chap ter title ver so 

J:@<EK@=@:
LE;<IJK8E;@E>
 pa rt title ver so
chap ter title ver so 

J:@<EK@=@:
LE;<IJK8E;@E>
Philosophical Perspectives

< ; @K < ;  9 P

?\ebN%[\I\^k#
JXY`eXC\fe\cc`#Xe[
BX`<`^e\i

LE@M<IJ@KPF=G@KKJ9LI>?GI<JJ
 pa rt title ver so

Published by the University of Pittsburgh Press, Pittsburgh, Pa., 

Copyright © , University of Pittsburgh Press
All rights reserved
Manufactured in the United States of America
Printed on acid-free paper
This paperback edition, 
         

ISBN : ----

ISBN : ---

Library of Congress Cataloging-in-Publication Data

Scientific understanding : philosophical perspectives / edited by Henk W. de Regt,

Sabina Leonelli, and Kai Eigner.
p. cm.
Includes bibliographical references and index.
ISBN-: ---- (cloth : alk. paper)
ISBN-: --- (cloth : alk. paper)
. Science—Philosophy. I. Regt, Henk W. de. II. Leonelli, Sabina. III. Eigner, Kai.
Q.S 
—dc

chap ter title ver so 

We dedicate this work

to the memory of
Peter Lipton (–)
 pa rt title ver so
chap ter title ver so 

:FEK<EKJ

Preface ix
. Focusing on Scientific Understanding
henk w. de regt, sabina leonelli, and kai eigner 

 EVgi>#JcYZghiVcY^c\!:meaVcVi^dc!VcY>ciZaa^\^W^a^in
. Understanding and Scientific Explanation
henk w. de regt 
. Understanding without Explanation
peter lipton 
. Ontological Principles and the Intelligibility of Epistemic Activities
hasok chang 
. Reliability and the Sense of Understanding
stephen r. grimm 
. The Illusion of Depth of Understanding in Science
petri ylikoski 

 EVgi>>#JcYZghiVcY^c\VcYBdYZah
. Understanding in Physics and Biology: From the Abstract
to the Concrete
margaret morrison 
. Understanding by Modeling: An Objectual Approach
tarja knuuttila and martina merz 
. The Great Deluge: Simulation Modeling and Scientific
Understanding
johannes lenhard 

 EVgi>>>#JcYZghiVcY^c\^cHX^Zci^ÒXEgVXi^XZh
. Understanding in Biology: The Impure Nature of
Biological Knowledge
sabina leonelli 
viii contents

. Understanding in Economics: Gray-Box Models

marcel boumans 
. Understanding in Physics: Bottom-Up versus Top-Down
dennis dieks 
. Understanding in the Engineering Sciences:
Interpretive Structures
mieke boon 
. Understanding in Psychology: Is Understanding a Surplus?
kai eigner 
. Understanding in Political Science: The Plurality of
Epistemic Interests
jeroen van bouwel 
. Understanding in Historical Science: Intelligibility and Judgment
edwin koster 

Contributors 
Index 
chap ter title ver so 

GI<=8:<

The idea for this volume was born in the aftermath of the conference “Philo-
sophical Perspectives on Scientific Understanding,” held in Amsterdam in Au-
gust . Our thanks go to all the conference participants and particularly to
Hasok Chang and Peter Lipton, whose engagement and vision were decisive
input for the making of this volume. Peter Lipton’s untimely death in Novem-
ber  is an irreparable loss for the philosophy of science community. He
submitted the final version of his contribution to our volume in the summer of
; it is sad that he cannot see it in print anymore. We dedicate this volume
to his memory.
Both the conference and the realization of this volume were part of the
research program “Understanding Scientific Understanding,” sponsored by the
Netherlands Organization for Scientific Research (NWO). We are grateful to
NWO and to the Faculty of Philosophy of the VU University Amsterdam for
the support given to the project and for enabling us to organize a conference to
share our interest with like-minded philosophers.
We would also like to thank our colleagues at VU University, especially Hans
Radder, for their enthusiastic support and for their intellectual contributions to
our thinking. Sabina Leonelli is grateful for the support of her colleagues at the
ESRC Centre for Genomics in Society, University of Exeter, and the project
“How Well Do ‘Facts’ Travel?” London School of Economics.
Finally, we thank Cynthia Miller and her staff at the University of Pittsburgh
Press for their cooperation.

ix
 pa rt title ver so
chap ter title ver so 

J:@<EK@=@:
LE;<IJK8E;@E>
 pa rt title ver so
chap ter title ver so

(
Focusing on Scientific Understanding
= : C @  L #  9 :  G : < I!  H 6 7 > C 6  A : D C : A A > !
6C9@6>:><C:G

 In the eyes of most scientists, and of educated laypeople, under-

standing is a central goal of science. In the past centuries scientific research
has enormously increased our understanding of the world. Indeed, it seems a
commonplace to state that the desire for understanding is a chief motivation
for doing science. But despite the prima facie plausibility of these claims, it is
not so clear what their precise content is. What do we mean when we say that
scientists understand, for example, global climate change? What is involved
in achieving scientific understanding of phenomena, be they the origin of the
universe, the structure of matter, the behavior of organisms, or economic and
social developments? These are fundamental philosophical questions, yet they
have received little philosophical attention. This volume constitutes the first
concerted effort by philosophers of science to explore these questions and pro-
vide a variety of possible answers.
Lord Kelvin famously declared, “It seems to me that the test of ‘Do we or
not understand a particular subject in physics?’ is, ‘Can we make a mechanical
model of it?’ ” (Kargon and Achinstein , ). What is expressed here is not
only the idea that understanding is an aim of science (in this case, physics) but
also the view that there is a preferred way of achieving that aim, namely, devis-
ing mechanical models of physical phenomena. Yet, while Kelvin’s approach
had strong appeal and was widely supported in the mid-nineteenth century,
it has proved to be untenable in the light of later developments in physics. In
particular, the advent of quantum theory refuted the universal applicability of


 de re gt, leo nelli, and eigner

mechanical modeling as a road to understanding. The failure of such models

led to a discussion among physicists about what kind of understanding phys-
ics could and should achieve. Thus, Erwin Schrödinger criticized the abstract
mathematical theory of matrix mechanics for its lack of Anschaulichkeit (intel-
ligibility, visualizability). Schrödinger (, ) argued that “we cannot really
alter our manner of thinking in space and time, and what we cannot compre-
hend within it we cannot understand at all.” Werner Heisenberg (, ), by
contrast, believed that nonvisualizable theories also could provide understand-
ing (anschauliches Verstehen) of quantum phenomena. These examples show
that ideas about what it takes to understand phenomena can change. Moreover,
they reveal that it is important to distinguish between understanding phenom-
ena and understanding the theories or models used to represent and/or ex-
plain phenomena. While scientists’ ultimate goal is to understand phenomena,
achieving that goal often involves the development and understanding of ap-
propriate theories or models.
A look at sciences other than physics confirms that there are many ways to
achieve understanding of phenomena and, consequently, that there are many
forms of understanding that play a role in scientific practice. Within experi-
mental biology, for instance, there is a strong emphasis on understanding phe-
nomena through the manipulation of biological entities such as organisms or
their components. Biologists have developed several forms of experimental in-
tervention ranging from the unobtrusive observation of animals in the field to
the dissection of specimens in a laboratory. Psychologists, who cannot resort to
the latter type of intervention on human subjects, rely on models and empathy
to acquire understanding of the human mind. Similarly, but on a very differ-
ent scale, economists favor the use of mathematical models and simulations to
capture the behavior of the market under changing conditions.
Given this diversity, it is not surprising that there is no satisfactory, gener-
ally accepted answer to the question of what precisely scientific understanding
consists in, and how it is achieved. It seems to be impossible to give a single,
universally valid definition of the notion of scientific understanding. But this
impossibility should not drive philosophers away from the study of under-
standing. On the contrary, the multifaceted nature of understanding testifies to
its central importance for practicing science and to the need for a philosophi-
cal analysis of its many aspects and forms. The essays in this book constitute a
wide-ranging and in-depth investigation of the nature of scientific understand-
ing. While a general account of understanding may prove to be unattainable,
much can be said, as this book shows, about the ways it becomes manifest
and about the specific tools and conditions that can promote understanding
of a typically scientific kind. In this introductory chapter we will introduce the
topic by reflecting on the most important features of scientific understanding,
fo c using on scientific under standing 

by reviewing what philosophers of science have said about it, and by outlining
the main themes of the book. Rather than defending specific claims in detail,
a task that we leave to individual chapters, we focus on what we perceive to be
the overall position emerging from this collection.
A key feature of understanding is the fact that it involves a cognizing sub-
ject. In the case of scientific understanding this subject is typically a scientist
who tries to understand a phenomenon, for example by developing a theory.
There is an important difference between understanding and explanation:
while one may legitimately say that theory T explains phenomenon P, one can
only speak about understanding P by means of T if one invokes a subject (in the
case above, the scientist). In other words, while explanation may be viewed as
a two-term relation between an explanans (for example, a theory) and an ex-
planandum (the phenomenon), understanding is always a three-term relation
involving explanans, explanandum, and a subject. This feature of understand-
ing has important implications.
First of all, it entails that understanding belongs, at least partly, to the do-
main of pragmatics: when we want to analyze how scientific understanding of
a phenomenon is achieved, we cannot restrict ourselves to analyzing the syn-
tactic structure and semantic content of the statements (for example, theories,
explanations) involved in this activity. In addition, we should consider the pur-
pose and effects of these statements for the particular persons who use them.
This applies to understanding as the end-product of explanatory activity (the
cognitive state achieved, the goal of explanation), because it is the state of a
cognizing subject. And it applies even more to understanding the theories and
models used in these explanatory activities: such an understanding concerns
the usefulness and tractability of the theories and models as evaluated by a
specific group of peers.
Philosophers in the logical empiricist tradition, most notably Carl Hempel,
claim that for this reason the notion of understanding falls outside the scope
of the philosophy of science. Their argument is that a philosophical account of
science should be objectivist and steer clear of pragmatic elements. Accord-
ing to this viewpoint, understanding is nothing but a subjective feeling evoked
by explanations, which is irrelevant for philosophy of science for it can have
no epistemic import. Nowadays, however, many philosophers of science have
more liberal viewpoints regarding the relevance of pragmatics for philosophy
of science and see no reason to reject the notion of understanding out of hand.
Obviously, the authors in this volume are among this group: they agree that
understanding is a philosophically interesting and relevant notion, despite or
by virtue of its pragmatic aspects.
Another implication of viewing understanding as a three-term relation is
accepting that it may be context-dependent. Whether a particular explanation
 de re gt, leo nelli, and eigner

makes scientists understand a phenomenon may depend on the context (both

social and material) in which those scientists find themselves. Among many
others, the case of Kelvin cited above suggests that criteria for understanding
vary in history and across disciplines. In other words, the historical and dis-
ciplinary context determines at least in part when and how understanding is
achieved. This offers a new perspective on the well-known thesis that there is a
fundamental distinction between the natural sciences, which aim at explaining
natural phenomena (Erklären), and the social sciences and humanities, which
aim at understanding human actions and cultural products (Verstehen). That
thesis hinges on an opposition between explanation and understanding that
we think should be abandoned. On the one hand, understanding plays a role
in the natural sciences, too. On the other hand, the understanding provided in
the social sciences and humanities may be the product of explanatory activi-
ties. Of course, it might be that the conditions for achieving understanding are
fundamentally different among the various sciences; this is just one example
of the contextuality of scientific understanding. This contextuality implies, in
turn, that it is difficult, if not impossible, to give a universal definition of (the
essence of ) scientific understanding.

JcYZghiVcY^c\^ci]ZE]^adhde]nd[HX^ZcXZ

Until quite recently philosophers of science have paid relatively little at-
tention to the topic of scientific understanding because, as noted above, un-
derstanding necessarily involves a subject. Many philosophers of science,
especially those raised in the objectivist tradition of logical empiricism, took
this to imply that understanding is thereby also subjective. They concluded that
it can be nothing more than a psychological by-product of scientific activity. In
Hempel’s words, “such expressions as ‘realm of understanding’ and ‘compre-
hensible’ do not belong to the vocabulary of logic, for they refer to the psycho-
logical and pragmatic aspects of explanation” (Hempel , ).
According to Hempel, every good scientific explanation is a logically valid
argument in which the explanandum is deduced from an explanans contain-
ing at least one universal law and relevant initial and background conditions.
This is the well-known deductive-nomological (D-N) model, which was first
proposed by Hempel and Oppenheim in .1 In this pioneering paper (re-
printed in Hempel ), the authors distinguished between “understanding in
the psychological sense of a feeling of empathic familiarity” and “understand-
ing in the theoretical, or cognitive, sense of exhibiting the phenomenon to be
explained as a special case of some general regularity” (Hempel , ).
They criticized the former as being neither a necessary nor a sufficient require-
ment for scientific explanations, and argued that explanations can only be said
fo c using on scientific under standing 

to provide understanding in the latter sense. Note, however, that it is unclear

whether the understanding lies in the objective relation of subsumption under
laws itself, or in the ability of “exhibiting” those relations. In his  “Aspects
of Scientific Explanation,” Hempel added that a deductive-nomological expla-
nation can provide understanding because “the argument shows that, given the
particular circumstances and the laws in question, the occurrence of the phe-
nomenon was to be expected; and it is in this sense that the explanation enables
us to understand why the phenomenon occurred” (). If Hempel wants to
speak about understanding at all, it is only in a sense that is closely tied to the
objective (deductive) relation between explanans and explanandum.
The D-N model dominated the scene until the s, even though it had
been criticized in earlier years. One influential early critic was Michael Scriven,
who argued for giving understanding a more prominent role in the analysis of
explanation than Hempel did: “Whatever an explanation actually does, in order
to be called an explanation at all it must be capable of making clear something
not previously clear, that is, of increasing or producing understanding of some-
thing” (Scriven , ). Scriven added that it depends on the context wheth-
er a particular piece of information appropriately fulfills this role, and hence
qualifies as an explanation. He emphasized, however, that “understanding is
not a subjectively appraised state anymore than knowing is; both are objectively
testable and are, in fact, tested in examinations” (), and he concluded that
Hempel’s theory is unsatisfactory for several reasons, among which, “It leaves
out of account three notions that are in fact essential for an account of scientific
explanation: context, judgment and understanding” ().
A decade later, in , Michael Friedman took another step toward a re-
habilitation of understanding. In his seminal paper, “Explanation and Scientific
understanding,” he argued that one of the requirements for a good philosophi-
cal theory of explanation is that it “should somehow connect explanation and
Understanding—it should tell us what kind of understanding scientific expla-
nations provide and how they provide it” (Friedman , ). Like Scriven,
Friedman criticized Hempel for not living up to this demand, but he followed
Hempel in insisting that a philosophical account of explanation (and accord-
ingly, of understanding) should be objective and noncontextual: “what counts
as an explanation should not depend on the idiosyncrasies and changing tastes
of scientists and historical periods. It should not depend on such nonrational
factors as which phenomena one finds somehow more natural, intelligible, or
self-explanatory than others” ().
Since Friedman’s paper, philosophers of science, including those in the ob-
jectivist tradition, have paid more attention to the topic of understanding. They
have in different ways tried to specify the (objective) kind of understanding that
science provides, and employed this specification as a requirement for good
 de re gt, leo nelli, and eigner

scientific explanations. There have been two major approaches in the objectiv-
ist tradition: the unificationist approach, advanced by Friedman, Philip Kitcher,
and others, and the causal-mechanistic approach, developed by Wesley Salmon
and others. In addition, more in line with Scriven’s position is a pragmatic ap-
proach to explanation advanced by Bas van Fraassen and Peter Achinstein.
Friedman’s idea was that objective understanding is provided by explana-
tions that unify the phenomena: “science increases our understanding of the
world by reducing the total number of independent phenomena that we have to
accept as ultimate or given” (Friedman , ). Because of technical problems
with Friedman’s account, Philip Kitcher (, ) developed an alternative
approach to unification in which argument patterns are central. An argument
pattern consists of a schematic argument with filling instructions and a clas-
sification of the inferential structure of the argument. According to Kitcher,
“Science advances our understanding of nature by showing us how to derive
descriptions of many phenomena, using the same patterns of derivation again
and again” (, ). Since Friedman’s and Kitcher’s essays, much more work
has been done on unification. A notable account that refers especially to the
understanding-providing power of unification is that of Gerhard Schurz and
Karel Lambert ().
Meanwhile, following the severe critiques launched against the D-N model,
causal approaches to explanation were developed that offered an alternative
conception of explanatory understanding. The most important proponent of
this view is Wesley Salmon, who presented an elaborate theory of causal-mech-
anistic explanation in his  Scientific Explanation and the Causal Structure
of the World. This theory was later refined and elaborated in discussion with
others (see Salmon , esp. chaps.  and ). According to Salmon, we need
a causal theory of explanation because “underlying causal mechanisms hold
the key to our understanding of the world” (Salmon , ; italics added).
This is because “causal processes, causal interactions, and causal laws provide
the mechanisms by which the world works; to understand why certain things
happen, we need to see how they are produced by these mechanisms” (). In
his later work, Salmon has put even stronger emphasis on the importance of
achieving understanding in science (see Salmon , , , , and chap. ).
Notable alternative causal approaches are defended by Paul Humphreys ()
and James Woodward (). Mechanistic explanation is nowadays very popu-
lar in the philosophy of biology (Machamer, Darden, and Craver ; Bechtel
and Abrahamsen ).
A different approach to explanation and understanding, which falls out-
side the objectivist camp, is provided by van Fraassen () and Achinstein
(). In van Fraassen’s account of science, explanation is a pragmatic virtue
fo c using on scientific under standing 

and does not constitute an epistemic aim of science. Van Fraassen analyzes
explanations as answers to why-questions, and argues that explanation is con-
text dependent. Although the issue of the character of scientific understand-
ing is not mentioned explicitly by van Fraassen, one may safely conclude that
his analysis of the nature of explanation as extrascientific holds a fortiori for
understanding. Achinstein () focuses more directly on understanding by
providing a critique of noncontextual approaches to explanation. In his view,
the act of explaining is an illocutionary act characterized by the intention to
make something understandable. Whether an explanation is “good” depends
on the values and interests of the audience to which it is directed. Achinstein
emphasizes that, due to this context dependency, there is a variety of different
forms of good explanation in science.

6heZXihd[HX^Zci^ÒXJcYZghiVcY^c\

The above overview shows that understanding has typically been viewed as
a by-product of scientific explanation. This volume provides a radically different
approach: the study of understanding is not only worthy of philosophical atten-
tion, but actually contributes new insights to a number of traditional debates
within the philosophy of science. The most important themes emerging from
this collection of papers are the following: the relation between understanding
and explanation; the feeling of understanding; the role of models and theories
as tools to achieve understanding; the notion of intelligibility; understanding as
tacit knowledge; and pluralism of understanding. This list highlights some key
results obtained through our philosophical investigation of scientific under-
standing; it also serves to guide the reader toward the specific chapters where
reflections on these themes are most thoroughly developed.

JcYZghiVcY^c\VcY:meaVcVi^dc
The objectivist position reduces understanding to an automatic conse-
quence of having a good explanation, where “good” indicates its unifying power
or causal-mechanistic nature. As discussed above, this view does not resolve
the issue of what understanding actually involves. We take issue with the facile
equation between possessing an explanation and possessing understanding of
a phenomenon. Gaining understanding through explanations is not an auto-
matic process, but rather a cognitive achievement in its own right—a point that
is stressed, in various contexts, by all contributors to this collection.
Once it is granted that deriving understanding from an explanation is a
matter of ability, the question arises of how that actually works. What are the
mechanisms through which scientists extract understanding from explanations
 de re gt, leo nelli, and eigner

that they already possess? Confronting this question in the case of historiog-
raphy, Edwin Koster emphasizes the role of personal judgment in determining
which ideas and events are selected as part of an explanation. Henk de Regt
and Sabina Leonelli both note how judgment in the natural sciences involves
the exercise of epistemic skills of various kinds, which help scientists to make
sense of the conceptual as well as the empirical significance of explanations for
the phenomena to which they refer. Peter Lipton strengthens the point by chal-
lenging the very idea that explanations are needed to acquire understanding: he
discusses cases where individuals acquire understanding without recourse to
explanations. Further, Lipton distinguishes four types of cognitive benefits that
derive specifically from understanding a phenomenon: causal information, as
acquired through activities such as observation, experimentation, manipula-
tion, and inference; a sense of necessity deriving from the knowledge that, in
given conditions, an event, object, or process could not have been otherwise;
a sense of what is possible, which can spring even from potential or false ex-
planations; and the unification obtained by comparing phenomena through
mechanisms such as analogies and classifications, as in the case of Kuhnian ex-
emplars. Lipton shows that each of these cognitive benefits can be obtained in
the absence of explanations, indicating that “understanding is more extensive
and more varied in its sources than those who would equate explanation and
understanding allow.”

I]Z;ZZa^c\d[JcYZghiVcY^c\
Petri Ylikoski explores cases where the acquisition of scientific understand-
ing is at least partly illusory, insofar as it springs from a feeling of understanding
(“Aha!”) that has no relation to whether that understanding is reliable or truth-
conducive. This cognitive phenomenon is recognized by several authors in the
volume, who all agree on the necessity to distinguish the feeling of understand-
ing and understanding itself, but who disagree quite radically on whether the
feeling of understanding has any epistemic value. De Regt does not think that
this is the case. In his view, “the phenomenology of understanding has no epis-
temic function”: what counts, rather, are the epistemic skills and values used
to construct tools for understanding, such as intelligible theories and models.
Ylikoski goes even further by characterizing this feeling, which he dubs “illu-
sion of depth of understanding,” as having a downright pernicious effect on
various aspects of scientific research. He notes how several features of contem-
porary science have evolved as measures against illusory understanding felt
(and promoted) by individuals. For instance, the peer review system exposes
personal insights to critique, while the development of external representations
makes it possible to express individual intuitions so that they can be discussed
fo c using on scientific under standing 

and tested. According to Ylikoski, scientists can improve their ability to dis-
cern reliable from illusory understanding by adopting strict standards for what
constitutes an adequate explanation—and philosophers of science are ideally
placed to contribute to this effort.
By contrast, Lipton and Stephen Grimm share a belief in the epistemic
value of the “aha” experience, which they both characterize as helpful rather
than detrimental to the search for scientific understanding. After illustrating
the differences between the feeling of understanding and other pleasant feel-
ings involved in research, such as surprise and expectation, Lipton notes how
the “aha” experience might plausibly work not only as an incentive, but also as
a guide to the human search for understanding. Lipton proposes that scientists
use their feelings of understanding as ground for their choice of the “loveliest”
explanation. This does not imply that understanding itself is hopelessly subjec-
tive, as Lipton makes clear by distinguishing the feeling of understanding from
actual understanding (as he states, “understanding is a form of knowledge; the
feeling of understanding is not”). Rather, it points to the unarticulated, and yet
epistemically relevant, grounds on which scientists judge their results. Grimm
complements this argument by pointing to the “conditional reliability” of the
feeling of understanding, which depends chiefly on the accuracy of the back-
ground beliefs of whoever experiences it. Scientists are most likely to feel that
they understand something when that bit of information coheres with the rest
of their beliefs about the world. This by no means guarantees the reliability of
the feeling of understanding, but it does imply that it can be checked by review-
ing one’s background knowledge (or, as Johannes Lenhard proposes, by employ-
ing simulations to check the empirical consequences of one’s assumptions).

BdYZah!I]Zdg^Zh!VcYJcYZghiVcY^c\
A crucial question raised in this volume concerns the relation between
models and theories as aids toward acquiring understanding. Models are the
subjects of part  and their role toward the acquisition of understanding is
highlighted by virtually all contributors. There are, however, various interpre-
tations of the relation between models and theories and of how scientists might
use either or both of them as tools to understand phenomena.
Some authors support the view that theories are only useful for under-
standing when they are represented through models. Lenhard argues along
these lines in his discussion of understanding through computer simulations.
In his view, simulations constitute surrogates for theories: they allow scientists
to explore the implications of adopting specific theoretical assumptions, yet
at the same time the development of simulations requires scientists to black-
box several aspects of the underlying modeling and theories so as to enable
 de re gt, leo nelli, and eigner

technical implementation (for example, programming and display). Thus,

simulations provide information on how to construct and use models to un-
derstand phenomena by allowing scientists to compare the implications and
advantages of models based on different theoretical assumptions. Tarja Kn-
uuttila and Martina Merz explore the implications of such an account: “the
understanding-providing features of models exist due to the interplay between
their materially embedded medium-specific and technological features, and
their multiple uses.” By pointing to models and their uses as the main locus for
scientific understanding, they limit the role of theories to their usefulness as
instructions for building models.
Other authors are instead interested in exploring the extent to which theo-
ries might play an independent role as tools for understanding. In his chapter
on understanding in physics, Dennis Dieks emphasizes the role of fundamen-
tal theory in developing different understandings of relativistic effects. Mieke
Boon also illustrates how, even in the supposedly “applied” engineering sci-
ences, theories may be seen to play a crucial role in fostering understanding.
She presents the notion of “interpretative structures” as involving the use of
fundamental theories to make sense of one’s experience of phenomena. These
structures serve the function assigned by Koster to personal judgment: that is,
they help scientists to determine which types of objects (and relations among
objects) to use when constructing an explanation of a phenomenon. Notably,
Boon leaves open the possibility that fundamental theories, causal explana-
tions, and models may all play the role of interpretative structures, depending
on the scientific context in which they are used.
Another take on the relation between theories and models derives from
studying the mechanisms, rather than the tools, through which understand-
ing is acquired. This is the task undertaken by Margaret Morrison in the case
of mathematical abstraction in physics and biology. As she notes, mathemati-
cal abstraction is an important mechanism through which understanding is
developed, especially in highly formalized fields such as quantum physics or
population genetics. Indeed, abstraction blurs the boundaries between what
is considered to be a model and what is considered to be theory—as in the
case of the Hardy-Weinberg equation, a model so abstract and general so as
to be regarded as a fundamental law in population genetics. Morrison argues
that mathematical abstraction can and often does generate understanding of
concrete systems that could not have been obtained otherwise: it allows scien-
tists to model those systems through assumptions that could never hold in the
physical world (such as the idea of infinite populations), and yet are necessary
to obtain new information about the potential behavior of those systems. Mor-
rison shares with several other contributors a willingness to shift philosophical
fo c using on scientific under standing 

focus from the structural components of scientific research (such as theories,

models, instruments, and phenomena) to the ways in which these components
are used to carry our research. What “used” actually implies is made clear when
reflecting on intelligibility and tacit knowledge.

>ciZaa^\^W^a^in
One well-known case of “using” instruments to understand phenomena
is Galileo’s use of the balance to understand principles of equilibrium. As ar-
gued by Machamer and Woody (), using the balance crucially improved
the intelligibility of Galileo’s views on motion to other scientists. Indeed, un-
derstanding unavoidably involves making something intelligible. Yet what de-
termines intelligibility and how can intelligibility be improved? Premodern
rationalist philosophers held that knowledge must be based on self-evident
“intelligible principles.” Modern science has rejected this methodology, and
empiricist philosophy leaves no room for intelligibility in this sense. Kant’s phi-
losophy revived the idea of intelligibility in a different form, namely via the
forms of intuition (Anschauungsformen) that structure sensory input and the
categories that transform it into knowledge. However, in the early twentieth
century the alleged refutation of the Kantian system by relativity and quantum
physics led again to a rejection of intelligibility as a philosophically reputable
concept, especially by logical-positivist and logical-empiricist philosophers
(Frank  provides a nice example of this tendency—this semipopular in-
troduction to the philosophy of science turns out to be one long attempt to
debunk intelligibility).
In his chapter, Hasok Chang gives new vitality to the Kantian insights by
viewing intelligibility as an epistemic virtue whose function is to harmonize
our actions with our basic beliefs about the world. An unintelligible action is
an action that cannot be performed because it betrays one of the principles
guiding all human epistemic activity—for instance, the principle that “a real
physical property can have no more than one definite value in a given situa-
tion,” which Chang calls the ontological principle of single value. The betrayal
of an ontological principle does not represent a mistake or a falsehood; rather,
it involves taking up a belief that makes no sense to us (such as the belief that a
physical property can have two different values in the same situation), because
we would not know how to act on its basis. Ontological principles constitute
the platform of common sense on which the whole scientific enterprise is built
through its various activities, including testing, experimentation, observation,
and the like. In a pragmatist fashion, intelligibility is thus defined as the per-
formability of an epistemic activity—and understanding as the ability to actu-
ally perform such an activity.2
 de re gt, leo nelli, and eigner

Another take on intelligibility as an epistemic virtue is defended by Kai

Eigner. On the basis of his analysis of the behaviorist movement in psychology,
Eigner argues that intelligibility has epistemic relevance as an essential virtue
of the models used to apply theoretical principles to the study of phenomena.
As he shows, using theoretical models to describe phenomena involves mak-
ing judgments about similarity and relevance that cannot be based on objec-
tive methodological rules. To make these skillful judgments, scientists need to
give “surplus meaning” to the theoretical terms in the models such that these
models are rendered intelligible to them. Due to the surplus meaning of the
theoretical terms, the models acquire virtues that allow a match with the skills
of scientist, which in turn enables them to establish the connection between
models and phenomena. De Regt generalizes this position by characterizing
intelligibility as “the value that scientists attribute to the cluster of virtues (of a
theory in one or more of its representations) that facilitate the use of the theory
for the construction of models.”

JcYZghiVcY^c\VhIVX^i@cdlaZY\Z
Objectivist objections to viewing understanding as a reliable source of
knowledge might perhaps seem plausible if one focuses only on theoretical
knowledge. However, a precious lesson taught by the philosophical literature on
models concerns the importance of tacit (or embodied) knowledge for obtain-
ing and interpreting theoretical knowledge. It is only through the use of mod-
els, or indeed any other kind of object or representation, that scientists acquire
understanding of the world. The unarticulated knowledge required for success-
ful interaction with models, phenomena, and experimental settings constitutes
not only an important source of knowledge in itself, but also a strong constraint
to the types of understanding that are actually developed in science.
As Knuuttila and Merz make clear in their chapter, scientists typically seek
to extract evidence from objects with very specific characteristics. Models
provide understanding insofar as they constitute “concrete constructed ob-
jects with which one can interact in various ways.” In their view, understand-
ing springs from the scientists’ interaction with these objects of knowledge.
Leonelli defines tacit knowledge more broadly as “the awareness of how to act
and reason as required to pursue scientific research.” Such awareness is ex-
pressed not only through the handling of objects as models, but also through
the implementation of experimental procedures and protocols, the calibration
of instruments, and the ability to work in specific research environments such
as the lab or the field. As Leonelli notes, in the life sciences the development of
explanations and interventions is inextricably intertwined—a result that is not
surprising in the light of the link proposed here between understanding and the
ability to apply theories and models.
fo c using on scientific under standing 

EajgVa^hbd[JcYZghiVcY^c\
As we noted, the history of science shows great variation in what is and
what is not deemed understandable. Even at one particular moment in his-
tory opinions about what is understandable often diverge. For example, in 
physicists in the Copenhagen-Göttingen circle believed that atomic phenom-
ena could be understood with the theory of matrix mechanics, while most
other physicists—notably Schrödinger—disagreed (Beller ; de Regt ).
Authors in this volume have several ways to account for scientific pluralism
among ways of understanding. One option is to distinguish between embodied
and theoretical knowledge as sources for the skills and commitments used by
scientists to understand phenomena. Theoretical skills and commitments are
the ones involved in reasoning through available concepts, theories, and ex-
planations, while performative skills and commitments are developed through
physical interaction with research objects and settings. Leonelli uses this dis-
tinction to argue for the coexistence of three types of understanding in biology:
one prioritizing recourse to theoretical commitments and skills, one where
performative skills are of primary importance, and one deriving from a bal-
anced coordination of theoretical and embodied knowledge. A second form of
pluralism concerns the diversity of understandings that can be acquired from
using the same tools. This point is supported by Knuuttila and Merz, who stress
the multiplexity of models as tools for acquiring understanding: the same mod-
el can be interpreted in a variety of ways depending on the background skills
and knowledge of their users.
Yet another form of pluralism concerns the epistemic interests and purposes
guiding the search for understanding. Jeroen Van Bouwel shows how “the plu-
rality of epistemic interests is better served by a plurality of theoretical per-
spectives than by a unified one.” In Van Bouwel’s view, understanding derives
from explanations that are both empirically adequate and adequate to the epis-
temic interests of whoever acquires it. Given the existing diversity of epistemic
interests among scientists, understanding is therefore acquired from several
explanations that cannot be reduced to one another. Unification is thus seen as
limiting, rather than aiding in, our understanding of the world.
This view is shared by Dieks in his analysis of physical explanations of rela-
tivistic effects. As he shows, the same physical theory may be used to con-
struct different arguments, depending on “exactly what one wishes to know,
and from which point of view”—that is, on which kind of understanding one
wishes to acquire. Dieks’s work dispels any suspicion that Van Bouwel’s claims
might be uniquely related to the area he studies, that is, political science. As
these two chapters jointly demonstrate, both in the natural and in the social
sciences epistemic interests vary considerably depending on the context and
group adopting them.
 de re gt, leo nelli, and eigner

The last, and possibly most obvious, form of pluralism in understanding

emerges from the discipline-specific analyses provided in part  of this volume.
Different disciplines focus on understanding diverse entities and processes.
The attempt to find methods perfectly suited to each subject matter has gener-
ated a vast array of tools to acquire understanding, which in turn signals great
variation in the types of understanding that can be obtained in each field. For
instance, Koster shows how historians acquire understanding through recourse
to empathy. Understanding a historical event involves judging the actions of
the people involved so as to select relevant factors and evaluate their relative
weight in the explanation of an event. The ability to make such judgments ben-
efits from the capacity of the historian to put himself or herself in the shoes of
the subjects in question.
Eigner also highlights the role of empathy in his analysis of understanding
in psychology. As he demonstrates, even behaviorist psychologists, notoriously
staunch believers in the experimental method and advocates of objective psy-
chology, rely on an empathic (“prescientific”) understanding of their research
subjects to construct intelligible models of their behavior. As argued by Marcel
Boumans, understanding in economics has very different features. Here the
favored tools are modular representations (or, in Boumans’s words, “gray-box
models”), which are efficient in capturing several features characterizing eco-
nomic practices: for instance, the recourse to “passive observations” as data in
the absence of other types of evidence; the calibration of models via mimick-
ing of actual economics; and the need to describe and predict the behavior of
extremely complex systems, which can only be fulfilled through partitioning
them into manageable subsystems.

I]ZE]^adhde]^XVa;jijgZd[HX^Zci^ÒXJcYZghiVcY^c\

The arguments presented in this volume bear wide implications for the phi-
losophy of science. All chapters in this volume firmly agree on linking under-
standing to cognition. Understanding is defined as a cognitive achievement, or
as an ability acquired through appropriate training and as the bearer of cogni-
tive benefits (such as, in Lipton’s words, “knowledge of causes, of necessity,
of possibility and of unification”) that scientists could not acquire solely from
explanations. This emphasis on the cognizant individual involves a reevalua-
tion of the epistemic role of human agency in producing, disseminating, and
using scientific knowledge. To understand scientific understanding, philoso-
phers must find ways to study and analyze scientific agency. This means taking
scientific practices seriously, for arguably a study of agency in science needs
to be based on knowledge of how scientists across various fields actually act.
The study of experimentation, the iconic scientific activity, has already yielded
fo c using on scientific under standing 

relevant insights, as demonstrated by the pioneering work on this topic by Ian

Hacking (), David Gooding (), and Hans Radder (), among oth-
ers.3 Research on scientific understanding can help to integrate these insights
into the philosophy of science and assess their impact on traditional notions of
inference, explanation, modeling, and theory-making.
Reflecting on scientific practices can also help to tackle another question
touched upon in this volume, yet deserving more explicit attention: the relation
between scientific understanding and other forms of understanding (technical,
humanistic, and so on). Any human being has the ability to engage in several
forms of understanding. One way to differentiate between scientific and non-
scientific understanding might be to analyze the social conditions under which
understanding is achieved in science. All scientific disciplines have developed
sophisticated ways to establish and communicate knowledge about nature,
including specific terminologies, representations, methods, and instruments.
Knowing how to use these tools usually requires years of training and profes-
sionalization within the relevant communities. An individual who lacks famil-
iarity with the social and material environment in which research is conducted
will not be able to use scientific tools and skills to understand phenomena—nor
to communicate to his or her peers what he or she understands. This makes
scientific understanding an intrinsically social, rather than individual, achieve-
ment, thus opening a potentially fruitful avenue for research in the social epis-
temology of science.
As we already suggested, defining understanding as pragmatic and contex-
tual goes hand in hand with emphasizing the pluralism in understandings of
phenomena that might be acquired by depending on the skills and commit-
ments of the individual(s) involved. The normative question then arises: Which
type of understanding is best suited for which type of research? In particular,
are there types of understanding (and thus specific combinations of tools and
mechanisms to understand) that are more valuable than others in specific re-
search contexts? One way to investigate this issue could be to construct a clas-
sification of scientific understandings and find criteria to establish how each of
them fulfills different goals and interests (where these goals and interests can
be scientific as well as economic, social, or ethical, as for instance when ac-
quiring a scientific understanding of stem cell research in order to evaluate its
ethical status). This research path could help to develop normative frameworks
to evaluate the quality of scientific understanding achieved in any given case, a
result of great interest to philosophers and scientists alike.
Finding ways to assess the quality of scientific understanding might also help
to assess the relation between scientific understanding and scientific knowl-
edge. Concerns about this relation are bound to remain the most fascinating
and contentious aspect of the approach proposed in this volume. Scientific
 de re gt, leo nelli, and eigner

understanding is the result of the tools, commitments, and skills available to

scientists at a particular point in time and space. Such tools are honed through
constant negotiation with the material world and are thus not simply the fruit
of cultural trends or social settings (as radical constructivists would like to be-
lieve). However, precisely because of the specificity of their context and of the
motivations and interests guiding their use, these tools and commitments are
fallible: they might lead to achieving an understanding of the world, but they do
not guarantee that such an understanding will be truthful. This volume sets no
sure path toward establishing what might work as a guarantee of truth-value in
this context. What we hope to offer is a framework to articulate further the no-
tion of understanding, with the aim of outlining a conception of scientific un-
derstanding that is not entirely dependent on the truth-value of the knowledge
used to understand, but rather incorporates the values, experiences, and skills
of the individuals and communities attempting to carve nature at its joints.

CdiZh
. Hempel later added inductive-statistical explanation to cope with explanations in
which the explanandum cannot be deduced with absolute certainty from the explanans,
but only with a high degree of probability.
. An important outcome of Chang’s analysis is a strong argument for disconnecting
intelligibility from truthfulness (or one of its measures, such as empirical adequacy). This
is a step that several other authors are reluctant to take.
. Paying attention to scientific practices such as experimentation also involves col-
laboration with historians and sociologists of science, whose work has already illuminated
several characteristics of experimental practice (for example, Shapin and Schaffer  and
Bruno Latour ).

