressivism promises to carry whatever attractions were in earlier views

38 For similar ideas, compare Blackburn (, pp. –), Silk (, p. ), Ridge (, chapter ), Perez
Carballo (). Charlow (forthcoming) develops a metasemantic form of expressivism in the context of a
dynamic “test” semantics for deontic and epistemic modals. This is importantly different in its rejection of
the semantic explanations of truth-conditional semantics, but it still shares in conceiving of expressivism
as a metasemantic thesis. Yalcin (, p. ) characterizes his expressivism as a pragmatic rather than a
semantic thesis. However, if that interpretation is going to carry the ontological benefits expressivism is
supposed to carry, then I think the pragmatics needs to be part of a broader metasemantic theory—perhaps
one that explains why sentences have the truth conditions they do in terms of what conversational moves
they are canonical means for making. I believe this is consonant with the characterization of metasemantics
in Yalcin ().

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

metanormative theory and the meaning of deontic modals 

in the expressivist tradition. Moreover, if one can make sense of evaluating directive
thoughts as being more and less reasonable, justified, or warranted, a proponent of this
view would be on the way to making sense of the way that normative beliefs seem to
be the objects of epistemic evaluations.39
So, at least one way to conceive of metanormative expressivism is as the combination
of the Humean psychology of motivation with an ideationalist metasemantic interpre-
tation of the truth conditions of normative sentences. More specifically, we will need to
think of the truth conditions of normative sentences such as () as the articulation of
directive rather than descriptive thoughts. We can give them truth conditions such as
[[()]] = T iff stealing is wrong
and think of this as an articulation of what one who asserts this sentence ought
to think, in order to conform to the core communicative rules associated with this
sentence. However, in their philosophy of mind, expressivists will argue that this
thought plays a directive rather than a descriptive role in the psychology of moti-
vation. By understanding expressivism as part of an overall metasemantic package,
rather than a controversial thesis in compositional semantics, we could then grant
expressivists access to all of the successes of standard truth-conditional approaches to
compositional semantics, while leaving room in the philosophy of mind for them to
make the Humean argument that normative thoughts are nondescriptive because they
are directive.
There is a lot more we might want to know about metasemantic forms of expres-
sivism, but the crucial question for me at this stage is whether this generates a
competitor to the inferentialist metasemantic view about ‘ought’ outlined above or a
possible partner for it in a coalition against representationalism about ‘ought.’ I think
that’s going to depend on what expressivists say about several further questions.
First Question: What sort of thoughts do metasemantic expressivists say are expressed
by logically complex sentences? Those versed in debate about the Frege-Geach prob-
lem might grant that cleaving to a truth-conditionalist semantics and now attempting
to develop expressivism in one’s metasemantics provides for a better answer to this
question than old-school versions of expressivism did. For the expressivist can now
begin by saying that a sentence such as
() Either stealing is common or stealing is wrong.
has the truth conditions of a disjunction, i.e. a logical form ensuring that it is true
if either of its disjuncts is true. Still, if our metasemantic expressivist says that it

39 Though we will have to reject Cuneo’s suggestion that it is platitudinous that “. . . propositional
attitudes display one or another epistemic merit (or positive epistemic status) such as being a case of
knowledge, being warranted, being an instance of understanding, insight or wisdom and the like, only
insofar as they are representative in some respect.” (, p. ) See Chrisman (, , c) for
discussion of the prospects of nonrepresentational attitudes’ achieving positive epistemic status.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 matthew chrisman

has these truth conditions in virtue of expressing the disjunctive thought that either
stealing is common or stealing is wrong, we can ask: what kind of thought is that? In the
simple Humean psychology of motivation we were working with above, there are only
descriptive and directive thoughts. And while we may want to say that this thought has
descriptive and directive elements, the expressivist metasemantics seems to require
us to be able to say what kind of mental state the whole thought represents. After
all, as I glossed it above, the expressivist’s way of securing the claim that normative
sentences don’t describe reality is by claiming that these sentences express directive
thought rather than descriptive thought.
In response, expressivists might try to argue that logically complex sentences such
as () are directive, but in an attenuated sense. The idea would be to treat them as
the expression of conditional dispositions to infer, where inferring is conceived as a
mental action. So, in this case, we might think of () as expressing a thought about
what to do: if I come to reject one of the disjuncts I shall embrace the other.40
As someone sympathetic with inferentialism, naturally I think there’s something
importantly right about understanding the meaning of logically complex sentences in
terms of their inferential interrelations with other sentences. However, I doubt using
this attenuated sense of “directive” is going to let us draw the line between represen-
tational and nonrepresentational thoughts anywhere near where we want to draw it
for the purposes of metanormative theory. For notice that even a paradigmatically
descriptive sentence such as () might be thought to direct inference (e.g. if you accept
this, do not infer any conclusion that entails that grass is not green!). Moreover, it
would be highly odd to place
() Grass is green or roses are red.
() Grass is green and roses are red.
on the directive side of the ledger, given that we place each of the atomic parts on the
descriptive side of the ledger.41
Because of this, the answer I think expressivists should give to the first question
is that the Humean psychology of motivation is too stark; there are more than just
descriptive and directive thoughts, and the thoughts expressed by logically complex
sentences are neither purely descriptive nor purely directive. But, if expressivists gives
that answer, then they owe us a new account of what distinguishes the class of sentences
that have their truth conditions in virtue of expressing a representational thought and