GZ[ZgZcXZh
Achinstein, P. . The nature of explanation. New York: Oxford University Press.
Bechtel, W., and A. Abrahamsen. . Explanation: A mechanistic alternative. Studies in
History and Philosophy of Biological and Biomedical Sciences :–.
Beller, M. . Quantum dialogue. Chicago: University of Chicago Press.
de Regt, H. W. . Spacetime visualisation and the intelligibility of physical theories.
Studies in History and Philosophy of Modern Physics B:–.
Frank, P. . Philosophy of science: The link between science and philosophy. Englewood
Cliffs, NJ: Prentice-Hall.
Friedman, M. . Explanation and scientific understanding. Journal of Philosophy
:–.
Gooding, D. . Experiment and the making of meaning. Dordrecht: Kluwer.
Hacking, I. . Representing and intervening: Introductory topics in the philosophy of
natural science. Cambridge: Cambridge University Press.
fo c using on scientific under standing 

Heisenberg, W. . Über den anschaulichen Inhalt der quantentheoretischen Kinematik

und Mechanik. Zeitschrift für Physik :–.
Hempel, C. G. . Aspects of scientific explanation and other essays in the philosophy of
science. New York: Free Press.
Humphreys, P. W. . The chances of explanation. Princeton: Princeton University Press.
Kargon, R., and P. Achinstein, eds. . Kelvin’s Baltimore lectures and modern theoretical
physics. Cambridge, MA: MIT Press.
Kitcher, P. . Explanatory unification. Philosophy of Science :–.
———. . Explanatory unification and the causal structure of the world. In Scientific
explanation, edited by P. Kitcher and W. C. Salmon, –. Minneapolis: University
of Minnesota Press.
Latour, B. . Science in action. Cambridge, MA: Harvard University Press.
Machamer, P., and A. Woody. . A model of intelligibility in science: Using Galileo’s
balance as a model for understanding the motion of bodies. Science and Education
:–.
Machamer, P., L. Darden, and C. Craver. . Thinking about mechanisms. Philosophy of
Science :–.
Radder, H. . In and about the world: Philosophical studies of science and technology.
Albany: State University of New York Press.
Salmon, W. C. . Scientific explanation and the causal structure of the world. Princeton:
Princeton University Press.
———. . Causality and explanation. Oxford: Oxford University Press.
Schrödinger, E. . Collected papers on wave mechanics. London: Blackie and Son.
Schurz, G., and K. Lambert. . Outline of a theory of scientific understanding. Synthese
:–.
Scriven, M. . Explanations, predictions, and laws. In Scientific explanation, space, and
time, edited by H. Feigl and G. Maxwell, –. Minneapolis: University of Min-
nesota Press.
Shapin, S., and S. Schaffer. . Leviathan and the air-pump. Princeton: Princeton Uni-
versity Press.
van Fraassen, B. C. . The scientific image. Oxford: Clarendon.
Woodward, J. . Making things happen: A theory of causal explanation. New York:
Oxford University Press.
 pa rt title ver so
chap ter title ver so 

G8IK@
JcYZghiVcY^c\!:meaVcVi^dc!
VcY>ciZaa^\^W^a^in
 pa rt title ver so
chap ter title ver so

)
Understanding and Scientific Explanation

= : C @  L#  9 :  G : < I

 In , physicist Erwin Schrödinger delivered the Shearman Lec-

tures at University College London. In , these lectures were published as
Nature and the Greeks. In this book Schrödinger argues that science, since
it is a Greek invention and is based on the Greek way of thinking, is “some-
thing special,” that is, “it is not the only possible way of thinking about Nature.”
Schrödinger then poses the following question: “What are the peculiar, spe-
cial traits of our scientific world-picture?” and he answers it immediately by
stating: “About one of these fundamental features there can be no doubt. It is
the hypothesis that the display of Nature can be understood. . . . It is the non-
spiritistic, the non-superstitious, the non-magical outlook. A lot more could
be said about it. One would in this context have to discuss the questions: what
does comprehensibility really mean, and in what sense, if any, does science give
explanations?” (Schrödinger , ). Schrödinger then observes that philoso-
phers from Hume to Mach to the positivists have not given positive answers to
these questions. On the contrary, they have argued that scientific theories are
merely economical descriptions of observable facts, which do not supply expla-
nations. This view, which was endorsed by most of Schrödinger’s physicist col-
leagues, seems to lead to a strange tension: the basic hypothesis of science, that
nature is understandable, does not seem to have any positive content—it con-
sists merely in a denial of supernatural, nonscientific worldviews. Schrödinger
(, –), however, suggests that “even from the positivist’s point of view
one ought not to declare that science conveys no understanding.” Scientific the-


 henk w. de r egt

ories not only summarize facts but also ipso facto uncover relations between
facts (sometimes from quite different domains of knowledge), and our grasping
of these relations may legitimately be called “understanding.”
From a hindsight perspective, Schrödinger’s words appear to be very in-
sightful. The early s were the time when the subject of explanation was be-
coming respectable again, after having been condemned by the nineteenth- and
early-twentieth-century positivist philosophers of science (see, for example,
Mach /, Pearson /). It was only in  that Carl Hempel
and Paul Oppenheim put the topic on the philosophical agenda again, with
their pioneering paper “Studies in the Logic of Scientific Explanation.”1
But Hempel, whose deductive-nomological model of explanation was to
dominate the debate for the next two decades, was still reluctant to talk about
understanding. The reason was that “such expressions as ‘realm of understand-
ing’ and ‘comprehensible’ do not belong to the vocabulary of logic, for they
refer to the psychological and pragmatic aspects of explanation” (Hempel ,
). In Hempel’s logical-empiricist view, the aim of philosophy of science is to
give an account of the objective nature of science by means of logical analysis
of its concepts. Its psychological and pragmatic aspects may be of interest to
historians, sociologists, and psychologists of science (in short, to those who
study the phenomenon of science empirically) but should be ignored by phi-
losophers. Since Hempel, philosophers of science have gradually become more
willing to discuss the topic of scientific understanding, but it has remained out-
side the focus of philosophical attention until quite recently.
In this chapter, I argue that understanding is an important philosophical
topic and I offer an account of the role of understanding in science. The outline
of the chapter is as follows. In the next section, I discuss objectivist views of the
relation between explanation and understanding, in particular those of Hempel
and J. D. Trout. I challenge these views by arguing that pragmatic aspects of
explanation are crucial for achieving the epistemic aim of science. I then offer
an analysis of these pragmatic aspects in terms of “intelligibility” and give a
contextual account of scientific understanding based on this notion. Finally, the
last section presents a brief case study from the history of physics as an illustra-
tion of the proposed account of scientific understanding.

DW_ZXi^k^hiK^Zlhd[:meaVcVi^dcVcYJcYZghiVcY^c\

While scientific explanation has been a legitimate topic for philosophers

of science since , the notion of scientific understanding has been largely
ignored until recently because logical empiricists such as Hempel emphasized
that only explanation is of philosophical interest. According to Hempel, notions
such as understanding and intelligibility are pragmatic and thereby subjective:
under standing and scientific e xpl anation 

Very broadly speaking, to explain something to a person is to make

it plain and intelligible to him, to make him understand it. Thus con-
strued, the word “explanation” and its cognates are pragmatic terms:
their use requires reference to the persons involved in the process of
explaining. In a pragmatic context we might say, for example, that a
given account A explains fact X to person P. We will then have to bear
in mind that the same account may well not constitute an explanation
of X for another person P, who might not even regard X as requiring
an explanation, or who might find the account A unintelligible or unil-
luminating, or irrelevant to what puzzles him about X. Explanation in
this pragmatic sense is thus a relative notion: something can be sig-
nificantly said to constitute an explanation in this sense only for this or
that individual. (Hempel , –)

The hallmark of scientific knowledge is, in Hempel’s view, its objective nature,
and philosophers of science should therefore try to give an objectivist account
of science, and of scientific explanation in particular—and therefore they should
ignore pragmatic aspects such as understanding and intelligibility.
It should be stressed that Hempel was right that the notion of understanding
is pragmatic in the sense that it is concerned with a three-term relation between
the explanation, the phenomenon, and the person who uses the explanation
to achieve understanding of the phenomenon. One can use the term “under-
standing” only with—implicit or explicit—reference to human agents: scientist
S understands phenomenon P by means of explanation E. That understanding
is pragmatic in this sense implies the possibility of disagreement and variation
based on contextual differences.2 But, as I will argue below, Hempel was wrong
when he stated that pragmatic notions are philosophically irrelevant.
Hempel sometimes used the term “scientific (or theoretical) understand-
ing,” which he associated with scientific explanation in the deductive-nomo-
logical sense.3 In the concluding section of his  essay he remarks that the
understanding conveyed by scientific explanation “lies rather in the insight that
the explanandum fits into, or can be subsumed under, a system of uniformities
represented by empirical laws or theoretical principles,” and that “all scientific
explanation . . . seeks to provide a systematic understanding of empirical phe-
nomena by showing that they fit into a nomic nexus” (Hempel , ).
In , Michael Friedman argued that philosophical accounts of scien-
tific explanation should make it clear how explanations provide us with under-
standing, and he claimed that Hempel’s D-N model of explanation failed to do
so. To be sure, Hempel had suggested how science may lead to understanding:
“The [D-N] argument shows that, given the particular circumstances and the
laws in question, the occurrence of the phenomenon was to be expected; and
 henk w. de r egt

it is in this sense that the explanation enables us to understand why the phe-
nomenon occurred” (Hempel , ; italics in original; compare ). In
this way, “understanding” is reduced to rational expectation. But, Friedman
(, ) observes, well-known examples, like the barometer prediction of a
storm, show that rational expectation is insufficient for understanding: we do
not understand the occurrence of a storm by merely referring to the barome-
ter’s indication.
While Friedman agrees with Hempel’s objectivism, he argues that Hempel
wrongly identified “pragmatic as psychological” with “pragmatic as subjective.”
Friedman (, –) holds that understanding can be pragmatic in the sense
of being psychological without also being subjective. Accordingly, an objec-
tivist conception of scientific understanding (connected with explanation as
its product) should be possible. Friedman defends an objectivist account of
scientific understanding based on the idea of unification (which was further
developed by Kitcher ). Since Friedman’s pioneering essay, philosophers
of science have been less reluctant to discuss the notion of understanding,
and particular models of explanation are typically defended by pointing to the
understanding-providing power of the favored type of explanation (for example,
Kitcher  on unification, and Salmon  and  on causal-mechanical
explanation). However, these philosophers typically refrain from presenting a
detailed analysis of the notion of understanding itself.
Moreover, some present-day philosophers still endorse the Hempelian view
that understanding (in the psychological, allegedly subjective sense) should be
banned from philosophical discourse. One of them is J. D. Trout (, ,
), who follows Hempel in arguing that philosophical analyses of scientific
explanation should be objectivist in the following sense: “What makes an ex-
planation good concerns a property that it has independent of the psychol-
ogy of the explainers; it concerns features of external objects, independent of
particular minds” (Trout , ). Therefore, Trout claims, philosophers of
science should eschew subjective notions such as intelligibility and understand-
ing. The latter are epistemically irrelevant if not dangerous (because hindsight
and overconfidence biases show that feelings of understanding are typically
misguided). Trout criticizes philosophers of explanation such as Friedman and
Salmon for (unwittingly) relating explanatory power to a subjective sense of
understanding, and he notes that even Hempel could not resist the tempta-
tion to give an account of how deductive-nomological explanations produce
a feeling of understanding (namely, as mentioned above, by showing us that
the explained event was to be expected). According to Trout, we should not
try to justify our theories of explanation by referring to alleged understanding-
providing features. The fact that a particular explanation of a phenomenon
under standing and scientific e xpl anation 

gives us a feeling (sense) of understanding is not a reliable cue to it being a cor-

rect (accurate) explanation. Even worse, the sense of understanding will lead
us to stop looking for better explanations, while it is typically a product of the
above-mentioned biases (Trout ). Thus, Trout conceives of understand-
ing as the phenomenology of explanation, and concludes that it should not be
regarded as epistemically relevant.4
Hempel and Trout favor an objectivist view of scientific understanding. On
this view, scientific explanations provide understanding only in the sense that
they show how the phenomena “fit into a nomic nexus” (Hempel , ).
According to Trout (, –), “scientific understanding is the state pro-
duced, and only produced, by grasping a true explanation,” where “one might,
for example, treat grasping as a kind of knowing.” But there is more to under-
standing than is admitted in Hempel and Trout’s objectivist view. A nonobjec-
tive kind of understanding does exist that has a crucial epistemic function and
is accordingly philosophically relevant.

I]Z:e^hiZb^XGZaZkVcXZd[H`^aah

I do agree with Trout that the phenomenology of understanding has no

epistemic function: the experience of a feeling of understanding (an “aha” ex-
perience) is neither necessary nor sufficient for scientific understanding of a
phenomenon. Furthermore, I agree with the basic Hempelian idea that explana-
tions are arguments that attempt to fit a phenomenon into a broader theoretical
framework. However, I will argue that the strict deductive-nomological model
of explanation fails as an account of the practice of scientific explanation, and
that it should be replaced by Nancy Cartwright’s simulacrum account. Actual
(real) scientific explanation involves a kind of understanding that is pragmatic
and thereby nonobjective. This type of understanding is associated with skills
and judgments of scientists and cannot be captured in objective algorithmic
procedures. It is therefore incompatible with the objectivist conception of ex-
planation and understanding favored by Hempel and Trout.
The pragmatic kind of understanding that I claim is crucial for scientific
explanation is not a product of explanation. There is a clear distinction between
three different ways in which the term “understanding” is used in connection
with scientific explanation:

;ZZa^c\d[jcYZghiVcY^c\;J I]Ze]ZcdbZcdad\n!ÆV]VÇ
   ZmeZg^ZcXZVXXdbeVcn^c\VcZmeaVcVi^dc
JcYZghiVcY^c\Vi]ZdgnJI 7Z^c\VWaZidjhZi]Zi]Zdgn
JcYZghiVcY^c\Ve]ZcdbZcdcJE =Vk^c\VcVYZfjViZZmeaVcVi^dc
   d[i]Ze]ZcdbZcdc
 henk w. de r egt

Understanding a phenomenon (UP) is generally regarded as an essential epis-

temic aim of science. It is not at odds with the Hempelian view; as we have seen
above, it corresponds to what Hempel calls scientific understanding. UP can
be accompanied by feeling of understanding (FU): explanations may produce
a sense of understanding, but that is not necessarily the case. Accordingly, as
Trout has rightly claimed, it is epistemically irrelevant whether or not UP is ac-
companied by FU. So far I agree with Hempel and Trout. However, contra their
objectivist view, I will argue that UP necessarily requires UT, and that since
UT is necessarily nonobjective, UP cannot be fully objective either. UT is prag-
matic: it concerns the understanding of the theory that is used in the explana-
tion. Below, I will analyze this with the help of the notion of “intelligibility.” The
main goal of this chapter is to argue that in order to achieve understanding of
phenomena by constructing explanations, one needs a nonobjective, pragmatic
kind of understanding. And the pragmatic condition UT is crucial for reaching
the epistemic aim UP.
Explanation is a central epistemic aim of science, but what is it? Explanations
are arguments that fit a phenomenon into a broader theoretical framework.
According to Hempel, they do so by deducing the explanandum from cover-
ing laws plus boundary conditions. For example, one can explain the fact that
jets fly by deducing it from Bernoulli’s principle and the relevant background
conditions (borrowing an example from Trout , ). Note, however, that
merely knowing Bernoulli’s principle and the background conditions does not
suffice for explanation: in addition, one should be able to use this knowledge in
the right way to derive the explanandum. Thus, a student may have memorized
Bernoulli’s principle and have all background conditions available but may still
be unable to use this knowledge to account for the fact that jets can fly—in that
case the student does not possess an explanation and accordingly no scientific
understanding of the phenomenon. The extra ingredient needed to construct
the explanation is a skill: the ability to construct deductive arguments from the
available knowledge.
Harold Brown shows this convincingly for the comparable case of deduc-
tive proofs in formal logic. Brown (, ; compare , –) argues
that although each step in such a proof is justified by appeal to an explicit rule,
“the process of constructing a proof—that is, the process of deciding which
rule to apply at each stage—is not governed by any comparable set of rules.”
One learns to construct proofs by developing a skill, by practicing while guided
by someone who already possesses the skill. In other words, one develops the
ability to make judgments about how to proceed without an algorithm.5 Brown
(, –) adds that this also holds for the case of checking proofs, which
shows that judgment does not merely belong to the context of discovery but to
the context of justification as well: “explicit following of rules is characteristic
under standing and scientific e xpl anation 

of an unskilled, rather than of a skilled, performance.” While the novice con-

sciously follows rules, the expert immediately recognizes which steps are valid
and which ones are not: this “leaps to the eye.”
Thus, deductive reasoning—and accordingly deductive-nomological expla-
nation—involves skills and judgment. This fact has two important implications.
First, a skill cannot be acquired from textbooks but only in practice (compare
Leonelli, this volume) because skills cannot be translated in explicit sets of
rules (Brown , –). Accordingly, to possess a skill is to have implicit
knowledge or “tacit knowledge” (Polanyi /). It has been argued, most
notably by Reber (), that tacit knowledge involves implicit learning, that
is, the unconscious and unintentional adoption of rule-following procedures.
However, Reber’s conclusion is not uncontroversial. According to Shanks (,
), “it has yet to be proved beyond doubt that there exists a form of learning
that proceeds both unintentionally and unconsciously.” Although experiments
have suggested that people can achieve unconscious knowledge of rules, there
are alternative interpretations of these experiments, in which rule knowledge
plays no role.
Brown (, –) rejects the idea that developing skills is equivalent
to internalizing rules, because it would require the (for naturalistic philoso-
phers of science unacceptable) assumption that there exists an unconscious
mind that follows the rules. Brown argues that learning cognitive skills is com-
parable to learning physical skills and involves training the nervous system.
Trout (, –) also emphasizes the role of skills and implicit knowl-
edge. He uses it as an argument against the relevance of the psychological
sense of understanding (FU): a conscious “aha” experience cannot be required
for achieving understanding (UP), which can allegedly be implicit. Trout is, of
course, right that FU is neither necessary nor sufficient for UP, and that (im-
plicit, tacit) skills are crucially important for achieving UP. It does not imme-
diately follow, however, that the process of achieving UP can be captured in an
objectivist account based on internalization of rule-following procedures, as
Trout suggests. On the contrary, the arguments of Brown and Shanks render
this claim implausible.
Second, the fact that particular skills of the subject are crucial for construct-
ing and evaluating explanations and for achieving scientific understanding en-
tails that explanation has a pragmatic dimension that is epistemically relevant
(because it involves not only an explanans and the explanandum but also an
explainer). Many philosophers of science hold that the pragmatic and the epis-
temic dimension of science can and should be kept separate. Thus, Hempel
(, –) argues that since explanation in a pragmatic sense is a relative
notion, it cannot be relevant to the philosophical analysis of science, because
the latter should be objectivist:
 henk w. de r egt

For scientific research seeks to account for empirical phenomena by

means of laws and theories which are objective in the sense that their
empirical implications and their evidential support are independent
of what particular individuals happen to test or to apply them; and
the explanations, as well as the predictions, based upon such laws and
theories are meant to be objective in an analogous sense. This ideal in-
tent suggests the problem of constructing a nonpragmatic concept of
scientific explanation—a concept which is abstracted, as it were, from
the pragmatic one, and which does not require relativization with re-
spect to questioning individuals. ()

In a similar vein, Bas van Fraassen (, , ) explicitly contrasts an epis-
temic and a pragmatic dimension of theory acceptance. The epistemic dimen-
sion contains the relevant beliefs concerning the relation between a theory and
the world, while the pragmatic dimension contains reasons scientists may have
for accepting a theory independently of their beliefs about its relation to the
world; these reasons typically pertain to the use and the usefulness of the the-
ory (van Fraassen , ). Both Hempel and van Fraassen see the epistemic
and the pragmatic as sharply distinguished domains. The pragmatic dimension
pertains to the relation between the theory and its users, that is, to the dimen-
sion that seems to be excluded from the epistemic dimension by definition. The
thesis that epistemic and pragmatic can and should be kept separate presup-
poses that the epistemic status of a theory only and uniquely depends on a di-
rect evidential relation with the phenomena it purports to describe or explain.
On this presupposition, pragmatic elements—such as virtues of a theory that
facilitate its use by scientists—are indeed epistemically irrelevant: they do not
carry any additional justificatory weight.
However, as the previous discussion of the role of skills in deductive rea-
soning makes clear, this presupposition is false. Establishing relations between
theories and phenomena, as happens in deductive-nomological explanation,
crucially depends on skills and judgment. It follows that although it is surely
possible and useful to distinguish analytically between the epistemic and the
pragmatic, the two are inextricably intertwined in scientific practice: epistemic
activities and evaluations (production and assessment of knowledge claims) are
possible only if particular pragmatic conditions are fulfilled.
Thus, the pragmatic dimension of explanation is epistemically relevant.
This is the case for D-N type explanations, but even more so for scientific ex-
planations that do not conform to the ideal D-N scheme. As mentioned above,
Hempel’s deductive-nomological model of explanation fails as a general ac-
count of scientific practice: actual explanations often are not of the D-N type
because usually there is no strictly deductive relation between explanans (the-
under standing and scientific e xpl anation 

oretical laws and initial conditions) and explanandum (the phenomenon or

phenomenological laws). Instead, as Cartwright (), Morgan and Morrison
(), and others have shown, in scientific practice the connection between
theory and phenomena is usually made through models (compare Knuuttila
and Merz; Boumans, this volume).
A model is conventionally taken as a representation of the object or system
that one wants to understand scientifically.6 In order to see what the role of
models is in achieving understanding we should first clarify the relation be-
tween models, theory, and empirical data. Morgan and Morrison (, –)
have argued for the “autonomy” of models: they are not derived from theory,
but neither do they follow from the empirical data. They typically contain both
theoretical and empirical information. Consequently, it is difficult to draw a
sharp distinction between theory and model (compare Morrison ). For
example, the kinetic theory of gases represents real gases (its target systems)
as aggregates of particles that behave according to the laws of Newtonian me-
chanics. Thus, the theory already provides a general model of gases. In order
to explain particular gas phenomena on the basis of the kinetic theory, more
specific models (such as of particle structure) have to be constructed.7 Given
a particular phenomenon (described in terms of a phenomenological law or
empirical data), the specific model required for explaining this phenomenon
cannot be simply derived from the kinetic theory. In this sense the models are
“autonomous.”8
So how are such models constructed and what precisely is their role in sci-
entific explanation? The function of a model is to represent the phenomenon
(target system) in such a way that the theory can be applied to it. In other
words, models replace the bridge principles that traditionally connect theory
to empirical phenomena. In the terminology of Morgan and Morrison (),
models “mediate” between theory and phenomena. The crucial difference be-
tween bridge principles and mediating models is that the former establish a
strict deductive relation between theory and phenomenon (explanans and ex-
planandum), while the latter connect the two in a looser way. The construction
of models and their role in scientific explanation has been analyzed extensively
by Nancy Cartwright (). She defends the simulacrum account of expla-
nation, which asserts that “to explain a phenomenon is to construct a model
that fits the phenomenon into a theory” (Cartwright , ). In the modeling
stage, the target system is presented in such a way that the theory can be ap-
plied to it: we decide to describe system S as if it is an M (where M is a model of
which the behavior is governed by the equations of the theory). The construc-
tion of models is not a matter of deduction but a complex process involving
approximation and idealization (compare Morrison, this volume). There are no
algorithms or formal principles that tell us how to get from the description of a
 henk w. de r egt

real system to a suitable model. But once we have constructed a model to which
the theory can be applied, equations for the behavior of the model system can
be deduced: “There are just rules of thumb, good sense, and, ultimately, the re-
quirement that the equation we end up with must do the job” (Cartwright ,
). Thus, the scientist must rely on skills and practical judgment, as has been
observed already by Hilary Putnam:

What the theory [in physics] actually describes is typically an idealized

“closed system.” The theory of this “closed system” can be as precise
as you want. And it is within this idealization that one gets the famil-
iar examples of the “scientific method.” But the application of physics
depends on the fact that we can produce in the laboratory, or find
in the world, open systems which approximate to the idealized sys-
tem sufficiently well to yield very accurate predictions. The decision
that conditions have been approximated well in a given case—that it
is even worthwhile to apply the idealized model to this case—typically
depends on unformalized practical knowledge. (Putnam , )

Putnam (, ) explicitly identifies such unformalized practical knowl-

edge with skills. Constructing a model, and accordingly constructing an ex-
planation, requires skill and judgment (and this extends to the context of
justification, as argued above). Understanding in the sense of UP (having an
explanation of the phenomenon) is an epistemic aim of science, but this aim
can only be achieved by means of pragmatic understanding UT (the ability to
use the relevant theory). This is the case for D-N explanations, and even more
so for model-based explanations that are typically not deductive. Achieving the
epistemic aim of science is a complex process that unavoidably has a pragmatic
dimension in which skills and judgment play crucial roles. Consequently, the
epistemic value of a theory cannot be determined in isolation from its use, and
successful use of a theory requires pragmatic understanding. In sum, pragmatic
understanding is epistemically relevant.

I]Z>ciZaa^\^W^a^ind[I]Zdg^Zh

I have argued that pragmatic understanding is a necessary condition for

scientific understanding. Pragmatic understanding UT is defined, somewhat
loosely, as the ability to use the relevant theory and as being based on skills
and judgment. In this section I will give a more precise characterization of the
nature of and conditions for pragmatic understanding. Pragmatic understand-
ing is required for constructing models that relate theories to phenomena. This
under standing and scientific e xpl anation 

activity involves making suitable idealizations and approximations: scientists

need to make the right judgments regarding idealization and approximation,
and possess the right skills to build a model on this basis.
But what are the right skills? This depends on which theory the scientist is
dealing with, in particular on which pragmatic virtues the theory possesses.
Particular virtues of theories, for example, visualizability or simplicity, may be
valued by scientists because they facilitate the use of the theory in constructing
models; in this sense they are pragmatic virtues. But not all scientists value the
same qualities: their preferences are related to their skills, acquired by train-
ing and experience, and are related to other contextual factors such as their
background knowledge, metaphysical commitments, and the virtues of already
entrenched theories.
Philosophers of science have listed and discussed the theoretical virtues
that generally play a role in the evaluation of scientific theories—for example,
accuracy, consistency, scope, simplicity, unifying power, and fertility.9 Many
historical studies have revealed that, in evaluations of theories, scientists em-
ploy other criteria than accuracy alone, and there is sometimes a trade-off
between accuracy and other virtues.10 Such evaluations cannot be the result
of algorithmic decision procedures, and accordingly judgment plays a crucial
role. In addition to the general virtues cited above, scientists in various fields
and historical periods have promoted other, more specific theoretical qualities;
for example, visualizability, causality, continuity, and locality.11
Scientists may prefer theories with particular pragmatic virtues because
they possess the skills to use such theories, that is, to construct models for
explaining phenomena by means of the theories. In other words, they have
pragmatic understanding (UT) of such theories. I would like to rephrase this
using the notion of intelligibility. If scientists understand a theory, the theory
is intelligible to them. I define the intelligibility of a theory (for a scientist) as
follows:

Intelligibility—The value that scientists attribute to the cluster of

virtues (of a theory in one or more of its representations) that fa-
cilitate the use of the theory for the construction of models.

It is important to note that intelligibility, thus defined, is not an intrinsic prop-

erty of a theory but a value that is projected onto the theory by a scientist
or group of scientists. Moreover, it is a pragmatic, context-dependent value
related both to properties (virtues) of a theory and its representation(s) and to
scientists’ skills.12 Phrased differently, one might say that intelligibility is a mea-
sure of the fruitfulness of a theory for constructing models. It is not an objec-
 henk w. de r egt

tive measure, however, because a theory can be more fruitful for one scientist
than for another.13 Which theories are deemed intelligible can vary through
time or across disciplines. Scientists prefer a more intelligible theory over a less
intelligible one, not because it gives them a psychological sense of understand-
ing, but rather because they have to be able to use the theory.
The arguments presented in the section “The Epistemic Relevance of Skills,”
above, can be summarized by the thesis that scientists need intelligible theories
in order to achieve scientific understanding of phenomena. Only an intelligible
theory allows scientists to construct models through which they can derive ex-
planations of phenomena on the basis of the given theory. In other words, un-
derstanding scientific theories is a prerequisite for understanding phenomena
scientifically. This can be stated as a criterion for understanding phenomena
(CUP):

CUP—A phenomenon P is understood scientifically if a theory T

of P exists that is intelligible (and the explanation of P by T meets
accepted logical and empirical requirements).14

First, note that understanding a phenomenon (UP) was defined as “having an

adequate explanation of it,” and explanations were defined as arguments that
fit a phenomenon into a broader theoretical framework. CUP states that hav-
ing an intelligible theory T is a condition for scientifically understanding P. Of
course, in addition we need to use T to actually construct the explanation that
fits P in the theoretical framework. Therefore, CUP also contains a requirement
about the actual explanation, in which accepted logical and empirical require-
ments determine what a good fit is.15 Accordingly, CUP does not entail that, for
example, astrologers understand personality traits of people if only they have
an intelligible astrological theory of these personality traits. In addition, expla-
nations on the basis of this theory have to meet accepted logical and empirical
requirements. Second, note that CUP is implicitly pragmatic (that is, it refers
to the scientists involved in the process of explanation) because it involves the
notion of intelligibility, which is pragmatic for reasons explained above.
Since intelligibility is not an intrinsic property of theories but a context-de-
pendent value ascribed to theories, one cannot specify universally valid criteria
for intelligibility. But this does not entail that intelligibility is merely a matter
of taste and that no criteria can be formulated at all. Since the intelligibility of
a theory can be regarded as a measure of how fruitful it is for the construction
of models (by scientists in a particular context), it is possible to determine this
measure by means of a criterion or test. There may be different ways to test
whether a theory is intelligible for scientists, and not all of them may be appli-
cable in all cases or for all disciplines. As one way of testing the intelligibility of
theories, I want to suggest the following criterion:
under standing and scientific e xpl anation 

CIT—A scientific theory T (in one or more of its representations)

is intelligible for scientists (in context C) if they can recognize
qualitatively characteristic consequences of T without perform-
ing exact calculations.

I argued above that skills and judgments involved in achieving understanding

cannot be reduced to rule-following procedures. Criterion CIT, which associ-
ates understanding with intuitively recognizing consequences of a theory, can
be regarded as a test of such tacit (implicit) skills. The criterion applies specifi-
cally to mathematical theories as used in the physical sciences. It is based on
a suggestion by Werner Heisenberg about understanding in physics (see de
Regt and Dieks  for a more elaborate defense). For theories in other disci-
plines, especially nonmathematical ones, other criteria may apply. Moreover,
there may be other ways to determine intelligibility of theories in the physical
sciences as well. CIT is a sufficient condition for intelligibility, not a necessary
one.
Note that CIT captures the pragmatic and contextual nature of intelligibil-
ity (and accordingly of understanding UT). CIT explicitly refers to the scien-
tists who use the theory for constructing models, and to the particular context
C in which they operate. Whether theory T is intelligible depends not only on
the virtues of T itself but also on such contextual factors as the capacities, back-
ground knowledge, and background beliefs of the scientists in C. Accordingly,
it can accommodate the variety of ways in which understanding is achieved in
scientific practice. Qualitative insight into the consequences of a theory can be
gained in many ways and therefore the criterion does not favor one particular
skill or theoretical virtue. Its fulfillment requires an appropriate combination of
skills and virtues. The preferred theoretical virtues, to which skills have to be
attuned, provide tools for achieving understanding. For example, a theory (or a
particular representation of it) may possess the virtue of visualizability, which
allows for visualization, which in turn may be a useful tool to construct mod-
els of the phenomena. Visualization is an important tool for understanding,
valued by many scientists, but it is not a necessary condition and there may be
scientists who prefer abstract theories over visualizable ones (see de Regt 
for examples).

>ciZaa^\^W^a^inVcYi]Z@^cZi^XI]Zdgnd[<VhZh

A case study from the history of physics can illustrate the above account of
scientific understanding. The case study concerns nineteenth-century attempts
to explain experimentally observed properties of gases on the basis of the ki-
netic theory of gases. The kinetic theory, developed in the s, was based on
 henk w. de r egt

the idea that heat is mechanical motion, and that gases consist of molecules
in motion behaving in accordance with the laws of Newtonian mechanics. In
order to explain and predict experimental facts, one needed specific models
of molecules. Among the phenomena to be explained were the well-known
experimental gas laws (Boyle-Charles, Gay-Lussac), transport phenomena (vis-
cosity, heat conduction, and diffusion), temperature independence of specific
heats, and ratios of specific heats at constant volume and pressure. Later, in the
s, still more experimental knowledge was accumulated, most importantly
regarding spectral emission of gases.
The first mature versions of the kinetic theory were published by Rudolf
Clausius (/) and James Clerk Maxwell (/). In both cases the
core of the theory consisted in an application of the laws of classical mechan-
ics to aggregates of particles (molecules), where the description of the system
did not consist in tracing the motion of each particle individually but in a sta-
tistical treatment. The elementary model used by both Clausius and Maxwell
was that of the molecule as a hard elastic sphere.16 Having modeled molecules
in this way, kinetic theory supplies equations for the behavior of the system.
One of these equations concerns the specific heat ratio γ, that is, the ratio be-
tween the specific heat at constant pressure and the specific heat at constant
volume. The observed variation in specific heat ratios for different gases can
be explained by modeling these gases as polyatomic molecules with different
energies of translation and/or rotation (see de Regt ). In , Maxwell
based his theoretical predictions on his equipartition theorem, which asserts
that there is a constant equal distribution of kinetic energy over the various
possible (translational, rotational, and/or vibrational) “modes of motion” of the
molecule. However, Maxwell discovered that for many common gases (oxygen,
nitrogen) the equation led to a discrepancy between theoretical predictions
and experimental results for specific heat ratios: theoretically, the ratio of total
kinetic energy to energy of translation was β = , while the experimental value γ
= . implied that β = . (Maxwell /, ). This problem became
known as the “specific heat anomaly.”
Some years later, in , Maxwell introduced the idea of “degree of free-
dom,” which provided a new way of modeling molecules.17 Maxwell assumed
that “the position and configuration of the molecule can be completely ex-
pressed by a certain number of variables,” and added: “Let us call this number
n. Of these variables, three are required to determine the position of the centre
of mass of the molecule, and the remaining n– to determine its configuration
relative to its centre of mass. To each of the variables corresponds a different
kind of motion” (Maxwell /, ). Subsequently, Maxwell () pre-
sented a new formula for the specific heat ratio:
under standing and scientific e xpl anation 

+n+e
γ=
n+e (1)

(where n is number of degrees of freedom, and e is a quantity depending on the

binding forces in the molecule; e accounts for the relative amount of heat that
is stored as potential energy in the molecule). Comparison with experimental
values of specific heats led Maxwell to conclude that n + e “for air and several
other gases cannot be more than .,” since the experiments have determined
that γ = .. Furthermore, these gases cannot consist of monatomic mol-
ecules: such molecules have no internal structure and therefore only three de-
grees of freedom (n = ), which yields γ = . +. Moreover, around  it was
well known that most gases emit spectral lines, the best explanation of which
seemed to lie in internal vibration of the molecules. These additional degrees
of freedom entail that n ≥  and thereby γ ≤ .. The observed ratio of γ =
. was thus unexplainable. Maxwell () concluded that this anomaly is
“the greatest difficulty yet encountered by the molecular theory.”
One year later, Ludwig Boltzmann proposed a solution of the problem.
Boltzmann (/, ) used the notion of “degrees of freedom” and
proposed to model the molecules as systems with five degrees of freedom.
He assumed that the anomalous gases were diatomic molecules that could be
represented as two rigidly connected elastic spheres. Such “rigid dumbbells”
possess only five degrees of freedom (three of translation and two of rotation),
hence n = . Moreover, since the molecules are absolutely rigid, they cannot
store potential energy, hence e = . Consequently, γ = ., almost precisely in
accordance with the measured value.

;^\jgZ'#&#7daiobVccÉhYjbWWZaa
  bdYZad[Y^Vidb^XbdaZXjaZh#

This simple model thus provides an explanation, and thereby scientific un-
derstanding, of the anomalous specific heat ratios. However, as Boltzmann
(/, ) admits, the dumbbell model cannot be a completely realistic
description of diatomic molecules because real gas molecules do have vibra-
tional degrees of freedom, as is already shown by the fact that they are capable
of emitting spectral lines.
Boltzmann’s dumbbell model nicely illustrates that model building is a
 henk w. de r egt

nondeductive process. The model does not follow deductively from the kinetic
theory, and it cannot be derived from experimental data either: approxima-
tions and idealization are necessary in order to achieve a fit between model
and data:

t
"QQSPYJNBUJPO
UIF
EBUB
EJE
OPU
QSPWJEF
UIF
FYBDU
γ = . value, from

which n + e =  could be deduced. Maxwell (/, /) calcu-
lated with the experimental value of γ = ., which entailed n + e = ..
Clausius (/) calculated with γ = ., which would have led to
n + e = ..
t
*EFBMJ[BUJPO
e = : the dumbbell model represents molecules as absolutely
rigid, which implies that heat cannot be stored as potential energy.
t
*EFBMJ[BUJPO
JOUFSOBM
WJCSBUJPOBM
EFHSFFT
PG
GSFFEPN
EP
OPU
DPOUSJCVUF
UP

specific heats.

Decisions to approximate and idealize in this way belong to the pragmatic

dimension of explanation. Boltzmann was the first who was willing and able
to make these pragmatic decisions and to construct the model. Why was
Boltzmann’s solution not proposed earlier? Clausius and the early Maxwell
possessed all relevant theoretical knowledge and experimental data, so why
did they not consider the possibility that the molecules of the anomalous gases
were rigid dumbbells? I submit that this was because the concept of “degrees of
freedom,” which was introduced only in , functioned as a crucial pragmat-
ic tool for constructing this model. Representing molecules in terms of degrees
of freedom allowed scientists to construct different kinds of molecular models
and to “see” new explanations and predictions on the basis of the kinetic theo-
ry. To be sure, the kinetic theory itself remained the same, and one might object
that the notion of degrees of freedom merely offered a new way of represent-
ing extant knowledge and did not add new knowledge. But it is precisely this
view that I want to challenge. The new tool facilitated the construction of new
models, mediating between the theory and the phenomena, and was accord-
ingly crucial to the production of new explanatory and descriptive knowledge.
It enhanced the intelligibility of the kinetic theory and paved the way, via equa-
tion (), for the construction of the dumbbell model.
Finally, even with the notion of degrees of freedom in hand, the dumb-
bell model (n =  solution) does not follow logically from theory and data: a
skillful scientist was needed to see the solution by making the necessary ap-
proximations and idealizations, and that scientist was Ludwig Boltzmann: for
him, the kinetic theory was sufficiently intelligible. (Today, all physics students
acquire these skills in the course of their education, as the dumbbell model
under standing and scientific e xpl anation 

has become an exemplary problem solution that is presented in undergraduate

textbooks).18

8dcXajh^dc

Understanding of phenomena (UP) is among the epistemic aims of science,

and such understanding is provided by explanations. This is generally acknowl-
edged, and granted by objectivists such as Hempel (as long as understanding is
equated with explanation). While I agree with the Hempelian view that expla-
nations give us scientific understanding UP, in this chapter I have argued, con-
tra Hempel’s objectivism, that there is a pragmatic type of understanding (UT)
that is a necessary condition for achieving UP. This pragmatic understanding
amounts to the ability of scientists to use relevant theories to construct expla-
nations. I have presented a further analysis of UT based on the notion of intel-
ligibility, and provided criteria for understanding phenomena (CUP) and for
(pragmatic) understanding of theories UT (the criterion for intelligibility, CIT).
Scientific understanding of phenomena (UP) requires intelligible theories,
where intelligibility is defined as the positive value that scientists attribute to
the theoretical virtues that facilitate the construction of models of the phenom-
ena. Intelligibility is not an intrinsic property of a theory but a value projected
onto the theory by scientists. It is a pragmatic, context-dependent value related
both to the theoretical virtues and to the scientists’ skills. The conclusion that
intelligibility has epistemic relevance is in direct opposition to the objectivist
view of Hempel and Trout: it implies that of two people who possess exactly the
same theories and background knowledge, one may achieve understanding of a
phenomenon while the other does not. Accordingly, scientific understanding is
not completely objective, that is, it is not fully independent of the subject.
Some may object that this implies that scientific understanding, and there-
by science itself, becomes a subjective and individual affair, but that worry is
unfounded. First of all, it should be noted that, as Kuhn () makes clear,
skills are acquired within a community and assessed by that community. This
guarantees that judgments based on skills are not arbitrary. Accordingly, reli-
ance on judgment does not entail relativism. According to Harold Brown, “the
accomplishments of science depend on the ability to exercise judgment, an
ability that scientists develop as they learn and practice their craft. Scientists
develop judgment in the specific fields that they have mastered, and this results
in the existence of a set of cognitive skills in the scientific community. These
skills are fallible, but their fallibility does not make them epistemically worth-
less” (Brown , –).
The thesis that pragmatic understanding of theories (UT, intelligibility) is a
 henk w. de r egt

necessary condition for understanding phenomena (UP) undermines tradition-

al objectivist conceptions of science and scientific explanation. But abandoning
such conceptions does not necessarily open the door to relativism: intelligibil-
ity is necessary but not sufficient for scientific understanding. Criterion CUP
includes accepted logical and empirical requirements to which explanations
have to conform. These requirements can be listed as a set of theoretical virtues
that generally play a role in the evaluation of scientific theories, which together
provide a shared basis for assessing explanations (compare Kuhn , ).
Intelligibility is a criterion for selecting theories that can be fruitfully employed
in the construction of explanatory models. Obviously, the resulting explana-
tions are scientific explanations if and only if they conform to the shared virtues
to a sufficient extent. Intelligibility is not the only requirement for scientific
theories, but it is an essential one if one wants to understand the phenomena.