40 This is one way to make sense of Blackburn’s (, pp. –; ) suggestion that logically complex
thoughts “tie one to a tree.”
41 I think a similar problem arises for Ridge’s (, , ) suggestion that logically complex
sentences express hybrid states. The idea is that sentences such as () express beliefs with a descriptive
content that is logically isomorphic to the normative sentence but they also express interlocking directive
attitude (e.g. the endorsement of an ideal adviser or a normative perspective). The problem I see for this idea
is that it ends up committing one to treating all logically complex sentences as expressive of a (partially)
desire-like state of mind, or else it involves an ad hoc treatment of the role of logical particles in various
logically complex sentences.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

metanormative theory and the meaning of deontic modals 

all of the other sentences that supposedly express nonrepresentational thoughts. If

being directive isn’t necessary for a thought to be nonrepresentational, what is?
I don’t think it’s obvious how expressivists should answer this question, but one way
to look for an answer is to think about the downstream inferential consequences of
the thoughts in question. For example, expressivists might insist that the thought that
grass is green directly implies something about how reality is (i.e. without the aid of
substantive auxiliary premises) but the thought that stealing is wrong does not directly
imply something about how reality is. If that’s right, there might be diverse reasons
why some thoughts do not directly imply something about how reality is, but it would
be the absence of this downstream inferential connection that marked out the class
of nonrepresentational thoughts. Obviously, if expressivists go in this direction, their
position begins to look like a coalition partner for rather than a competitor with the
inferentialist view sketched above.

Second Question: What do metasemantic expressivists say about ‘ought’? By asking

this question, I don’t mean to reopen the issue of whether a modal analysis of the truth
conditions of ought-sentences is right (say, in comparison with attempts to reductively
analyze ‘ought’ in terms of values or reasons). Indeed, since the sort of expressivists I
have in mind are now on board with truth-conditional semantics, I shall assume they
accept something like the enhanced modal rule outlined above. But what do they say
it is in virtue of which the truth conditions predicted by this rule are (approximately)
correct?
The initial response is clear: it’s in virtue of the fact that ought-sentences are
conventional and canonical means for expressing ought-thoughts. But what are ought-
thoughts, descriptive or directive or do they fall in some yet-to-be-specified third
category? Again, if we work with the stark Humean division between descriptive
and directive thoughts, it may seem as if expressivists are going to say that ought-
thoughts are uniformly directive. However, when we consider the data, that doesn’t
seem plausible. At least it seems pretty implausible to say that all ought-sentences have
the truth conditions that they do in virtue of expressing a thought whose primary
functional role is to provide the goal-setting rather than means-providing element in
a Humean explanation of action. Consider:
() I ought to have begun my higher education at a different college.
() NATO ought not to expand any further.
I won’t ever be in a position to begin my higher education again; I won’t ever be in
the position of NATO; so why think that the thoughts I would express by asserting
these sentences have the motivational efficacy characteristic of directional thoughts
in a Humean explanation of action? To be sure, an expressivist might insist that these
sentences express something like plans or preferences conditioned on being in exactly
the situation of my former self or of NATO. But that looks like a patch: something
expressivists have to say rather than an intuitive gloss of the core communicative

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 matthew chrisman

function of sentences. So, this is a cost for the overall plausibility of the theory.
Moreover, as we have already observed, there are non-agentive but at least weakly
normative ought-claims, such as the teleological and evaluative ought-sentences ()–
() from above. Whatever connection there is between having the thought expressed
by these sentences and being motivated to perform some specific act is going to
be rather indirect and highly defeasible. We are right, I think, to question whether
someone understands what she is saying if she says something of the form “All things
considered, I ought to φ,” and yet has no motivation to φ. However, it is much
less plausible to think that competence with ‘ought’ more generally—even in many
paradigmatic normative uses—requires any particular motivational propensities.
This is why I think expressivists should instead answer the question by arguing that
many ought-thoughts fit into some third category of thoughts, neither purely descrip-
tive nor purely directive.42 Again, however, we should ask: what are they like and
what about them makes them nonrepresentational? And again I think the inferentialist
story sketched above might provide some helpful resources for expressivists. That is
to say that expressivists could treat some thoughts (paradigmatically ones expressed
with necessity modals) as not themselves representational but rather ways of encoding
metaconceptual manipulations on more basic pieces of semantic content, resulting in
thoughts explicating the structure of a space of inferential relations. Maybe there are
other way expressivists could answer this question,43 but obviously if they answer it
this way, the view is again a plausible partner to rather than a competitor with the
inferentialist view sketched above.
Third Question: If semantic content of sentences is to be explained in terms of the
content of the mental states (or “thoughts”) they express, as expressivists maintain,
then what explains the content of these mental states? The ideationalist conception of
truth conditions sketched above takes for granted mental content in order to explain
semantic content, which means that anyone endorsing it owes a further account of
mental content.

42 Although Blackburn (, ) is deeply influenced by Hume, he finds in Hume a much richer
diversity of types of mental states besides the descriptive and the directive, including all sorts of different
projectable stances and mental dispositions.
43 Another way, which I do not quite know how to categorize, is to follow Yalcin’s treatment of claims of
probability and epistemic modality (see especially Yalcin, , ). He replaces the traditional Humean
distinction between descriptive and directive thoughts with a distinction between mental states representing
the world as being one way or another (modelable as conditions on worlds) and second-order properties of
one’s first-order representational states (not modelable conditions on worlds but rather as a function from
sets of worlds). In a similar vein but within a dynamic framework, Charlow (forthcoming) distinguishes
between information-carrying thoughts and thoughts that are instructions to “test” one’s overall belief and
preference state for certain structural features. The more general idea behind Yalcin’s and Charlow’s versions
of expressivism is that there may be declarative sentences whose function is not to express representational
beliefs, nor to express desire-like states, but rather to coordinate on a higher-order property of one’s global
mental state, e.g. credence distributions and preference orderings. Depending on how these are related
to inferential commitments, this too might constitute a useful ally with rather than competitor with the
inferentialist view sketched above.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

metanormative theory and the meaning of deontic modals 

Early expressivists were, I think, tempted to say that the mental states expressed by
normative sentences are not thoughts with normative content; rather, all content is
representational but we can have nonrepresentational attitudes towards such content
(e.g. we can desire that stealing be less common). Then normative sentences were
treated as conventional and canonical means for expressing such attitudes. If we pro-
ceed this way, expressivists could insist on a purely representationalist conception of
mental content, but argue that it’s not the content of but the type of attitude that makes
a mental state nonrepresentational. However, I think it is wildly implausible to say that
there are no thoughts with normative content (what then is one doubting, assuming,
hoping, etc., when one doubts, assumes, hopes, etc. that stealing is wrong?).44
Another way expressivists could go at this point is to apply something akin to
the sort of inferentialist account of content I sketched above to the case of mental
content.45 If this is right, the ideationalist conception of truth conditions might be
correct as far as it goes, but the inferentialist view would be providing a deeper
explanation of what content is in virtue of which sentences have the meaning that
they do.

 Conclusion
As I see things, the key issue between representationalists, inferentialists, and ideation-
alists is one about order of theoretical priority. Everyone should agree that some
sentences describe things in our environment, that most sentences have inferential
connections to other sentences, and that all sentences are vehicles by which we express
our minds. The interesting question is which of these is most theoretically basic for
understanding that in virtue of which sentences have the contents that they have.
Here I have argued that each of these views can accept the predictions that truth-
conditionalist approaches to compositional semantics give for ‘ought’; in effect that
it is a modal operator that quantifies over possibilia at which it evaluates embedded
contents. Given this, I think we can take representation as the master concept and
argue that these truth conditions reveal that ought-claims don’t describe how things
actually are but rather how things are with respect to possible worlds and possible
norms. But we might also take inference as the master concept and argue that these
truth conditions reveal that ought-claims are not ways of encoding a representation
of modal space but rather ways of explicating inferential relations standing between
other things we can say. It’s tempting to see the resulting inferentialist view in the
lineage of expressivism. And I’ve suggested that there may be sophisticated forms

44 Because of this, I think appealing to off-the-shelf Fodorian (, ) or Millikanian () or
Dretskean (, ) accounts of mental content is not going to work for expressivists, since these all
proceed in terms of representation relations between concepts and reality, whereas expressivists need
something different for normative concepts. Moreover, I suspect going in this direction opens expressivists
up to Dorr’s () wishful-thinking challenge.
45 See Blackburn () for an expressivist who ends up moving in the direction of inferentialism.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 matthew chrisman

of expressivism that are naturally allied with inferentialism about ‘ought.’ However,
it’s important to the way I am thinking of inferentialism that inferential connections
rather than conventional expressive potentials are most fundamental. This is the idea
that I want to recommend as an interesting metasemantic account of that in virtue of
which ought-sentences have the truth-conditional content that they do, which I see as
providing new foundations for a form of antidescriptivism that has often been ignored
in metanormative theory.46

Acknowledgments
For helpful feedback on earlier versions of this material, I am grateful to audiences at the
University of Edinburgh, University of North Carolina at Chapel Hill, University of Szczecin,
Robert Beddor, Nate Charlow, Graham Hubbs, David Plunkett, Michael Ridge, Alex Silk, and
two anonymous referees.

References
Bar-On, Dorit and Matthew Chrisman () Ethical Neo-Expressivism. In Oxford Studies in
Metaethics. Volume , edited by Russ Shafer-Landau. pp. –. Oxford; New York, NY:
Oxford University Press.
Björnsson, Gunnar, and Stephen Finlay () Metaethical Contextualism Defended. Ethics. .
. pp. –.
Blackburn, Simon () Spreading the Word: Groundings in the Philosophy of Language. New
York, NY: Oxford University Press.
Blackburn, Simon () Essays in Quasi-Realism. New York, NY: Oxford University Press.
Blackburn, Simon () Ruling Passions: A Theory of Practical Reasoning. New York, NY:
Oxford University Press.