CdiZh

Thanks to Kai Eigner, Frank Hindriks, Tarja Knuuttila, Sabina Leonelli, and Hans
Radder for helpful discussion.
. “Studies in the Logic of Scientific Explanation” was reprinted with a new
postscript in Hempel ().
. For example, in  physicists in the Copenhagen-Göttingen circle believed that
atomic phenomena could be understood on the basis of matrix mechanics, while most
other physicists—notably Schrödinger—disagreed. Such differences may be traced back
to different scientific, philosophical, or social backgrounds (see de Regt , ).
. See especially his discussions of explanation in the social and historical sciences,
where he contrasts it with empathic understanding. Hempel deems the latter neither
necessary nor sufficient for scientific understanding (see Hempel , –, –).
. In response to my discussion of his  essay (de Regt ), Trout (, )
claims that he does not want to eliminate talk of understanding completely: it is only
allusions to the (psychological) sense of understanding that he criticizes. He allows
for “understanding proper” in contrast to “the mere sense of understanding,” and for
“genuine understanding” in contrast to “counterfeit understanding.” However, he does
not explain what this purely objectivist conception of understanding consists in. He
refers to “objectivist accounts of implicit understanding” () and associates this with
Kitcher’s internalization of argument patterns, but does not give a positive account of
genuine understanding ().
. Compare Koster, this volume, for an in-depth analysis of the role of judgment in
the field of history.
. Note, however, that there are widely different accounts of representation, and that
the target systems of models may be real, fictional, or ideal (Knuuttila ).
under standing and scientific e xpl anation 

. One might argue that the kinetic theory consists only of the laws of Newtonian
mechanics (and some additional statistical laws and assumptions), and accordingly does
not contain a model of gases. But then it is hard to understand the kinetic theory as a
distinct theory and as a theory of gases—it seems that the particle model of gases forms
an essential part of the kinetic theory.
. Strictly speaking, the term “autonomous” is too strong: the model is not fully
independent of the theory.
. See Kuhn (, ); McMullin (, –); Longino (, ); and Lacey
(). These theoretical virtues are called “values” by Kuhn (, ), “epistemic
values” by McMullin () and “constitutive values” by Longino (). I use the term
“virtues” in order to avoid confusion: the virtue is a property of the theory, which may
be valued by scientists. Furthermore, it is difficult to draw a sharp distinction between
epistemic and pragmatic values/virtues (McMullin), or between constitutive and
contextual values/virtues (Longino). For my purposes, this is not necessary because my
account of scientific understanding implies that there is interaction between pragmatic
and epistemic dimensions of science.
. The production of accurate theories is more often stimulated than hindered by
reliance on other (allegedly nonepistemic) theoretical virtues. For example, McAllister
() argues, on the basis of historical evidence, that scientists usually rely on
established aesthetic canons (containing accepted aesthetic criteria for theory choice),
and that this conservatism is conducive to empirical progress.
. See de Regt () for examples from modern physics.
. It often happens that a theory can be represented in different ways (think of the
various formulations of classical mechanics), and each of these representations may have
its own specific virtue, which may be relevant to the intelligibility of the theory.
. There may be a discrepancy between scientists’ perception of the fruitfulness
of a theory and its actual fruitfulness for these scientists. Such errors of judgment will
hamper their work. I assume that they are the exception rather than the rule.
. This formulation differs slightly from that given in de Regt and Dieks ().
. To be sure, these requirements are included in the list of (scientific, epistemic,
constitutive) virtues that are subject to valuation by scientists. Accordingly, there may be
some variation in the way they are valued and applied in specific cases. However, they
are generally accepted among scientists, and explanations that completely fail to meet
them will be rejected as unscientific.
. As mentioned above, the boundary between theory and model is vague and
arbitrary to some extent. In the case of the kinetic theory there is good reason for
drawing the distinction as follows: the representation of gases as aggregates of molecules
in motion is part of the theory, while the representation of molecules as hard elastic
spheres is part of the modeling stage. This is because all kinetic theories share the former
 henk w. de r egt

model of a gas, while the latter model of a molecule is not an essential part of the theory
(kinetic theory can operate on different molecular models).
. Maxwell introduced the concept but did not yet use the phrase “degrees of
freedom,” which was introduced by H. W. Watson (/, vi).
. One might ask why it was Boltzmann and not Maxwell who proposed the
dumbbell model. Maxwell had introduced the concept of degrees of freedom, and
surely had the skills to construct the model. When Boltzmann advanced his dumbbell
model, Maxwell rejected it because he objected to the idealizations regarding rigidity
and vibrational degrees of freedom. The difference between Maxwell’s and Boltzmann’s
attitudes toward these idealizations can be traced back to differing philosophical
commitments (see de Regt ).

GZ[ZgZcXZh

Boltzmann, L. /. Über die Natur der Gasmoleküle. In Wissenschaftliche

Abhandlungen, :–. Chelsea: New York.
Brown, H. I. . Rationality. London: Routledge.
———. . Judgment, role in science. In A companion to the philosophy of science,
edited by W. H. Newton-Smith, –. Oxford: Blackwell.
Cartwright, N. . How the laws of physics lie. Oxford, NY: Oxford University Press.
Clausius, R. /. Über die Art der Beweging welche wir Wärme nennen. English
translation reprinted in Kinetic theory, vol. , The nature of gases and heat, edited by
S. G. Brush, –. Oxford: Pergamon Press.
de Regt, H. W. . Philosophy and the kinetic theory of gases. British Journal for the
Philosophy of Science :–.
———. . Erwin Schrödinger, Anschaulichkeit, and quantum theory. Studies in
History and Philosophy of Modern Physics B:–.
———. . Spacetime visualisation and the intelligibility of physical theories. Studies in
History and Philosophy of Modern Physics B:–.
———. . Discussion note: Making sense of understanding. Philosophy of Science
:–.
de Regt, H. W., and D. Dieks. . A contextual approach to scientific understanding.
Synthese :–.
Friedman, M. . Explanation and scientific understanding. Journal of Philosophy
:–.
Hempel, C. G. . Aspects of scientific explanation and other essays in the philosophy
of science. New York: Free Press.
Kitcher, P. . Explanatory unification and the causal structure of the world. In
Scientific explanation, edited by P. Kitcher and W. C. Salmon, –. Minneapolis:
University of Minnesota Press.
under standing and scientific e xpl anation 

Knuuttila, T. . Models, representation, and mediation. Philosophy of Science

:–.
Kuhn, T. S. . The structure of scientific revolutions. Chicago: University of Chicago
Press.
———. . Objectivity, value judgment, and theory choice. In The essential tension:
Selected studies in scientific tradition and change, –. Chicago: University of
Chicago Press.
Lacey, H. . Is science value free? Values and scientific understanding. nd ed.
London: Routledge.
Longino, H. . Science as social knowledge. Princeton: Princeton University Press.
Mach, E. /. The guiding principles of my scientific theory of knowledge and its
reception by my contemporaries. In Physical reality, edited by S. Toulmin, –.
New York: Harper & Row.
Maxwell, J. C. /. Illustrations of the dynamical theory of gases. In Maxwell on
molecules and gases, edited by E. Garber, S. G. Brush, and C. W. F. Everitt, –.
Cambridge, MA: MIT Press.
———. /. On the dynamical evidence of the molecular constitution of bodies.
In Maxwell on molecules and gases, edited by E. Garber, S. G. Brush, and C. W. F.
Everitt, –. Cambridge, MA: MIT Press.
McAllister, J. W. . Beauty and revolution in science. Ithaca: Cornell University Press.
McMullin, E. . Values in science. In PSA , vol. , edited by P. D. Asquith and
T. Nickles, –. East Lansing, MI: Philosophy of Science Association.
Morgan, M. S., and M. Morrison, eds. . Models as mediators: Perspectives on
natural and social science. Cambridge: Cambridge University Press.
Morrison, M. . Where have all the theories gone? Philosophy of Science :–.
Pearson, K. /. The grammar of science. Reprint of rd ed. New York: Meridian
Books.
Polanyi, M. /. Personal knowledge: Towards a post-critical philosophy. London:
Routledge and Kegan Paul.
Putnam, H. . Meaning and the moral sciences. London: Routledge and Kegan Paul.
Reber, A. S. . Implicit learning and tacit knowledge: An essay on the cognitive
unconscious. Oxford: Oxford University Press.
Salmon, W. C. . Scientific explanation and the causal structure of the world.
Princeton: Princeton University Press.
———. . Causality and explanation. Oxford: Oxford University Press.
Schrödinger, E. . Nature and the Greeks. Cambridge: Cambridge University Press.
Shanks, D. R. . Implicit learning. In Handbook of cognition, edited by K. Lamberts
and R. Goldstone, –. London: Sage.
Trout, J. D. . Scientific explanation and the sense of understanding. Philosophy of
Science :–.
 henk w. de r egt

———. . Paying the price for a theory of explanation: De Regt’s discussion of Trout.
Philosophy of Science :–.
———. . The psychology of scientific explanation. Philosophy Compass :–.
van Fraassen, B. C. . The scientific image. Oxford: Clarendon Press.
Watson, H. W. /. A treatise on the kinetic theory of gases. nd ed. Oxford, NY:
Clarendon Press.
chap ter title ver so 

*
Understanding without Explanation

E:I:GA>EIDC

 Explaining why and understanding why are closely connected. In-

deed, it is tempting to identify understanding with having an explanation. Ex-
planations are answers to why questions, and understanding, it seems, is simply
having those answers. Equating understanding with explanation is also attrac-
tive from an analytic point of view, since an explanation is understanding in-
carnate. The explanation is propositional and explicit. It is also conveniently
argument shaped, if we take the premise to be the explanation proper and the
conclusion a description of the phenomenon that is being explained. So we are
on the way to specifying the logic of understanding.
Specifying the structure of explanation would not, however, tell us every-
thing we want to know. Why should things of that shape be identified with
understanding? To understand why a phenomenon occurs is a cognitive
achievement, and it is a greater cognitive achievement than simply knowing
that the phenomenon occurs—we all know that the sky is blue, but most of
us do not understand why. What, then, are the extra cognitive benefits that
a good explanation provides? How does having an explanation take us from
merely knowing that something is the case to understanding why it is that way?
From this point of view, it is more natural to identify understanding with the
cognitive benefits that an explanation provides rather than with the explana-
tion itself. These benefits are various and include conspicuously four kinds of
knowledge: of causes, of necessity, of possibility and of unification. Explana-
tions may give information about the causes of the phenomenon in question,


 peter lip ton

they may show that the phenomenon had to occur, they may show how the
phenomenon could have occurred, and they may show how the phenomenon
fits into a broader pattern.
The switch from identifying understanding with explanation to identifying
it with some of the cognitive benefits of an explanation may not seem dramatic,
but it makes this essay possible. For by distinguishing explanations from the
understanding they provide, we make room for the possibility that understand-
ing may also arise in other ways. It is that space that I will begin to survey here:
the possibilities for understanding without explanation. I will look briefly in
turn at the four cognitive benefits of explanation—causation, necessity, pos-
sibility, unification—and suggest how each of them might also be acquired by
routes that do not pass through explanations. If these suggestions are along the
right lines, then understanding is more extensive and more varied in its sources
than those who would equate explanation and understanding allow. This is of
considerable intrinsic interest, and it may also have repercussions for other is-
sues in the philosophy of science, for example, through the connection between
understanding and inference. I will go on to consider a fifth and rather different
consequence that consumers of good explanations enjoy. This is the distinctive
and satisfying feeling of understanding, the “aha” feeling. I will suggest that
although the “aha” feeling is not itself a kind of understanding, it may play an
important role in the way understanding is acquired.

8VjhVi^dc

Explanations often bridge the gap between knowing that a phenomenon

occurs and understanding why it does by providing knowledge about the phe-
nomenon’s causal history (Lewis ). Causal information is the conspicuous
cognitive benefit of many explanations: it is the form of understanding that
many explanations provide. Our question, then, is whether this benefit also
has other sources. The answer seems clearly yes, since causal information does
not come exclusively from explanation. Indeed, much of empirical inquiry
consists in activities—physical and intellectual—that generate causal informa-
tion, activities such as observation, experimentation, manipulation, and infer-
ence. And these activities are distinct from the activity of giving and receiving
explanations.
But showing that there are sources of causal information distinct from
causal explanations is not yet to show that there are routes to understanding
that do not pass through explanations. For it may be that although the route
to understanding begins with something other than an explanation, it passes
through an explanation along the way. We need cases that, in addition to not
under standing withou t e x pl anation 

being explanations themselves, do not work by means of generating explana-

tions that are then the proximate cause of the consequent understanding.
It is not immediately clear that the diverse sources of causal information
satisfy this condition. Thus, while an experiment is not itself an explanation,
the causal information it provides seems tantamount to one. The case is per-
haps even clearer if we focus on causal inference. Here we often begin with in-
formation about effects and infer what would explain them. Such an inference
to the best explanation begins from propositions that do not explain but only
describe the phenomenon under investigation and ends with an understand-
ing of that phenomenon (Lipton ). But this is not an interesting case of
the disassociation of understanding from explanation, since the understanding
arises by means of the explanation inferred from the description. The process
of acquiring understanding does not begin with an explanation, but the under-
standing is nevertheless a product of an explanation, which is not what we are
looking for here.
Explanation and causal information may thus seem too close to each other
to prise apart. But this line of thought only seems compelling if we focus ex-
clusively on causal information that takes the form of propositions explicitly
entertained. As soon as we consider the possibility that causal information may
be the content of tacit knowledge, the prospect for routes to understanding that
do not pass through explanations becomes bright, since causal explanation, I
suggest, requires that the information be given an explicit representation. In
short, there is such a thing as tacit understanding, but not tacit explanation,
and this provides the space we are looking for, where there can be understand-
ing without explanation.
One natural place to look for sources of causal information that may yield
only tacit knowledge is in the use of images and physical models. For example,
I never properly understood the why of retrograde motion until I saw it dem-
onstrated visually in a planetarium. A physical model such as an orrery may
do similar cognitive work. These visual devices may convey causal information
without recourse to an explanation. And people who gain understanding in this
way may not be left in a position to formulate an explanation that captures the
same information. Yet their understanding is real.
Manipulation provides an even more obvious and pervasive example of the
acquisition of causal information without the acquisition of an explanation.
Thus, a scientist may gain a sophisticated understanding of the behavior of a
complicated piece of machinery by becoming an expert at using it, and that
understanding consists in part in the acquisition of causal information that the
scientist may be in no position to articulate. I do not suggest that all the causal
information thus acquired provides understanding. The causal history behind
 peter lip ton

any phenomenon is long and wide, and most of it will not provide real under-
standing in the context of a scientist’s particular cognitive interests. To account
for this, one needs a model of causal selection that shows which causes pro-
vide understanding and which do not (Lipton , –). But that is another
project. For my present purposes, all I suggest is that some of the tacit causal
knowledge gained from manipulation provides understanding.
Images, models, and manipulation are thus plausible sources of understand-
ing without explanation. How might they be resisted? One might say that, even
if the causal knowledge gained in these cases is partially tacit, the knower will
nevertheless be able to say something, however oblique, about the causes. This
may well be so, but what the knower can say in these cases does not exhaust
what the knower understands. The cases could also be resisted by either ex-
panding our notion of explanation or contracting our notion of understanding.
But neither maneuver is plausible here. A tacit explanation would be an expla-
nation one could not give, even to oneself, and this does violence to the notion
of explanation. And to deny that this tacit knowledge constitutes understand-
ing would be Procrustean. We want a way of marking the difference between
someone who knows that the phenomenon occurs but has no inkling why, and
someone who has a deep and subtle appreciation of its causes, though he or
she is not in a position to articulate that knowledge. “Understanding” seems to
be the right word for the difference. It is not ad hoc to insist that explanations
must be verbal but that understanding need not be.

CZXZhh^in

Manipulation, images, models, and other sources of tacit knowledge may

generate understanding without passing through explanations. In these cases,
the understanding consists in tacit knowledge of causal information. Moreover,
in these cases, the route to understanding often essentially involves nonverbal
processes, as in the case of manipulation. But can there also be cases of explicit
information generating understanding without explanation, cases where ev-
erything is propositional and out in the open? Such cases seem unlikely where
the understanding consists in causal knowledge, because the causal informa-
tion then is just an explanation. But there are other forms of understanding
that suggest routes that are propositional and explicit but still do not involve
explanations.
Consider the necessity conception of understanding (Glymour ). On
this view, to understand why something occurs is to see that, in some sense, it
had to occur. Some explanations do seem to work this way. I came to under-
stand why my class had four students whose last birthdays fell on the same day
of the week when it was explained to me that since there are only seven days
under standing withou t e x pl anation 

in a week and twenty-two students in my class, there is no way to arrange the

birthdays to avoid this result. That argument explains an outcome not by giv-
ing causes but by showing that it could not have been otherwise. Just how the
relevant notion of necessity should be characterized is an interesting question.
We could limit the necessity conception so that it can only apply to phenom-
ena that occur in all possible worlds. The birthday phenomenon does not meet
this condition as it stands, though perhaps it would if reconfigured to be the
fact that a class of twenty-two students has four birthdays on the same day of
the week. This is a necessary truth, and the day-filling considerations explain
it by showing that it is necessary. On the other hand, any explanation that can
be couched in terms of a deductively valid argument will show the conditional
necessity that the phenomenon in question must occur given the explanation;
but we do not wish to say that every deductive argument provides understand-
ing, or even that every deductive argument that is explanatory provides under-
standing by showing necessity (rather than, for example, by providing causal
information mediated by causal laws that support the deduction). Fortunately,
for our present purposes we can leave this issue unresolved, since in what fol-
lows, the necessity that provides the understanding is unconditional.
The cases I seek are arguments that are not explanations but do generate
understanding by showing necessity. There are a number of plausible candi-
dates, including various symmetry and optimization arguments in physics.1
Thought experiments are another promising source. For example, Galileo’s
thought experiment that demonstrates that gravitational acceleration is inde-
pendent of mass gives one an understanding of why masses must fall with the
same acceleration even though his experiment may not offer an explanation
for that phenomenon. Suppose that heavier things accelerate faster than light
things, and consider a heavy and a light mass connected by a rope. Considered
as two masses, the lighter one should slow down the heavier one, so the system
should accelerate slower than the heavier mass alone. But considered as one
mass, the system is of course heavier than the heavy mass alone and so should
accelerate faster than the heavier mass alone. But the system cannot accelerate
both slower and faster, so acceleration must be independent of mass (Galileo
/, –; Brown , –).
This beautiful thought experiment does not itself seem to be an explana-
tion. Nor does it appear that the route to understanding passes subsequently
though an explanation that the thought experiment supports. Having absorbed
Galileo’s argument, I understand why acceleration must be independent of
mass; but if you ask me to explain why acceleration is independent of mass, I
can’t do it: the best I can do is to give you the thought experiment so that you
can see for yourself (though more sophisticated physicists may be able to give
an explanation). So, if we maintain as I would that the thought experiment is
 peter lip ton

not itself an explanation, then we seem to have an interesting case of under-

standing without explanation where everything is explicit and propositional.
But why not say that the Galilean thought experiment in itself is an explana-
tion after all? Certainly there are a priori arguments that do seem explanatory.
The birthday argument above may be one example. Another appears in Plato’s
Meno, where Socrates teaches the slave boy (and many readers of the dialogue)
that building one square on the diagonal of another doubles its area. Plato’s
argument shows why the area of the big square must be twice that of the small
one by showing the big square to be made up of four triangular halves of the
small square. That thought experiment shows necessity, provides understand-
ing, and, moreover, in my judgment this argument is also an explanation. Gali-
leo’s argument, by contrast, though it gives the necessity and the understanding,
seems to me not an explanation. Why this difference? It cannot be because the
Galilean argument is noncausal, giving no cause of the fact that acceleration is
independent of mass, since Meno’s argument and the birthday argument are
not causal either, yet those two are explanations. There are noncausal explana-
tions, but the Galilean argument is not one of those. It does not provide a direct
answer the question, “Why is acceleration independent of mass?”
Why not? Perhaps it is because the Galilean argument is a reductio ad ab-
surdum. Rather than saying directly why acceleration must be independent of
mass, the argument works by showing that the contrary assumption would en-
tail a contradiction. My suggestion, then, is that the Galilean argument provides
understanding by showing necessity, but does not provide an explanation, be-
cause it works by reductio. We may find the same contrast between explana-
tory and unexplanatory demonstrations in the case of mathematics. Meno’s
argument is explanatory, but mathematical proofs by reductio may not be. It
seems that to explain by showing necessity requires a kind of constructive ar-
gument that not all proofs or thought experiments supply. Perhaps we require
of an explanation that it show one of the right kinds of “determination” of the
phenomenon. The right kinds are various: it may be cause but it need not be, as
mathematical explanations show. But not everything we might call determina-
tion will do. Thus, mere deductive determination is not sufficient for explana-
tion, as the many valid arguments that are not good explanations show. My
suggestion is that reductio arguments may fail to provide explanations because
they fail to show the right kind of determination. But this does not mean they
provide no understanding. Galileo’s thought experiment is a wonderful dem-
onstration of the necessity of the mass independence of gravitational accelera-
tion and, since seeing necessity is one form of understanding, it is a source of
understanding, but it is not an explanation. If you ask me why acceleration is
independent of mass, I cannot say, though I can show you that it must be so.
Reductio arguments are thus one plausible route to understanding that work
under standing withou t e x pl anation 

by showing necessity but without providing an explanation, and there are likely
to be others as well.

Edhh^W^a^in

My strategy so far has been to look at two characteristic benefits of expla-

nation that we plausibly identify with understanding, and to see whether these
benefits may also accrue without a prior explanation. I have suggested that the
possession of causal information and apprehension of necessity are both forms
of understanding that may be acquired without the help of explanation. Causal
information may be acquired by various nonverbal mechanisms, such as ma-
nipulation and use of images and models, and necessity may be apprehended
by means of thought experiments and other arguments that are not themselves
explanations. I want now to consider a different sort of gap between explanation
and understanding, one that arises because we can gain actual understanding
from merely potential explanation. As in the case of thought experiments, the
kind of understanding gained is modal, but here what is gained is knowledge of
possibility rather than of necessity.
Descartes provides a striking example of one way actual understanding
might come from merely potential explanations:

For there is no doubt that the world was created right from the start
with all the perfection which it now has. . . . This is the doctrine of the
Christian faith, and our natural reason convinces us that it was so. . . .
Nevertheless, if we want to understand the nature of plants or of men,
it is much better to consider how they can gradually grow from seeds
than to consider how they were created by God at the very beginning
of the world. Thus we may be able to think up certain very simple and
easily known principles which can serve, as it were, as the seeds from
which we can demonstrate that the stars, the earth and indeed every-
thing we observe in this visible world could have sprung. For although
we know for sure that they never did arise in this way, we shall be able
to provide a much better explanation of their nature by this method
than if we merely describe them as they now are or as we believe them
to have been created. (Descartes /, )

Was Descartes sincere? He seems to have been serious about the existence and
power of God, or else he would not have given him a lynchpin role in the cen-
tral positive argument of his Meditations. But for our purposes the example
Descartes gives here is only illustrative. One might perhaps prefer a purely nat-
uralistic case, say the ability of a Newtonian explanation of planetary motion to
 peter lip ton

provide actual understanding, even though from the general relativistic point
of view that explanation is fundamentally mistaken about what is really going
on. The general suggestion is just that an explanation known to be false may
provide actual understanding.
Here again we need to compare our standards for explanation and for un-
derstanding. Explanations known to be false will only occupy the gap between
explanation and understanding if they are not themselves actual explanations,
and yet the benefit they confer is actual understanding. So why not rather say
that these explanations provide only potential understanding? A natural gloss on
the notion of potential understanding is that you would have understood why
the phenomenon occurred if the world had been as the explanation describes
it to be. Thus, supposing that God did actually create all living things, the seed
explanation provides only potential understanding; it would only have given you
actual understanding if living things had in fact come from seeds. On this view,
these false explanations only provide understanding of a possible world.
But this does not in my view do justice to the power of potential explana-
tions: they may provide understanding of the actual world—which is to say ac-
tual understanding—not just of possible worlds where they are true. Descartes
is right about this because understanding is a modal notion. This is obvious
from the signal roles of causation and necessity in understanding, themselves
modally laden notions. Information about other worlds illuminates the actual
world. The fact that my computer would not have overheated if the cooling fan
had not broken helps to explain why my computer overheated, and the fact that
my computer would not have overheated if the fan had not broken is, I take it, a
fact about nonactual worlds. Now in this case, we also have a fact that is in some
sense about the actual world, namely that the breaking caused the overheating,
but it seems that information about the merely possible may contribute to un-
derstanding too. As Descartes wrote, we may improve our understanding as to
why a phenomenon is as it is not just by being given a more or less accurate ac-
count of how it actually came about, but by being given even a wildly divergent
account of how it could have come about. Actual understanding may arise from
being shown how something is possible (Nozick , –).
So we cannot plausibly exclude potential explanations from the gap between
explanation and understanding by saying that they provide no actual under-
standing. But we might try to squeeze from the other end. That is, we might ad-
mit that the understanding gained from these explanations is actual, but insist
that this is so only insofar as the explanation itself is actual too, or contains a
part that is actual, or approximates an actual explanation. On this view, there is
not really any gap exposed between explanation and understanding: the actual
understanding that merely potential explanations produce is simply a product
under standing withou t e x pl anation 

of those aspects of the explanation that are true. After all, few if any high-level
scientific explanations are strictly true throughout, but the understanding we
get arises from their approximate truth, or from the parts that are true.
But cases such as Descartes’ explanation do not work in this way. The seed
explanation (on Descartes’ assumption that the real cause is divine) does not
work because it approximates the actual etiology of life. Rather, it works be-
cause, as I have already suggested, it gives modal information, showing how
the phenomenon in question could have arisen. This may generate understand-
ing in many different ways. Information about a possible mechanism may give
oblique information about the actual mechanism (telling us something about a
class into which it falls) by describing a mechanism in the same class, namely
the class of mechanism that can generate this phenomenon.
Perhaps Darwin’s account of the domestic selection of animals provides an
example. “Intelligent selection” illuminates natural evolution even though it is
false of it, since it shows how natural selection can be understood as a mecha-
nism in a larger class.2 Again, possible explanations may provide improved un-
derstanding of actual capacities, powers, and causal relations. Merely possible
explanations may also be failures to describe the actual mechanisms. In such
common cases, we may again gain oblique information about why the phenom-
enon occurred by learning that a particular way it could have occurred was not
the way it did occur.
A merely possible explanation may also give information about the modal
status of the actual explanation. Thus, it may show a degree of contingency in
the actual explanation by showing an alternative route to the same result, or
it may indicate the opposite, showing a degree of necessity to the actual ex-
planation by showing how far from actuality one would have to travel to find
an alternative mechanism. It may also show necessity in another way, by re-
vealing fail-safe overdetermination. Thus, if the merely possible mechanism is
such that it would have been actual, had the actual mechanism not been in
operation, then showing how the phenomenon would have been produced by
the alternative mechanism may help to show that the phenomenon had to oc-
cur, one way or another. Even though we already know how the phenomenon
actually came about, the merely possible explanation thus improves our under-
standing. Suppose that a boxing match between Able and Baker is rigged so
that Baker—though in fact the far better boxer—would take a dive in the tenth
round. Knowing this helps us to understand why Able won, even if as a matter
of fact Able floors Baker with a lucky uppercut in the fifth.
Potential explanations may provide actual understanding without approxi-
mating an actual explanation, and so seem to earn their place in the present
survey of routes to actual understanding that do not pass through actual ex-
 peter lip ton

planation. The understanding involves a kind of cognitive gain about the actual
phenomenon, even though the proffered explanation is not true of the actual
phenomenon.
There is still a way for skeptical readers to insist that the explanations are
actual after all: they can change the explanandum. They may take it that what
is explained in these cases is not the original phenomenon, but rather its pos-
sibility. The effect of this explanandum switch is to convert a merely possible
explanation of actuality into an actual explanation of possibility. My short an-
swer to this maneuver is that it is not to the point. For even if it provided an
acceptable gloss on how these explanations work, it still supports the claim of
a gap between explanation and understanding. Since now explanation of one
thing provides understanding of another, a counterfactual explanation of how
the phenomenon could have come about (but did not) provides understanding
of how it actually happened. To deny the gap in these cases, one would have to
deny that these explanations of how the phenomenon is possible provide any
understanding of the actual phenomenon.

Jc^ÒXVi^dc

I have so far considered how understanding in the form of causal and modal
knowledge might be acquired without relying on actual explanations. Another
equally familiar conception of explanatory benefit is unification (Friedman
). It is an ancient idea that we come to understand the world by seeing the
unity that underlies the apparent diversity of superficial phenomena, and while
philosophers have found this idea difficult to analyze, it is clear that one of the
ways science improves our understanding of the world is by showing how di-
verse phenomena can share underlying similarities.
Unification is a signal achievement of many great scientific explanations;
can it also be achieved in other ways? Thomas Kuhn’s account of the dynamics
of normal science provides a important instance (Kuhn ). His central ex-
emplar mechanism suggests that scientific understanding through unification
yet without explanation is ubiquitous. The formative power of exemplars is a
central theme in Kuhn’s account of science, and appreciating its force shows
not just that the acquisition of theoretical know-how is a route to knowledge of
what a theory says and so to the nature of the world that the theory describes,
but also that what is gained from this process goes beyond the explicit content
of the theory.
One of Kuhn’s central projects is to account for the coordinated behavior of
researchers in a given specialty during periods of normal science. How can one
explain rulelike behavior in the absence of rules? What gives the problem its
bite is the extent of that coordination, involving not just agreement on accepted
under standing withou t e x pl anation 

theory, but also agreement on which new problems to address, how to address
them, and what would count as solving them. Kuhn found the rough functional
equivalent of the coordinating rules of research that were not there in shared
exemplars. These are concrete problem solutions in the specialty—very much
like the solutions to the problems in the student’s problem sets—that are shared
by members of the community and serve as a guide to future work. Unlike
rules, exemplars are clearly available to the members of the relevant scientific
specialty. And although they are, again unlike rules, specific in their explicit
content, they are general in their import. For shared examplars set up perceived
similarity relations, and normal scientists go on to choose new problems that
seem similar to the exemplars, to attempt solutions that seem similar to those
that worked in the exemplars, and to judge the adequacy of proposed solutions
by reference to the standards the exemplars represent. Shared exemplars thus
perform the same function as would shared rules.
The exemplar mechanism provides a plausible example of a route to un-
derstanding without explanation. Kuhn is here providing an account of sci-
entific knowledge, of the scientists’ ability to select problems and to generate
and evaluate solutions. And these abilities correspond also to a knowledge that
goes beyond the explicit content of the theory. The exemplars provide knowl-
edge of how different phenomena fit together. The unarticulated similarity
relations that the exemplars support provide a taxonomy that gives informa-
tion about the structure of the world. They thus have the effect of unifying the
phenomena and so provide understanding, and they do this by analogy, not by
explanation.
For Kuhn himself, the unifying structure that exemplars forge is not some-
thing antecedently present in nature, waiting for scientists to disclose. Kuhn is
an antirealist. Like Kant, he holds that the world that scientists investigate is
a joint product for the things in themselves and the organizing activity of the
investigators. But where Kant holds that the human contribution to the empiri-
cal world can only take one form, Kuhn maintains that the contribution varies
across scientific revolutions, as the exemplars change. Kuhn is Kant on wheels.
But this does not spoil the appropriation of the Kuhnian mechanism as an ex-
ample of scientific understanding of the world. For again like Kant, Kuhn insists
that he is no idealist. Scientists do not make the world up any way they please;
indeed, anomalies—nature’s recalcitrance—play a central role in the Kuhnian
account of the dynamic of scientific development. Resolving anomalies is the
central puzzle-solving task of normal science, and the continued resistance of
certain anomalies is what prompts the transition from normality to crisis, and
sometimes to scientific revolution.
Thus, even in the context of Kuhn’s antirealism, the unification understand-
ing that exemplars provide is a genuine understanding of the world. Moreover,
 peter lip ton

in my view, we need not see this as a package deal. The exemplar mechanism
does not demand an antirealist reading: it is compatible with the view that nat-
ural kinds mark mind-independent divisions in nature and that the exemplar
mechanism provides a way for us to explore those structures. The idea that
we may unify the phenomena (and so improve our understanding of them) by
constructing schemes of classification that do not in themselves provide expla-
nations is compatible with a conventional realist conception of science as being
in the truth business. The same holds for the other routes to understanding I
have highlighted in this chapter. The understanding we acquire may not pass
through an explanation, but it may nevertheless be understanding of an inde-
pendent and objective reality.
One natural way to show that a purely “explanationist” conception of un-
derstanding is excessively narrow is to point out the extent to which scientific
understanding is understanding how rather than understanding why, and thus
the extent to which understanding flows from abilities or skills rather than from
explanations. Abilities appear early in this chapter in my example of causal
knowledge coming from manipulative skills. They appear again in this section
since Kuhn’s exemplar mechanism is in the first instance an account of scien-
tific abilities, such as the ability to select appropriate problems and appropriate
candidate solutions. But while abilities appear prominently in my story, they are
not what my story is about. I am sympathetic to the broader conception of un-
derstanding that encompasses understanding how in addition to understanding
why, but my present purpose is to show that, even on a narrow conception of
understanding as understanding why, we may nevertheless get understanding
without actual explanation. That narrow conception takes understanding why
to be a form of propositional knowledge, and lets explanations set the standard
for what kind of knowledge counts as understanding. For my present purposes,
what is important about abilities is not the sui generis forms of understanding
they provide, but rather the conventional forms of understanding they support,
as conventional as knowledge of causes and unification.
In sum, I have argued in this section that understanding may come from ex-
emplar rather than from explanation. If Kuhn is right about exemplars, the abil-
ity to use a theory in research involves the construction of similarity relations.
This is excess content over what the theory explicitly says, and it yields a kind of
unification, a linking of cases, that is plausibly identified with understanding.

;ZZa^c\

The consequences of explanation I have so far considered—causal, modal,

and unifying—are commonly identified with understanding. The final conse-
quence I will consider has a more contentious status. This is the “aha” feeling,
under standing withou t e x pl anation 

the feeling of understanding.3 What is controversial about this feeling is not its
existence but its status. How does the “aha” feeling, which I take to be a genuine
and distinctive consequence of some explanations, fit into the economy of un-
derstanding? At one extreme, we might say that it is itself a form of understand-
ing; at the other we might take it to be purely epiphenomenal, the irrelevant
phenomenal steam that the brain lets off in the course of its serious cognitive
work. Both of these positions are incorrect: the feeling is neither a form of
understanding nor irrelevant to understanding (compare de Regt ).
How does this issue of the role of the feeling of understanding bear on our
general theme of the scope of understanding without explanation? The familiar
philosophical literature that has approached understanding entirely through
models of explanation has ignored the feeling. This is not surprising. Certainly
the phenomenology is unlikely to be foregrounded if one’s focus is on the logic
or the metaphysics of explanatory arguments. But once understanding is disas-
sociated from explanation, the “aha” feeling becomes more salient. It is certain-
ly associated with the alternative routes to understanding we have considered.
The acquisition of information about causes, necessity, and possibility and the
exercise of the similarity relations that Kuhn describes are all sources of the
feeling of understanding.
I am not the person to attempt an articulated analysis of the phenomenol-
ogy of the feeling of understanding, but I would like to begin by making a few
brief comparisons between it and other feelings—of surprise, familiarity, and
expectation—comparisons that may begin to locate the “aha” feeling in phe-
nomenal space. A feeling of surprise often prompts a why-question. Indeed,
even when what we explain is at one level familiar, a why-question is often
motivated by a process of defamiliarization. Thus, initially we find it unsurpris-
ing that the same side of the moon always faces the Earth, until we realize that
this requires not that the moon not spin on its own axis, but that it spin with a
period exactly the same as the apparently unrelated period of its orbit around
the Earth. (Try this out with a salt and pepper shaker if you need convincing.)
The phenomenon comes to seem very surprising, and this prompts us to ask
why those periods should be the same.
Surprise does have a phenomenological aspect, and in some ways that feel-
ing is a rough counterpart to the feeling of understanding, but the feeling of
understanding is not merely the absence of a feeling of surprise, as is shown by
the many mundane, unsurprising things around us that we do not understand.
Moreover, the two feelings seem compatible. Recall Galileo’s thought experi-
ment. One of the reasons for its extraordinary intellectual beauty is that it gen-
erates the feelings of “aha” and surprise together. Typically, however, the feeling
of understanding supplants the feeling of surprise: once we understand why
something happens, we are no longer surprised that it does. It is sometimes
 peter lip ton

supposed that understanding involves making the surprising familiar, but of

course there are many familiar things many of us do not understand, such as
that blueness of the sky. Contrariwise, it is also possible to have the feeling of
understanding without the feeling of familiarity. The Galilean thought experi-
ment is one example; the use of high theory to explain mundane phenomena
provides others (Friedman ).
What is the relationship between the feeling of understanding and the feel-
ing of expectation? This comparison is prompted by David Hume’s discussion
of the idea of necessary connection (Hume /, sec. ). Although he de-
nies that we could conceive of any “glue” or necessary connection in the world
linking causes to their effects, surprisingly Hume insists that we do have an idea
of necessary connection. This obliges him in the context of his empiricism to
provide a source impression, and so he does, though the impression he finds
is unusual, quite different from the sense impressions he typically cites as the
sources of our ideas.
According to Hume, the source impression for our idea of necessary con-
nection is nothing perceived but is rather a feeling of expectation, the feeling
we have that a familiar effect is coming when we see a familiar cause. On this
view, the reason I have the idea that there is a connection between dropping the
glass and it breaking is because I expect it to break when I see it fall. This is a
typical Humean inversion. It is not that we expect the effect because we believe
that the cause will bring it about; rather, we believe that the cause will bring
about the effect because we expect it. And Hume’s admission that we do have
an idea of necessary connection turns out to be compatible with his insistence
that we cannot conceive of any external connection between cause and effect
since the idea of necessary connection turns out to be an idea of something
entirely internal and subjective—the feeling of expectation.
Hume here appeals to causes generating a feeling toward their effects, and it
is natural to compare this to the feeling of understanding since understanding so
often consists in the appreciation of causes. In Hume’s case, the feeling in ques-
tion is associated with looking forward, from cause to effect, whereas in causal
understanding we look backwards, from effect to cause; but the feelings could
be similar. Carl Hempel makes this sort of Humean connection when he writes
that “any rational rationally acceptable answer to the question ‘Why did event X
occur?’ must offer information which shows that X was to be expected” (,
–). By the time we ask why about an effect it is typically too late to expect
it, but we might still, as Hempel suggests, be looking for information that would
have led us to expect the effect, had we had that information earlier.
The feeling of understanding thus could not be quite the same as the feel-
ing of expectation, since you cannot expect what you already know to have
occurred, but it could perhaps be a kind of subjunctive expectation, roughly
under standing withou t e x pl anation 

the feeling that one would have expected it. In fact, however, this does not
capture the feeling of understanding. For one thing, unlike the indicative feel-
ing of expectation itself, it is not clear that seeing that something would have
been expected qualifies as a feeling. In any event, it is certainly not a feeling of
expectation. On seeing my younger son drop a balloon filled with water from
the upstairs window, I expect that my older son below will very soon be in for a
surprise; but when I am myself the unsuspecting victim of this prank, my sub-
sequent thought that I would have expected it if I had been looking up at the
right time is not itself a form of expectation.
The feeling of understanding is thus distinct from a feeling of expectation,
and when you have the former you do not typically have the latter. (The obvi-
ous exception is understanding why something will occur in the future.) And
vice versa too: you may have an expectation without any understanding or feel-
ing of understanding, as many of the counterexamples to the deductive-nomo-
logical model of explanation show. The falling barometer leads me to expect a
storm, but provides no understanding nor feeling of understanding as to why
the storm occurs.
In sum, while the feeling of understanding is interestingly related to the
feelings of surprise, familiarity, and expectation, it is different from all of them.
What then is the relationship of the feeling of understanding to understanding
itself? Clearly the two cannot be identified. The feeling may be genuine even
when the understanding is not, and one may have the understanding without
the feeling. This is particularly plausible in cases of one’s own intentional ac-
tions. When I decide on a course of action, for example, I have an explanation
for what I will do before the fact, and in mundane cases of this sort there seems
to be no distinctive “aha” feeling. When I decide that I will leave a great pile of
roses on the kitchen table for my wife on Valentine’s Day, I know in advance
why those flowers will be there, but at no stage in the operation do I seem to
enjoy that distinctive “aha” feeling (though perhaps my wife will). Perhaps the
feeling requires a transition between ignorance and understanding that cases
of this sort do not supply. Understanding why is a kind of knowledge—whether
of causes or modal facts or connections—and a feeling is not a kind of knowl-
edge. So it is not a kind of understanding.
Thus, I endorse the intuitive position that the “aha” feeling is not under-
standing itself but rather a consequence of understanding. But it does not
follow that the feeling is epiphenomenal: this effect may serve an important
function. Here it is irresistible to introduce Alison Gopnik’s suggestion from
her memorably titled essay “Explanation as Orgasm”:

My hypothesis is that explanation is to theory formation as orgasm is

to reproduction—the phenomenological mark of the fulfillment of an
 peter lip ton

evolutionarily determined drive. From our phenomenological point of

view, it may seem to us that we construct and use theories to achieve
explanation or that we have sex to achieve orgasm. From an evolu-
tionary point of view, however, the relation is reversed, we experience
orgasms and explanations to ensure that we make babies and theories.
(Gopnik , )

Gopnik’s point is that although the pleasant “aha” feeling is an effect of ac-
tivities that leads to understanding, it is also a motivating cause of those activi-
ties. Just as orgasm is nature’s way of encouraging reproduction, the feeling of
understanding is nature’s way of encouraging cognitive activities that generate
understanding. (Gopnik, in fact, identifies explanation with the “aha” feeling,
whereas on my gloss here I distinguish them. The way I see it, explanations are
one of the causes of feeling rather than the feeling itself.)
Gopnik, in fact, appeals to two feelings—a “hmm” feeling corresponding
to the why-question and the “aha” feeling corresponding to finding the an-
swer—and she emphasizes the role of those feelings as forces driving the in-
quirer to engage in the practices of intervention and inference needed to infer
underlying causal structure.4 But I want to focus on three points concerning the
general and tantalizing idea that the “aha” feeling’s function is to motivate the
search for understanding.
First, whatever one thinks about evolutionary psychology, it is plausible to
say that the desire for this pleasant sensation serves to motivate the search for
understanding. Scientific understanding in the various forms I have touched
upon in this chapter has instrumental value. Knowing about causal history,
about necessity in nature, about how things might have come about, and about
how to work one’s way around a theory may all make us better able to cope
with our environment. But the payoffs are often highly indirect and only in the
long term. So it is just as well that there is also such an immediate and personal
phenomenological payoff to keep us going.
The second point is that this picture of the role of the feeling of understand-
ing helps to justify my decision not to count the feeling as a part of the un-
derstanding itself. By giving the feeling a different, plausible, and conspicuous
role—one of understanding motivator rather than of understanding compo-
nent—it seems even more natural to distinguish understanding from the feel-
ing it causes.
The third point is that there is a natural though substantial extension of
Gopnik’s hypothesis concerning the causal role of the “aha” feeling that I
would like to promote. Her point is that the feeling motivates the activity of
hypothesis construction. My suggested addition is that the feeling may also
play a role in guiding that activity. For it is natural to link Gopnik’s hypothesis
under standing withou t e x pl anation 

to the popular characterization of scientific inference as Inference to the Best

Explanation (IBE).5
According to IBE, when scientists are faced with competing explanations of
the same data, they decide which hypothesis to infer partially on explanatory
grounds. Leaving all qualifications and complications to one side, they prefer
the best explanation. But what does “best” mean here? It had better not mean
likeliest, because then the account descends to the triviality that scientists pick
as likeliest the explanation that they judge to be . . . likeliest. An interesting
version of IBE will maintain instead that “best” means loveliest. The idea is that
scientists prefer the hypothesis which, if correct, would provide the greatest
understanding (Lipton , chap. ).
Insofar as this version of IBE is along the right lines, then it may be that the
feeling of understanding, or the intensity of that feeling, is an important part of
the mechanism by which scientists decide which explanatory hypothesis to in-
fer. Again, without any qualifications, the suggestion is that scientists consider
which of the competing hypotheses would, if inferred, provide the strongest
feeling of understanding, and that is the one they infer. In this model, the feel-
ing is a guide as well as a spur to inquiry. I hesitate to develop the sexual analogy
by appeal to selection of sexual practices in terms of the degree of satisfaction
they provide, but it is difficult to resist. The crude idea is that just as we select
activities by considering how much pleasure they might produce, we may select
hypotheses by considering how strong a feeling of understanding they might
produce.
There is a great deal that could be done to articulate, to defend, and to criti-
cize this thought that scientists decide which hypothesis to infer in part by
considering which hypothesis would provide the greatest satisfaction, but here
I will consider briefly just one natural objection: that this mechanism would
make science too subjective—theory choice governed by feelings! I am, how-
ever, relatively untroubled by this objection, in spite of my village realist orien-
tation toward science. There are at least two things that reassure me. The first
is that the feeling, on my showing, is no part of the understanding itself. We do
not understand because we have a certain feeling; we have that feeling because
we understand. This leaves plenty of room for misleading feelings: we may find
a bad explanation very satisfying because we do not realize how bad it is. In
these cases the feeling is perfectly genuine, but the understanding is not. This
preserves as much objectivity for the understanding as you like, while still al-
lowing the feeling a signal epistemic role.
The second reassuring thought arises from a broadly reliabilist perspective
on inference. Our perceptual systems are mediated by our subjective experi-
ences, but this does not mean that perceptual knowledge is itself impossible or
dangerously subjective. What counts is whether our on-board instrumentation
 peter lip ton

is in fact a reliable way of detecting features of our environment. And so it may

be, even though it is dependent on how things seem. Similarly, although the
feeling of understanding is subjective—it is a feeling, after all—it may never-
theless be well calibrated and a reliable guide to theory choice. These feelings
are highly sensitive to background belief, but one need not draw an antirealist
conclusion from that either. Quite the contrary: the high sensitivity of feeling to
background may show that this mechanism is highly attuned to relevant fact. It
enables scientists to make exquisitely sensitive evidential judgments, an ability
that unsurprisingly goes far beyond their ability to articulate the full range of
grounds on which those judgments are based.
I am not saying that the feeling of understanding is an inferential guide; I
am only suggesting how it could be. Such an argument would require more
than arguing for the basic IBE package. That is simply that explanatory consid-
erations are a guide to inference, that we are helped to decide which hypoth-
esis to infer on the basis of judgments about which hypothesis would provide
a better explanation if it were inferred. That view requires no distinctive role
for feelings. The mechanism could presumably be run entirely on an explicit
propositional basis. But given how much of scientific judgment is tacit, that
does not seem likely.
The case for tacit mechanism seems especially strong in the instance of
judgments about explanatory quality, where the contrast between our ability
to make them and our inability to describe the basis on which we make them is
particularly stark. And such a view would fit particularly well with the sources
of understanding I have considered in this chapter that involve tacit knowledge.
If understanding sometimes consists of tacit knowledge, then it is natural to
suppose that some judgments of understanding will be tacitly made.

8dcXajh^dc

It is one thing to know that something is the case; it is something else to

understand why. Explanation is one way to bridge this gap. How does it do this?
We get an important clue from the “why-regress.” As many of us learned as
young children, and to our parents’ cost, you can almost always respond to an
explanation by asking why about the explanation itself. A moral of that lesson
is that something does not itself need to be explained before it can be used to
explain other things. If I ask my student why she did not hand in her essay on
time, she may explain that her computer crashed. This is, in principle, a good
explanation, even if neither of us have any idea why the computer crashed.
What the why-regress shows is not that we never really explain anything, but
that understanding is not like a kind of substance that gets transmitted from
premise to conclusion in an explanation. Instead, to understand is simply to
under standing withou t e x pl anation 

have the right kind of knowledge. Understanding why a phenomenon occurs is

more than knowing that it occurs, but it is just more knowledge. Explanations
give us that knowledge, but the knowledge may also be acquired in other ways.
These alternative routes to understanding are what I have been exploring in
this chapter.
I have looked at five familiar benefits of explanation: knowledge of causes,
necessity, possibility, and unification, and also the feeling of understanding. All
of these consequences of explanation, I have suggested, may also be acquired in
other ways. Manipulation and other scientific activities may give causal infor-
mation, thought experiments may show necessity, false explanations may give
information about possibility, exemplars and the theoretical skills they sup-
port may provide unification, and all of these activities may produce the “aha”
feeling.
The first four of these benefits are unproblematically identified with under-
standing, so if I am right in saying that they do not only derive from explanation,
then it looks like understanding does not only derive from explanation. I have
treated the fifth case differently. Although the “aha” feeling is a consequence
of explanation that my other routes may also generate, it does not seem to me
sufficiently cognitive to count as itself a form of understanding. Understanding
is a form of knowledge; the feeling of understanding is not. But the feeling may
nevertheless promote understanding. Gopnik suggests that it may promote the
search for explanations that provide understanding, and in the context of my
discussion, that point generalizes to the claim that it promotes all the activities
I have flagged that provide understanding. I have also suggested an extension
of Gopnik’s proposal. The feeling of understanding may serve not only to mo-
tivate activities that promote understanding, but may also play a role in select-
ing explanations in the context of inferences to the best explanation, since the
strength of the feeling may be used as a symptom of explanatory quality.
My discussion has appealed to the inarticulate at two levels. The first is
tacit knowledge. The causal understanding that comes from manipulation and
the unification understanding that comes from exemplars are forms of tacit
knowledge. These are nevertheless forms of actual understanding, I have sug-
gested, because they are knowledge with the right sort of content, even though
that content is tacit. Explanations may have to be explicit to do their work,
but tacit knowledge can give actual understanding. The second appeal to the
inarticulate has been to the feeling of understanding. This is not itself a form of
understanding, because it is not a form of knowledge, but I have suggested that
it may be a prod and a guide to the acquisition of understanding.
If understanding does have a tacit dimension, and understanding is a guide
to inference, then it is not surprising that inference should have a tacit dimen-
sion as well. There is no reason to suppose that epistemic judgments are entirely
 peter lip ton

transparent to the judge. Insofar as judgments of potential understanding play

a role in scientific inference, these judgments may depend on the scientists’
“good sense” (Duhem /, –), and the feeling of understanding
may be one of the ways scientists make contact with what Pierre Duhem, quot-
ing Pascal, referred to as those “reasons which reason does not know.”

CdiZh
I am grateful to my fellow participants at the conference on Philosophical Perspectives
on Scientific Understanding that took place at the Faculty of Philosophy, Vrije Univer-
sity Amsterdam, in August . I learned a great deal there from talks, comments, and
questions, and I am particularly grateful to the Henk W. de Regt, Kai Eigner, and Sabina
Leonelli for their hard organizational work and to Henk and Sabina for exceptionally help-
ful suggestions for revising my essay. I presented some of the material in this chapter to the
Department of Science and Technology Studies at University College London in March
 and to the Division of Humanities and Social Sciences at the California Institute of
Technology in May , and I thank the audiences there. In particular, I would like to
thank Jed Buchwald, Robert Bud, Hasok Chang, Moti Feingold, Marcus Giaquinto, Alison
Gopnik, Christopher Hitchcock, Jeremy Howick, Michela Massimi, Alex Pine, and Jim
Woodward for very helpful comments and suggestions.
.In the examples below, the first class was suggested to me by Christopher Hitchcock,
the second by Jed Buchwald.
. I owe this suggestion to Sabina Leonelli.
. The “aha” feeling is also discussed in the essays by Grimm and Ylikoski in this
volume.
. Gopnik’s idea is also discussed in Grimm’s essay in this volume.
. This latter view is one I have discussed at length elsewhere (Lipton ). Here I will
just say enough to make the bridge to Gopnik’s proposal.

GZ[ZgZcXZh
Brown, J. R. . The laboratory of the mind. London: Routledge.
de Regt, H. W. . Discussion note: Making sense of understanding. Philosophy of Sci-
ence :–.
Descartes, R. /. Principles of philosophy. In The philosophical writings of Des-
cartes, vol. , trans. J. Cottingham, R. Stoothoff, and D. Murdoch, –. Cam-
bridge: Cambridge University Press.
Duhem, P. /. The aim and structure of physical theory. Trans. P. P. Wiener. Prince-
ton: Princeton University Press.
Friedman, M. . Explanation and scientific understanding. Journal of Philosophy
:–.
Galileo Galilei. /. Dialogues concerning two new sciences. Trans. H. Crew and A.
de Salvio. New York: Macmillan.
under standing withou t e x pl anation 

Glymour, C. . Explanations, tests, unity, and necessity. Noûs :–.

Gopnik, A. . Explanation as orgasm and the drive for causal knowledge: The function,
evolution, and phenomenology of the theory formation system. In Explanation and
cognition, edited by F. C. Keil and R. A. Wilson, –. Cambridge, MA: MIT Press.
Hempel, C. G. . Aspects of scientific explanation. In Aspects of scientific explanation
and other essays in the philosophy of science, –. New York: Free Press.
Hume, D. /. An enquiry concerning human understanding. Edited by T. Beau-
champ. Oxford: Oxford University Press.
Kuhn, T. S. . The structure of scientific revolutions. nd ed. Chicago: University of
Chicago Press.
Lewis, D. . Causal explanation. In Philosophical papers, :–. New York: Oxford
University Press.
Lipton, P. . Inference to the best explanation. nd ed. London: Routledge.
Nozick, R. . Philosophical explanations. Cambridge, MA: Harvard University Press.
 hasok chang

+
Ontological Principles and the
Intelligibility of Epistemic Activities

=6HD@8=6C<

 My main goal in this essay is to establish intelligibility as an epis-

temic virtue that is meaningful and desirable independently of any connection
it might or might not have with truth.1 In brief, my argument is that intelligibil-
ity consists of a kind of harmony between éour ontological conceptions and our
epistemic activities. I will modify and broaden that formulation in the course
of the discussion.
I will start with an example to illustrate briefly the direction of my thinking.
After that, I will make a more careful consideration of the nature of intelligibil-
ity. Then I will finish with some brief reflections on the implication of this con-
cept of intelligibility for our understanding of scientific understanding. I will
be building on an earlier article (Chang ), in which I tried to reformulate
scientific realism as a search for intelligibility rather than truth. Although the
subject of the present essay is not realism, some aspects of that earlier article
are quite pertinent to the discussion of intelligibility.
The discussion of intelligibility here constitutes the rudiments of a much
larger philosophical project, which I call “epistemic humanism.” The first part
of the project is to reach a full phenomenological understanding of direct hu-
man epistemic experience. The second part is to understand other aspects of
knowledge as metaphorical extensions of the direct human experience. The ex-
plicit focus on the human is intended to combat the prevalence of alienating
abstractions in today’s analytic philosophy.


ontol o gic al principle s and episte mic ac ti vitie s 

AZ^Wc^oVhVc>che^gVi^dcVcYVAVYYZgid7Z@^X`ZY6lVn

In my earlier work (Chang ), I noted some historical cases in which

realist scientists such as Albert Einstein, Humphry Davy, and Gottfried Wil-
helm Leibniz rejected even empirically adequate scientific theories because
they considered them absurd or unintelligible. Here I will rehearse the case of
a little-known work by Leibniz, in which he advanced a devastating critique of
René Descartes’ physics;2 in particular, I will focus on Leibniz’s attack on Des-
cartes’ laws of collision.
Descartes had formulated seven rules governing the elastic collision of two
bodies. Take the first two rules (illustrated in fig. .):

C B

C B

9ZhXVgiZhÉ;^ghiGjaZd[8daa^h^dc
Descartes’s First Rule of Collision

C B

C B

9ZhXVgiZhÉHZXdcYGjaZd[8daa^h^dc
Descartes’s Second Rule of Collision

;^\jgZ)#&#9ZhXVgiZhÉildgjaZhd[Xdaa^h^dch#

Rule : if these two bodies (for example, B and C) were exactly equal
[in size] and moved with an equal velocity towards one another, then
they would rebound equally, and each would return in the direction it
came, without any loss of speed.
Rule : if B were somewhat larger than C, and if they met each other
at the same speed, then C would be the only one to rebound in the
 hasok chang

direction from which it came, and henceforth they would continue

their movement in the same direction; B having more force than C,
the former would not be forced to rebound by the latter. (Descartes
/, –)

Now, it might seem that Descartes’ second rule is patently false, and anyone
can refute it empirically by doing a very simple experiment. Leibniz the meta-
physician, however, chose to criticize it without recourse to experience. He put
Descartes’ two rules side-by-side and asked what would happen in the limiting
case in which the object approaching from the right-hand side (B in the figure)
was bigger than the left-hand one (C), but the difference approached zero. Ac-
cording to Descartes, as long as the difference is nonzero, however small, the
object coming from the right (B) should push straight through without any
change in its speed, but as soon as the difference becomes zero it is bounced
back in a U-turn.
Leibniz (/, –) thought this was a metaphysical absurdity,
violating the principle of continuity: how would a continuous change in the
circumstance of the experiment make such a discontinuous difference in the
outcome?3 Leibniz articulated his principle of continuity at least in two places.
In the Specimen Dynamicum of , he stated, “If one case continually ap-
proaches another case among the givens, and finally vanishes into it, then it is
necessary that the outcomes of the case continually approach one another in
that which is sought and finally merge with one another” (Leibniz , ). A
slightly different formulation can be found in a letter to Pierre Varignon from
: “The Law of Continuity demands that when the essential determinations
of one being approximate those of another, as a consequence, all the properties
of the former should also gradually approximate those of the latter” (Leibniz
, ). According to Leibniz, Descartes’ laws of collision could be rejected
with certainty, without a single experiment being performed, or even imagined
as a thought experiment. Descartes’ physics was not intelligible to Leibniz (pro-
ducing a “monstrous and inconsistent” result), and that was sufficient ground
for rejection (Leibniz /, ).
This case shows an eminent philosopher-scientist using the lack of intel-
ligibility as a cogent reason for rejecting a major piece of theory (part of the
physical foundation of Cartesian mechanical philosophy). And that was done
without any regard to empirical adequacy. The contrast to standard empiricist
schemes could not be clearer: in Bas van Fraassen’s (, –) view, for ex-
ample, something like intelligibility would probably be treated as a “pragmatic
virtue,” a merely subsidiary criterion that might function as a tiebreaker if a
choice were desired between two theories of equal empirical adequacy. With
Leibniz, it is clear that intelligibility was just as important as empirical adequa-
ontol o gic al principle s and episte mic ac ti vitie s 

cy, if not more important. What is also nice about Leibniz’s thinking is that the
verdict of unintelligibility is based on a clearly articulated ontological principle
that seems beyond dispute at first glance. This is rationalist physics at its best.
On the other hand, the Leibniz story serves as a cautionary tale as well.
Why was he so sure about the principle of continuity? How was the principle
justified? At least in light of certain developments in modern science, Leib-
niz’s principle of continuity is unjustified. Granted, there might be some room
for debate about whether or not quantum mechanics (or chaos theory) really
shows the principle of continuity to be false. But the real point is that we can
imagine the universe to be a place in which the principle of continuity is vio-
lated. I would like to craft a concept of intelligibility that supports the positive
aspects of what Leibniz was doing, while avoiding the pitfalls in his thinking.

L]Viid9dVWdjii]ZJcXZgiV^cind[BZiVe]nh^Xh

The troubling thing about Leibniz’s thinking is that he was so sure about his
principle of continuity seemingly without having an argument to back up that
certainty. Subjective certainty, even for such an eminent thinker as Leibniz,
may not be worth very much.4 This will be a problem with rationalistic thought
in general, and also with any attempt to take intelligibility as an indication of
truth. Einstein pleased himself in his objection to the Copenhagen Interpreta-
tion of quantum mechanics by declaring that “God does not play dice,” but how
did he have this access to God, which was not only direct but apparently ex-
clusive, since Niels Bohr would seem to have been denied the direct telephone
line? Even Kant fell into the trap of thinking that things like Euclidean geometry
and Newtonian mechanics were necessarily true. Looking at Kant from a little
distance makes one thing regrettably clear: his location in the Newton-enrap-
tured eighteenth century must have exerted a strong hold on his imagination.5
Philipp Frank gives a sober assessment of the poverty of philosophical op-
position to new scientific ideas: what parades as a deep metaphysical truth is
often simply the petrified remains of an outdated scientific theory. In Frank’s
view, science needs to educate metaphysics, not vice versa (, –). Paul
Feyerabend (, ) describes the same phenomenon as “contamination by
older cosmologies,” in the context of his discussion of Galileo’s defense of Co-
pernicanism. In order to render the “tower argument” harmless for Coperni-
canism, Galileo had to dismantle the implicit assumption inherited from earlier
physics that apparent motion was real motion.
There are many other cases from the history of science showing the ques-
tionable effect of metaphysical beliefs that are leftovers from earlier scientific
theories: Einstein’s insistence on determinism against the Copenhagen inter-
pretation of quantum mechanics, Erwin Schrödinger’s objection to quantum
 hasok chang

jumps, the Cartesians’ objection to Newtonian action at a distance, and the

like. The scientific reactionaries protested that the new theory did not make
sense, that it was unintelligible. These cries of unintelligibility largely ceased to
resonate with new generations of scientists, who grew up with the new theory
and found it just as intelligible or unintelligible as any other theory.6
If we are going to use metaphysics as a guide in scientific work, we must try
to do something about the lack of certainty displayed by metaphysics. The prob-
lem is that uncertainty tends to destroy the regulative force of principles. The
starting point of the solution I would like to propose here is to note that some
bits of metaphysics are more secure than others.7 We can distinguish between
mere ontological opinions (or prejudices), and what might properly be called on-
tological principles. There is much more to the difference between opinion and
principle than relative degrees of certainty. The hallmark of an ontological prin-
ciple is that its denial should strike us as nonsensical rather than false; that is to
say, ontological principles provide the basis of intelligibility, not truth. Leibniz’s
“principle” of continuity is a mere ontological opinion, even though it may be
one that many of us share. Its denial may sound strange, but it is intelligible. So,
in my view, it does not really qualify as an ontological principle.
An example of what does qualify as an ontological principle is what I call
the principle of single value. This principle states that a real physical property
can have no more than one definite value in a given situation; or, that two cor-
rect value ascriptions about the same situation cannot disagree with each other
when we are concerned with a real physical property.8 Consider what we get
when we violate the principle of single value. Someone says, “The length of this
stick is both  meters and  meters.” “No,” we would respond, “the value has to
be one or the other, not both at once.” (Even quantum mechanics, even in its
Copenhagen interpretation, stops short of saying that a particle can have mul-
tiple positions at once; a particle can only have nonzero probabilities of detec-
tion in various places.) If I can successfully make and maintain the distinction
between ontological opinion and ontological principle, then it will point to a
way of establishing intelligibility as an aspect of scientific knowledge that is not
merely subjective. That is something I will try to achieve in the course of the
discussion in the next two sections.

I]Z8dccZXi^dcWZilZZcDcidad\^XVaEg^cX^eaZhVcY
:e^hiZb^X6Xi^k^i^Zh

=dlH^c\aZKVajZ>hCZXZhhVgn[dgIZhi^c\"Wn"DkZgYZiZgb^cVi^dc
Let me now take a closer look at the nature of ontological principles. By an
“ontological” principle I do not mean a principle that dictates whether some-
thing exists. Rather, the term “ontological” indicates a concern with the basic
ontol o gic al principle s and episte mic ac ti vitie s 

nature of the entities that we are thinking about. The entities in question may
be of any type (for example, physical or abstract), but in any case they will
have some basic properties, and the ontological principles dictate what kind of
properties they can have.
Coming back to the principle of single value, we need to ask: what are the
grounds of the validity of this principle? The first thing to note is that it is not an
empirical generalization. To see that point, imagine trying to test the principle
by experiment. If we went around taking measurements of various quantities
in order to check that we get a single value each time, that would seem as silly
as testing “All bachelors are unmarried” by collecting a bunch of bachelors and
checking each one carefully to see if he is married. Worse, if someone reported a
counterinstance (say, that his weight is both  kilograms and  kilograms ), our
reaction would be total incomprehension. We would have to say that such an ob-
servation in “not even wrong,” and we would feel compelled to engage in a meta-
physical discourse to persuade this person that he is not making any sense.9
The principle of single value is not a logical truth, either. One can easily
imagine variables that have multiple values, especially in the realm of non-
physical quantities and designations: for example, names of persons or places,
or multivalued functions in mathematics. Still, where we do recognize it as
valid, the principle clearly seems to have a necessity about it. What could be the
source of this necessity? In terms of conceptions that are commonly known, I
think the closest thing to ontological principles as I conceive them is the Kan-
tian synthetic a priori. But there is a significant difference from the Kantian
conception, as I will explain later. Especially given the presence of exceptions,
what we must conclude is that the necessity of the principle of single value is
not universal but conditional, holding only in some situations.
I would like to propose that the necessity of the principle of single value
comes from the requirements of testing.10 Let me elaborate on that idea. Take
an ordinary situation in which a theory is tested, in the simplified view given
by the hypothetico-deductive model. We derive a prediction from the theory
and check it against observation. If the predicted and observed values differ, we
decide that there is something false about the theory. This is usually regarded
as a matter of simple deductive logic, but we are actually making a tacit reliance
on the principle of single value. We predict that the temperature here at  a.m.
today will be ° C; our observation says it is °C. If the temperature could be
both °C and °C, there would be no refutation here. To put it formally: there
is no logical contradiction between “T = °C” and “T = °C,” until we also put
in the principle of single value, which gives “If T = °C, then T ≠ °C.” Now
we do have a contradiction, because we have “T = °C” and “T ≠ °C.” With-
out the principle of single value, we cannot engage in this kind of testing.
More generally, the principle of single value is required for what I call test-
 hasok chang

ing-by-overdetermination, in which we determine the value of a quantity in two

different ways. If the values match, that gives credence to the basis on which
we made the two determinations; if the values do not match, then we infer that
there is some problem in our starting assumptions. The two determinations
could be the familiar prediction-observation pair, or two observations, or even
two theoretical determinations. So, what we engage in is attempted overdeter-
mination, which is really just an elaborate way of saying “checking.”
To summarize, we need to subscribe to the principle of single value if we
want to engage in testing-by-overdetermination. In other words, the necessity
of the principle of single value springs from our commitment to testing-by-
overdetermination. Or, single valuedness is necessary for enabling the activ-
ity of testing-by-overdetermination. What we have is a pragmatic necessity—a
necessity arising from the requirements of action, not some kind of hypertruth
that pertains to a proposition.11

=dl8djci^c\GZfj^gZh9^hXgZiZcZhh
The above idea about the necessity of ontological principles can be articu-
lated more carefully with the help of an even simpler example. The ontological
principle that enables the activity of counting is that of discreteness. If we want
to engage in the activity of counting, then we have to assume that the things we
are trying to count are discrete. Let me highlight several aspects of this situa-
tion, which we can later try to generalize to other pairs of ontological principle
and epistemic activity.12
. It is obvious that the necessity here is conditional: it is only if we
count things that discreteness has to be presumed. (Whether or not we
ought to be engaged in counting is a separate question.)
. The principle of discreteness says nothing about whether the world
in itself is discrete or not, only that we need to take it as such if we are go-
ing to count things.
. What underlies the sense of necessity here is impossibility: we
can try, but we just cannot count anything in an utterly undifferentiated
continuum. The nature of that impossibility is interesting. It is a pragmatic
impossibility, which is a notion I will take as a primitive rather than try to
analyze it further in terms of other notions, which I think are less basic.13
. The difference between empirical truth and pragmatic necessity
should be clear. A given bit of the universe may or may not contain dis-
crete pieces; if it does, then that domain is, in fact, discrete, which would
be an empirical truth; that would be a fact, with no necessity about it. The
necessity of discreteness only comes with our commitment to count.
. The denial of an ontological principle, while we are engaged in the
activity that requires it, generates a sense of unintelligibility. If we are try-
ontol o gic al principle s and episte mic ac ti vitie s 

ing to count things that we cannot grasp as discrete, that makes no sense;
we are engaged in an unintelligible activity.
. This sense of intelligibility pertains to activities in the first instance,
and only indirectly to statements, propositions, representations, or mod-
els. A formal system in itself is neither intelligible nor unintelligible; it is
what we do with it that may or may not have intelligibility.

Di]ZgEg^cX^eaZ"6Xi^k^inEV^gh
So far we have examined two epistemic activities and found an ontological
principle that is required for each activity (single value for testing, discreteness
for counting). This picture can now be broadened: to each well-defined, basic
epistemic activity, there is an associated ontological principle that enables it
and makes it intelligible.
The proposed one-to-one pairing of epistemic activity and ontological
principle might seem too neat, artificial, and contrived. But it will seem more
natural if we recognize two things. First, the one-to-one correspondence only
applies to the most basic epistemic activities; complex ones will require mul-
tiple principles, as I will discuss further below. Second, in the principle-activ-
ity pairing, the ontological principle and the epistemic activity are mutually
constitutive. A general link between ontology and epistemology will be easily
granted, in the sense that the appropriate method of studying something is
surely linked with the nature of that something. Here I am pointing to the pur-
est and strongest version of that link: a distinct type of epistemic activity has
its particular type of object, the essential characteristic of which is defined by
the fundamental ontological principle associated with the activity. At the same
time, the nature of the activity is shaped by the basic character of its objects.
Table . gives a list of basic epistemic activities and their associated onto-
logical principles that I have started compiling. Even a brief consideration of
these pairings will reveal some quite interesting consequences of this way of
thinking. About the first two activity-principle pairs I have already said enough.
There is just one brief note I must add on the second pair: there are other modes
of “testing,” especially in the sense of a “trial run,” which can be made without a
specific prior prediction of the outcome. This is why I must always specify what
I have in mind here as “testing-by-overdetermination.” The remaining activity-
principle pairs deserve some further comments.

(Rational) Prediction ' The Principle of Uniform Consequence (“Prin-

ciple of Induction”)
In order to make any rational predictions (as opposed to random guessing,
or prophecy by oracle or divine illumination), we must take it for granted that
the same initial circumstance will always have the same outcome. Induction
 hasok chang

IVWaZ)#&#6eVgi^Vaa^hid[eV^g^c\hd[
Ze^hiZb^XVXi^k^i^ZhVcYdcidad\^XVaeg^cX^eaZh

:e^hiZb^XVXi^k^in Dcidad\^XVaeg^cX^eaZ

8djci^c\ 9^hXgZiZcZhh
IZhi^c\"Wn"dkZgYZiZgb^cVi^dc H^c\aZkVajZ
GVi^dcVaEgZY^Xi^dc Jc^[dgbXdchZfjZcXZ
8dcigVhi^kZ:meaVcVi^dc Hj[ÒX^ZcigZVhdc
CVggVi^dc HjWh^hiZcXZ
A^cZVgDgYZg^c\ IgVch^i^k^in

is just another name for this “principle of uniform consequence” in action. As

Wesley Salmon points out in his critique of Popper, an attempt to do away with
induction disables the activity of rational prediction.14 This is why we cannot
seem to get away from induction, no matter what its logical deficiencies are.
Induction is not merely a custom, nor an optional piece of scientific methodol-
ogy, but a necessary condition for rational prediction. (I do not claim to have
solved the problem of induction, but I do think this way of looking at the prob-
lem is instructive.)
One might object that quantum mechanics shows that the principle of uni-
form consequence is false: two particles may be in exactly the same initial state,
but behave differently. But even here, to the extent we engage in rational predic-
tion, we can only do so within the schema dictated by the principle of uniform
consequence, with the help of the concept of probability: “for each and every
particle in quantum state S, the probability of observable O taking the value X is
P.” So, as long as we are engaged in rational prediction, even indeterminism has
to be conceived in terms of uniformity within a given descriptive category: all
particles in that state have the same probability of doing X, and the unpredict-
ability is conceived as randomness rather than willfulness. Thus, quantum me-
chanics turns out to be an exception that proves the rule here. It is instructive
to view indeterminism or determinism as more than a mere ontological belief.
Determinism is a commitment to engage in the activity of rational prediction in
every individual case. That commitment was one important aspect of Einstein’s
insistence that quantum mechanics was incomplete, and it is shared by those
who are still pursuing hidden-variable theories.
Leibniz’s principle of continuity can be seen as an overly strong version of
the principle of uniform consequence. The proper principle says that the same
circumstances must lead to the same outcome. Leibniz’s principle overreach-
es by saying that proximate circumstances must lead to proximate outcomes,
without specifying a convincingly relevant sense of proximity.

(Contrastive) Explanation 'The Principle of Sufficient Reason

ontol o gic al principle s and episte mic ac ti vitie s 

The activity of explaining why something happened (as opposed to some-

thing else) requires the assumption that when there is an observed difference,
there is a reason behind it. I will call this the principle of sufficient reason, since
it can be considered a weak version of that old metaphysical principle.15 This
principle of sufficient reason provides a very broad schema of explanation (“X
rather than Y, because C,” or, “X rather than Y, because A rather than B”). This
schema is fully compatible with the subsumption model of explanation, and it
is more general in that it does not require the explanans to contain any laws
or even to be more general than the explanandum. And the explanatory factor
may be causal, or anything else. (As with testing, I should note that there are
other modes of explanation than the contrastive.)

Narration 'The Principle of Subsistence

Next, consider the activity of narration. What is required for telling a story?
It might be thought that something to do with time would be the most fun-
damental principle of narrative. But, contrary to expectation, it is not time-
ordering that provides the most basic ontological principle behind narration. A
narrative requires entities or persons whose identities last through time, which
“house” the changes that are narrated. Otherwise, we cannot even formulate
narrative strands like “someone ran down the street,” or “on earth, first this hap-
pened and then that happened.” Paradoxically, without postulating something
that lasts, it is impossible to describe any change. In the context of physical
science, this principle of persistence (or better, subsistence) manifests itself as
more specific assumptions, such as that fundamental particles persist without
alteration (“atomism”), or that there is a substance or quantity that is conserved
(mass, energy, and the like).
This, incidentally, is why the Big Bang makes an unintelligible narrative
about the origin of the universe. I am not saying that the science behind Big
Bang cosmology is unintelligible, but only that it is impossible to make an intel-
ligible narrative about the Big Bang itself. The book of Genesis at least offers an
intelligible narrative about the origin of the universe since there is God, whose
identity lasts through the act of creation. But if we demanded a story about how
God came to be, we would run into the same unintelligibility as with the Big
Bang. The story of ultimate origin is untellable, much as Kant taught us in his
antinomies of reason, though for a slightly different reason from what he gave.

(Linear) Ordering 'The Principle of Transitivity

If we want to put a set of entities in an ordered sequence, we must assume
that the method of ordering (or, the relation that forms the basis of ordering)
is transitive. This applies not only to numerical ordering, but all other kinds
 hasok chang

of linear sequencing such as time-ordering, status-based hierarchies, chemi-

cal affinity tables, and completely nonquantitative cases such as alphabetical
ordering.

EgV\bVi^XGddihd[CZXZhh^inVcYDcidad\n
An ontological principle derives its necessity from its rootedness in an
epistemic activity. An ontological opinion, in contrast, has no such grounding
in the requirements of an activity, and may be discarded with no pragmatic
consequences. These reflections actually throw a new light on the concept of
necessity in general. Necessity is a concept rooted in human action, which we
project metaphorically onto the rest of nature (sometimes sensibly and some-
times not). The ignoring of this human, pragmatic context is the main reason
why we have such trouble understanding the notion of necessity, as we try to
render “necessary truth” as a kind of truth stronger than mere actual truth,
with the help of possible-worlds semantics and such. Rather, if something is
“necessarily true,” that means it needs to be taken as true for the purpose of
some epistemic action.
Necessity also does not imply universality. Ontological principles are not
truly universal, as their necessity remains conditional on our commitment to
engage in their corresponding epistemic activities. We may not see this usually,
but there are situations in which we would decline to engage in certain basic
epistemic activities, such as rational prediction, or counting, or explanation.
This is the key difference I see between my way of thinking and the legacy of
Kant. In my scheme there are no eternal and universal categories since we are
not compelled to engage in all the same basic epistemic activities all the time.
It is a mistake to treat the activity-dependent conditions of intelligibility as un-
conditional metaphysical truths. Even compared to William Whewell’s scheme,
in which the categories evolve with the general development of science (see
Kockelmans , –), or Michael Friedman’s system of the relativized a
priori in which revolutions are allowed (, part , chap. ), my scheme is
distinct in allowing more freedom to epistemic communities and even indi-
viduals in their choice of epistemic activities, and in taking the pragmatic per-
formability of activities as the ultimate basis of necessity.
It is satisfying to see how our basic ontological conceptions arise from the
way we engage with the world. The most fundamental part of ontology is not
abstracted from what we passively observe; rather, it emerges and becomes es-
tablished as an essential ingredient of our successful epistemic activities.17 A
whiff of scientific realism can be derived from that insight: if some of our epis-
temic activities are successful, that must be some sort of indirect vindication of
the ontological principles presumed in those activities.18 But this vague sense
of vindication is all we can have, and it only points to an inarticulable harmony
ontol o gic al principle s and episte mic ac ti vitie s 

between the state of the world and our ontological principles. There is a Kan-
tian insight here: nature in itself (not what Kant called “nature,” which consists
of phenomena) is not the object of our epistemic activities; our activity-based
ontology is not a direct representation of nature (or, the Kantian “thing in it-
self ”). As Henri Bergson put it (/, ), “The bodies we perceive are,
so to speak, cut out of the stuff of nature by our perception, and the scissors
follow, in some way, the marking of lines along which action might be taken.”
When nature “speaks” to us, it is only through the outcomes of our epistemic
activities. That is not to deny that nature enters and rules our life by deter-
mining which epistemic activities are pragmatically possible. But ontological
principles do not give us the kind of direct representation of nature that the
correspondence theory of truth would prompt us to seek.