46 I say “foundations” because I have not tried to motivate this idea for all claims one might want to
classify as nondescriptive—not even for all normative claims, let alone evaluative claims. I see three general
ways the foundations I have laid here with respect to ‘ought’ could be developed to address other normative
and evaluative terms. First, one might pursue a priority view, according to which ‘ought’ is the fundamental
normative/evaluative term, perhaps because when one is deciding what to think, feel, or do, all other
considerations should be thought to bear on a question most perspicuously posed with an ‘ought.’ Compare
Broome (, pp. – and chapter ). Along these lines, one might then try to analyze other terms,
insofar as they have normative uses, in terms of more fundamental normative ‘ought’s which one treats
as metaconceptual devices along the lines I have sketched above. Secondly, one might pursue a no-priority
view, according to which other terms with normative or evaluative uses deserve a different treatment, which
is not somehow derivative of the treatment of ‘ought’ presented here, but which also does not provide a
reduction going the other way for ‘ought.’ Along these lines, maybe the sort of treatment of ‘ought’ as a
metaconceptual device presented here works for what is clearly one of our core normative terms; and we
simply have to dig deeper to untangle the bilateral relationship between ought-claims and other normative
and evaluative claims. Finally, one might want to maintain that other terms with normative import (e.g.
‘right’ and ‘good’) are themselves intensional operators. Along these lines maybe they are metaconceptual
devices for endorsing different sorts of inferential commitments.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

metanormative theory and the meaning of deontic modals 

Blackburn, Simon () The Semantics of Non-Factualism. In The Blackwell Guide to the
Philosophy of Language, edited by Michael Devitt and Richard Hanley. pp. –. Malden,
MA, Oxford: Blackwell.
Blackburn, Simon () Pragmatism: All or Some? In Expressivism, Pragmatism, and Repre-
sentationalism, edited by Huw Price. Cambridge: Cambridge University Press.
Block, Ned () Functional Role and Truth Conditions. In Proceedings of the Aristoteian
Society. Supplementary Volume LXI. pp. –.
Brandom, Robert () Making It Explicit: Reasoning, Representing, and Discursive Commit-
ment. Cambridge, MA: Harvard University Press.
Brandom, Robert () Articulating Reasons: An Introduction to Inferentialism. Cambridge,
MA: Harvard University Press.
Brandom, Robert () Reason in Philosophy: Animating Ideas. Cambridge, MA: Harvard
University Press.
Brandom, Robert () Perspectives on Pragmatism. Cambridge, MA: Harvard University Press.
Brandom, Robert () Global Anti-Representationalism? In Expressivism, Pragmatism and
Representationalism, edited by Huw Price. pp. –. Cambridge: Cambridge University
Press.
Bronfman, Aaron, and Janice Dowell () Contextualism About Deontic Conditionals. In
Deontic Modality, edited by Nate Charlow and Matthew Chrisman. Oxford: Oxford Univer-
sity Press.
Broome, John () Rationality Through Reasoning. London: Wiley-Blackwell.
Cariani, Fabrizio () ‘Ought’ and Resolution Semantics. Noûs. . . pp. –.
Carr, Jennifer () Deontic Modals Without Decision Theory. Proceedings of Sinn Und
Bedeutung : Chemla, E., Homer, V., and Winterstein, G. (eds.) École Normal Supérieure,
Paris, – September. pp. –.
Carr, Jennifer () The If P, Ought P Problem. Pacific Philosophical Quarterly. . . pp. –.
Castañeda, Hector-Neri () Thinking and Doing: The Philosophical Foundations of Institu-
tions. Berlin: Springer.
Charlow, Nate () What We Know and What to Do. Synthese. . . pp. –.
Charlow, Nate (a) The Problem with the Frege-Geach Problem. Philosophical Studies. .
pp. –.
Charlow, Nate (b) Logic and Semantics for Imperatives. Journal of Philosophical Logic.
. . pp. –.
Charlow, Nate (forthcoming) Prospects for an Expressivist Theory of Meaning. Philosophers’
Imprint.
Chrisman, Matthew () Expressivism, Truth, and (Self-)Knowledge. Philosopher’s Imprint.
. . pp. –.
Chrisman, Matthew () Constructivism, Expressivism and Ethical Knowledge. International
Journal of Philosophical Studies. . pp. –.
Chrisman, Matthew () Ethical Expressivism. In The Continuum Companion to Ethics, edited
by Christian Miller. Chapter . London: Continuum.
Chrisman, Matthew (a) ‘Ought’ and Control. Australasian Journal of Philosophy. . .
pp. –.
Chrisman, Matthew (b) On the Meaning of ‘Ought’. In Oxford Studies in Metaethics. Volume
, edited by Russ Shafer-Landau. pp. –. New York, NY: Oxford University Press.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 matthew chrisman

Chrisman, Matthew (c) Epistemic Expressivism. Philosophy Compass. . . pp. –.