>ciZaa^\^W^a^inVcYHX^Zci^ÒXJcYZghiVcY^c\

Now I am ready to treat the concept of intelligibility more definitively. I

propose: intelligibility is the performability of an epistemic activity. Here I take
pragmatic performability as a primitive, just as I take pragmatic impossibility
as a primitive. This is a very minimal conception of intelligibility, but it is not
entirely vacuous. Performability requires a certain harmony within the activity.
For example, any statement made within it needs to conform to the ontological
principle associated with it.
As I conceive it, intelligibility is ultimately a property of an epistemic activ-
ity as a whole. But depending on what we are focusing on in our analysis, we
could easily say that various components or aspects of an epistemic activity
are intelligible or unintelligible. In an earlier essay I said that statements were
unintelligible if they violated an ontological principle (Chang , , –).
That would be the situation if we had fixed on the intention to perform a cer-
tain epistemic activity, therefore accepting the relevant ontological principle
as necessary. In that circumstance, we should regard any statement violating
that principle as unintelligible. In a different kind of circumstance, however,
unintelligibility may be assigned to other aspects of the situation. Suppose we
have an already established belief that a certain domain is utterly continuous. If
we believe that, and we still attempt to count things in it, then it is our action,
or our desire to count, that is unintelligible.
Will my concept of intelligibility help us to understand what scientific un-
derstanding is? What I would like to offer is a pragmatist concept of under-
standing to go with my pragmatist concept of intelligibility: understanding is
knowing how to perform an epistemic activity. Intelligibility is essential for un-
derstanding because we will not be able to perform an unintelligible epistemic
activity.19 (This is, of course, not to say that intelligibility is sufficient for under-
 hasok chang

standing, or to deny that there are other key factors involved in understand-
ing.) All this may strike the reader as an anticlimax, but if so, I think I will have
made a healthy deflationary move. There are so many things people mean by
“understanding” that it is going to be difficult to lay down any firm definition
of it. I think the best nonvacuous general conception we can have of it is what
I have just given.
Understanding, as I see it, is not some distinct quality or state of mind that
exists separately from the sense of knowing how to do epistemic things. Under-
standing is simply knowledge taken in the active sense. The feeling of under-
standing is the sense of performing an epistemic activity well: “I know what I
am doing.” The degree of understanding is the degree of success with which we
engage in an epistemic activity, as assessed by ourselves or by others.
It is futile to try to tie understanding down to any particular activities. For
example, as Peter Lipton argues convincingly in this volume, it is a mistake
to think that understanding only comes from explanation (unless one defines
“explanation” as whatever increases understanding, in the manner of Michael
Scriven).20 Henk de Regt and Dennis Dieks (, –) offer a seemingly
more liberal notion of understanding: “A phenomenon P can be understood if
a theory T of P exists that is intelligible (and meets the usual logical, method-
ological and empirical requirements).” The intelligibility of theories, in turn, is
defined as follows: “A scientific theory T is intelligible for scientists (in context
C) if they can recognize qualitatively characteristic consequences of T without
performing exact calculations.”21 Even this notion of understanding, I think, is
too narrow, in two senses: it requires the existence of an intelligible theory, and
the criterion of intelligibility is tied to the activity of qualitative prediction.
With these insights, I can finally tackle something that has been an endur-
ing puzzle to me: what is the difference between simply applying an algorithm
to solve a problem, and doing the same with a sense of understanding? Simply
following an algorithm provides no relevant understanding to someone who
is interested in some other epistemic activity, for example, visualizing what is
going on, or giving a mechanical explanation. But for someone whose goal is
to derive a prediction, there is surely the relevant sense of understanding in
knowing how to apply the right tricks to derive the answer. (When I was an
undergraduate, learning physics in the standard way gave me no sense of un-
derstanding because I went into the enterprise expecting something else. It was
my mistake to think that my classmates who were not dissatisfied like me were
in some way shallow or unreflective; they were simply happy with the epistemic
activity they were engaged in.)
Any well-performed epistemic activity can generate a sense of understand-
ing. Since people have different preferences about what kind of activities they
ontol o gic al principle s and episte mic ac ti vitie s 

appreciate and want to engage in, what they consider “real” understanding will
vary. This creates an ineliminable dimension of subjectivity in the concept of
understanding, which is due to the differences in people’s epistemic aims. I
believe that this essential subjectivity is closely related to the contextuality in
scientific understanding that de Regt stresses.

I]ZLVn6]ZVY

In lieu of a conclusion, I will present some preliminary thoughts, which will

be worth considering more carefully if the way of thinking I have laid out in this
paper is solid and fruitful.

Di]ZgEg^cX^eaZ"6Xi^k^inEV^gh
 First of all, I think there are many other pairs of basic epistemic activity and
ontological principle yet to be discovered. The following are worth considering,
though I cannot yet offer them with confidence.

Voluntary Action ' Agency. Engaging in voluntary action makes no

sense unless we presume that our will directs certain parts of our body to move
in certain ways.22 That presumption may be called the principle of agency.

Intervention ' Causality. Similarly, intervening in the world with our

own actions requires a presumption that our actions do make things happen;
I am calling that the principle of “causality” for want of a better word, but this
needs to be distinguished from the notion of causality that putatively obtains in
the interaction of things between each other, which I think is only a metaphori-
cal projection of the causality in our own direct actions. Agency and causality
together provide the most fundamental schemata for our bodily actions.

Empathy 'Other Minds. In the activity of understanding other people’s

intentions and emotions (which I have called “empathy” for lack of a better
word), one needs to presume that there are other minds who possess intentions
and emotions like one’s own.

Observation 'Externality/Objectivity. Similarly, the activity of observa-

tion makes no sense unless we presume that there are entities “out there” whose
manifestations we cannot control at will. The last two activity-principle pairs
give us the most fundamental constituents of the world: external objects and
other minds.

Identification 'Identity of Indiscernibles. The identification of an object

 hasok chang

(or a property) as something of any description must be based on something

like Leibniz’s principle of the identity of indiscernibles.23 The principle itself
can be divorced from various uncertain uses that Leibniz himself (and others)
made of it.

Deductive Inference 'Noncontradiction. Even the principle of noncon-

tradiction may be an ontological principle, associated with the epistemic ac-
tivity of deductive inference. It may seem odd to say that a logical principle is
an ontological principle, but we can think of noncontradiction as the essential
nature of propositions involved in deductive inference; this may also help us
get beyond the idea that ontology is about the question of what “exists” in the
world. A new light may be thrown on the nature of logical necessity as well: the
principle of noncontradiction is necessary because of a pragmatic impossibil-
ity; we find it impossible to engage in inferential reasoning without it.

8dbeaZm:e^hiZb^X6Xi^k^i^Zh
So far I have only discussed very simple and elementary epistemic activi-
ties. But we hardly ever engage in such epistemic activities. What we do in real
life, even in the practice of mathematics or physics, are mostly very complex
and specific activities, such as, say, predicting the frequencies of molecular
radiation by solving the Schrödinger equation using certain techniques of ap-
proximation. How can we make sense of intelligibility and necessity pertaining
to these real activities? I think it is going to be a very interesting and informa-
tive challenge to extend my analysis to complex epistemic activities in various
areas. A complex activity will involve multiple basic activities, each with its
attendant ontological principle. Therefore, all of those ontological principles
have to be satisfied for the complex activity to be intelligible. Each particu-
lar epistemic activity will be shown to have a particular set of conditions of
intelligibility.
There are some very interesting and common conjunctions of basic epis-
temic activities that play essential roles in science and everyday life. I will only
discuss a few of them briefly here to give a flavor of the insights we might gain
from a thorough analysis of complex epistemic activities. The activity of giving
mechanical explanations can be seen as a synthesis of contrastive explanation,
narration, imagined intervention, and perhaps visualization as well. Visualiza-
tion, I think, is supported by a set of principles that form the basis of Euclidean
geometry. (This explains why non-Euclidean geometry is deeply unintelligible
to those of us who try to visualize what is going on there.)24 The construction
of physical objects is another crucial activity, but it cannot be conceived fully in
terms of the basic activities and principles I have discussed so far in this paper.
It involves first of all the assumption that every object is “located” in space and
ontol o gic al principle s and episte mic ac ti vitie s 

time (“physicalism” as meant by Otto Neurath). We also presume that all sen-
sations attributed to one and the same object are harmonious with each other
(for example, a big sharp edge felt in gently running a finger along a completely
smooth-looking surface is not intelligible). Quite possibly there is a presump-
tion of causality, too, embodied in the notion of a physical object. We can see
that the analysis of a complex epistemic activity is a challenging task that will
require imagination as well as care.
It would be wrong to think of a complex activity as made up of basic activi-
ties in an atomistic way. All distinct activities to which we give definite char-
acterization are abstractions from what we actually do. Although it is often
instructive to think of a complex abstraction as constructed from simpler ab-
stractions, it would be a mistake to build up a whole universe of activities from
basic activities.25 Rather, the important thing to note is the relationships hold-
ing between the domains of applicability of various activity-descriptions. If an
act that qualifies as an instance of activity A always qualifies as an instance
of activity B as well, but not vice versa, we can say that activity B is more ba-
sic than activity A. Applying that idea to the basic epistemic activities that I
have discussed so far, I find that some of them are more basic than others. For
example, testing-by-overdetermination always involves inference (as well as
attempted overdetermination itself ). Therefore, inference is a more basic activ-
ity than testing-by-overdetermination, and the latter requires the principle of
noncontradiction in addition to the principle of single value. Similarly, hardly
any other activity can take place without identification.

JcYZghiVcY^c\!Igji]!VcYi]Z6^bhd[HX^ZcXZ
I will close with a very brief remark on intelligibility and understanding in
relation to the aims of science. In my earlier essay I was at pains to emphasize
that understanding was a distinct goal from description and prediction, and
that it was an important aim of science in its own right (Chang , –).
Now, with the broadened sense of understanding that I am proposing in this
chapter, understanding becomes part of all the aims of science and all other
epistemic activities since understanding simply comes from performing any
epistemic activity well.
What about truth, then? I will limit myself to a few suggestive sentences for
now. It would be nice if we had a simple activity of truth-finding, but frankly
I do not know what that would look like. What scientific realists should hope
for is that there might be some complex activities the successful performance
of which would allow us to have a reliable sense of attaining truth. I used to
believe that truth and understanding were separate and equal aims. Now I can
see truth is only a possible by-product of certain types of understanding.
 hasok chang

CdiZh
I would like to thank Henk de Regt, Sabina Leonelli, and Kai Eigner for inviting me
to speak at the Amsterdam conference on scientific understanding, in , and Dennis
Dieks for chairing the session and allowing a great deal of discussion. I cannot list here all
of the conference participants who gave me highly encouraging and stimulating feedback.
I thank Roberto Torretti for highlighting in my previous work the idea that formed the
central core of this chapter, and for some detailed comments on an earlier draft. I would
also like to thank the following people for their helpful criticism and suggestions, most of
which I regret not having been able to make full use of: Henk Procee, Mieke Boon, Rasmus
Winther, Peter Morgan, Eric Schliesser, Maarten Van Dyck, Marcel Boumans, Bill Brewer,
Hans Radder, Matt Lund, Michela Massimi, and Grant Fisher.
. I use the term epistemic broadly to mean “in relation to knowledge,” without tying
it to the concept of truth, in order to avoid prejudging various important issues in favor of
standard scientific realism. This marks one of the few disagreements I have with Henk de
Regt, on his concession that being “epistemically relevant” means being “truth-conducive”
(de Regt ,  n. ).
. This text, titled “Critical Remarks Concerning the General Part of Descartes’ Prin-
ciples,” was composed in . It was a paragraph-by-paragraph refutation of Descartes’
Principles of Philosophy, intended for publication in an edition that juxtaposed it against
Descartes’ text. It only appeared posthumously.
. Descartes may have sought to avoid the continuity problem by the word “somewhat”
in the statement of Rule . In that case, Descartes’ problem would be incompleteness.
. Descartes should at least get due credit for being painfully aware of the difficulty of
establishing certainty anywhere. I am not going to debate here the correctness of Des-
cartes’ epistemology and metaphysics.
. The Kantian project could be interpreted in a way that avoids a commitment to the
necessary truth of the basic theories of science, and that is something I hope to achieve in
future work. See Chang (), for a start.
. See de Regt and Dieks (, –), for an instructive discussion of the context-
dependence of intelligibility.
. What I present here is a superior solution to the one I proposed in Chang (),
–.
. See Chang (, , ). In experimental practice, “disagreeing” means differing so
much as to go beyond the admitted margin of error. I have shown how this principle has
done useful work, although it might seem entirely vacuous (Chang , –).
. The nonempirical nature of the principle of single value allows it to support thought
experiments; for an example, see the discussion of Galileo’s thought experiment showing the
independence of acceleration on mass in free fall in Peter Lipton’s chapter in this volume.
. See Chang (, –). I thank Roberto Torretti for getting me to think about
this point in a fresh way (Torretti, private communication, March  and , , and April
, ).
. Belatedly I discovered the work of C. I. Lewis, a system of “conceptual pragmatism,”
in which he advances a “pragmatic conception of the a priori”; see Lewis (/) for
ontol o gic al principle s and episte mic ac ti vitie s 

the most definitive summary. For my reflections on Lewis and ontological principles, see
Chang ().
. By an “epistemic activity” I mean a system of mental or physical actions contribut-
ing to the production or improvement of knowledge in a particular way, and following
some discernible rules (though the rules may be unarticulated). This is not intended as a
precise definition. For a more extensive characterization of epistemic activities, see Chang
(forthcoming).
. I want to argue that other types of impossibility are actually grounded in pragmatic
impossibility, being metaphorical extensions of the latter. This is why it would be futile to
try to analyze pragmatic impossibility further. In working out the notions of necessity and
possibility sketched here, I would like to build connections to Roberto Torretti’s (,
chap. ) ideas on the subject.
. See Salmon (), from which I also take the phrase “rational prediction.”
. Depending on the exact formulations, it could be that the principle of sufficient
reason is just the contrapositive of the principle of uniform consequence.
. My thinking about this particular activity-principle pair was inspired by the work
of my former Ph.D. student Georgette Taylor on affinity tables in eighteenth-century
chemistry.
. Compare Carlson () for an insightful argument that there were signifi-
cant Kantian elements in William James’s thinking, despite the latter’s renunciation of
Kantianism.
. This is reminiscent of Duhem’s (/) argument that when we witness the
extensive organization of phenomena through a successful theory, “it is impossible for us
to believe that this order and this organization are not the reflected image of a real order
and organization” (–); he also cites Poincaré in that discussion.
. The last statement is really a tautology (and therefore certainly true!), since I have
defined intelligibility as performability.
. Scriven (, ) states: “The question we have to answer is how and when cer-
tain descriptions count as explanations. . . . Explaining therefore sometimes consists simply
in giving the right description. . . . The right description is the one which fills in a particular
gap in the understanding of the person or people to whom the explanation is directed.”
. For an earlier articulation of the intelligibility criterion, see de Regt (, ).
. Whether the will itself is free or not is a separate question. Along vaguely Kantian/
Jamesian lines, I would imagine that the freedom of the will is an ontological principle
concomitant to the activity of making moral judgments (compare Carlson , –).
. I thank Roberto Torretti for this suggestion.
. This is not to say that every Euclidean situation is visualizable.
. The limitations are similar for any other kind of compositional reductionism. See,
for example, the antireductionist arguments in Dupré (, chap. ).

GZ[ZgZcXZh
Bergson, H. /. Creative evolution. Translated by A. Mitchell. London: Macmillan.
Carlson, T. . James and the Kantian tradition. In The Cambridge companion to Wil-
 hasok chang

liam James, edited by R. A. Putnam, –. Cambridge: Cambridge University Press.

Chang, H. . How to take realism beyond foot-stamping. Philosophy :–.
———. . Contingent transcendental arguments for metaphysical principles. In Kant
and philosophy of science today, edited by M. Massimi, –. Cambridge: Cam-
bridge University Press.
———. Forthcoming. The philosophical grammar of scientific practice. Paper delivered
at the first biennial conference of the Society for Philosophy of Science in Practice
(SPSP), University of Twente, August .
de Regt, H. W. . Discussion note: Making sense of understanding. Philosophy of Sci-
ence :–.
de Regt, H. W., and D. Dieks. . A contextual approach to scientific understanding.
Synthese :–.
Descartes, R. /. Principles of philosophy. Translated by B. Reynolds. Lewinston,
NY: Edwin Mellen Press.
Duhem, P. /. The aim and structure of physical theory. Translated by P. P. Wiener.
New York: Atheneum.
Dupré, J. . The disorder of things: Metaphysical foundations of the disunity of science.
Cambridge, MA: Harvard University Press.
Feyerabend, P. . Against method. London: New Left Books.
Frank, P. . Modern science and its philosophy. Cambridge, MA: Harvard University
Press.
Friedman, M. . Dynamics of reason. Stanford: CSLI Publications.
Kockelmans, J. J., ed. . Philosophy of science: The historical background. New York:
Free Press.
Leibniz, G. W. /. Critical remarks concerning the general part of Descartes’ Prin-
ciples. In Monadology and other philosophical essays, edited by P. Schrecker and A. M.
Schrecker, –. New York: Macmillan.
———. . Leibniz selections. Edited by P. P. Wiener. New York: Charles Scribner’s Sons.
———. . Philosophical essays. Translated and edited by R. Ariew and D. Garber. India-
napolis: Hackett.
Lewis, C. I. /. Mind and the world-order: Outline of a theory of knowledge. New
York: Dover.
Salmon, W. C. . Rational prediction. In The limitations of deductivism, edited by A.
Grünbaum and W. C. Salmon, –. Berkeley and Los Angeles: University of Califor-
nia Press.
Scriven, M. . Explanations, predictions, and laws. In Scientific explanation, space, and
time, edited by H. Feigl and G. Maxwell, –. Minneapolis: University of Min-
nesota Press.
Torretti, R. . Creative understanding: Philosophical reflections on physics. Chicago:
University of Chicago Press.
van Fraassen, B. C. . The scientific image. Oxford: Clarendon Press.
reliabilit y and the sense of under standing

,
Reliability and the Sense of Understanding

HI:E=:CG#<G>BB

 If we are fortunate, at some point while pursuing the answer to one

of our explanation-seeking why-questions we will experience a sense of un-
derstanding. In other words, we will seem to “grasp” or “see” what it is that
accounts for the thing we want to explain, a moment of “grasping” or “seeing”
that is often accompanied by a distinctive phenomenology—perhaps even a
phenomenology along the lines of the celebrated “aha” experience.
In a series of articles J. D. Trout (, , ) argues that the sense of
understanding that we enjoy at these moments is deeply unreliable.1 In other
words, he argues that our sense that we can “grasp” or “see” why things are one
way rather than another is poorly connected with the truth about why things
are one way rather than another. If correct, this would have wide-ranging im-
plications for our ability to understand the world. For instance, if Trout is right,
then on the uncommon occasion when our sense of understanding actually
does manage to grasp the truth, the grasp will presumably be too accidental or
tenuous to have much if any positive epistemic status. For comparison, suppose
you claim to have the ability to determine whether someone is honest just by
looking at them. If it turns out that most of the people that you classify as hon-
est are inveterate liars, then on the few occasions when you identify a genuinely
honest person we would not say that your identification had much going for it
in the way of positive epistemic status. To earn that a more secure link to the
truth is required.
In the following two passages, Trout suggests that the history of science is
replete with evidence against the reliability of the sense of understanding.


 stephen r . gr imm

Peirce identifies the distinctive cognitive experience of explanatory

understanding by isolating the final moment of acceptance; the good
explanation “is turned back and forth like a key in a lock.” This de-
scription alone should supply little solace to those holding that good
explanations are epistemically reliable. After all, alchemists surely felt
the key turn, but once inside we find only false descriptions of causal
mechanisms. And when Galen arrived at a diagnosis of melancholy
due to black bile, his sense of understanding was so gratifying it must
have balanced his humors. (, )

Later, he is even more emphatic:

The fact is, our history is littered with inaccurate explanations we con-
fidently thought were obviously true: the explanation for mental ill-
ness in terms of demonic possession, the humoral theory of illness,
and so on. The sense of understanding would be epistemically idle
phenomenology were it not so poisonous a combination of seduction
and unreliability. It actually does harm, sometimes making us squea-
mish about accepting true claims that we don’t personally understand,
and more often operating in the opposite direction, causing us to over-
confidently accept false claims because they have a kind of anecdotal
or theoretical charm. (, –)

By Trout’s lights, apparently, the unreliability of our sense of understanding

is therefore just obvious; at any rate, even a casual look at the history of sci-
ence confirms it.2 The list of those who were thoroughly, albeit mistakenly, con-
vinced that they could “grasp” or “see” why things were one way rather than
another—not just past figures such as Galen, Ptolemy, and the alchemists, but
also more recently astrologers, CIA conspiracy theorists, and so on—goes on
and on.
In this chapter I will argue, first, that a bit of armchair empirical research
suggests that Trout dramatically exaggerates the de facto unreliability of our
sense of understanding. Second, I will point out that current clinical research
reinforces what our armchair empirical research suggests: namely, that as a
simple matter of record (and by and large), we are quite good at identifying why
things are one way rather than another. One tempting conclusion to draw from
this research is that typically, and contra Trout, our sense of understanding
simply is reliable. Call this the “optimistic conclusion.” I will suggest below that
the optimistic conclusion too is mistaken. What we need is a compromise posi-
reliabilit y and the sense of under standing 

tion, one that recognizes the presumed positive epistemic status of many of the
exercises of our sense of understanding while also making sense of the fact that
astrologers, conspiracy theorists, and the like can go so systematically wrong.
Before we begin, it will help to bring a bit more order to the notion of “the
sense of understanding.” When Trout claims that the sense of understanding
is unreliable, what he seems to mean by this is that the phenomenological re-
sponse we feel while entertaining certain explanatory stories rather than others
is poorly correlated with the truth of those explanatory stories. In other words,
that more often than not (or at least, not often enough for some other reliability
threshold Trout has in mind) when a certain explanatory account “feels right”
to us, it turns out to be wrong.
Like its cousins the “sense of sight” and the “sense of hearing,” however, it is
important to see that at a more basic level the “sense of understanding” seems
to refer to the exercise of an ability or faculty that undergirds the phenomenol-
ogy. Thus, when we seem to “grasp” or “see” what it is that accounts for the
thing we want to explain, even when this grasp turns out to be mistaken, the ex-
perience of grasping is nonetheless distinctively active. Perhaps, for example, as
James Woodward () suggests, it is active in the sense that it brings with it
the ability to answer what-if-things-had-been-different questions. Or perhaps,
as Henk de Regt and Dennis Dieks () claim, it brings with it the ability to
recognize the consequences of certain theories without needing to perform
exact calculations (see also de Regt in this volume). However we ultimately
characterize this ability, the basic point that bears emphasizing is that there is
some sort of competence in play here, and that its operation is what undergirds
the phenomenology emphasized by Trout.
Given this more layered view of the sense of understanding, we can there-
fore think of Trout’s claim along the following lines: to say that our sense of
understanding is unreliable is to say the connections or correlations that it
“grasps”—where the graspings are in turn accompanied by a distinctive phe-
nomenology—turn out to be false more often than not.
One final preliminary: since the purpose of this chapter is to evaluate
whether and in what sense our sense of understanding is reliable, it is impor-
tant to bear in mind how we normally assess reliability. If we are evaluating
whether a certain thermometer is reliable, for instance, it is no strike against
the reliability of the thermometer to point out that it would fail to give an ac-
curate reading on the sun.3 Or again, if we are evaluating whether Ichiro Suzuki
is a reliable hitter of baseballs, then it is no strike against Ichiro’s reliability to
point out that he would fail to make contact with the ball while in the dark
or while blindfolded.4 Judgments about reliability, in other words, are always
 stephen r . gr imm

made relative to a certain environment. I will simply assume throughout this

chapter that the environment that is relevant to the assessment of our sense of
understanding is the environment in the “normal” world—which is to say, at
least, not a Cartesian demon world, or a world where all the walls have been
replaced with funhouse mirrors, and so on.5

6gbX]V^g:kVajVi^dc

One immediate worry about Trout’s complaint is that he fails to specify a

reference class for his unreliability claim. In pointing out that astrologers, al-
chemists, and the like have “seen” or “grasped” connections that are not there,
he then immediately attributes the unreliability of these apparent graspings to
the unreliability of the sense of understanding, as if our sense of understanding
were in some way unreliable at the root. But that does not follow. Just as from
the fact that the visual system of cataract sufferers is unreliable nothing follows
about the typical reliability of “the” visual system, so too from the fact that the
sense of understanding of a certain subgroup of people is unreliable nothing
follows about the typical reliability of “the” or “our” sense of understanding.
A further problem is that even at the individual level the reliability of some-
one’s sense of understanding is apt to vary. Consider Ptolemy again. Trout is
right to point out that Ptolemy’s sense of understanding led him astray on oc-
casion, sometimes spectacularly. One thing that focusing on the spectacular
failures overlooks, however, is that there were presumably countless other oc-
casions throughout Ptolemy’s life when his sense of understanding served him
well. We can only speculate about what exactly these occasions might have
been, but we can imagine that there were many: his sense that the fire was fail-
ing to catch because the wood was wet, for example, that the horse was limping
because it just suffered a bad fall, that Ptolemy Jr. was slurring his words be-
cause he just drank a bowl of wine, and so on. Assuming this is right, then the
correct thing to say about Ptolemy’s sense of understanding is not that it was
unreliable simpliciter but rather that, at most, it was unreliable with respect to
some domains and reliable with respect to others.

8a^c^XVa:kVajVi^dc

Does any non-armchair, clinical research support this suggestion? Evidently,

yes. In a series of publications, the cognitive psychologist Alison Gopnik (along
with a number of research partners) argues that human beings are built with
a distinct cognitive system for identifying causal dependencies in the world.6
According to Gopnik, the system basically functions as an input-output device:
reliabilit y and the sense of under standing 

it takes as its input patterns of observed contingency and correlation among

phenomena, and then mines these observations to identify apparent depen-
dencies among the phenomena in order to construct what Gopnik calls “causal
maps” of the world.
So how reliable is this system—are we—in terms of constructing accurate
causal maps? What clinical studies on young children suggest is that the system
is, in fact, typically quite reliable. As described by Gopnik and Glymour,

In several studies, we have begun to show that children as young as

two years old, in fact, do swiftly and accurately learn new causal rela-
tions—they create new causal maps. They do so even when they have
not themselves intervened to bring about an effect, and when they
could not have known about the relation through an evolutionarily
determined module or through prior knowledge. We present children
with a machine, the “blicket detector,” that lights up and plays music
when some objects but not others are placed in it. Children observe
the contingencies between the objects and the effects and have to in-
fer which objects are “blickets.” That is, they have to discover which
objects have this new causal power. Children as young as two years
old swiftly and accurately make these inferences. They identify which
objects are blickets and understand their causal powers. . . . Even very
young children use a form of “screening off ” reasoning to solve these
problems.7 (, )

In other words, what studies by Gopnik and her colleagues suggest is that not
only do we have a cognitive system dedicated to constructing causal maps of
the world that is apparently hard-wired (given our ability to construct these
maps from a very young age), but that we also have a natural aptitude for the
task. Constructing accurate causal maps is the kind of thing, apparently, that
we are naturally good at doing.
Alongside Gopnik’s research with children, Mark Steyvers and his col-
leagues () have reproduced similar findings with adults. In order to weed
out potential differences in background information (a point that will reap-
pear as significant later), Steyvers and his coauthors designed an experiment
in which subjects were asked to identify causal relationships among three hy-
pothetical “alien mind-readers.” At the beginning of the experiment subjects
were shown a picture displaying the thoughts of three aliens (of the Hollywood
variety), and they were asked to identify which alien was reading the minds of
the others.8
What Steyvers found was that not only were subjects quite good at identify-
 stephen r . gr imm

ing how the alien mind-reading system worked based on passive observation of
the system, but that when they were allowed to manipulate the interactions their
reliability increased dramatically. As Steyvers and his colleagues summarize,

Faced with the challenge of inferring the structure of a complex causal

network, and given no prior expectations of what causes what, people
bring to bear inferential techniques not so different from those com-
mon in scientific practice. Given only passive observational data, peo-
ple attempt to infer a system’s underlying structure by comparing what
data they see to what they would expect to see most typically under
alternative causal hypotheses. Given the opportunity to learn actively
from observing the effects of their own interventions on the system,
they make more accurate inferences. ()

Combined with Gopnik’s work, both studies suggest that the very meth-
odology that a mature science uses to identify dependencies in the world—of
searching for correlations in the data, of using control groups, and so on—is in
some sense the built-in method that we use in our attempts to understand the
world. Given that we think the scientific method is a reliable way to identify
whatever it is that accounts for what we want to explain, it should therefore
come as no surprise that our sense of understanding turns out to be as reliable
as it is.
How, if at all, does the phenomenology of understanding that Trout empha-
sizes enter into this picture? Here again Gopnik (, ) provides a nice
insight. As she points out, once one appreciates the potential power that comes
from grasping how things depend on one another in the way specified above, it
is not particularly surprising that, over the course of time, a special kind of phe-
nomenology should have come to be associated with the exercise of our sense of
understanding. As Gopnik suggests, just as it “made sense” for Mother Nature
to associate sexual orgasm with its own pleasurable phenomenology in order to
encourage reproduction, in the same way it would presumably have made sense
for graspings or seeings to acquire their own distinctive, and distinctively desir-
able, phenomenology: a phenomenology that Gopnik characterizes in terms of
the “aha” experience. Since the ability to grasp dependencies in this way—and
thus to potentially manipulate one’s environment—is obviously a fitness-en-
hancing characteristic, it is natural to suppose that, over time, evolution would
have found a way to impress the goodness of such graspings upon us.9
Whether or not Gopnik is correct in characterizing the phenomenology
that accompanies the sense of understanding in terms of the “aha” experience is
reliabilit y and the sense of under standing 

something we can set to one side here. My point is much more modest: namely,
that there is something that is distinctively like to “see” or “grasp” (or, at least,
to seem to see or to seem to grasp) why things are one way rather than another.
Whether one prefers to describe this phenomenology in terms of something
as dramatic as the “aha” experience, or whether instead one prefers to think of
the phenomenology in more subdued terms, the basic idea is that at least some
phenomenology accompanies the exercise of our understanding system, a sys-
tem that is apparently well connected with how things stand in the world.

8dggjei^dch

One possible conclusion to draw from these studies is therefore that, contra
Trout, our sense of understanding simply is reliable at identifying dependen-
cies—at least characteristically and under normal conditions. We might think
of this as the “optimistic conclusion.” But the optimistic conclusion is mislead-
ing; at the very least, it needs to be qualified significantly. The optimistic con-
clusion suggests that even though our sense of understanding is susceptible to
unreliability (the astrologers and the alchemists have to be accounted for, after
all), the unreliability is the result of some sort of external corruption rather than
the result of a problem inherent in the normal operation of the system—just as,
we might think, although our visual system is susceptible to unreliability via ex-
ternal corruption (cataracts, mind-bending drugs, and the like), in the absence
of such external corruption it is basically reliable. To put the point another way,
what the optimistic conclusion suggests is that when our sense of understand-
ing is functioning properly in its normal environment, it will be largely reliable,
just as when our visual system is functioning properly in its normal environ-
ment, it too will be largely reliable.
What I want to propose instead is that even a properly functioning sense
of understanding in its normal environment might be thoroughly unreliable at
identifying dependencies. In addition to its intrinsic interest, another reason
this would be an important result, if true, is that it would help to explain why
eminent thinkers such as Ptolemy and Galen did identify and affirm extremely
strange dependencies, in a way that would not have to attribute to them some
sort of malfunction in their sense of understanding.
In the remainder of this chapter I will consider two reasons—or, as I will
put it below, “types” of reasons—why even a properly functioning sense of un-
derstanding might systematically produce bad results. The first of these has to
do with problems with intellectual vices and the second with problems with
inaccurate background beliefs.
 stephen r . gr imm

IneZ&HdjgXZhd[JcgZa^VW^a^in/>ciZaaZXijVaK^XZh

The basic idea behind Type  sources of unreliability is that even in normal
environments, and even when the information we have at our disposal is quite
good, the influence of so-called “intellectual vices” can significantly diminish
the reliability of our sense of understanding.10
By way of illustration, suppose that Lisa is tracking the progress of one hun-
dred cold sufferers who are trying a new product, Cold Away!11 As it happens,
all one hundred of the sufferers recover from their cold within a week. Based
on the strong correlation between the taking of Cold Away! and the timely relief
from their colds, Lisa concludes that it was the Cold Away! that made the differ-
ence to the group’s improvement.
This will be a bad inference, naturally, and the problem (roughly put) can
be traced to the lack of anything resembling a control condition. At a mini-
mum, in order to determine whether it was the Cold Away! that made a dif-
ference to their recovery, Lisa should have checked to see how cold sufferers
who did not take Cold Away! fared. In this case, what she would have found is
that even those who failed to take Cold Away!—that is, even those who took
nothing—would have likewise recovered within a week. Colds are simply the
kinds of things that, given a little time, run their course. But then, presum-
ably, it was not the Cold Away! that made the difference to the cold sufferers
after all.
Crucially, however, the source of the failure in this case does not seem to be
Lisa’s sense of understanding—it is not as if, for instance, her sense of under-
standing was misled by the equivalent of cataracts or occlusions—but rather
Lisa’s handling of her sense of understanding.12 To draw again on an analogy
with the visual system, the more time you give a properly functioning visual
system to take in the world, the more likely it will be to represent the world ac-
curately. By contrast, the more you rely on quick glimpses and the like, the less
likely it will be to represent the world accurately—hence the less likely that be-
liefs formed on the basis of such information will be accurate. In the same way,
however, the more information Lisa feeds into her sense of understanding—in
the form of wide and varied samples, control groups, and so on—the more like-
ly it will be that she identifies genuine dependencies. By contrast, the more she
fails to provide her sense of understanding with this information—the more
she jumps to conclusions based on a small amount of data, for example—the
less likely it will be able to identify genuine dependencies.
I have characterized the source of Lisa’s unreliability in terms of a kind of in-
tellectual carelessness, perhaps even a kind of intellectual laziness: she jumped
to conclusions without bothering to acquire all of the relevant data. Needless to
say, however, there are dozens of other intellectual faults that likewise can con-
reliabilit y and the sense of under standing 

tribute to the unreliability of someone’s sense of understanding. For example,

George might have a wide and varied set of information at his disposal, but he
might nevertheless be guided by his prudential interests (his interest in seeing
certain patterns, for whatever reason) to emphasize certain parts of the data
and to deemphasize other parts. Or again, because of intellectual arrogance
someone might fail to consider alternative hypotheses or might reject as faulty
data that fail to support his or her own hypothesis. In all of these cases, how-
ever, the problem will lie not with the faculty itself but rather with someone’s
careless (or perhaps simply incompetent) handling of the faculty.

IneZ'HdjgXZhd[JcgZa^VW^a^in/7VY7VX`\gdjcY7Za^Z[h

Another reason—perhaps the more likely reason—why even those with

properly functioning senses of understanding might systematically go wrong is
because of bad (that is, false) background beliefs. Even when everything else is
ideal—even when, for example, one has the benefit of a rich and varied amount
of data, even when one pours through the data in the most careful and atten-
tive way—one’s background beliefs will still have an enormous influence on the
sorts of possibilities that one takes seriously in the first place.
The following case helps to bring this out. Suppose Albert is an atheist and
Paul a theist, and both are wondering why a certain group of patients, Group A,
recovered from their illnesses while another group, Group B, failed to recover.13
Albert and Paul subsequently learn that all of the patients in Group A were be-
ing prayed for by a group of pious nuns several thousand miles away and that
none of the patients in Group B were being prayed for by the nuns. Moreover,
they learn that the nuns’ prayers for the patients in Group A were, in fact, part
of a rigorous, double-blind experiment. None of the patients in either group
was aware that they were being prayed for, and other possible countervailing
influences were carefully screened out.
As an atheist who believes that petitionary prayers are useless, learning
about the results of the study leads to no graspings or seeings for Albert—no
sense that the prayer-healing connection “feels right.” By his lights, the thing
that is identified as a difference maker here—petitionary prayer—simply is not
qualified to be a difference maker.14 He therefore comes to think not that the
recovery of the patients in Group A depended on the nuns’ petitionary prayers,
but rather than there was something wrong with the study.
Paul, by contrast, not only believes in God but also believes in the power of
petitionary prayer. While he may not immediately accept that it was the nuns’
prayers that made the difference between Group A and Group B, he will none-
theless find the results worth taking seriously. If he learns of further studies
that also indicate the existence of such a correlation, moreover, he will then
 stephen r . gr imm

likely conclude that it was, in fact, the prayers that were making the difference
on these occasions.
Both Albert and Paul are therefore interestingly like Ptolemy in one respect
at least. Presumably Ptolemy’s belief that it was the earth that stood still and
that it was the sun that moved was so well entrenched (after all, just take a
look! which one seems to be doing the moving?) that this belief automatically
ruled out any explanatory story that portrayed the earth as a mover. What is
more, one of the two—either Albert or Paul—is also like Ptolemy in the fol-
lowing sense: not only is one’s belief about God false (as Ptolemy’s geocentric
belief was false), but it will be false in a way that negatively affects many of the
other conclusions he draws. If God does not exist, then Paul will be identifying
countless dependencies that simply are not there (petitionary prayer being only
one of them), not to mention (perhaps) overlooking the dependencies that are
there. If God does exist, however, then it will be Albert who will be missing out
on countless dependencies: he will be failing to grasp what actually makes the
difference in all of these cases.15

8dcY^i^dcVaGZa^VW^a^in

Bearing all this in mind, our answer to the question, “Is our sense of under-
standing reliable?” should therefore be: it depends. More precisely, our answer
should be: the faculty is conditionally reliable, and this in several respects.16
One respect in which the faculty is conditionally reliable is clear-cut: if you
get bad data in you will almost inevitably get a bad grasping out. Specifically, if
the information that you are mining is full of inaccuracies, then the dependen-
cies you identify on the basis of that information will likewise be inaccurate. As
we saw with respect to Type  cases, this is also true in a more subtle sense: if
you have bad background beliefs, then scores of accurate explanatory stories
will not even receive consideration from you. In both of these cases, however, it
would be a mistake to say that the problem lies with the faculty or ability per se.
Instead, the problem has something to do with the quality of the information
that is being fed into the faculty.
With this in mind, moreover, when we look back to the Gopnik-Steyvers ex-
periments above, we can now appreciate one of the reasons why the subjects in
the experiment, first the children reasoning about blickets and then the adults
reasoning about alien mind-reading, performed so uniformly well. For notice
that in both experiments the problem of false background beliefs was essentially
factored out. For the children, the background beliefs were doubly factored out:
first, because children lack a large stock of background beliefs to begin with,
and second because the novelty of the blickets experiment required them to
approach the situation without any preconceptions. With the adults, the pe-
reliabilit y and the sense of under standing 

culiar alien mind-reading experiment was likewise chosen precisely because of

its novelty. Lacking any preconceived notions of alien mind-reading skills, the
adults were able to evaluate the relationships among the variables with few if
any of the preconceptions that usually shape our view of the world.
We can also now appreciate how our sense of understanding is condition-
ally reliable in a still further respect: namely, the reliability of our sense of un-
derstanding depends not just on the quality of our information but also on how
the information is used. Intellectual character is relevant to reliability because,
as Type  cases show, even if one’s background beliefs are true and the infor-
mation one has about a particular situation is large and varied, there is always
a temptation to train one’s focus on some parts of the data rather than others,
either out of a vain desire to protect one’s own views or out of a prudential de-
sire to find certain patterns rather than others.17
So far, I have argued for what I take to be a moderate conclusion: not that
“the” sense of understanding is reliable (whatever exactly that might entail), but
rather that it is conditionally reliable, or reliable so long as our background be-
liefs are more or less sound, our intellectual practices are more or less virtuous,
and so on. More optimistically, it seems plausible to say that many of us in fact
do satisfy these further conditions—hence that, when all goes well, we do enjoy
a certain understanding of the world.
Still, Trout might object that to this point my account has blithely over-
looked a crucial element of his case against the de facto reliability of our sense
of understanding.18 In particular, Trout might claim the account so far simply
fails to address his discussion of recent studies concerning hindsight bias and
overconfidence bias to the effect that our sense of understanding is deeply and
pervasively unreliable. Before turning to the next section, I should therefore say
a word about why I am skeptical about these studies.
Suppose, by way of illustration, that you ask me a blunt question: whether
or not I think my daughter is above average. Considering it for a moment, it
seems to me—more or less irresistibly—that, yes, she is above average. I then
become aware of the literature pointing out that this kind of Lake Wobegon
effect is pervasive: the vast majority of parents think their children are above
average, but they cannot all be right. Indeed, I then find further evidence that
these sorts of mistakes are alarmingly common. It sometimes seems to me
more likely that Linda is a bank teller and a feminist than simply a bank teller.
It sometimes seems to me that when my lucky lottery numbers have not ap-
peared for a while they are due. And so on. On the basis of this evidence, I then
infer the following conclusion: sometimes when certain propositions seem true
to me, they are not (in fact) true.
But what exactly is supposed to follow from this conclusion? In keeping
with Trout’s line of thought, the idea seems to be that, in light of these studies,
 stephen r . gr imm

when a certain proposition seems true to me, and when I am thereby inclined
to assent to it, I should resist this inclination. If what I want is knowledge, then
I should only put my faith in reliable indicators of truth, and the “seems right”
test is demonstrably not such an indicator.
But is that really good advice? More to the point, is that advice that it is even
possible to follow? Suppose you tell me that I should distrust the way things
seem to me and instead depend strictly on some other sort of evidence (per-
haps, statistical modeling). This is the only reliable way to get to the truth, you
say. As it happens, moreover, suppose that what you say seems right to me. I
feel like you are on to something. But now the problem seems obvious: Should
I distrust this feeling of seeming right? This feeling that I am on to something?
It seems, in sum, that relying on how things seem to us is simply our lot
in life—a maze from which we cannot escape—and that understanding seems
to have it no worse than everyday forms of belief on this score. I would also
continue to insist that the faults that we sometimes find in particular exercises
of the sense of understanding are, for most of us, less pervasive than Trout
suggests: perhaps, for many people, about as pervasive as my temptation to
be duped by something like the gambler’s fallacy. In any case, not pervasive
enough to underwrite the deep concerns about the reliability of the sense of
understanding that worry Trout.