Chrisman, Matthew () The Meaning of ‘Ought’: Beyond Descriptivism and Expressivism in
Metaethics. New York, NY: Oxford University Press.
Chrisman, Matthew and Graham Hubbs. unpublished. The Pratitional Approach to Imperatives.
Cuneo, Terence () The Normative Web. New York, NY: Oxford University Press.
Davis, Wayne () Meaning, Expression, and Thought. Cambridge: Cambridge University
Press.
Dorr, Cian () Non-Cognitivism and Wishful Thinking. Noûs. . pp. –.
Dowell, Janice () Flexible Contextualism About Deontic Modals: A Puzzle About
Information-Sensitivity. Inquiry. . –. pp. –.
Dreier, James () Expressivist Embeddings and Minimalist Truth. Philosophical Studies. . .
pp. –.
Dretske, Fred () Knowledge and the Flow of Information. Bradford, MA: MIT Press.
Dretske, Fred () Explaining Behavior: Reasons in a World of Causes. Bradford, MA: MIT
Press.
Finlay, Stephen () Oughts and Ends. Philosophical Studies. . . pp. –.
Finlay, Stephen and Justin Snedegar () One Ought Too Many. Philosophy and Phenomeno-
logical Research. . .
Fintel, Kai von () The Best We Can (Expect) to Get? Challenges to the Classic Semantics
for Deontic Modals. Prsented at the Central APA,  February.
Fintel, Kai von and Anthony S Gillies () Might Made Right. In Epistemic Modality, edited
by Andy Egan and Brian Weatherson. pp. –. Oxford: Oxford University Press.
Fintel, Kai von and Sabine Iatridou () How to Say Ought in Foreign: The Composition of
Weak Necessity Modals. In Time and Modality, edited by Jacqueline Gueron and Jacqueline
Lecarme. pp. –. Berlin: Springer.
Fodor, Jerry () Psychosemantics: The Problem of Meaning in the Philosophy of Mind. Cam-
bridge, MA: MIT Press.
Fodor, Jerry () A Theory of Content. Bradford, MA: MIT Press.
Frege, Gottlob () Begriffsschrift, Eine Der Arithmetischen Nachgebildete Formelsprache Des
Reinen Denkens. Halle a/S.: L. Nebert.
Geach, Peter () Whatever Happened to Deontic Logic? Philosophia. . –. pp. –.
Gibbard, Allan () Wise Choices, Apt Feelings: A Theory of Normative Judgment. Cambridge,
MA: Harvard University Press.
Gibbard, Allan () Thinking How to Live. Cambridge, MA: Harvard University Press.
Grice, H. P. () Studies in the Way of Words. Cambridge, MA: Harvard University Press.
Hale, Bob () Can There Be a Logic of Attitudes. In Reality, Representation, and Projection,
edited by Haldane John and Crispin Wright. Oxford: Oxford University Press.
Hanks, Peter () Propositional Content. Oxford: Oxford University Press.
Hauser, Roland () Surface Compositionality and the Semantics of Mood. In Speech Act
Theory and Pragmatics, edited by Searle, J., Kiefer, F. and Bierwisch, M. pp. –. Dordrecht:
Reidel.
Heim, Irene () The Semantics of Definite and Indefinite Noun Phrases. Ph.D. thesis.
University of Massachusetts, Amherst.
Kalderon, Mark () Moral Fictionalism. Oxford: Oxford University Press.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

metanormative theory and the meaning of deontic modals 

Kamp, Hans () A Theory of Truth and Semantic Representation. In Formal Methods in the
Study of Language, Part I, edited by Jeroen Groenendijk, Theo Janssen, and Martin Stokhof.
pp. –. Amsterdam: Mathematisch Centrum.
Kant, Immanuel () Kritik Der Reinen Vernunft. Riga: Johann Friedrich Hartknoch.
Kolodny, Niko and John MacFarlane () Ifs and Oughts. Journal of Philosophy. .
pp. –.
Kratzer, Angelika () The Notional Category of Modality. In Words, Worlds, and Contexts,
edited by Eikmeyer, H. J. and Reiser, J. Berlin: de Gruyter.
Kratzer, Angelika () Conditionals. In Handbuch Semantik/Handbook of Semantics, edited
by von Stechow Arnim and Dieter Wunderlich. Berlin, New York, NY: de Gruyter.
Macarthur, David, and Huw Price () Pragmatism, Quasi-Realism and the Global Challenge.
In New Pragmatists, edited by Cheryl Misak. . Oxford: Oxford University Press.
MacFarlane, John () Epistemic Modals Are Assessment-Sensitive. In Epistemic Modality,
edited by Andy Egan and Brian Weatherson. pp. –. Oxford: Oxford University Press.
Millikan, Ruth G. () Language, Thought and Other Biological Categories. Bradford, MA: MIT
Press.
Peacocke, Christopher () Thoughts: An Essay on Content. Oxford: Blackwell.
Peacocke, Christopher () Understanding Logical Constants: A Realist’s Account. Proceed-
ings of the British Academy. . pp. –.
Perez Carballo, Alejandro () Hermeneutic Expressivism. In New Essays in Metasemantics,
edited by Burgess, A. and Sherman B. New York, NY: Oxford University Press.
Portner, P. () The Semantics of Imperatives Within a Theory of Clause Types. In Proceedings
of SALT editor: Young, R. B. Northwestern University, – May. : pp. –.
Portner, Paul and Aynat Rubinstein () Extreme and Non-Extreme Deontic Modals. In
Deontic Modality, edited by Nate Charlow and Matthew Chrisman. Oxford: Oxford Univer-
sity Press.
Price, Huw () Expressivism for Two Voices. In Pragmatism, Science and Naturalism, edited
by Knowles, J. and Rydenfelt, H. pp. –. Frankfurt am Main: Peter Lang.
Price, Huw () Expressivism, Pragmatism and Representationalism. Cambridge: Cambridge
University Press.
Richard, Mark () When Truth Gives Out. New York, NY: Oxford University Press.
Ridge, Michael () Ecumenical Expressivism: Finessing Frege. Ethics. . . pp. –.
Ridge, Michael () Anti-Reductionism and Supervenience. Journal of Moral Philosophy. .
. pp. –.
Ridge, Michael () Impassioned Belief. Oxford: Oxford University Press.
Rosen, Gideon () Blackburn’s Essays in Quasi-Realism. Noûs. . . pp. –.
Schroeder, Mark () Being for. New York, NY: Oxford University Press.
Schroeder, Mark () Noncognitivism in Ethics. New York, NY: Routledge.
Schroeder, Mark () Oughts, Agents, and Actions. Philosophical Review. . pp. –.
Segerberg, Krister () Validity and Satisfaction in Imperative Logic. Notre Dame Journal of
Formal Logic. . pp. –.
Sellars, Wilfrid () Inference and Meaning. Mind. . . pp. –.
Sellars, Wilfrid () Empiricism and the Philosophy of Mind. In Minnesota Studies in the
Philosophy of Science. Volume , edited by Feigl, H. and Scriven, M. pp. –. Minneapolis,
MN: University of Minnesota Press.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 matthew chrisman