8d]ZgZcXZGZ\V^cZY4

Studies concerning hindsight bias and the like aside, it therefore seems—to
pick up again on our earlier line of thought—that our sense of understanding
will be reliable to the extent that (among other things) our background beliefs
are accurate and varied and to the extent that we have a good cognitive charac-
ter that uses these beliefs appropriately.
We can now make sense, moreover, of one final point about the nature of
our graspings or seeings—and thus about the nature of understanding that sev-
eral commentators working at the intersection of epistemology and the philos-
ophy of science—including Jonathan Kvanvig (), Wayne Riggs (), and
Catherine Elgin ( and )— stress: namely, that coherence considerations
have a significant role to play in the exercise of our sense of understanding.19
According to Riggs, for example, understanding a particular bit of information
is essentially, perhaps even exclusively, a matter of seeing how it “fits together”
or “coheres with” the rest of one’s beliefs:

An important difference between merely believing a bunch of true

statements within subject matter M, and having understanding of M
(or some part of M), is that one somehow sees the way things “fit to-
reliabilit y and the sense of under standing 

gether.” There is a pattern discerned within all the individual bits of

information or knowledge. . . . The epistemological notion of “coher-
ence” and the idea of “explanatory coherence” in particular seems to
be getting very close to something characteristic of understanding.
(, )

According to Kvanvig, too,

The central feature of understanding, it seems to me, is in the neigh-

borhood of what internalist coherence theories say about justification.
Understanding requires the grasping of explanatory and other coher-
ence-making relationships in a large and comprehensive body of in-
formation. One can know many unrelated pieces of information, but
understanding is achieved only when informational items are pieced
together by the subject in question. (, )

I do not want to endorse—in fact, elsewhere (for example, Grimm ) I

actively dispute—the suggestion that understanding is fundamentally an inter-
nal affair, a matter of grasping how one’s beliefs cohere with one another. That
said, we can now see the sense in which coherence considerations do have a
crucial role to play in the exercise of our sense of understanding. For one thing,
as we mine our information about the world in a search for dependencies, our
mining is necessarily influenced by our background beliefs. As we have seen, if
you are convinced that it is the sun that moves and the earth that is still, then
when you are trying to explain (say) stellar parallax, you will focus on explana-
tory stories that do not portray the earth as a mover and rule out the rest. Or
again, if you have been brainwashed into believing that the CIA is responsible
for everything suspicious in the world, then when you wake up one morning
to find your tap water tasting funny, you will take seriously explanatory stories
that involve the CIA and disregard the rest. Our background beliefs therefore
constitute a kind of first filter (compare Lipton , chap. ), and possible de-
pendencies do not even get a hearing at this stage if they fail to pass the test.
After the first (possible/impossible) filter, moreover, we evaluate possible
dependencies in terms of how well they fit with the correlations that are already
familiar to us, either via testimony or direct observation. Whether or not we
accept that a certain dependency exists will therefore be a matter of how well
it makes sense of the rest of what we believe. If our background beliefs are
accurate, however, if we have a rich and varied amount of information to sift
through (whether consciously or not), and if we have good intellectual habits,
then, finally, we have reason to be optimistic: our chances of identifying genu-
ine dependencies (with any luck) would seem to be quite good.
 stephen r . gr imm

CdiZh
Thanks to Michael DePaul, Henk de Regt, Armond Duwell, John Greco, Sabina
Leonelli, Peter Lipton, J. D. Trout, and Fritz Warfield for helpful comments on earlier ver-
sions of this essay.
. De Regt , to which Trout  is a response, likewise addresses the question of
the reliability of the sense of understanding. Although my approach differs from de Regt’s,
we share similar concerns about Trout’s view.
. See also Trout (, , –), where “the epistemically unreliable relation
between the sense of understanding and genuine (that is, factive) understanding” () is a
central theme.
. Compare Alston (, ).
. Compare Greco (, –).
. As we will see, however, the reliability of someone’s sense of understanding will
depend heavily on the accuracy of the person’s background beliefs, and I will make no at-
tempt to specify what counts as normal with respect to background beliefs.
. See especially Gopnik and Glymour  and Gopnik et al. ().
. For more on the blickets experiment, see Gopnik and Sobel () and Gopnik et al.
().
. The alien “thoughts” were displayed in thought bubbles, and were quite basic. Thus,
in one experiment the first alien was thinking “TUS,” the second “POR,” and the third
“POR.”
. See also Lipton (this volume) for a similar appeal to Gopnik on this issue.
. Several epistemologists, especially those working in the virtue epistemology tradi-
tion, have stressed the connection between the intellectual vices and unreliability (see, for
example, Zagzebski ; Hookway , ; and Baehr ).
. The basic example is from (Salmon , sec. ); see also Bishop and Trout (,
–).
. In reality, of course, it would be very hard to pin the blame on the character flaw
rather than the faculty or ability; for all we know she might have some cataract-equivalent
problem with her faculty. In this case, though, we can simply stipulate that she does not.
. This case is loosely based on real life. In , the American Medical Association’s
Archives of Internal Medicine published a better-designed study of nearly a thousand
consecutive patients who were newly admitted to the coronary care unit of a hospital
in Kansas City. The researchers created a thirty-five-item score sheet that was used to
measure what happened to the patients during a twenty-eight-day period in which fifteen
groups of five persons (“intercessors”) prayed individually for about half the patients. The
intercessors were given the patients’ first names and were asked to pray daily for “a speedy
recovery with no complications.” The prayed-for group had a – percent reduction
in total scores even though their average length of hospital stay was similar to that of the
“usual-care” group. For further the details see Harris et al. ().
. At least, not from a distance; the fact that patients who pray for their own health
tend to improve as a result has been widely documented, and is now generally accepted.
reliabilit y and the sense of under standing 

Whether the improvement is due to the patient’s sense of optimism, or whether instead to
divine intervention, is an open question.
. Naturally, the unreliability of someone’s sense of understanding can be attributable
to more than one source. Thus, someone might be unreliable at identifying dependencies
both because he or she is intellectually careless and because he or she has bad background
beliefs.
. In Alvin Goldman’s () words, “A process is conditionally reliable when a suf-
ficient proportion of its output beliefs are true given that its input beliefs are true” ().
Deductive reasoning therefore counts as an example of a conditionally reliable process,
for example, in the sense that if you start with garbage (falsehoods), you will end up with
garbage (more falsehoods), whereas if you start with truth you will end up with truth.
. The opposite sort of person—namely, someone with a good and reliable cognitive
character—will therefore be someone who takes all of the data seriously, does not restrict
their attention to a limited range of information that might favor a certain pattern, does
not rush to conclusions from small sample sizes, and so on. In short, someone with a good
cognitive character will be someone who has essentially internalized the scientific method.
What’s more, as Gopnik () notes in the following passage, even if one has not internal-
ized the method in this way, one might still use the method as a kind of external corrective
to compensate for the unreliable tendencies in one’s intellectual character, much as one
corrects the results of an unreliable visual system by using glasses:
It appears that one of the differences, perhaps the most important cognitive
difference, between organized science and spontaneous theory formation is
precisely that science contains additional normative devices that are designed
to supplement the basic cognitive devices of the theory formation system, and
to protect them from error. We might think of science as a kind of cognitive
optometry, a system that takes the devices we usually use to obtain a veridical
picture of the world and corrects the flaws and distortions of those devices. The
fact that most people over forty wear glasses is not, however, usually taken as an
indictment of the visual system. In fact, the analogy might be taken even further,
perhaps science compensates for our deteriorating adult theory formation abili-
ties the way optometry compensates for our deteriorating adult vision” (;
italics mine).
. Indeed, Trout has objected in this way, in his comments on an earlier version of
this chapter at the  Central APA meeting. Thanks to Trout both for his comments,
and for the ensuing helpful discussion.
. So too, to a lesser extent, with Ernest Sosa:
Beyond “animal knowledge” there is a better knowledge. This reflective
knowledge does require broad coherence, including one’s ability to place
one’s first-level knowledge in epistemic perspective. But why aspire to any
such thing? What is so desirable, epistemically, about broad coherence?
Broad coherence is desirable because it yields integrated understanding, and
also because it is truth-conducive, even if in a demon world broad coherence
fails this test and is not truth conducive. (, ; italics mine)
 stephen r . gr imm

If I understand Sosa correctly, the suggestion is therefore that understanding (or at

least “integrated understanding,” if that is a distinct type of understanding) comes from
seeing, in some sense, how one’s beliefs cohere with one another. It also bears pointing out
that Kvanvig, Riggs, and Elgin might be reluctant to speak of a grasping “faculty,” as I have
done here; but I think the basic point remains the same.

GZ[ZgZcXZh
Alston, W. . How to think about reliability. In Epistemology: An anthology, edited by E.
Sosa and J. Kim. Oxford: Blackwell.
Baehr, J. . Character, reliability, and virtue epistemology. Philosophical Quarterly
:–.
Bishop, M., and J. D. Trout. . Epistemology and the psychology of human judgment.
New York: Oxford University Press.
de Regt, H. W. . Discussion note: Making sense of understanding. Philosophy of Sci-
ence :–.
de Regt, H. W., and D. Dieks. . A contextual approach to scientific understanding.
Synthese :–.
Elgin, C. . Considered judgment. Princeton: Princeton University Press.
———. . True enough. Philosophical Issues :–.
Goldman, A. . What is justified belief? In Epistemology: An anthology, edited by E.
Sosa and J. Kim, –. Oxford: Blackwell.
Gopnik, A. . Explanation as orgasm. Minds and Machines :–.
———. . Explanation as orgasm and the drive for causal knowledge: The function,
evolution, and phenomenology of the theory formation system. In Explanation and
cognition, edited by F. C. Keil and R. A. Wilson, –. Cambridge, MA: MIT Press.
Gopnik, A., and C. Glymour. . Causal maps and Bayes nets: A cognitive and com-
putational account of theory formation. In The cognitive basis of science, edited by P.
Carruthers, S. Stich, and M. Siegal. New York: Cambridge University Press.
Gopnik, A., C. Glymour, D. Sobel, L. Schulz, T. Kushnir, and D. Danks. . A theory
of causal learning in children: Causal maps and Bayes nets. Psychological Review
:–.
Gopnik, A., and D. Sobel. . Detecting blickets: How young children use informa-
tion about novel causal powers in categorization and induction. Child Development
:–.
Greco, J. . Putting skeptics in their place. New York: Cambridge University Press.
Grimm, S. R. . Is understanding a species of knowledge? British Journal for the Phi-
losophy of Science :–.
Harris, W. S., et al. . A randomized, controlled trial of the effects of remote, interces-
sory prayer on outcomes in patients admitted to the coronary care unit. Archives of
Internal Medicine :–.
Hookway, C. . Epistemic akrasia and epistemic virtue. In Virtue epistemology: Essays
reliabilit y and the sense of under standing 

on epistemic virtue and responsibility, edited by A. Fairweather and L. Zagzebski. New

York: Oxford University Press.
———. . How to be a virtue epistemologist. In Intellectual virtue: Perspectives from
ethics and epistemology, edited by M. DePaul and L. Zagzebski. Oxford: Oxford Uni-
versity Press.
Kvanvig, J. . The value of knowledge and the pursuit of understanding. Cambridge:
Cambridge University Press.
Lipton, P. . Inference to the best explanation. nd ed. London: Routledge.
Riggs, W. . Understanding “virtue” and the virtue of understanding. In Intellectual
virtue: Perspectives from ethics and epistemology, edited by M. DePaul and L. Zagzeb-
ski. Oxford: Oxford University Press.
Salmon, W. C. . Statistical explanation. In The nature and function of scientific theo-
ries, edited by R. Colodny. Pittsburgh: University of Pittsburgh Press.
Sosa, E. . Reflective knowledge in the best circles. In Epistemology: An anthology,
edited by E. Sosa and J. Kim. Oxford: Blackwell.
Steyvers, M., J. Tenenbaum, E. Wagenmakers, and B. Blum. . Inferring causal net-
works from observations and interventions. Cognitive Science :–.
Trout, J. D. . Scientific explanation and the sense of understanding. Philosophy of Sci-
ence :–.
———. . Paying the price for a theory of explanation: De Regt’s discussion of Trout.
Philosophy of Science :–.
———. . The psychology of scientific explanation. Philosophy Compass :–.
Woodward, J. . Making things happen: A theory of causal explanation. New York:
Oxford University Press.
Zagzebski, L. . Virtues of the mind. New York: Cambridge University Press.
 pe tri ylikoski

-
The Illusion of Depth of Understanding
in Science

E:IG>NA>@DH@>

 Philosophers of science have a long tradition of making a connection

between explanation and understanding, but only lately have they started to
give the latter notion a substantial role in their theories. The reason is because
understanding is an even more difficult notion than explanation. To my mind,
the recent interest in understanding (exemplified by this volume), springs from
the fact that explanation is a cognitive activity, and for too long theories of ex-
planation have dismissed the cognitive dimension with the weak excuse of its
being a too “subjective” ingredient for a theory of scientific explanation. Expla-
nation is connected with understanding, and therefore philosophy of science
needs an analysis of scientific understanding.
In this chapter I will employ a well-known scientific research heuristic that
studies how something works by focusing on circumstances in which it does
not work. Rather than trying to describe what scientific understanding would
ideally look like, I will try to learn something about it by observing mundane
cases where understanding is partly illusory. My main thesis is that scientists
are prone to the illusion of depth of understanding (IDU), and as a consequence
they sometimes overestimate the detail, coherence, and depth of their under-
standing. I will analyze the notion of understanding and its relation to a sense
of understanding. In order to make plausible the claim that these are often
disconnected, I will describe an interesting series of psychological experiments
by Frank Keil and his coauthors that suggests that ordinary people routinely
overestimate the depth of their understanding. Then I will argue that we should


t h e i l lusi o n o f d e p th o f un d e r sta n d i n g i n s c i e n c e 

take seriously the possibility that scientific cognition is also affected by IDU
and spell out some possible causes of explanatory illusions in science. I will
conclude this chapter by discussing how scientific explanatory practices could
be improved and how the philosophy of science might be able to contribute to
this process.

JcYZghiVcY^c\VcYi]Z>aajh^dcd[9Zei]

What is understanding? Wittgenstein argues that understanding should

not be understood as a sensation, an experience, or a state of mind. It is not
primarily a process: coming to understand something is a process, but not un-
derstanding itself. More generally, understanding is not a special moment or
phase, but a more permanent attribute. It is an ability. When a person under-
stands something, he or she is able to do certain things (Wittgenstein ,
§§–, –, –; Baker and Hacker , –.)
This does not mean that understanding is some sort of special skill. Un-
derstanding consists of knowledge about relations of dependence. When one
understands something, one can make all kinds of correct inferences about it,
many of which are counterfactual: What would have happened if certain things
had been different? What will happen if things are changed in a certain man-
ner? To get a better grasp of these counterfactual inferences, it is useful to con-
sider the ways in which we can demonstrate that we understand something.
In favorable circumstances, understanding allows for successful interaction
with an object. Our knowledge of the relevant relations of dependence allows
us to make inferences about the consequences of our interventions. We know
what can be done and how. We also know what cannot be done. In this way,
understanding gives us an ability to control the phenomenon. Understanding
comes in degrees, and the amount of control can be used as its measure: other
things being equal, the wider the range of control, the deeper is our under-
standing. With enough understanding of how something works, we can repair
it when it fails or even build a new one, and the more failures we can repair,
the better is our understanding of it. A similar point applies to understanding
human beings. The better we understand a person, the more successful we are
in cooperating and communicating with him or her.
We are not always in a position to causally interact with the phenomenon,
so the amount of control we have cannot be regarded as the ultimate crite-
rion for understanding. If the fact to be understood is in the past, or if we lack
means of intervention, we cannot demonstrate our understanding in this way.
But there are alternatives.
If causal interaction with an object is not possible, we might still be able to
demonstrate our understanding by anticipating how the object behaves. Here,
 pe tri ylikoski

the relevant inferences would be about the future consequences of some event
or series of events. In this case we would not make inferences about the conse-
quences of our imagined or real interventions, but would predict what will hap-
pen. Some of these predictions are about the real world, and others are about
counterfactual situations. Again, the amount of understanding can be roughly
measured by the scope of our predictive abilities. Other things being equal, the
broader the range of circumstances our anticipatory understanding covers and
the more precise our predictions, the better is our understanding of the phe-
nomenon. However, other things are not always equal, so we cannot equate the
amount of explanatory understanding with the ability to predict. Anticipatory
understanding is just one criterion for understanding.
We might be able to interact with an object or predict how it will behave
without making this knowledge explicit, which is required in science. For this
reason, the above criteria are not sufficient in the case of scientific understand-
ing. Something more is required—the ability to explain, that is, to communicate
our understanding. This constitutes the third way to demonstrate understand-
ing. Tacit understanding might give the ability to make predictions and to in-
teract with an object, but science requires us to make our knowledge explicit
by stating the underlying principles. I will call this theoretical understanding.1
Scientific understanding is demonstrated by giving explanations, that is, by
communicating one’s understanding. Explanations are answers to questions of
why; more specifically, they are answers to questions of what if things had been
different (Woodward ), so again we are dealing with an ability to make
counterfactual inferences. We also have a rough measure of the amount of un-
derstanding: the more, and more diverse, explanation-seeking questions we are
able to answer, the better (or deeper) is our understanding. In this case, ex-
planatory knowledge and understanding go hand in hand.
These three ways to demonstrate one’s understanding constitute criteria
according to which people attribute understanding to one another.2 They are
concerned with external displays of their knowledge. This knowledge might be
tacit or explicit; the crucial thing is that it gives an ability to do certain things.
Some authors suggest that understanding involves having an internal mental
model of the object of understanding (see, for example, Waskan ). This is
an interesting suggestion, but it cannot be the whole story about understand-
ing. When we evaluate someone’s understanding, we are not making guesses
about his or her internal representations, but about the person’s ability to per-
form according to set standards. The concept of understanding allows that the
ability can be grounded in various alternative ways, as long as the performance
is correct. Furthermore, the correctness of the internal model is judged by the
external displays of understanding, not the other way around. This makes un-
derstanding a behavioral concept.
t h e i l lusi o n o f d e p th o f un d e r sta n d i n g i n s c i e n c e 

It is plausible that having an internal mental model is a causal condition for

understanding, but it should not be equated with understanding. It would also
be a mistake to commit oneself too rigidly to the idea of internal representa-
tion since much of scientific cognition employs various sorts of external rep-
resentations. Successful scientific cognition combines internal representations
with external representations and other tools. This fact is easily forgotten when
one concentrates on mental models. The distinctive feature of scientific cogni-
tion might not be some special mental models but the extensive use of various
external representations, as work on distributed cognition suggests (Donald
; Hutchins ; Clark ; Giere ).
If understanding is an ability rather than a mental state or an experience,
why do many people think otherwise? The reason is that there is a mental ex-
perience that is closely related to understanding: the sense of understanding,
which is a feeling that tells us when we have understood or grasped something.
Now it is an open question whether this sensation is always the same, but in
any case we recognize it when we have it. This sense of confidence and the
feeling that (often) comes with it can be easily confused with what we think it
indicates: understanding. Ideally these two things would go hand in hand, and
assimilating them should not create any trouble. Real life is different. The sense
of understanding is only a fallible indicator of understanding (Grimm, this vol-
ume). Sometimes one has a false sense of understanding and sometimes one
understands without having any associated feelings or experiences.3
The existence of the sense of understanding should not be regarded as any
kind of oddity. It is a special case of feeling of knowing, a much discussed meta-
cognitive notion. According to Koriat (), the feeling of knowing serves as
an important source of information in our cognitive activities. For example, it
helps to regulate our learning and recall activities. Frank Keil () suggests
that the sense of understanding has a similar metacognitive role. It gives us
confidence to try things, and when it is lacking we can sensibly abstain from
the activity in question. It also guides the search for new knowledge. It tells us
when to stop the search for new information; it signals when we know enough.
A stopping device like this is very useful. A comprehensive understanding of
something would require huge amounts of time and effort, and might still not
be achievable. In everyday life, this kind of ideal knowledge is impractical. It
is better to have some signal that tells us when we have enough knowledge
to function effectively, even if this indicator is not wholly reliable. In addition
to this, the sensation associated with the sense of understanding can have a
motivational role. Satisfying curiosity is highly rewarding (Schwitzgebel ;
Gopnik ; Lipton, this volume); it provides motivation for learning and
other cognitive activities and in this way provides a psychological mechanism
that has a key role in human cognition. The desire to satisfy one’s curiosity also
 pe tri ylikoski

provides an important psychological motivation for doing scientific research.

Although the phenomenology of the sense of understanding can mislead
one into thinking that understanding is an on-off phenomenon (“Now I get
it!”), it actually comes in degrees. First, the understanding can be about dif-
ferent aspects of the phenomenon. Second, those aspects may be understood
in various degrees. Consider an ordinary object like a personal computer. Dif-
ferent individuals understand to varying degrees how their computer works.
Some might know about the software, or some specific piece of software, and
others the hardware. Most people just use the software without any under-
standing of the internal workings of a computer. Despite these differences, they
all understand something about their PC. The crucial question is what aspects
of it they understand. By asking what has been understood, the extent of under-
standing can always be specified. There is nothing mysterious in the idea that
understanding comes in degrees, nor in the idea that there are clear differences
in the degree of understanding that different individuals possess.
From the point of view of this chapter, the crucial thing about the sense
of understanding is that it is a highly fallible source of metacognitive infor-
mation. It does not give us direct access to knowledge that is the basis of our
understanding. Like any other metacognitive judgment, it can misfire: a false
sense of understanding is a real possibility. In such cases we overestimate our
understanding. We can also underestimate our understanding. In these cases
we think we understand less than we actually do. In fact, it would be very sur-
prising if the sense of understanding would turn out to be perfectly calibrated
with our understanding.
The fallibility of the sense of understanding can be demonstrated experi-
mentally. People often overestimate the detail, coherence, and depth of their
understanding. Keil calls this effect the illusion of depth of understanding
(IDU). Together with his associates (Rozenblit and Keil ; Mills and Keil
), Keil designed an experimental setup where participants were first taught
to use a seven-point scale that rated their knowledge. They were then asked to
rate how well they knew how various devices or systems worked. After rating
their understanding of a large set of these items, the participants were then
asked to explain in detail the actual workings of these systems. After giving
each explanation, they were asked to re-rate the depth of their initial knowl-
edge. The participants were then asked to answer diagnostic questions that ex-
perts considered to be central to understanding the system, after which they
were asked to rate their knowledge again. Finally, they were presented with a
concise but thorough expert explanation and were asked to rate their initial
knowledge once more in light of that expert explanation.
The results across several studies show a strong drop in ratings of knowl-
edge after each re-rating, and often the participants were shocked and sur-
t h e i l lusi o n o f d e p th o f un d e r sta n d i n g i n s c i e n c e 

prised by their own ignorance. The general conclusion is that most people are
prone to feel that they understand the world with far greater detail, coherence,
and depth than they really do. According to Keil, this effect is distinct from the
general overconfidence effect found in many psychological studies. Within the
experimental setup described above, the illusion of having detailed and coher-
ent knowledge occurs primarily for explanatory understanding. In contrast,
people’s ratings of how well they know facts, procedures, or narratives are quite
well calibrated, and they are not surprised at what they actually know (Rozen-
blit and Keil ; Keil ).
What is behind this interesting phenomenon? Keil suggests four possible
contributing factors. One factor is confusion between what is represented in
the mind with what can be recovered from a display in real time. People may
underestimate how much of their understanding exists in relations that are ap-
parent in the object as opposed to being mentally represented. A second factor
may be a confusion of higher and lower levels of analysis. For example, while
explaining how a car works, one might describe the function of a unit, such as
the brakes, in general terms, and then describe the functions of subcompo-
nents, such as brake lines and brake pads, which in turn can be broken down
even further. The iterative structure of explanations of this sort may lead to an
illusion of understanding when a person gains insight into a high-level func-
tion and, with that rush of insight, falsely assumes an understanding of further
levels down in the hierarchy of causal mechanisms. A third possible factor is
that many explanations have indeterminate end states. One usually has little
idea of what the final explanation will look like, and the end state is largely
indeterminate from the posing of the question. This makes self-testing one’s
knowledge difficult. The final factor in Keil’s list is the rarity of production: we
rarely give explanations and therefore have little information on past successes
and failures (Rozenblit and Keil , –.)
Keil’s thesis about IDU looks quite similar to the claims made by J. D. Trout
(), so it is important to see their differences. Trout argues that the sense
of understanding is often influenced by the overconfidence bias and that this
makes it a highly unreliable source of information. However, he does not cite
any studies about explanatory cognition to support his argument. His argu-
ment against the epistemic relevance of the sense of understanding is based
on the idea that humans are generally biased toward overconfidence. This idea
of general overconfidence bias has been criticized on theoretical and meth-
odological grounds (Juslin, Winman, and Olsson ). If these criticisms are
right, Trout’s argument is in trouble. However, these criticisms do not apply to
Keil’s experiments. He claims that there is a specific overconfidence effect in
the assessment of understanding, but the general overconfidence effect does
not have any role in his argumentation. Similarly, his experimental set-ups are
 pe tri ylikoski

not based on the assumptions that the critics of overconfidence research have
found problematic. Finally, the conclusions he draws are different: he does not
suggest that we should give up the notion of understanding, as Trout does.
Keil’s studies show that understanding and the sense of understanding do
not always go hand in hand. Could the sense of understanding be calibrated to
be a more reliable indicator of understanding? We simply do not know. There
are no empirical studies of possible ways of improving explanatory practices.
It is possible that if we focus more on our explanatory practices, and make our
explanations more explicit, the calibration of our sense of understanding would
improve. This would help to address two causes of IDU suggested by Keil: the
indeterminacy of the end state and the rarity of production. However, there are
some grounds for being skeptical of our prospects in calibration. The extensive
experimental literature on reading comprehension shows that the calibration
of comprehension is not easily achieved (Lin and Zabrucky ).

I]Z8VhZ[dg:meaVcVidgn>aajh^dch^cHX^ZcXZ

The studies by Keil and his associates concentrate on ordinary people’s illu-
sions of understanding, and they avoid discussing this in the context of scientif-
ic enquiry. I maintain that IDU is possible and indeed common in the sciences
as well, and I will give a number of reasons for this suspicion. My claims will
be empirical hypotheses about factors affecting scientists. I will not assert that
these problems are unavoidable; rather, I will argue that they are prevalent
enough to be grounds for taking seriously the possibility of IDU in science.
The first reason to suspect that IDU might be relevant to science is the
continuity between scientific and everyday cognition. Scientists use the same
cognitive mechanisms as everybody else. Although scientific standards for
evidence, instruments, and social practices make scientific cognition different
from lay cognition, we should not assume that their use automatically makes
the problems related to evaluating explanatory understanding disappear. Most
of the things that contribute to the differences between scientists and ordinary
people are not related to the assessment of explanations.
The second reason IDU may be relevant is the level of attention in the sci-
entific community about the articulation and evaluation of explanations. Al-
though many scientists and philosophers are enthusiastic about explanation as
a proper cognitive aim of scientific enterprise, it is surprising how little explicit
concern explanatory practices receive in the sciences. As an example, scientific
journals provide many guidelines for presenting data and methods; however,
they do not provide any direction for presenting explanations. Typically, scien-
tific journals are structured to report empirical results, and as a consequence
t h e i l lusi o n o f d e p th o f un d e r sta n d i n g i n s c i e n c e 

they do not put much emphasis on explanatory practices. The explanatory

claims are located in the discussion section, and they are often left implicit or
vague, and the editors are often happy with this. Of course, not all scientific
publications follow this format; nor do they all concentrate on reporting exper-
iments or observations, but in general they still share the same lack of explicit
focus on explanation. There are no specialized scientific forums to discuss or
evaluate explanatory claims. The situation is not any better in science educa-
tion. What students learn about presentation and evaluation of explanations,
they learn implicitly. Scientific education does not include formal instruction
on explanatory practices. Although optional philosophy of science courses
might have some discussion about explanation, the content of these courses
might still be more or less irrelevant from the point of view of the explanatory
challenges the students will face.
My third reason to expect IDU to be relevant to science is based on the fact
that scientists’ circumstances are more difficult than those of the participants
in Keil’s experiments. Scientists cannot use correct expert opinion as a bench-
mark, and the criteria for the correctness of the explanations are themselves
ambiguous. This means that the end state is even more indeterminate, making
self-testing more difficult.
The fourth reason is the role of the division of cognitive labor in science. It
is commonly recognized that science is a collective enterprise. The subject of
scientific knowledge is not an individual scientist but a community character-
ized by an extensive division of cognitive labor. In this context it is not possible
to spell out the various ways to understand this idea, but fortunately this is not
needed to make my point. No matter whether the scientist is trying to assess
what she herself, her research group, her discipline, or science as a whole re-
ally understands, she is involved in evaluating understanding that is based on a
division of labor. This means that the scientist must evaluate not only her own
knowledge, but also that of other people whose competence is different from
her own. This creates additional difficulties for assessing understanding. First,
the scientist cannot rely on her sense of understanding, as it does not give any
kind of access to other people. Second, she is forced to evaluate the people (and
theories) outside her narrow, special field based on relatively shallow under-
standing; basically, she is not competent to evaluate them. For this reason, the
scientist must employ various indirect indicators of competence, like reputa-
tion and the general standing of the field. Furthermore, she is not often in a
position to test the understanding of the other persons, so she is more or less
forced to simply accept their word for their competence. The situation is fur-
ther complicated by the fact that there are various incentives for scientists to
overstate their understanding to others. Everybody wants to impress, and intel-
 pe tri ylikoski

lectual authority is a valuable resource. For all these reasons, a scientist might
be too impressed about the understanding of others and consequently might
be prone to overestimate the collective level of understanding of the scientific
community.
The fifth reason IDU is relevant to science is the future-oriented nature of
scientific evaluation. Most of the time scientists evaluate their theories, mod-
els, and methods in the context of making decisions about future research.
Although most of the philosophical discussion about the problem of theory
choice is set in the context of acceptance, scientists make their choices mostly
in the context of pursuit. They are not primarily assessing the current achieve-
ments of their theories, but making informed guesses about their fruitfulness
in the future. In other words, they are not choosing which theory to accept, but
choosing which theory they are going to work with. In the context of pursuit
they do not have to make up their minds about acceptance, but they must place
their bets on the most fruitful approach.
This observation is also relevant to our discussion. Scientists assessing ex-
planations are not simply evaluating their current status in terms of explana-
tory virtues and evidential support; they are also making guesses about the
future prospects of these explanatory hypotheses. The real issue is the promise
of future understanding, not the things that have been delivered thus far. This
future orientation makes the product to be evaluated extremely ambiguous;
the point in time when the evaluation is supposed to take place is not speci-
fied. Furthermore, the future versions of hypotheses to be evaluated cannot be
spelled out, so one does not really know what one is evaluating, nor does one
know what one is evaluating it against.4 If assessing existing explanations is
difficult, the assessment of future explanatory potential just might be impos-
sible.5 Of course, the evaluation is not completely arbitrary. The compatibility
with other relevant theories is one criterion; another is the past track record of
the candidate. But these are still indirect criteria, and people can reasonably
disagree on how to use them.
I submit that, most of the time, scientists make their choices on much more
pragmatic criteria than the promise of explanatory excellence. They choose op-
portunistically—a line of work, for example, that promises doable problems
that can be solved by the expertise, skills, and research resources they have at
their disposal. These primary reasons do not have much to do with explana-
tion, but scientists might still attempt to justify their choices (for example, to
possible funding sources) by appealing to future epistemic virtues. This would
make explanatory argumentation more or less mere window-dressing. The
real reasons for the choices might be different from the ones that are publicly
presented.
t h e i l lusi o n o f d e p th o f un d e r sta n d i n g i n s c i e n c e 

HZkZcHdjgXZhd[>9J

The previous section already indicated reasons to expect that IDU is a real
danger in science. In this section I will investigate some factors that might make
scientists prone to miscalibrated assessments of the depth of their understand-
ing. My suggestions will be a bit speculative, but this is unavoidable, as I have
not found much empirical research that would be relevant to this issue. How-
ever, the individual hypotheses that I will present are all empirically testable. I
do not claim that the following list is exhaustive; there are probably also other
relevant sources of IDU. It is also possible that their relevance may vary in dif-
ferent scientific fields. I will not make guesses about how influential these fac-
tors might be, nor am I making judgments about their relative importance. As
well, I leave the direction of the miscalibration open: given the point made in
the previous section, it is probable that scientists are disposed to overestimat-
ing their understanding, but it is also possible that some of the following cause
an underestimation of the understanding (in some circumstances).
The first possible source of IDU is the simple failure to find things puzzling. It
is typical in everyday cognition that only anomalous or otherwise surprising ob-
servations raise the need for an explanation. Only the unexpected or abnormal
events challenge one’s existing scheme of categorization. The same is probably
also true for the cognitive life of scientists, at least most of the time. Of course,
it is an important characteristic of science that it also asks why-questions about
the normal (Ylikoski ), but the investigative focus is always directed at a
limited number of issues, so this difference does not change the basic point I
wish to make. How does a failure to find things puzzling contribute to IDU?
A puzzling phenomenon forces one to come up with an explanation and
at the same time to face the limits of one’s understanding. This situation is in
sharp contrast to familiar things that behave as they are expected to. They do
not call for explanation, and as a consequence do not challenge a person’s confi-
dence in the depth of his or her understanding. The ability to find a categoriza-
tion for the initially surprising observation brings about a sense of confidence
in one’s conceptual scheme. This restoration of confidence in the conceptual
scheme might be confused with an increase in understanding and regarded
as a consequence of explanatory insight. In this way, familiarity can sustain an
instance of IDU. But it can also give rise to one. The puzzling things that are
not understood are contrasted with the familiar and the expected. This contrast
can easily be mistaken for a contrast between being understood and not being
understood. Furthermore, an explanation aims to remove the mystery about
the phenomenon, so understanding an anomaly makes it appear normal. From
this it is easy to infer that the observations regarded as normal already have the
 pe tri ylikoski

property of being understood. Of course, this inference is not valid, but that
does not mean that we can easily escape making it.
The second possible source of IDU is the confusion between explanation
and description. Quite often one gets the impression that people think that
they automatically contribute to the explanatory understanding of an event
simply by finding out and describing facts about its causal history.6 Further-
more, they might think that the more facts about the causal history you have,
the better is your understanding of the explanandum. This is an interesting
confusion. Causal history is naturally relevant for explanation, but the crucial
point is that not all facts about the causal history are relevant to a given ex-
planandum. Their relevance depends on the aspects of the explanandum one is
trying to make sense of.7 It might be that every single fact about the causal his-
tory is relevant to some explanation-seeking question about the explanandum
event. But we are not interested in all possible explanation-seeking questions
related to the explanandum. We are not looking for questions that would make
our findings explanatory, but for facts that would answer our explanation-
seeking questions. Despite this, some scientists seem to be involved in this kind
of “explanatory fishing”: they think that their findings are important because
they might be crucial in explaining something. What constitutes that something
is conveniently left unarticulated. This ambiguity helps to maintain the illusion
of gained understanding. The illusion itself might be the product of the hard
work of research, even if the findings turn out to be irrelevant: one likes to feel
that one has achieved something, in this case, explanatory understanding.
The third possible source of miscalibration of the sense of understanding is
the ambiguity of the notion of explanation. The first piece of evidence for the
ambiguity comes from the fact that despite extensive debate, philosophers of
science have not reached a consensus about the nature of explanation. This is
weak evidence, as typically philosophers cannot reach a consensus about any-
thing, but in this case the philosophical disagreement reflects a more general
uncertainty. Discussions with scientists show that they have quite different
ideas about the criteria or characteristics of explanatory understanding. Some-
times they present an account that sounds quite strange to an interlocutor
who is familiar with various philosophical theories, but sometimes scientists
present ideas and concepts that they (or their teachers) have learned during a
philosophy of science class. It seems that the notion of explanation is a kind of
metascientific notion that is not explicitly present in everyday scientific prac-
tice: the scientists manage to do their work without having an articulated or
even shared notion of explanation.
When comparing explanations, scientists and philosophers often appeal to
metaphors of explanatory power and depth. Very few have tried to articulate
t h e i l lusi o n o f d e p th o f un d e r sta n d i n g i n s c i e n c e 

what these notions actually mean. However, there are grounds for thinking that
these notions do not refer to a single attribute of explanations. For example, the
notion of explanatory power can refer to a number of different properties of
explanations. An explanation could be considered powerful when it () is cog-
nitively salient (that is, is easy to understand and use); () is factually accurate;
() gives a precise characterization of the explanandum; or () is robust with
respect to changes in background conditions (Ylikoski and Kuorikoski ).
These virtues of explanation sometimes can be in conflict, which means that
the evaluation of explanatory power is not one-dimensional. The important
point in this context is that these different dimensions are not usually articu-
lated in scientific practice and that it is probable that scientists confuse them,
at least sometimes. Consequently, this confusion can lead to a miscalibration of
the assessment of understanding.
The fourth source of IDU is another kind of ambiguity. One cannot explain
anything completely; rather, one always explains some aspects of the explanan-
dum phenomenon. But which aspects are addressed with a given explanation?
Quite often it is very hard to see what the precise explanandum is that scientists
are trying to address with their theories. The ambiguity of the explanandum is a
real problem that surfaces in scientific controversies. A similar point applies to
the ambiguity of the explanans. All explanations take some things for granted
and treat them as background assumptions. But which are dispensable parts of
the background and which are the essential ingredients of the hypothesis? This
is a difficult problem, even for the person suggesting the hypothesis. Scientists
usually have intuitive ideas about the explanandum and the explanans, but they
cannot fully articulate them. In these circumstances it is extremely difficult to
say what has actually been achieved with the explanatory hypothesis and how
much it was the hypothesis (in contrast to the background assumptions) that
did the explanatory work. This gives more room for our wishful thinking and
egocentric bias to operate.
The fifth source of IDU is circular reasoning. The identification of cases of
circular reasoning is difficult outside formal logic (Rips ), and it is plau-
sible that sometimes people can mistake a restatement or a presupposition of
the explanandum for the explanans. In highly theoretical (and philosophical)
contexts this fallacy might be quite common, as the concepts are often difficult
to define precisely. Circular inferences are logically valid, and identifying them
is quite hard without making the whole argument explicit. Furthermore, as
definitions of the concepts and structure of the argument are often ambiguous,
judgments concerning circular reasoning are often controversial. This makes
their identification and avoidance even more difficult.
The sixth possible source of IDU is the confusion between the explanatory
 pe tri ylikoski

power of the hypothesis and evidential support for it. When people are evalu-
ating an explanation they are not only making judgments about its explana-
tory virtues, they also assess the reasons to believe it. Psychologists are quite
confident that there exists an “explanation effect” according to which people
estimate the probability of an event much higher when they have an explana-
tory story about it. A task, like giving an explanation, which requires that a
person treat a hypothesis as if it were true, strengthens the confidence with
which that hypothesis is held. People also seem to have difficulty distinguish-
ing between explanation and evidence (Koehler ; Brem and Rips ). Of
course, this is a philosophical issue, as will be clear to anyone who has followed
the discussions around Peter Lipton’s () distinction between likeliness and
loveliness. Loveliness might be an indicator of likeliness, as many friends of
inference to the best explanation believe. I do not wish to take a stand on this
issue here. My point is that the mere existence of this debate is evidence for
the claim that explanatory power and evidential support are difficult to keep
separate. It is not always clear to which group a given criterion of explanatory
assessment belongs.
I am suggesting here that the evidential considerations might influence the
evaluation of explanatory goodness. This claim differs from the psychologists’
claim that explanatory considerations influence the evaluation of evidential
support. If people do not explicitly distinguish between their confidence in the
understanding provided by an explanation and their confidence in the truth of
the hypothesis, it is possible that these will influence each other. As a conse-
quence, one might think that a hypothesis provides understanding because one
is so confident that it is true.
The last source of IDU arises from scientists who often use complex repre-
sentations to make sense of the phenomena of interest. In order to understand
the phenomenon, one must also understand the epistemic tools that are used
as the medium of representation. In cases like this, understanding is mediated:
an object is understood via a tool that also needs to be understood. How might
this contribute to IDU? Learning to use the relevant mathematics, modeling
assumption, and other ingredients of the representative medium requires a lot
of work; acquiring the relevant abilities is not a trivial operation, but a real
achievement for an aspiring scientist. At the same time, many of the objects
that scientists are attempting to understand are difficult to access. They quite
often are abstracted, idealized, or stylized in a manner that makes them difficult
to grasp without the representative means provided by the theory. (Think, for
example, of the phenomena most economic theories are trying to deal with.) In
these circumstances a scientist can confuse his or her understanding of the theo-
ry with his or her understanding of the phenomena that the theory is intended to
t h e i l lusi o n o f d e p th o f un d e r sta n d i n g i n s c i e n c e 

deal with. In other words, the scientist confuses the intelligibility of the theory
(de Regt and Dieks ) with the intelligibility of the phenomenon.
The confusion is understandable, because in this situation the intelligibil-
ity of the theory is a necessary condition for understanding the phenomenon.
Both involve understanding; however, understanding a theory is different from
understanding a phenomenon when one considers the ways the understanding
can be demonstrated. The understanding of a theory is displayed by making the
right sorts of inferences from it, by knowing what kinds of problems it can deal
with and by building models, whereas the understanding of a phenomenon is
ultimately displayed by making right predictions, successful interventions, and
by answering explanation-seeking questions about it. The latter cannot be done
without a theory, but it requires more than understanding the theory. After
all, one can also understand theories that are considered false, outdated, or ir-
relevant. Making sense of a theory does not guarantee that one makes sense of
the phenomenon.