Sellars, Wilfrid () Counterfactuals, Dispositions, and the Causal Modalities. In Minnesota
Studies in the Philosophy of Science. Volume II, edited by Herbert Feigl, Michael Scriven, and
Grover Maxwell. pp. –. Minneapolis, MN: University of Minnesota Press.
Sellars, Wilfrid () Meaning as Functional Classification. Synthese. . pp. –.
Silk, Alex () Modality, Weights, and Inconsistent Premise Sets. Proceedings of SALT
Chereches, Anya (ed.) University of Chicago, Chicago, IL, – May . pp. –.
Silk, Alex () Truth Conditions and the Meaning of Moral Terms. In Oxford Studies in
Metaethics. Volume , edited by Russ Shafer-Landau. pp. –. New York, NY: Oxford
University Press.
Silk, Alex () Evidence Sensitivity in Weak Necessity Deontic Modals. Journal of Philosoph-
ical Logic. . . pp. –.
Speaks, Jeff () Theories of Meaning. In Stanford Encyclopedia of Philosophy, edited by
Edward N. Zalta, Winter,  edn.
Starr, William () Mood, Force and Truth. Protosociology. . pp. –.
Suikkanen, Jussi () Metaethics, Semantics, and Metasemantics. Pea Soup Blog. [Online]
Available from: http://peasoup.typepad.com/peasoup///metaethics-semantics-and-
metasemantics.html. [Accessed  November .]
Swanson, Eric () On the Treatment of Incomparability in Ordering Semantics and Premise
Semantics. Journal of Philosophical Logic. . . pp. –.
Thomasson, Amie () Modal Normativism and the Methods of Metaphysics. Philosophical
Topics. .  and . pp. –.
Thomasson, Amie () The Nancy D. Simco Lecture: Norms and Necessity. Southern Journal
of Philosophy. . . pp. –.
Unwin, Nicholas () Quasi-Realism, Negation and the Frege-Geach Problem. Philosophical
Quarterly. . pp. –.
Veltman, Frank () Logics for Conditionals. Ph.D. thesis. University of Amsterdam.
Veltman, Frank () Defaults in Update Semantics. Journal of Philosophical Logic. . . pp.
–.
Wedgwood, Ralph () Conceptual Role Semantics for Moral Terms. The Philosophical
Review. . . pp. –.
Wedgwood, Ralph () The Nature of Normativity. Oxford, New York, NY: Oxford University
Press.
Wedgwood, Ralph () Objective and Subjective ‘Ought’. In Deontic Modality, edited by Nate
Charlow and Matthew Chrisman. Oxford: Oxford University Press.
Wertheimer, Roger () The Significance of Sense. Ithaca, NY: Cornell University Press.
Williams, Michael () Meaning and Deflationary Truth. The Journal of Philosophy. . .
pp. –.
Williams, Michael () Pragmatism, Minimalism, Expressivism. International Journal of
Philosophical Studies. . . pp. –.
Williams, Michael () How Pragmatists Can Be Local Expressivists. In Expressivism, Prag-
matism and Representationalism, edited by Huw Price. Cambridge: Cambridge University
Press.
Yalcin, Seth () Epistemic Modals. Mind. . . pp. –.
Yalcin, Seth () Nonfactualism About Epistemic Modality. In Epistemic Modality, edited by
Andy Egan and Brian Weatherson. New York, NY: Oxford University Press.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

metanormative theory and the meaning of deontic modals 

Yalcin, Seth () Bayesian Expressivism. Aristotelian Society. . pp. –.
Yalcin, Seth () Semantics and Metasemantics in the Context of Generative Grammar. In
Metasemantics: New Essays on the Foundations of Meaning, edited by Alexis Burgess and Brett
Sherman. Oxford: Oxford University Press.
Zvolenszky, Zsófia () Is a Possible-Worlds Semantics of Modality Possible? Proceedings of
SALT Jackson, B. (ed.) University of California, San Diego, CA, – March. . pp. –.