=dlid6kd^Y>aajhdgnJcYZghiVcY^c\

The above list of possible sources of illusions of explanatory understanding

raises the question of how seriously we should take explanatory ambitions of
science. Is it possible to calibrate scientists’ sense of understanding to a level
that would make us confident enough in it? If it were possible, how could we
improve our sense of understanding? Or if the answer is negative, is there any
way around the bottleneck? I will not make any guesses about the answer to the
first question. I hope future research may provide this. We should, however,
not be too optimistic: the research on metacognitive self-assessment (Lin and
Zabrucky ; Davis et al. ) shows that the task is difficult and the case
of explanatory understanding is probably more challenging than the cases of
self-assessment these studies usually evaluate. I do not have ready answers to
the two other questions, but I would like to suggest that the ways in which we
can try to improve our sense of understanding also help us to gain some inde-
pendence from it.
The key objective is to take seriously the idea that science is an intersub-
jective enterprise. The way science has succeeded in improving itself beyond
our everyday individual cognitive performance has largely been due to three
factors: () the attempt to make the claims more explicit, () the use of external
representations in this process, and () the development of social practices that
allow critical debate on these representations. The quality of epistemic prac-
tices has improved as we have given up the idea that everything must be done
solely in one’s own head (Donald ; Hutchins ; Clark ; Giere ).
 pe tri ylikoski

Maybe we could do something similar in the case of explanatory understand-

ing. There is no reason to suppose that current scientific explanatory practices
are as perfect as they can get. Most of the possible sources of IDU presented
above are related to the intuitive manner in which judgments about explanato-
ry adequacy are made. If the explanatory claims were made more explicit, their
correct assessment would also be easier. This would provide better feedback for
the calibration of the sense of understanding; moreover, it would make us less
dependent on this quite fallible source of metacognitive insight. The crucial
question is: how do we make our explanatory practice more explicit?
Here lies an opportunity for philosophers of explanation. They could adopt
the aim of providing constructive suggestions for the improvement of scientific
practices. In this way philosophers could prove their relevance to scientific en-
terprise and acquire an external standard for evaluating the adequacy of their
proposals. Currently, philosophers of science judge the competing theories of
explanation mostly against their “intuitions.” This is a standard procedure in
analytical philosophy, but it is not a very useful standard of evaluation. Phi-
losophers tend to have very different “intuitions” about explanation. Although
our intuitions are strong and not easily changed, they seem to be products of
an array of slightly suspect causes: the philosophy of science literature we were
exposed to during some sensitive period, the intuitions of people (philosophers
or scientists) we feel are authoritative, explanatory practices in the fields of sci-
ence we are familiar with, our personality, and so on. It would be better to have
some external standard that is better connected with science practice.
It would not be an improvement to simply replace the philosopher’s intu-
itions with intuitions of practicing scientists. Instead of fitting with intuitions,
the philosophical theories should be judged on the basis of their adequacy
in solving explanation-related controversies in the sciences and their useful-
ness in science education and communication, in short, in fighting IDU. Here
the final judgment would rest with the scientists: they would evaluate which
philosophical suggestions have proved fruitful in improving scientific practice.
Philosophical theories should not simply describe current scientific practices,
but instead suggest improvements. This methodological recasting could help to
reformulate some old issues in the theory of explanation.
The first example of methodological recasting is the view Wesley Salmon
() has dubbed “deductive chauvinism” and according to which explana-
tions are deductive arguments. Although Salmon failed to show that this view is
tied to the assumption that the world is deterministic, the philosophical debate
over the last decades has established that deduction does not have a constitu-
tive role in explanation (Ylikoski ). People had been giving explanations
long before the invention of formal logic; there are plenty of nonpropositional
explanations (for example, using diagrams and pictorial representations); and
t h e i l lusi o n o f d e p th o f un d e r sta n d i n g i n s c i e n c e 

most propositional explanations do not have a deductive structure (although

they might have some deductive passages). In this standard sense, deductive
chauvinism is dead, but I think we should not give up this idea so easily.
We should not regard deductivism as a claim about what makes explana-
tions explanatory, but as a practical suggestion for improving explanatory prac-
tice. Explanations are not deductive arguments, but they can be reconstructed
as such (Ylikoski ). The idea is that an (even partial) attempt at deductive
reconstruction leads to improvements in the process of articulating explana-
tions by forcing one to explicate both many of the background assumptions and
the intended explanandum. I think that this is the main reason some people
are still intuitively attracted to the deductive ideal. Of course, deductivism is
not a foolproof procedure for fighting IDU. If the source of illusion is circular
reasoning, the deductive test is not helpful. (Naturally, deductive reconstruc-
tion might still make it more visible.) A more serious problem is formed by
the various ways one can fudge the explanation by using filler terms and other
placeholders. Deductive reconstruction does not help with these.
The other example of methodological recasting is the idea that the ex-
planandum is contrastive (Woodward ; Lipton ; Ylikoski ). This
old concept has usually been interpreted as a thesis about the pragmatics of
explanation and as a claim about what people have in mind when they put
forward an explanation-seeking question. These are not the best ways to use
the contrastive idea. The contrastive thesis can be regarded as a thesis about
fruitful ways of explicating the intended explanandum. All explananda can be
reconstructed in contrastive terms and the suggestion is that this is helpful
both from the point of view of enquiry and from the point of view of evaluating
explanations. Understood in this way, the contrastive proposal includes two
theses: one about reconstructability and another about fruitfulness. Clearly the
latter is more interesting, as it might hold even if the first one does not hold
universally. The idea is that the contrastive reconstruction helps to be more
specific about the intended explanandum and in this way it makes the com-
parison of apparently competing explanations possible: quite often it turns out
that these explanations address a slightly different aspect of the explanandum
phenomenon.
These two suggestions are only examples. There are probably many other
ways in which scientific explanations could be made more explicit in a manner
that prevents us from becoming victims of IDU. Finding out these practical
measures should be one of the chief aims of the theory of explanation. In this
way it could contribute to the truly philosophical enterprise of figuring out
the limits of our (current) scientific understanding. After all, explicitly spelling
out our understanding often shows that we understand less than we originally
thought.
 pe tri ylikoski

8dcXajh^dc

In this chapter I have argued for a number of different theses about explan-
atory understanding. I first argued that understanding should be characterized
as an ability and that it should be distinguished from the sense of understand-
ing. This is a distinction that is not commonly made in philosophy of science
literature. When understanding is analyzed as an ability to make counterfactual
inferences in the contexts of manipulation, prediction, and explanation, its re-
lation to knowledge can be clarified. Understanding is only knowledge about
the relations of dependence, not something mysterious added to knowledge.
This characterization allows for the possibility of tacit understanding, but this
chapter focuses mainly on scientific understanding, that is, explanatory un-
derstanding. One has explanatory understanding when one is able to answer
explanation-seeking questions. The more questions one can answer about a
phenomenon, the better one’s understanding, which makes the connection be-
tween scientific understanding and explanation quite close. However, this does
not mean that these notions are equivalent. Explanations are best analyzed
as answers to explanation-seeking questions. The notion of understanding is
more appropriate for characterizing the epistemic aims of science: organized
knowledge that allows one to answer to a whole series of what-if-things-had-
been-different questions about the world. In short, understanding is an ability
to give explanations.
The second main theme in this chapter is the possibility of illusory under-
standing in science. The key here is the sense of understanding that has an
important metacognitive role in our cognition. Despite its importance, there
is no reason to assume that the sense of understanding is perfectly calibrated
to our understanding. This makes it possible for people to overestimate the
detail, coherence, and depth of their understanding. Following Keil, I call this
phenomenon the illusion of depth of understanding. My main claim has been
that IDU is a real possibility in science. Due to the continuity of lay and scien-
tific cognition, the lack of explicitness in explanatory practice, the lack of clear
benchmarks, the division of cognitive labor, and the future-oriented nature of
scientific cognition, there is every reason to suspect that scientists might over-
estimate the depth of their understanding.
What does the possibility of illusion of understanding tell us about scien-
tific understanding? The main message, I believe, is that we should take more
care in assessing the level of our understanding. We might understand less
than we think. This is not a reason for skepticism concerning the possibilities
of scientific understanding, but a call for an improvement in our explanatory
practices. The simple reliance on the sense of understanding is not a sufficient
t h e i l lusi o n o f d e p th o f un d e r sta n d i n g i n s c i e n c e 

criterion. In order to escape the limitations of our built-in metacognitive ap-

paratus, we should make the assessment of explanatory understanding more
public by making our explanatory practices more explicit. This is one of the
main goals of the philosophy of scientific explanation.

CdiZh
I would like to thank the editors of this volume and Tarja Knuuttila, Tomi Kokkonen,
Jaakko Kuorikoski, Aki Lehtinen, Uskali Mäki, Päivi Oinas, and Samuli Pöyhönen for their
useful comments.
. Knowledge cannot be made completely explicit; even scientific knowledge is based
on the foundation of some basic skills. For similar reasons, theoretical understanding is
often not easily transformable into a more practical form of understanding. The acquisi-
tion of theoretical understanding might not be accompanied by the relevant practical
knowledge and skills.
. These criteria of understanding raise many questions: How are they related to one
another? Are the different measures of amount of understanding comparable? Could
these criteria be different in different cultures or historical periods? How does one spell
out the crucial notion of “favorable circumstance” employed in the above discussion? I do
not pretend to be able to answer all of these questions. Here my modest aim is to describe
what kind of an ability understanding is by describing some criteria according to which it
is usually attributed to people.
. The psychological research (Cleeremans et al. ) on implicit learning testifies to
the plausibility of acquiring understanding without a feeling of understanding. Although
the boundary between learning with and without awareness is tricky to draw, the research
at least shows that one can learn without the metacognitive feeling that one is acquiring
new knowledge.
. Kyle Stanford () draws from the history of science and argues it is always pos-
sible that there are equally well-confirmed and scientifically serious alternatives to current
best theories that are just not conceived.
. A couple of psychological factors might make the evaluation of the future potential
of theories even more difficult. First, the assessment of the fruitfulness of one’s pet theory
is probably often tainted by wishful thinking. From a motivational point of view, this source
of optimism might be instrumental in bringing about bold hypotheses, but it is not helpful
in increasing the reliability of the scientist as a source of explanatory evaluation. The other
factor is the egocentric bias in evaluation. If one is tempted to be too optimistic about one’s
own theories, the contrary might hold for the competing accounts. They tend to be seen as
less promising. One way to bias the evaluation is to evaluate one’s own theory by its future
promise, and the competition by its achievements thus far. (There are various ways to
rationalize this asymmetry to oneself in such a manner that the imbalance seems only fair.)
. A similar point can be made about noncausal explanations, like constitutive ex-
planation. In this case, people think that findings about the parts of a system contribute
automatically to the understanding of the capacities of the system.
 pe tri ylikoski

. The accumulation of irrelevant details makes the explanation worse, as its recipient
might not see which pieces of information are relevant for understanding and which are
not. But this has at least one positive consequence: at least the recipient is not prone to
IDU since he or she fails to get the point.

GZ[ZgZcXZh
Baker, G. P., and P. M. S. Hacker. . Wittgenstein: Understanding and meaning, part ,
Essays. Oxford: Blackwell.
Brem, S. K., and L. J. Rips. . Explanation and evidence in informal argument. Cognitive
Science :–.
Clark, A. . Being there: Putting brain, body, and the world together again. Cambridge,
MA: Massachusetts Institute of Technology Press.
Cleeremans, A., A. Destrebecqz, and M. Boyer. . Implicit learning: News from the
front. Trends in Cognitive Sciences :–.
Davis, D. A., P. E. Mazmanian, M. Fordis, R. Van Harrison, K. E. Thorpe, and L. Perrier.
. Accuracy of physician self-assessment compared with observed measures of
competence. Journal of the American Medical Association :–.
de Regt, H. W., and D. Dieks. . A contextual approach to scientific understanding.
Synthese :–.
Donald, M. . Origins of the modern mind. Cambridge, MA: Harvard University Press.
Giere, R. N. . Scientific cognition as distributed cognition. In The cognitive basis of sci-
ence, edited by P. Carruthers, S. Stich, and M. Siegal, –. Cambridge: Cambridge
University Press.
Gopnik, A. . Explanation as orgasm and the drive for causal knowledge: The function,
evolution, and phenomenology of the theory formation system. In Explanation and
cognition, edited by F. C. Keil and R. A. Wilson, –. Cambridge, MA: Massachu-
setts Institute of Technology Press.
Hutchins, E. . Cognition in the wild. Cambridge, MA: Massachusetts Institute of
Technology Press.
Juslin, P., A. Winman, and H. Olsson. . Naïve empiricism and dogmatism in confi-
dence research: A critical examination of the hard-easy effect. Psychological Review
:–.
Keil, F. C. . Folkscience: Coarse interpretations of a complex reality. Trends in Cogni-
tive Sciences :–.
Koehler, D. J. . Explanation, imagination, and confidence in judgment. Psychological
Bulletin :–.
Koriat, A. . The feeling of knowing: Some metatheoretical implications for conscious-
ness and control. Consciousness and Cognition :–.
Lin, L.-M., and K. M. Zabrucky. . Calibration of comprehension: Research and
implications for education and instruction. Contemporary Educational Psychology
:–.
Lipton, P. . Inference to the best explanation. nd ed. London: Routledge.
Mills, C. M., and F. C. Keil. . Knowing the limits of one’s understanding: The devel-
t h e i l lusi o n o f d e p th o f un d e r sta n d i n g i n s c i e n c e 

opment of an awareness of an illusion of explanatory depth. Journal of Experimental

Child Psychology :–.
Rips, L. J. . Circular reasoning. Cognitive Science :–.
Rozenblit, L., and F. C. Keil. . The misunderstood limits of folk science: An illusion of
explanatory depth. Cognitive Science :–.
Salmon, W. C. . Causality and explanation. Oxford: Oxford University Press.
Schwitzgebel, E. . Children’s theories and the drive to explain. Science and Education
:–.
Stanford, K. . Exceeding our grasp: Science, history, and the problem of unconceived
alternatives. Oxford: Oxford University Press.
Trout, J. D. . Scientific explanation and the sense of understanding. Philosophy of Sci-
ence :–.
Waskan, J. A. . Models and cognition. Cambridge, MA: Massachusetts Institute of
Technology Press.
Wittgenstein, L. . Philosophical investigations. Oxford: Basil Blackwell.
Woodward, J. . Making things happen: A theory of causal explanation. New York:
Oxford University Press.
Ylikoski, P. . The third dogma revisited. Foundations of Science :–.
———. . The idea of contrastive Explanandum. In Rethinking explanation, edited by J.
Persson and P. Ylikoski, –. Dordrecht: Springer.
Ylikoski, P., and J. Kuorikoski. . Dissecting explanatory power. Philosophical Studies,
forthcoming.
 pe tri ylikoski
chap ter title ver so 

G8IK@@
JcYZghiVcY^c\VcYBdYZah
 pa rt title ver so
chap ter title ver so

.
Understanding in Physics and Biology
From the Abstract to the Concrete

B6G<6G:IBDGG>HDC

 It is commonly thought that the greater the degree of abstraction

used in describing phenomena the less understanding we have with respect to
their concrete features. I want to challenge that myth by showing how math-
ematical abstraction—the characterization of phenomena using mathematical
descriptions that seem to bear little or no relation to concrete physical enti-
ties/systems—can aid our understanding in ways that more empirically based
investigations often cannot. What I mean here by “understanding” is simply
having a theoretical account of how the system is constituted that enables us to
solve problems, make predictions, and explain why the phenomena in question
behave in the way they do. Although much of my discussion will focus on the
role of mathematics in biology, specifically population genetics, I also want to
look at some examples from physics to highlight the ways in which understand-
ing via mathematical abstraction in these two fields can be strikingly similar.
Many philosophical accounts link understanding with explanation. While
some (Hempel , van Fraassen ) claim that understanding is a pragmat-
ic notion and hence epistemically irrelevant to assessing the merit of a theory,
most others see understanding as an important goal of science but define the
notion in a variety of different ways. One option is to link understanding with
unification (Friedman ; Kitcher ), where the explanation that best
unifies the phenomena is the one that produces the greatest understanding.
Salmon (), on the other hand, sees knowledge of causal mechanisms as the


 m ar g ar et mo r r iso n

feature that furnishes understanding of the physical world. What both of these
accounts have in common is that they privilege a particular type of explanation
as the vehicle for producing understanding. De Regt and Dieks () criticize
these monolithic accounts of understanding and instead define it in terms of
a criterion of intelligibility that involves having a theory where one can recog-
nize qualitatively characteristic consequences without performing exact calcu-
lations. This is achieved using a variety of conceptual tools that are relevant to
the problem at hand. Their point is that in some cases causal knowledge will be
relevant to understanding, and in some cases it will not—the tool kit contains a
number of different resources that, depending on the context, will produce un-
derstanding. An important feature of their account is that its pragmatic quality
in no way compromises the epistemic relevance of understanding for scientific
investigation.
My own sympathies lie with the type of contextual analysis provided by
de Regt and Dieks, but I would put the point in slightly more negative terms.
While I agree that what it means to “understand” depends on contextual fac-
tors such as the nature of the problem and the resources available to solve it,
I want to claim that it is neither possible nor desirable to formulate a “theory”
of understanding. That does not mean that we cannot explain what it is to un-
derstand why a phenomenon behaves as it does or what we mean when we say
someone understands a mathematical concept.1 Rather, there is no canonical
account of what it means to understand. I would suggest the same holds for
explanation. Very often in science we desire explanations that are causal, but
that is not to say that this is the only form of explanation that is acceptable, or
that we can give a “theory” of explanation that centers on causal relations. On
the contrary, whether something has been successfully explained will depend
on the question and the stock of available answers, answers that come from our
background theories.
We simply cannot specify in advance what qualifies as an explanation or
what form it will take. However, I do think explanation and understanding are
linked in the intuitive sense that one often accompanies the other; that is, we
demonstrate our understanding by being able to offer explanations of the ob-
ject/concept in question and the success of our explanations is typically a func-
tion of how well we understand what we are trying to explain. My position is a
minimalist one insofar as neither understanding nor explanation is capable of
being codified into a philosophical theory.2
Bound up with the question on how mathematical abstraction enables us
to understand physical phenomena is the role played by models. In Morrison
 and , I articulate how models can provide knowledge by focusing on
what I call their mediating role. That is, models function as autonomous media-
tors between theory and applications (the model provides simplifications of the
under standing in ph ysic s and biol o gy 

theory’s equations so that they can then be applied) or between theory and the
world (the model is an idealized or abstract representation of some phenom-
enon or physical system). In this latter case, the representation can then be
compared with a concrete object/system.
The distinction between idealization and abstraction I am adopting for
the purposes of this chapter is roughly the following: abstraction is a process
whereby we describe phenomena in ways that cannot possibly be realized in
the physical world (for example, infinite populations). In this case, the math-
ematics associated with the description is necessary for modeling the system
in a specific way. Idealization, on the other hand, typically involves a process of
approximation whereby the system can become less idealized by adding cor-
rection factors (such as friction to a model pendulum). In the latter case, the
idealization is used primarily to ease calculation since the factors that have
been idealized will not, in general, be relevant for the type of problem at hand.
In their original state both abstraction and the idealization make reference
to phenomena that are not physically real; however, because the latter leaves
room for corrections via approximations, it can bear a closer relation to a con-
crete physical entity.3 This distinction will become important in showing how
mathematical abstraction enables us to understand certain features of empiri-
cal phenomena in ways that are different from the more straightforward fea-
tures associated with idealization.
The role of mathematical abstraction in model construction figures impor-
tantly in characterizing the autonomy of models.4 For example, in cases where
the system or phenomena under investigation may be inaccessible due to large
distances (galactic phenomena) or size (microstructures), the model occupies
center stage and takes on the task of representing how we assume the system
is constructed. And, because direct comparison between the model and the
system is not possible, the model supplants the system as the object under
investigation.
Models of stellar structure in astrophysics fit this category. In these cases,
the model fulfills its mediating role by serving as a source of mediated knowl-
edge, which is less direct than knowledge of real systems. In these cases, as well
as in more general modeling instances, there is often a good deal of mathemati-
cal abstraction (and idealization), structure that itself becomes the focus of in-
vestigation. Hence, our knowledge of real world situations is typically mediated
to some degree by this structure. In the cases I discuss below the mathematical
model supplants the physical phenomena as the object of investigation, and
the degree to which the model produces understanding comes via its abstract
mathematical structure. The interesting question is how we should understand
the relation between this abstract structure and the concrete physical systems
that this structure purportedly represents.5
 m ar g ar et mo r r iso n

While this type of scenario is common in theoretical physics and many of

the natural sciences, its acceptance as an appropriate method for biology is
much more recent. In other words, the kind of mathematical abstraction com-
monly found in physical models was not typically associated with investiga-
tions in biology. Here, we associate the experimental context with fieldwork
and the mathematical foundations of biological theory with statistical meth-
ods used to analyze experimental findings. But, the increasing use of abstract
mathematical models in biology is not simply due to the rise of subdisciplines
like molecular biology. Since the development of population genetics in the
late teens and early s, the use of abstract mathematical techniques like
diffusion equations have resulted in new ways of characterizing populations.
Instead of the natural or “real” populations studied by many biologists at the
time, populations were now mathematical constructs that could be manipu-
lated using sophisticated mathematics. In that context, the model, rather than
the real world environment or system (that is, populations of living things),
became the object of inquiry, supplanting the system under investigation. Be-
cause conclusions about the role of selection in Mendelian populations were
simply not available using empirical methods, this new form of mathematical
modeling marked an important turning point for understanding the genetic
features of natural populations.
As I noted above, one of the difficulties with this kind of “mediated” knowl-
edge is that it often involves a good deal of mathematical abstraction, taking
us further away from the kind of empirical situations we are frequently inter-
ested in. Moreover, we typically would like models to provide at least a partially
realistic representation of the system under investigation. In fact, it is partly
because theories function at such an abstract level that we require models for
the application of theory to concrete situations and to fill in the details where
theory is silent. However, as we shall see and contrary to what we might expect,
increasing levels of mathematical abstraction often provide the only way to un-
derstand concrete physical and biological systems.
R. A. Fisher was able to explain human variation in his work on population
genetics by invoking a very unrealistic and abstract model of a population. This
turns on its head what we think of as the typical modeling scenario; instead of
models providing a realistic account of a situation, they produce understand-
ing by introducing a level of abstraction that seemingly takes us away from the
very phenomena we want to understand. Moreover, as we shall see below, it
is sometimes only by virtue of this type of abstraction that we come to some
understanding of the system itself.
My discussion begins with a brief example from physics, specifically the ab-
straction involved in taking the macroscopic limit in thermodynamics, a tech-
nique necessary for understanding phase transitions. From there I discuss the
under standing in ph ysic s and biol o gy 

Hardy-Weinberg law and how the mathematical abstraction built into this law
allows us to understand fundamental features about biological populations. I
conclude with an account of how the abstract mathematics used by R. A. Fisher
produced tremendous advances in our understanding of the genetics of natural
populations.

Ldgg^ZhVWdjiA^b^ih

Even without bringing mathematics into the equation the very notion of
what it means to “understand” is by no means straightforward. As I noted
above, philosophical attempts to understand understanding are usually linked
to the equally controversial notion of explanation, more specifically, the task
of conveying information in a way that will answer a why- or how-question.
Typically, scientific understanding, especially in physics, involves the applica-
tion of both general principles and more specialized models that enable us to
describe, predict, and explain specific types of behavior. This is not to say that
an appropriate answer to every question must invoke both strategies but simply
to recognize that certain kinds of questions demand certain kinds of answers.
In other words, the idea that explanation can be achieved solely on the basis
of fundamental laws and initial conditions, as described by the D-N model, is
now thought to be insufficient for understanding much of the physical world.
We need a variety of things— fundamental theory, different kinds of models, as
well as laws that constrain the kinds of behaviors that systems display.
The important question here is how the abstraction (as opposed to idealiza-
tion) built into theories, models, and laws enables or indeed prevents us from
understanding physical systems. Initially one might think that understanding is
inhibited by abstraction, but we only need to look to scientific textbooks to see
that this is clearly not the case. The kind of models that serve as exemplars in
the Kuhnian sense (for example, the quantum harmonic oscillator, the infinite
potential well, and so on), as well as entities like virtual particles, embody a
great deal of abstraction and yet are the cornerstone of understanding essential
features of certain kinds of physical systems. This is because they enable us to
conceive of systems as being of a particular type, exhibiting certain kinds of
behavior that allow us to classify them in terms of their general features. All of
these factors enhance our understanding.
But, understanding via abstraction is not simply limited to general features
of physical systems. In many of these cases, abstract representations also pro-
vide the kind of detailed knowledge required to answer causal questions. For
example, if we ask why a particular metal exhibits superconducting properties,
we can explain it in terms of zero resistivity and the accompanying thermo-
dynamic phase transition, something that is only possible in certain kinds of
 m ar g ar et mo r r iso n

metals. And if we want to know the details of what happens at the subatomic
level, we can invoke the BCS model complete with its description of electron-
phonon interactions, Cooper pairing, and the BCS wave function. Although
these models involve abstraction and approximation, they allow for a rather
peculiar situation: from them we can derive exact results associated with su-
perconductivity—infinite conductivity, exclusion of magnetic fields, flux quan-
tization, and zero resistivity.
The question is how one can get these kinds of exact consequences from
models that are approximations. In other words, we want to understand why
the models work so well. The reason is because we can also understand a su-
perconductor as a material in which electromagnetic gauge invariance is spon-
taneously broken and these models contain, as a fundamental assumption, the
breakdown of electromagnetic gauge invariance. The detailed dynamic theories
and models like BCS are required to explain why and at what temperatures this
symmetry breaking occurs, but not to derive the kinds of consequences men-
tioned above. In other words, these properties can be derived directly from the
assumption of the spontaneous breaking of electromagnetic gauge invariance
and so are consequences of general principles rather than specific approxima-
tions embodied in the model.6
But how, exactly, is mathematical abstraction involved here? The situation
is far from straightforward. The spontaneous breakdown of electromagnetic
gauge invariance involves a phase transition that is associated with the super-
conducting state. The occurrence of phase transitions requires a mathematical
technique known as taking the “thermodynamic limit,” N ∞; in other words,
we need to assume that a system contains an infinite number of particles in
order to understand the behavior of a real, finite system. Very briefly, the situa-
tion is as follows: in thermodynamics (TD), phase transitions are accounted for
in terms of discontinuities in the thermodynamic potentials. However, once we
move to statistical mechanics (SM), the equations of motion that govern these
systems are analytic and hence do not exhibit singularities. As a result, there
is no basis for explaining phase transitions in SM. In order to recover the TD
explanation, singularities are introduced into the equations, and thus far the
only way to do this is by assuming the number of particles in the system is in-
finite. Note that the problem here is not that the limit provides an easier route
to the calculational features associated with understanding phase transitions;
rather, the assumption that the system is infinite is necessary for the symmetry
breaking associated with phase transitions to occur. In other words, we have a
description of a physically unrealizable situation (an infinite system) that is re-
quired to explain a physically realizable phenomenon (the occurrence of phase
transitions).
The question is whether this procedure yields the kind of explanation that
under standing in ph ysic s and biol o gy 

produces understanding. Although one might want to claim that within the
mathematical framework of SM we can causally account for (explain) the oc-
currence of phase transitions by assuming the system is infinite, it is neverthe-
less tempting to conclude that this explanation does not help us to physically
understand how the process takes place since the systems that SM deals with
are all finite. Similar doubts have been expressed by Earman (), who argues
against taking the thermodynamic limit as a legitimate form of idealization:
“a sound principle of interpretation would seem to be that no effect can be
counted as a genuine physical effect if it disappears when the idealisations are
removed” (). In other words, we should not assume that phase transitions
have been explained and thereby understood if their occurrence relies solely on
the presence of an idealization. Initially this seems an intuitive and plausible
objection, but if we reflect for a moment on the way mathematical abstraction
is employed for explanatory purposes in scientific models or theories, it be-
comes clear that this line of reasoning quickly rules out explanations of the sort
we deem acceptable in other contexts. Here the distinction I introduced at the
beginning between abstraction and idealization becomes especially important.
Specifically, we need to distinguish between the kind of mathematical abstrac-
tion required for a theoretical representation and the more straightforward
kinds of mathematical idealizations that are used simply to facilitate calcula-
tion. In the former case, the abstraction becomes a fundamental part of how
the system is modeled or represented and consequently proves crucial to our
understanding of how it behaves.
For example, consider the intertheoretical relations that exist in fluid me-
chanics between Navier Stokes equations and the Euler equations, or between
theories like wave optics and ray optics, and classical and quantum mechanics.
Because of the mathematical nature of physical theories the relations between
them will typically be expressed in terms of the relations between different
equations or solutions. In each case, we are interested in certain kinds of limit-
ing behavior expressed by a dimensionless parameter δ. In fluid dynamics δ
is equal to /Re (Reynolds number), and in quantum mechanics it is Planck’s
constant divided by a typical classical action (ћ/S). But, in fluid mechanics (as
in the other cases listed above) the limit δ  is singular and it is this singular-
ity that is responsible for turbulent flows. Similarly, in the ray limit where geo-
metrical optics accurately describe the workings of telescopes and cameras the
wavelength λ. Because ψ is nonanalytic at λ=, the wave function oscillates
infinitely fast and takes all values between + and – infinitely often in any
finite range of x or t.
A good deal of asymptotic behavior that is crucial for describing physical
phenomena relies on exactly these kinds of mathematical abstractions. What
we classify as “emergent” phenomena in physics such as the crystalline state,
 m ar g ar et mo r r iso n

superfluidity, and ferromagnetism, to name a few, are the result of singularities,

and their understanding depends on just the kinds of mathematical abstrac-
tions described above. Moreover, it is equally important to mention that in
many of these cases it is possible to note changes in the system’s behavior as it
approaches a critical point, gradually approximating the kinds of changes asso-
ciated with the phase transition. In that sense, we do have physical “indicators”
for the kind of behavior the mathematics predicts.
Consequently, subscribing to Earman’s account would be tantamount to
ignoring large portions of the mathematical foundations of our theories, so
we need to ask ourselves why cases like the thermodynamic limit ought to be
considered illegitimate as opposed to sending distances and time intervals to
zero in the use of differential equations. Here the distinction between idealiza-
tion and abstraction offers some help. We typically think of an idealization as
resembling, in certain respects, the concrete phenomenon we are trying to un-
derstand. We usually know how to correct or compensate for what we have left
out of, or idealized in, the description. That is, we know how to add back things
like frictional forces that may not have been needed for the problem at hand.
Or, we know how to change the laws that govern a physical situation when we
introduce more realistic assumptions, as in the move from the ideal gas law
to the van der Waals law. These cases are best thought of as idealizations that
represent a physical situation in a specific way for a specific purpose.
Contrast this with the kinds of mathematical abstractions described above.
In the case of the thermodynamic limit, we do not introduce abstractions sim-
ply as a way of ignoring what is irrelevant to the problem or as a method for
calculational expediency. Instead, the mathematical description functions as a
necessary condition for explaining and hence understanding the phenomena in
question.7 Thinking of them in this way sheds a different light on the problem
Earman mentions. If we think of idealizations as calculational devices, then yes,
we should not associate them with genuine physical effects; but the case of ab-
straction is different in kind. Undoubtedly it also requires us to think differently
about what constitutes “understanding”; but, given the mathematical character
of our theories, that consequence is unavoidable. If physical phenomena are
described in terms of mathematical abstractions, then it seems reasonable to
expect that their explanations be given in similar terms.