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

i i

i i
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i
Index
actualism n
agglomeration , , , n, , ,
–
Allais paradox 
analyticity –
analytic/synthetic , n
anaphora , , , n
assertion n, n, –, n, n,
n, , , –, , , , ,
–, , –, –, n, 
attitude ascription n, , 
Bayesian semantics n, –, –,
–, , –, 
belief , , , n, , n, , –, , ,
, , –, –, , , ,
–, –, –, , –, –,
–, , , n, , , –,
n
belief ascription , , 
belief revision , 
partial belief , 
binding –, –, –, 
Blackburn, S. , n, , , nn, ,
n, n, n, n, n
BNC See British National Corpus
Brandom, R. , –, , –
British National Corpus n, nn, ,
n
Cariani, F. –, –, n, n, n,
n, n, nn, , n, , ,
, , , –, –, –, –,
–, n, n, –, , n,
n, 
Carr, J. n, n, n, , , –, –,
–, , , –, n, –, –,
n, , , , n, n
causal decision theory , –, , –, ,
–, –, . See also evidential
decision theory
CCP See context-change potentials
CDT See causal decision theory
Charlow, N. –, –, –, , , –, –,
–, –, n, , n, n,
–, n, n, n, n,
n, n, –n, n, n,
n, n, n, n, n
Chisholm, R. , , 
i i
i i
Chisholm’s Paradox , n, , , ,
, 
circumstance of evaluation . See also world
of evaluation
COCA See Corpus of Contemporary American
English
common ground , –, , , –n,
, –
communication n, , , , , –,
, , –, , –, , 
competence , –, , –, , , –,
n, –, , –, , , ,
n, , –, 
competent acceptance , –, n
compositional semantics , –, n, , ,
, –, n, , , , , –, ,
–, –, , , n, , ,
, , , , , , 
conditionals –, , , , , , n, ,
, , , –, , , , , , ,
–, , , n, , –, ,
, , n, , –, 
anankastic conditionals , n
deontic conditionals , , –, , n,
, , n, , n, –, –, ,
–, –, –, –, n, n,
, n
indicative conditionals n, , , n,
, , n, , , –, 
instrumental conditionals –, 
probabilistic conditionals , –, –,
n
context-change potentials , , 
context-sensitivity , n, –, ,
, 
contextualism , , n, –, –, ,
, –, n, n
non-indexical contextualism n
conversation , n, , , –, –, ,
, , , , , , , , ,
–, –, , , , , n
conversational background –, ,
–, –, 
conversational context , , , –,
–, , , 
conversational dynamics –
conversational scoreboard n, 
Corpus of Contemporary American
English n, n
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i
 index
credence , , , n, , , , , ,
, n, , –, , n
deliberation , , –, –, , , ,
, . See also practical
reason/rationality
demonstratives 
direction of fit , –, , 
disagreement n, –, , –, n, ,
, , –, , –, , n, ,
n, , , , n
dominance , , , , , –, , –,
n
dynamic semantics , n, n, n,
n, , , , , –, –,
, –, –, , , n,
–, n, n, n
EDT See evidential decision theory
ellipsis n, n
Ellsberg Paradox , h
evidential decision theory –, , –, ,
n. See also causal decision theory
exclamitives , , –, –, n
expected utility , n, , , –, n,
, , , –, , n, , ,
–, , 
expected value –, –, –, n, ,
n, n, , –, –, , , ,
, , , –, , n, –, ,
, –
expressivism , –, n, , –, –, ,
, n, , –, –, –,
–, –, n, n, –,
–
Finlay, S. –, , , , , n,
n, n, n
Fodor, J. –, n
Frege , , , , n
Frege-Geach problem , –, 
Geach, P. nn, , , , n, ,
, , n
Gentle Murder Parodox –
Gibbard, A. , n, , , , , –,
, –, , n
Gillies, A. n, n, n, , –,
, , 
gradability , , , , , , , , , ,
, –, , , 
Grice, P. , , n
Há –
Harman, G. , , 
Harper, W. L. , , 
Horty, J. F. , 
i i
i i
hyperplan n, 
Iatridou, S. –, –, –, –, , n,
, n, –
ideationalism , , –, –
illocutionary force/function , , 
imperatives , , –, , n, ,
–, –, , –, –, ,
n, , n, n, –,
–
categorical imperative 
hypothetical imperative , , 
implicature n, , n, , n, n,

inferentialism –, –, –, –
information-sensitivity –, n, , , ,
, , , –, , –, , , –,
, –, n, , , , , –,
, , , –, –, n, n,
, –, –, , –
inheritance , n
intention , , , –, n, , n,
, , –, –, –, –, ,

irrationality , , –, n, –,
–, , . See also rationality
Jackson, F. n, , , , –, , 
Kant, I. n, n, n, n, n
Kaufmann, M. –, –, –, –,
n, 
Kaufmann, S. –, –, –, –,
n, 
Kennedy, C. –, 
Kolodny, N. , –, , –, , –, ,
, n, , , n
Kratzer, A. –, , –, n, n, –,
–, , , , n, , –, –,
, , –, , –, , , –,
–, –, , , –, , ,
, n, , , n, n
Lassiter, D. –, , n, n, n, n,
n, , –, , , 
Lepore, E. n, , –
Lewis, D. , , , , , –, ,
nn, , , , , , , , ,
–, –, –, , , nn,
, n, , , n, n, 
lexical ambiguity n, , , , , ,
–
lexical semantics –, , –, 
limit assumption n, , n, n,
–, n, , n
logical form , n, n, n, , ,
–, n, , 
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i
MacFarlane, J. , –, –, , , , n,
–, –, , , , n, n
McNally, L. –, 
Medicine Puzzle , , 
metanormativity –, , –, , –,
n, –, 
metasemantics , –, n, , , n,
, , , , n, , , , ,
n, , , , –, , –,
–, 
Miner Puzzle , , –, –, –, ,
, , , –, –, –, n,
, –, –, –, –, ,
n, 
modal base n, –, n, –, , n,
, , n, n, n, –,
–, –, , –, –, , ,
–, n, , –, –. See also
ordering source
mode of presentation –, , –
monotinicity –, n, –, –, ,
, –. See also nonmonotinicity
Moore’s Paradox 
neutrality –, , , , , , n,
–, , , 
Newcomb Problem –
nonmonotinicity –, , , , , ,
–, , , , –, –.
See also monotinicity
normativity –, –, , n, , –, ,
, –, , , –, –, , , ,
, –, –, –, , –,
–, , –, , , , , ,
–, n, , –, n, –,
–, –, , –, –,
n
Open Question Argument n, , 
ordering source n, , –, , , , n,
n, , n, , –, –,
–, –, –, –, , ,
–, , –, –, , n.
See also modal base
perlocutionary force/function –, 
Plunkett, D. n, , n, n, 
polysemy n, 
Portner, P. –, , , n, , n,
n, , , n, n
possible world , , , , , n, ,
n, , , –, n, , , , ,
, –, , , –, –, ,
, , , , –, , , , ,
, , –, –, n, –, ,
–, , –, –, –,
–, 
i i
i i
index 
practical reason/rationality , , –, , , ,
, –, , , , , , n,
–, 
pragmatics , , –, , , n, , , ,
, –, , , , , , –,
, , –, , , –, , ,
, , –, n, , n, ,
, , –, , , , ,
n, –, –, n, ,
–, , n, n, n
pragmatism 
preference , –, –, –, , –, ,
–, , –, , , , –, , –,
, , n, n, –, –,
–, n, , , , n
proposition , –, , , –, , , ,
–, n, –, , –, , , ,
–, –, –, –, –, ,
, –, , –, –, , , ,
, n, , –, –, –, ,
, , , –, n, –,
–, –, –, , , n,
–, –, –, , –, , ,
–, n, , , –, –,
n, , n, –, n
Fregean proposition –, –
Russellian proposition –
propositional attitude , n, n, ,
n
Putnam, H. , 
quantifier n, , , , –, , , ,
, , –, , n, 
Quine, W. V. O. –, n
Ramsey, F. , , 
rationality , , , , n, . See also
irrationality
relativism n, –, –, , –,
n, n
relevant alternative , –, , , , n,
–, , n, 
representationalism –, , n, –,
–, –, –, 
restrictor n, , n, , , 
Ross’s Paradox , –, , –, ,
–, , , 
Rubinstein, A. –, , , , n
scale , –, , , , , , ,
–, –, –
Schroeder, M. n, n, –, –, ,
n, n, , n, , ,
n, n, n
Sellars, W. , , , n, , –, 
Silk, A. n, n, n, n, –,
–, –, n
 i i
OUP CORRECTED PROOF – FINAL, //, SPi
i i

 index

Spanish  uncertainty , –, n, , , , ,

speech act , , , , , , –,
–, 
St’át’imcets , 
Stalnaker, R. , , n, , n, ,
, n, , 
Stone, M. –, –, , 
strong necessity –, , n, , , , ,
, , –, , –, , ,
–, , 
Sundell, T. n, , n, n, 
Swanson, E. , n, n, n,
, 
syntax , , , , , , 
truth-condition –, –, n, , , –, ,
–, –, , , , , –, –,
, , n, –, , , n,
–, , , , n, , , ,
, , –, –, –
truth-conditional semantics –, , n, ,
, , –, , , , , –,
, n, , , , , n, 
Tsafiki , 
n, n, n, n, n, , ,
, , , , , n, , ,
, –, , , , 
Veltmann, F. , n, , n, , –,
n, –, , n, 
Von Fintel, K. n, n, , , n, ,
–, n, , n, , n, n,
n, –, , , , , n,
, , –, , , , n, ,
n, –, n, n
weak necessity –, n, n, , , –,
–, , , , , , , ,
–, –, –, , , –,
, , n
world of evaluation n, , , –, ,
, n, –, –, , n, 
Yalcin, S. –, n, –, n, –,
n, n, n, n, , n,
n, n, n, , , ,
–, , n, n, n

i i

i i