:migVXi^c\i]ZIgji][gdbÆ;VahZÇAVlh

While we might reluctantly accept this way of looking at things in phys-

ics, one might be even less enthusiastic when it comes to explanation in biol-
ogy. The degree of complexity associated with biological phenomena is largely
under standing in ph ysic s and biol o gy 

thought to rule out explanations that embody abstractions as simply uninfor-

mative; they do not tell us the kinds of things we need to know in order to
understand the behavior of biological phenomena. Moreover, the reduction-
ist assumptions that underscore many explanations in physics are thought to
be inappropriate for biology. By ignoring the individual nature of biological
phenomena, we draw the wrong kinds of conclusions about complex living
systems, which in turn can be used to justify undesirable social and political
practices. Instead, biological explanations need to be “closer to the ground” and
appeal to the specific nature of individuals or species and their interaction with
their environment.
A notable exception to this is population genetics, a highly mathematical
branch of biology that deals with the genetic basis of evolution. The objects of
study are primarily the frequencies and fitnesses of genotypes in natural popu-
lations. Some population geneticists even define evolution as the change in the
frequencies of genotypes over time, something that may be the result of their
differences in fitness. Or to put it more accurately, we can model evolution by
assigning fitnesses to genotypes and then following the changes in allele fre-
quencies. The problem is that while genotype or allele frequencies are easily
measured, their change is not. Most naturally occurring genetic variants have
a time scale of change that is on the order of tens of thousands to millions of
years, making them impossible to observe.8 Fitness differences are likewise very
small, less than . percent, also making them impossible to measure directly.
What this means is that although the state of a population can be observed,
the evolution of a population cannot be directly studied—hence the need for
mathematical models.
One investigates most systems that are not directly accessible by construct-
ing mathematical models of the evolutionary process and then comparing their
predictions with the behavior or states of the system, in this case populations,
that can be directly observed. But because one cannot know the genetic struc-
ture of a species (not only would we need a complete description of the genome
but also the spatial location of every individual at one instant in time which
would then change in the next instant), the models need to incorporate cer-
tain idealizing assumptions that ignore the complexities of real populations.
They focus on one or a few loci at a time in a population that mates randomly
or has a simple migration pattern. The success of these idealized models has
been remarkable, and indeed the birth of population genetics resulted from the
application of mathematical techniques and idealizations that invoked infinite
populations and ideas from statistical physics. To that extent, then, the meth-
ods of population genetics look very much like the methods of physics. And
indeed the claim by many population geneticists is that allele-based models are
 m ar g ar et mo r r iso n

what constitutes our understanding of much of the mechanics of evolution.9

What this seems to suggest is that, at least in the domain of population genet-
ics, understanding in physics and biology is very similar.
In fact, one only needs to look to the very early developments in popula-
tion genetics to see how mathematical abstraction plays a role in enabling us
to understand certain features about populations. The Hardy-Weinberg law is
a simple example of a fundamental law that makes assumptions not realized
in any natural population. So, in what sense do we want to say that this law is
explanatory? One of the things that renders laws explanatory, as highlighted by
the D-N model, is the fact that they are general enough to apply to a diverse
number of phenomena. In other words, they enable us to understand specific
features of phenomena as similar in certain respects; for example, universal
gravitation shows that both terrestrial and celestial bodies obey an inverse
square force law. Cartwright () claims that this generality is a reason for
thinking fundamental laws like these are false; their generality results in their
being unable to fully describe the situations they reportedly cover, or they de-
liberately omit aspects of the situation that are not relevant for the calculation
at hand. In that sense they do not accurately describe concrete situations. The
problem, then, is how could they possibly impart understanding?
Part of Cartwright’s reason for claiming that covering laws are false is to
contrast them with phenomenological laws (or models) that supposedly give
us more accurate descriptions of the physical world. But, since all scientific
description embodies a certain amount of abstraction/idealization, it is diffi-
cult to know where to draw the line. Ignoring the things that are seemingly
irrelevant and idealizing objects and relations that are too complex to describe
in detail are necessary features of science and part and parcel of laws as well
as theories and models. While it is certainly true that we distinguish between
phenomenological and more fundamental laws, it is usually the former that
are thought of as giving an incomplete description of the relevant phenomena.
This is because they fail to appeal to underlying microprocesses of the kind
necessary for understanding much of the behavior of physical systems. What
the Hardy-Weinberg law shows is that embedded in what Cartwright would
call a “false” law is a great deal of accurate information about biological popu-
lations, information that is embedded in the synthesis of Mendelian heredity
and Darwinian natural selection. To that extent, it serves as an example of how
mathematical abstraction can enhance our understanding far beyond simple
predictive capabilities.10
As I mentioned above, the question of whether the kinds of models used
in population genetics provide explanation and understanding of biological
processes is intimately linked to the question of whether biology is similar to
physics. Does the mathematical nature of models and explanation in biology
under standing in ph ysic s and biol o gy 

provide us with understanding of populations in the same sense that math-

ematical laws and models in physics provide information about physical sys-
tems? As we shall see below, the Hardy-Weinberg law enables us to understand
fundamental features of heredity and variation by establishing a mathemati-
cal relation between allele and genotype frequencies that embodies the very
gene-conserving structure that is the essential feature of Mendelism. Hence,
the claim that the law is false in some sense misses the point if our concern is
understanding and conveying information. What is important for my purposes
here is to show why the unrealistic nature of its assumptions in no way affects
the significance of either the conclusions it provides or the information implicit
in its formulation. And once again we can differentiate the mathematical ab-
stractions from other types of idealizing assumptions that figure in the formu-
lation of the law and show why the former are essential for understanding the
basis of the mechanics of evolution.
The Hardy-Weinberg law is often described as a consequence of Mendel’s
law of segregation, or a generalization of Mendel’s laws as applied to popula-
tions. It relates allele or gene frequencies to genotype frequencies and states
that in an infinite, random mating population in the absence of external fac-
tors such as mutation, selection, sampling error, and migration, one generation
of random mating will produce a distribution of genotypes that is a function
solely of allele frequencies and does not change over subsequent generations,
provided all conditions are held constant. In other words, if we have a pair of
alleles A1 A2 at a particular gene locus and the initial ratio of A1 to A2 is p to q,
then for every succeeding generation the ratio will be p to q, and regardless of
the distribution of genotypes in the initial generation, the distribution for all
succeeding generations will be

p2 is just the probability of getting an A1 A1 homozygote, which is the probabil-

ity that the egg is A1 times the probability that the sperm is A1 (by the product
rule for independent events). Both of these probabilities are p because in its
simplest form the law assumes that the species is hermaphroditic. Since the
heterozygote can be formed in two different ways, the probability is  pq (by
the addition rule for mutually exclusive events). So, if you know the value for p,
then you know the frequencies of all three genotypes.
Since random mating does not change allele frequencies, all one needs to
calculate the genotype frequencies after a round of random mating are the al-
lele frequencies before random mating. In populations where each individual
is either male or female with different allele frequencies, it will take two gen-
 m ar g ar et mo r r iso n

erations to reach Hardy-Weinberg equilibrium. One can see, then, the relation
between the stability of the frequencies and Mendel’s law of segregation. With
random cross-fertilization there is no disappearance of any class whatever in
the offspring of the hybrids; each class continues to be produced in the same
proportion.11
But, and here is the important point, what is significant about the Hardy-
Weinberg law is not so much the binomial form of the genotype frequency and
the prediction of genotypes based on the stability of the population, but rather
what the stability actually shows or presupposes. Despite the unrealistic as-
sumptions, the stability allows us to comprehend something about Mendelian
populations that is significant for comprehending heredity and variation. In
other words, certain conditions must be present for the stability to be possible.
Thus, the predictive success of the law is intimately connected with certain ba-
sic claims about genetic structure. What the Hardy-Weinberg law says is that
if no external forces act, then there is no intrinsic tendency for the variation
caused by the three different genotypes that exist in a population to disappear.
It also shows that because the distribution of genotype frequencies is inde-
pendent of dominance, dominance alone cannot change genotype frequencies.
In other words, there is no evidence that a dominant character will show a
tendency to spread or allow a recessive one to die out. Instead, the genotype
frequencies are maintained in constant proportions. The probabilistic genetic
structure is conserved indefinitely; but should it be influenced by an outside
force, for example, mutation, the effect would be preserved in a new, stable
distribution in the succeeding generation.
This was crucial for understanding the problems with blending inheritance
as advocated by the Darwinians, and to that extent the claim that the law is false
in some sense misses the point if our concern is understanding and conveying
information. Under blending inheritance, variation was thought to decrease
rapidly with each successive generation, but Hardy-Weinberg shows that under
a Mendelian scheme it is maintained. This pointed to yet another fundamental
aspect of Mendelism, namely, the discontinuous nature of the gene and why
it was important for the preservation of variation required for selection. How
was it possible for the genetic structure to be maintained over successive gen-
erations? The reason for the stability could be traced directly to the absence of
fusion, which was indicative of a type of genetic structure that could conserve
modification. This condition was explicitly presupposed in the way the law was
formulated and how it functioned. In that sense, then, one can see the Hardy-
Weinberg law as the beginnings of a completely new understanding of the role
of mutation and selection and how they affect our understanding of evolution.
Appealing to the abstraction/idealization distinction can also clarify our
under standing in ph ysic s and biol o gy 

understanding of how deviations from the conditions or assumptions specified

by the law affect its applicability. Essentially we can divide the assumptions
associated with the Hardy-Weinberg law into two groups. The first involves as-
sumptions that do not allow for relaxation without violating Hardy-Weinberg
equilibrium, such as infinite population size and random mating. The second
includes the absence of selection, migration, and mutation. These assump-
tions affect allele frequencies but not random mating. Violations of these latter
conditions will not rule out Hardy-Weinberg proportions; instead, the allele
frequencies will change in accordance with the changing conditions. In other
words, these conditions function as idealizations that may or may not hold but
whose effect on the system can be straightforwardly calculated. Put differently,
we can think of them as external factors that isolate basic features of a Mende-
lian system that allow us to understand how variation could be conserved.
Contrast that situation with the requirements of infinite populations and
random mating. Infinite populations are crucial in that one must be able to rule
out genetic drift, which is a change in gene frequencies that results from chance
deviation from expected genotypic frequencies. That is, we must be able to
determine that detected changes are not due to sampling errors. Although ran-
dom mating seems like the kind of restriction that is typically violated, we can
see how its violations affect gene frequencies: in the case of assortative mat-
ing, there will be an increase in homozygosity for those genes involved in the
trait that is preferential, such as height or eye color. Traits such as blood type
are typically randomly mated. Similarly in the case of inbreeding there will be
an increase in homozygosity for all genes. Because both of these assumptions
are necessary for Hardy-Weinberg equilibrium they cannot, in general, be cor-
rected for and hence are fundamental for the formulation of the law. In other
words, they ought to be considered abstractions rather than idealizations be-
cause they are necessary for explaining how the law enables us to understand
particular features of Mendelian populations. We can see here that the sig-
nificance of the Hardy-Weinberg law extends beyond understanding the role of
dominance. By defining the relation between allele frequencies and genotype
frequencies it laid the foundation for modeling evolution: given certain condi-
tions it is possible to show that a distribution of genotypes is a function solely
of allele frequencies.
While the Hardy-Weinberg law is based on rather simple mathematics (ba-
sically just multiplication and probability) the conceptual revolution it began
was extended by R. A. Fisher, who introduced not only the analysis of vari-
ance but also the more abstract mathematics of diffusion theory, Markov chain
models, and branching processes into genetics. While some results in popula-
tion genetics do not require abstract mathematics there are some that could
 m ar g ar et mo r r iso n

not have been arrived at nonmathematically, such as the results in the multilo-
cus and the stochastic theories. Moreover, the fact that selection could operate
in Mendelian populations could only be shown using the kind of mathematical
abstraction that we frequently see in physics.

>cÒc^iZEdejaVi^dch/;gdbHiVi^hi^XVaE]nh^Xhid
BVi]ZbVi^XVa7^dad\n

Despite remarkable theoretical success, some eminent biologists and phi-

losophers have voiced criticism about the mathematical nature of the models
used in population genetics. For example, in responding to R. A. Fisher’s ()
first technical paper on the topic, a paper which marks the origin on the syn-
thesis of Mendelism with Darwinism, the referee Reginald Punnett complained
that it displayed the hallmarks of treating weightless elephants on frictional
surfaces.12 Ernst Mayr () complained that population genetics, to the ex-
tent that it treats evolution as mere changes in gene frequency (as an input or
output of genes), is like beanbag genetics involving the addition and removal of
beans from a bag. Here Mayr was echoing an earlier remark made by Wadding-
ton (), who claimed that the mathematical theory of evolution had not led
to any noteworthy quantitative statements nor revealed any new type of rela-
tions or processes that could explain phenomena that were previously obscure.
Mayr () also remarked that one cannot really understand the workings
of natural selection unless one understands the uniqueness of the individual,
something that population genetics clearly ignores. Moreover, Provine ()
has claimed that the models of the s, still widely used today, are an impedi-
ment to the understanding of evolutionary biology.
As I mentioned above, what motivates many of these criticisms is the view
that the methods of population genetics, with their emphasis on reduction and
gene selection, simply ignore many of the important factors that figure in evo-
lutionary development. They just are not explanatory in the sense that provides
us with a proper understanding of the evolutionary process. All of this seems to
presuppose that the uniqueness of biological individuals must be taken account
of in explanations in the way that the uniqueness of physical objects need not.
Idealization and abstraction can be informative and aid in our understanding
of physical systems, but not in the case of biological populations. A closer look
reveals that not only is this not the case, but it is difficult to see how it could
be the case. The fact that the mathematics of population genetics has estab-
lished results that were impossible using direct observation and other types of
empirical methods is certainly sufficient for claiming that it has increased our
understanding of evolutionary processes. Fisher’s mathematization of selection
under standing in ph ysic s and biol o gy 

created a new framework in which its operation was understood as an irreduc-

ibly statistical phenomenon, a reconceptualization that emerges in conjunction
with the application of specific mathematical, as opposed to purely statistical,
techniques.13
Karl Pearson’s statistical account of selection (biometry) formed the basis
for much of Fisher’s thinking about Darwinism. However, because the Darwin-
ian foundations of biometry provided no understanding of variation, Fisher’s
goal was to unite the two approaches into a consistent theory of genetical in-
heritance. What Fisher () wanted to do was determine the extent to which
characteristics such as stature were determined by a large number of Men-
delian factors. Studies had shown that in the case of brothers the correlation
coefficient is around . (amount of variance due to ancestry), which leaves 
percent of the variance to be accounted for in some other way. Fisher wanted
to determine how much of the total variance was due to dominance, how much
resulted from other environmental causes, and how much from additive genet-
ic effects. If one could resolve observed variance into these different fractions
(that is, expressing these fractions as functions of observed correlations), then
one could easily determine the extent to which nature dominated over nurture.
Using the analysis of variance, Fisher succeeded in distinguishing between ge-
netic and environmental variance but also between the different components
of genetic variance itself (additive and dominance).
In order to perform this kind of statistical analysis, Fisher made a number
of explicit assumptions that were clearly at odds with some of Pearson’s (;
a, b) earlier investigations regarding a possible compatibility of Mendel-
ism and Darwinism. Although Pearson believed that biology differed from
physics in the sheer number of variables one needed to take account of in any
single case of inheritance, he did think that there were certain features of a
population that needed to be specified in order to arrive at a proper statisti-
cal description.14 Perhaps the most important difference between Fisher and
Pearson was Fisher’s assumption of an indefinitely large number of Mendelian
factors (genes) which seemed not only out of the region of experiment using
Mendelian methods but was also mathematically intractable for biometrical
statistics. The mathematical difficulties arose because Pearson assumed that
for each Mendelian factor one needed to know certain specific information,
such as which allelomorph was dominant and the extent to which dominance
occurred. In addition, one needed to know what the relative magnitudes of the
effects produced by different factors were, whether the factors were dimorphic
or polymorphic, to what extent they were coupled, and in what proportion the
allelomorphs occurred in the general population.
One can begin to see, then, that an analysis involving a large number of
 m ar g ar et mo r r iso n

genes was virtually impossible using biometrical techniques. In addition to the

information above there were more general considerations that needed to be
taken into account, such as the effects of homogomy (preferential mating) as
opposed to random mating and selection versus environmental effects—all of
which needed to be treated separately if one was to determine the genetic basis
of the inheritance of particular characteristics.
Pearson thought that if one assumed an indefinite number of Mendelian
factors then the nature of the population could not be specified in any complete
sense, thereby undermining any statistical result that might follow. In other
words, we need certain kinds of information about individuals that make up a
population in order for our methodology to give us reliable results.
However, if one takes as a model the velocity distribution law in statistical
physics (which Fisher did), then just as a sufficiently large number of indepen-
dent molecules would exhibit a stable distribution of velocities, a sufficiently
large number of Mendelian factors or genes in a population will enable one to
establish general conclusions about the presence of particular traits. Contra
Pearson, Fisher did not assume that different Mendelian factors were of equal
importance, so all dominant genes did not have a like effect. In order to simplify
his calculations Fisher also assumed random mating as well as the indepen-
dence of the different factors. Finally, because the factors were sufficiently nu-
merous, some small quantities could be neglected. So, not only did Fisher differ
from Pearson with respect to specific assumptions about the nature of Men-
delian factors (that all were equally important, for example), but assumptions
necessary to characterize a Mendelian population were much more general. By
assuming an indefinite number of factors it was possible to ignore individual
peculiarities and obtain a statistical aggregate that had relatively few constants.
Underwriting these results is, of course, the central limit theorem.
Once the causes of variance were determined, Fisher () went on to
specify the conditions under which variance could be maintained. This was es-
pecially important since in a Mendelian system loss of genetic variability would
be infinitely less than with blending inheritance and due only to finite popula-
tion size and consequential stochastic losses. So, how would the rate of loss
compare with the gains through mutation under differing assumptions about
selection? How would gene frequencies change under selection pressures and
environmental conditions?
To answer these questions, Fisher began with a discussion of equilibrium
under selection. He first demonstrated that the frequency ratio for the alleles
of a Mendelian factor was a stable equilibrium only if selection favored the
heterozygotes. He then showed that the survival of an individual mutant gene
depended on chance rather than selection. Only when large numbers of indi-
under standing in ph ysic s and biol o gy 

viduals were affected would the effect of selection override random survival,
and even then only a small minority of the population would be affected. To
do this he introduced stochastic considerations and examined the survival of
individual genes by means of a branching process analyzed by functional itera-
tion and then set up the “chain-binomial” model and analyzed it by a diffusion
approximation.15 He was able to calculate the amount of mutation needed to
maintain the variability given a specific amount of selection and found that to
maintain variability in the case of equilibrium in the absence of selection, the
rate of mutation had to be increased by a very large quantity. So the presence
of even the slightest amount of selection in large populations had considerably
more influence in keeping variability in check than did random survival. Con-
sequently, the assumption of genotypic selection balanced by occasional muta-
tions fit the facts deduced from the correlations of relatives in humans.
So, by making assumptions about the large size of the population and its
high degree of genetic variability, Fisher was able to demonstrate how his
stochastic distributions led to the conclusion that natural selection acting on
genes (rather than mutation, random extinction, epistasis, and so on) was the
primary determinant in the evolutionary process. He found that mutation
rates significantly higher than any observed in nature could be balanced by
very small selection rates. The distribution of the frequency ratios for different
factors was calculated from the assumption that the distribution was stable.
The kind of statistical independence that figured prominently in the velocity-
distribution law was applicable to the effects of selection in Mendelian popula-
tions. In keeping with the idealization/abstraction distinction I have been mak-
ing we can see why Fisher’s assumptions and techniques fall into the category of
mathematical abstraction. Without the requirement of an infinitely large popu-
lation of genes and the mathematical techniques necessary for dealing with it in
the context of other aspects of the population, Fisher would have encountered
the kind of problems that Pearson envisaged using straightforward statistical
analysis. In that sense, mathematical abstraction became a necessary feature
for understanding basic features of Mendelian populations.
What can we say therefore about the understanding of selection that
emerges from Fisher’s analysis? The distribution of the gene ratio provides the
ultimate expression of selective effects because the gene remains the only trace
of the existence of an individual in a population.16 Given that selection acts
on the heritable, what is important is the mean effect of each allele. Although
we cannot know the state of each of the genes in the population, we can know
the statistical result of their interaction in the same way that gas laws can be
deduced from a collection of particles. Selection is mass selection, taking into
account only the additive effects of genes; stochastic factors can be ruled out
 m ar g ar et mo r r iso n

because of population size. Selection became irreducibly statistical because: ()

it applies only at the level of large populations and is defined in terms of gene
frequencies; () this kind of (large) population is necessary if one’s statistical
methodology is to be objective and guarantee the appropriate kinds of results.
Consequently, evolution could be understood as the modification of genetic
structure with the gene ratio constituting the real state of the species. Indeed,
one can think of all observable processes of evolution as described in the lan-
guage of the statistical distribution of genes. The natural population is simply
an aggregate of gene ratios.
Fisher’s differences with Pearson are, to be sure, ones that result from
bringing Mendelism into the equation, but these differences go well beyond
any dispute about its status as a theory of heredity. For Pearson, selection is
still about the kind of individual differences that were important for Darwin;
however, because we cannot have a proper science of individuals, selection has
to be established by using statistical techniques to analyze the occurrence of
traits in a population. Because the existence of selection is established statisti-
cally and because statistical knowledge is the only kind of knowledge that can
justify the principle of natural selection, a redefinition of the concept is neces-
sary. Although Pearson defines selection statistically it is still thought to func-
tion at the level of individuals. But this is largely a methodological issue about
what constitutes proper scientific investigation rather than a substantive mat-
ter about the way selection operates. That is to say, although Pearson’s account
of selection has become absorbed into the theory of correlation, we still have
the fundamental notion that its task is to explain why individuals, understood
as members of a population, have the traits they do. However, in the hands
of Fisher, selection becomes irreducibly statistical because the mathematics
used to describe it no longer allows it to explain the occurrence of individual
traits. It is now understood as a population-level phenomenon explained in
terms of changing gene frequencies. In that sense, the mathematics used to
formulate an account of how selection operates determines the way it should
be understood.
A good deal of Fisher’s methodology, including his understanding of se-
lection, forms the foundation of modern population genetics. There remain,
however, debates about whether selection operates most effectively in the large
populations, and the role that random drift (sampling error), migration, and ge-
netic interaction play in affecting gene frequencies. However, what is most im-
portant is that selection is now defined in terms of changes in gene frequencies
and that this reconceptualization resulted largely from the impact of mathemat-
ics in describing that process. A full discussion of the details surrounding the
nature of selection would require a more in-depth study of the claims of people
under standing in ph ysic s and biol o gy 

like Mayr who argue that the methods of population genetics provide little in
the way of understanding evolutionary development. However, what I take my
analysis to have shown is that this is at least prima facie incorrect. While it is
true that our understanding of selection has departed from the traditional Dar-
winian one, it is also the case that a properly quantitative account of selection as
the mechanism of evolutionary change necessitated this departure.17

8dcXajh^dc

What I have tried to do here is first to highlight similarities in the use of

mathematical abstraction for characterizing and understanding features of
physical and biological systems. Part of my purpose in telling that story was to
highlight the point, made at the outset, that what it means to understand and
explain certain types of behavior and phenomena can be different in different
contexts; but what is common to each case—the use of abstract mathematical
models—often defines what it means to understand and explain the phenom-
ena in question. In keeping with that goal I also wanted to show how abstract
mathematics can function in ways that go well beyond simply representing
phenomena for the purposes of exact calculation and prediction. I hope the
examples illustrate how mathematical representations provide a level of under-
standing and explanation that is sometimes impossible using more empirically
based forms of description and investigation.
This point was made not only with respect to phase transitions in phys-
ics but also in both the Hardy-Weinberg law and the models used by Fisher
in the development of population genetics. In the case of Hardy-Weinberg,
both the abstraction and idealization present in the law actually embody a
great deal of information that allows us to understand why the law functions
as a successful predictive device. Even in biological contexts, where details
of complex phenomena are thought to be important, mathematical abstrac-
tion can function as the source of concrete information about the behavior of
these systems.
Because much of the knowledge that we have of both physical and biologi-
cal systems comes to us via abstract mathematical models used to represent
the phenomena in question, the “mediated” knowledge those models provide
needs to be properly positioned in the epistemic hierarchy. To do this, one
needs to recognize the role that mathematical abstraction plays in producing
explanation and understanding. To rule out this kind of knowledge simply be-
cause it relies on mathematically abstract descriptions and principles is to ren-
der incomprehensible much of what contemporary science currently claims to
understand.
 m ar g ar et mo r r iso n

CdiZh
I would like to thank the Social Sciences and Humanities Research Council of Canada
for research support. I would also like to thank the editors for helpful comments.
. For example, we understand a mathematical concept if we can apply it to solve prob-
lems in ways other than simply reproducing the example used for learning the concept;
and we understand a phenomenon like the weather if we are able to make predictions
and explain the reasons for those predictions. The point here is simply that in some cases
prediction is a feature of understanding and not in others.
. I realize, of course, that de Regt and Dieks may not agree with my interpretation of
their position. My intention here is not to argue for my view of explanation or understand-
ing but simply to state how I see the relation between the two for the purposes of this
chapter.
. My use of the terms abstraction and idealization differs from many of the accounts
in the literature, for example, Cartwright (). Instead of defining abstraction as leaving
out properties that belong to the system in question and idealization as the distortion of
certain properties (for example, point particle), I have chosen, instead, to focus on how
certain types of abstract mathematical frameworks are necessary for representing physical
systems/phenomena. My reason for changing the terms of reference here relate to what
I see as a deficiency in the standard account where the motivation for idealization and
abstraction is typically related to modeling assumptions or ease of calculation. In cases
of both idealization and abstraction, we are able to add back properties or de-idealize in
order to make the system more realistic or concrete. However, it says nothing about how
we should understand cases where this is not possible—where the abstract mathematical
representation is necessary for representing the target system. Because of the importance
of this latter type of abstraction for mathematical modeling, we need to have some account
of the way it functions in our understanding of the model systems. For more on this way of
characterizing the distinction, see Morrison (b).
. In Morrison () I claim that what makes a model autonomous is not simply that
it is not derived from theory or phenomena but involves elements from a variety of dif-
ferent sources that, taken together, make the model a unique kind of entity. It is by virtue
of this uniqueness that they can and do play a mediating role. With models, and in other
situations as well, mediators need to occupy a relatively independent position if they are to
properly fulfill the function they are assigned. But, their role as a mediator can vary widely.
Indeed, in some cases the structures comprising the model fail to provide any clues about
the physical nature of the system in question, and the model also fails to point to specific
relations/reduction to some background theory. Instead, models may be used simply for
their predictive power or to account for specific kinds of behaviors without interpreting
them as having any realistic representational power whatsoever. In these latter cases, the
models do not really provide us with any understanding of the system/phenomena in ques-
tion; they save the phenomena without answering questions about why the system behaves
as it does. An example of this kind of model is the aether models developed by James Clerk
Maxwell in his early formulations of the electromagnetic theory. While I still believe these
considerations are important in characterizing a model’s autonomy, other features also
play a crucial role—one of which is the way the model represents the system of interest.
under standing in ph ysic s and biol o gy 

My concern in this chapter is how mathematical abstraction represents systems in a way

that enables us to understand fundamental features about their behavior.
. Space prevents me from dealing with the often thorny issues surrounding what con-
stitutes a representation in the context of models and theories. For an extended discussion
of this, see Morrison ( and a). For my purposes in this chapter I do not neces-
sarily require a well-worked-out notion of representation in order to address some of the
concerns surrounding the relation between understanding and mathematical abstraction.
. The existence of these “universal” properties indicates that super conductivity is a
thermodynamic phase. For more on this, see Morrison (a).
. For an excellent discussion of issues surrounding the notion of reduction and the
thermodynamic limit, see Batterman ().
. As Sabina Leonelli has pointed out to me (personal communication, ), there
are several types of fast evolving populations (such as viruses) whose evolution can be
observed and measured, but modeling these systems does not provide the basis for general
models of evolving populations in the way that more mathematically based approaches do.
. My claim here is not that population genetics captures all of mathematical model-
ing in biology; rather, that many population geneticists see their methods as providing the
foundations for understanding the genetical basis of evolutionary change. There are many
areas of evolutionary biology that consist of loosely connected models that are not part
of any unifying theoretical framework, but that point is irrelevant for how we understand
what happens when we combine population level processes (such as selection) with the
rules of transmission genetics.
. Why this is an instance of abstraction as opposed to idealization will be discussed
below.
. The law of segregation refers to the fact that the characters that differentiate hybrid
forms can be analyzed in terms of independent pairs; that is, each anlagen acts sepa-
rately—they do not fuse. We can also understand this as stating that any hybrid for a given
character produces an offspring distributed according to definite proportions. If the pure
parental forms are A1 and A2 and the hybrid A1 A2 , then the offspring of the hybrid will be
distributed according to the ratio 1A1 : 2 A1 A2 :1A2 . Pearson () was probably the first
to show the relation between the law of segregation and the stability of a population in the
absence of selection.
. As quoted in Norton and Pearson ().
. As I have argued elsewhere (Morrison , ), the synthesis of Mendelism
and Darwinian selection was accomplished through the employment of mathematical
techniques that allowed its founders (R. A. Fisher, Sewall Wright, and J. B. S. Haldane) to
establish the operation of selection in Mendelian populations. To that extent, the synthesis
produced an enhanced understanding of selection as something compatible rather than
at odds with Mendelism. But, the interesting aspect of the synthesis was that while its
authors, particularly Fisher () and Wright (, ), agreed on the general conclu-
sion, each had a very different account of how selection functioned and the conditions
under which is would be most effective.
. “The causes of any individual case of inheritance are far too complicated to admit
of exact treatment” (Pearson , ).
. For more on these techniques, see Edwards () and Ewens ().
 m ar g ar et mo r r iso n

. That is, the Mendelian mechanism ensures that although a population may be said
to have continued existence, the individuals that comprise it do not. The variability that
passes from one generation to the next through reproduction is related to but not identical
to phenotypic variability.
. For a longer discussion of these issues, see Morrison (b).

GZ[ZgZcXZh
Batterman, R. . Critical phenomena and breaking drop: Infinite idealizations in phys-
ics. Studies in History and Philosophy of Modern Physics :–.
Cartwright, N. . How the laws of physics lie. Oxford: Clarendon Press. .
———. . The dappled world. Cambridge: Cambridge University Press.
. How the laws of physics lie. Oxford: Clarendon Press.
de Regt, H. W., and D. Dieks. . A contextual approach to scientific understanding.
Synthese :–.
Earman, John. . Curie’s principle and spontaneous symmetry breaking. International
Studies in the Philosophy of Science :–.
Edwards, A. W. F. . The fundamental theorem of natural selection. Biological Review
:–.
Ewens, W. . Mathematical population genetics, vol. , Theoretical introduction. New
York: Springer.
Fisher, R. A. . The correlation between relatives on the supposition of Mendelian
inheritance. Transactions of the Royal Society of Edinburgh :–.
———. . On the dominance ratio. Proceedings of the Royal Society of Edinburgh
:–.
Friedman, M. . Explanation and scientific understanding. Journal of Philosophy
:–.
Hempel, C. G. . Aspects of scientific explanation. In Aspects of scientific explanation
and other essays in the philosophy of science, –. New York: Free Press.
Kitcher, P. . Explanatory unification and the causal structure of the world. In Scientific
explanation, edited by P. Kitcher and W. C. Salmon, –. Minneapolis: University
of Minnesota Press.
Mayr, Ernst. . Where are we? Reprinted in Ernst Mayr, Evolution and the diversity of
life. . Cambridge, MA: Harvard University Press, –.
———. . The growth of biological thought: Diversity, evolution, and inheritance. Cam-
bridge, MA: Belknap Press.
Morrison, M. . Modelling nature: Between physics and the physical world. Philosophia
Naturalis :–.
———. . Models as autonomous agents. In Models as mediators: Perspectives on natu-
ral and social science, edited by M. S. Morgan and M. Morrison, –. Cambridge:
Cambridge University Press.
———. . Unifying scientific theories: Physical concepts and mathematical structures.
Cambridge: Cambridge University Press.
———. . Modelling populations: Pearson and Fisher on Mendelism and biometry. Brit-
ish Journal for the Philosophy of Science :–.
under standing in ph ysic s and biol o gy 

———. a. Emergence, reduction, and theoretical principles: Rethinking fundamental-

ism. Philosophy of Science :–.
———. b. Scientific understanding and mathematical abstraction. Philosophia
:–.
———. . Where have all the theories gone? Philosophy of Science :–.
———. a. Models as representational structures. In Nancy Cartwright’s philosophy of
science, edited by S. Hartmann, L. Bovens, and C. Hoefer, –. London: Routledge.
———. b. Fictions, representations, and reality. In Fictions in science, edited by M.
Suarez, –. London: Routledge.
Norton, B. J., and E. S. Pearson. . A note on the background to and refereeing of R. A.
Fisher’s  paper “On the Correlation of Relatives on the Supposition of Mendelian
Inheritance.” Notes and Records of the Royal Society :–.
Pearson, K. . Mathematical contributions to the theory of evolution, III: Regression,
heredity and panmixia. Philosophical Transactions of the Royal Society of London A
:–.
———. . Mathematical contributions to the theory of evolution, XII: On a generalized
theory of alternative inheritance with special reference to Mendel’s laws. Philosophical
Transactions of the Royal Society of London A :–.
———. a. The theory of ancestral contributions in heredity. Proceedings of the Royal
Society B :–.
———. b. On the ancestral gametic correlations of a Mendelian population mating at
random. Proceedings of the Royal Society B :–.
Provine, William B. . The origins of theoretical population genetics. Chicago: University
of Chicago Press.
Salmon, W. C. . Scientific explanation and the causal structure of the world. Princeton:
Princeton University Press.
van Fraassen, B. C. . The scientific image. Oxford: Clarendon Press.
Waddington, C. H. . The strategy of the genes. London: George Allen & Unwin.
Wright, S. . The effects of inbreeding and crossbreeding on guinea pigs. U.S. Depart-
ment of Agriculture Bulletin :–, :–.
———. . Evolution in a Mendelian population. Anatomical Record :.
 tarja knuu t til a and m artina mer z

/
Understanding by Modeling
An Objectual Approach

I6 G ? 6  @ C J J II > A 6  6 C 9  B 6 G I > C 6  B : G O

 Scientific understanding has, so far, mainly been discussed in terms

of the intelligibility of theories and the relation between understanding and
explanation. In the course of this debate several theoretical virtues have been
identified as being conducive to such understanding. Interestingly, the virtues
addressed are the ones that have also occupied a prominent place in the discus-
sion of models. The epistemic value of models, for example, has been ascribed
to their providing us with visualizations (Nagel ; Griesemer ), speci-
fying relevant causal mechanisms (Glennan ; Darden ), equipping us
with mathematically tractable systems (Humphreys ), describing different
kinds of phenomena by means of common templates (Black ; Humphreys
), and telling “plausible” stories (Hartmann ; Morgan ). All this
suggests that there is a close link between scientific understanding and model-
ing, which makes one ask what it is about models that forges such a link.
Our aim in the following is to account for this intimate relationship be-
tween modeling and scientific understanding. Since there are many different
approaches to models in the philosophy of science we will first engage in a dis-
cussion concerning what kinds of entities models are and how they provide us
with knowledge. We will argue for approaching models as productive entities
that assume many roles in scientific research and are valued for the results they
produce. From this perspective, models appear as concrete constructed objects
with which one can interact in various ways. We suggest that such an objectual
approach can accommodate their understanding-providing features.


under standing by modeling 

Moreover, an objectual approach offers a solution to several philosophical

puzzles concerning scientific understanding that have undermined the cred-
ibility of that very concept. For example, the notions of skill, experience, and
familiarity have occupied an important role in discussions of scientific under-
standing. At the same time, they have appeared suspect from the philosophical
point of view because of their psychological, individual, and everyday char-
acter. It is our contention that, if we consider experience, skill, and (the feel-
ing of ) familiarity as grounded in the interactive relations between researchers
and their research objects, the concepts lose their overly mental and subjective
character. It also seems that the recent approaches to scientific understanding
that liken it to the abilities of scientists in one way or another (see de Regt and
Dieks ; Ylikoski, this volume; Kuorikoski ) could be fruitfully com-
pared and combined with the objectual approach we are putting forward.
We will proceed by outlining the recent philosophical discussion on the
subject of models and argue that instead of conceiving of them as represen-
tations, one should rather consider them from a productive perspective. We
will then introduce an approach that relates scientific understanding to the
scientists’ abilities that are manifested in the interaction between them and
their research objects—in which models play an increasingly important role in
contemporary science. This close connection between modeling and scientific
understanding is then exemplified by two cases based on our earlier research:
in a discussion concerning event generators in particle physics and parsers in
language technology, we will show how scientific understanding emerges in
and is intertwined with the development and use of models.

I]ZGZegZhZciVi^dcVakZghjhi]ZEgdYjXi^kZ6XXdjcid[BdYZah

At the outset, understanding seems to be akin to a special kind of knowl-

edge, which is not restricted to “knowledge that,” being, rather, knowledge
concerning the hows and whys (see Lipton, this volume; Grimm ). If we
accept that scientific understanding and knowledge are inherently bound to-
gether, then one approach to the understanding-providing features of models
is to question how knowledge is attained through them. The philosopher’s gen-
eral answer to this question has been that models provide us with knowledge
because they represent their target systems more or less accurately.1 In line
with this view, a substantial part of the recent discussion on models has indeed
focused on the question: “In virtue of what do models represent their target
systems?” (Morrison ).
The consensus of philosophers of science concerning the representational
character of models seems rather superficial, however, given the fact that the
accounts given to representation differ widely. Moreover, these differences have
 tarja knuu t til a and m artina mer z

far-reaching consequences that have gone largely unnoticed. What is at stake?

Whereas the more traditional conceptions of representation have focused on
the properties by virtue of which models represent their target objects, the
recent pragmatic approaches view it as an accomplishment of representation
users. This, in turn, amounts to a minimalist account that somewhat paradoxi-
cally paves the way for a productive understanding of models that does not tie
their epistemic value to representation alone. In what follows, we claim that
approaching models as concrete constructed objects—as multiplex epistemic
artifacts—provides such a productive account. This move also explains why,
in scientific practice, knowledge, scientific understanding, and the scientists’
skills are often intertwined to such a degree that it is difficult to tell them apart:
both the use and development of models generates new knowledge and cre-
ates the abilities, which we often take as the criteria for understanding. Why,
then, does the pragmatist approach to representation empty this very concept
of much of its explanatory content with regard to the epistemic value of mod-
els, thus motivating the search for new approaches?2
The pragmatist approach to representation could be seen as a critique of
the structuralist notion that is part and parcel of the semantic conception of
models (or rather of theories, as it was originally formulated). The semantic
conception provides a straightforward answer to the question of how models
give us knowledge of the world: they specify structures that are posited as pos-
sible representations of either the observable phenomena or, even more ambi-
tiously, the underlying structures of the real target systems. This relation of
representation between a model and its target system has been formulated in
terms of isomorphism or—less frequently—of similarity (see, for example, van
Fraassen ; French and Ladyman ; Giere ). Thus, according to the
semantic view, the structure specified by a model represents its target system if
it is either structurally isomorphic or somehow similar to it.
Pragmatist critics of the semantic conception argue, among other things,
that isomorphism—being a symmetrical, reflexive, and transitive relation—
does not satisfy the formal and other criteria we might want to affirm of repre-
sentation (see, for example, Suárez ; Frigg ). Suárez has extended this
critique to similarity. For instance, both isomorphism and similarity denote
a symmetric relation, whereas representation does not: we want a model to
represent its target system, but not vice versa. Moreover, the isomorphism ac-
count does not accept false representations as representations. The idea that
representation is either an accurate depiction of its object—which is interpret-
ed in terms of isomorphism within the structuralist conception—or it is no
representation at all does not fit our actual representational practices. These
problems appear to be solved once the pragmatic aspects of representation are
taken into account. The users’ intentions create the directionality needed to
under standing by modeling 

establish a representative relationship: something is being used and/or inter-

preted as a model of something else, which makes the representative relation
triadic, involving human agency. This also introduces indeterminateness into
the representative relationship: human beings as representers are fallible.
In stressing the importance of human agency for what representation is all
about, the pragmatic approaches criticize the assumption in the semantic con-
ception that representation is a dyadic relation of correspondence between the
representative vehicle (a model) and its target (Suárez ; Giere ; Bail-
er-Jones ; Frigg ). The dyadic conceptions attempt, as Suárez ()
puts it, “to reduce the essentially intentional judgments of representation-users
to facts about the source and target objects or systems and their properties,”
whereas the point made in the pragmatic approaches is that no thing is a rep-
resentation of something else in and of itself. Consequently, what is common
among these approaches is a focus on the intentional activity of representation
users and the denial that representation may be based only on the respective
properties of the representative vehicle and its target object.
The gesture of relating representation to the intentional activity of its users
solves many problems of the semantic notion, but it comes at a price: if repre-
sentation is grounded primarily in the specific goals and the representing activ-
ity of humans as opposed to the properties of the representative vehicle and its
target, nothing very substantial can be said about it in general. This has been
explicitly admitted by proponents of the pragmatic approach (compare, for ex-
ample, Giere ), of whom Suárez () has gone furthest in arguing for a
“deflationary” account of representation that seeks not to rely on any specific
features that might relate the representative vehicle to its target. This has rather
radical consequences for how one should conceive of the epistemic value of
models. Namely, if we take models to be primarily representations and accept
the pragmatic notion of representation, we are also left with a minimal explana-
tion of the epistemic value of models. Thus, when it comes to representation we
seem to face the following dilemma: either we choose the strong structuralist
notion with all its shortcomings, or we opt for the pragmatist alternative in-
stead, which is too weak to do any significant philosophical work. One obvious
way out of this dilemma is not to attribute the knowledge-bearing properties of
models to (successful) representation alone. Indeed, recent practice-oriented
studies differ significantly from the traditional approach that portrays mod-
els as ready-made representations of some predefined target phenomena. In
focusing on their multifunctional and experimental nature, these studies, we
suggest, emphasize in one way or another their productive character.
The productive character of models is nicely spelled out by Keller () in
her claim that models are “models for” something in addition to being “models
of.” She claims in her study of modeling in contemporary molecular biology—
 tarja knuu t til a and m artina mer z

where the design of new kinds of organisms on the one hand and new kinds of
computers on the other have converged—that the resulting “genetic comput-
ers” are not mere models in the traditional sense of the word: they are tools for
both material intervention and thinking. In much the same vein, Darden (,
) points out, “The model of DNA served as a model for investigating the
functions of the genetic material.” What is interesting about this remark is its
insight that “modeling of ” and “modeling for” proceed in tandem in scientific
research. The idea of models as tools for thinking is also present in concep-
tions that consider them to be surrogate systems (Suárez ). Investigating
economic models, which have often been considered to give notoriously false
depictions of their target systems, Mäki () characterizes them as “mini-
worlds” that are easier to study than the complex real world in which we live.
Consequently, as models enable further investigation of different phenomena,
a large part of their scientific and cognitive value is due to their workable and
tool-like character rather than to their success in representing some target phe-
nomenon more or less accurately.3
Considering models from a productive perspective requires one to address
them as autonomous, concrete objects. This is (partly) recognized by Morrison
and Morgan (), who focus in their account of models as investigative in-
struments on how we learn from them by constructing and manipulating them.
This, they claim, is made possible by the model’s partial autonomy, which they
interpret in terms of its relative independence from the traditional theory-data
framework. However, it seems that their view on modeling permits, and to be-
come efficient in fact requires, a more radical reading. If our aim is to stress how
models enable us to learn from the processes of constructing and manipulating
them, it is not sufficient that they are considered autonomous: they also need to
be concrete in the sense that they must have a tangible dimension that can be
worked on. This concreteness is provided by a material embodiment that gives
them the spatial and temporal cohesion that enables their workability, and that
preserves their identity across different sites and users. Consequently, once we
approach models as concrete things, the possible mediation that they provide
is no longer restricted to mediation between theory and data. Instead, scien-
tific models can then be viewed as concrete, human-made objects that draw
together numerous activities and widely different actors.4
The idea that models are concrete objects that have a material dimension
appears to run counter to the established philosophical view, which, apart from
treating models as representations, also tends to conceive of them as abstract
or conceptual things. However, it is difficult to see how one could work and
experiment with models and communicate one’s findings about them to oth-
ers without their concretization—in the form of diagrams or various symbols
under standing by modeling 

on paper or on screen. This also applies to so-called abstract models: when

working with them we typically construct and manipulate external diagrams
and symbols. Thus, even abstract entities need to have a material dimension if
they are to be able to mediate. Without materiality, mediation is empty. In sum,
the productive approach to models presumes that they are concrete objects
with which we can interact in various ways. This concrete objectual dimension
may, in turn, explain how scientific understanding is brought about by means
of model building and use.

6cDW_ZXijVa6eegdVX]idHX^Zci^ÒXJcYZghiVcY^c\

The idea that models are productive, concrete objects resonates nicely with
views that link scientific understanding to the abilities of scientists. Whereas
some writers claim that understanding is a kind of ability (Woodward ;
Ylikoski, this volume; Kuorikoski ), for others certain kinds of abilities
serve as a criterion for scientific understanding (de Regt and Dieks ). For
instance, Kuorikoski combines the Wittgensteinian nonmentalist conception
of understanding with the manipulative theory of causation and explanation
presented by James Woodward in his aim to show that understanding need not
be seen as any “special or privileged mental state” (Kuorikoski ). Instead,
he claims that it is constituted of our perceived abilities to infer and intervene,
which are due to our knowledge about the relevant causal dependencies. Thus,
the rationale behind likening understanding to abilities is that we typically tend
to attribute it to others to the extent that they are able to use their knowledge
successfully (de Regt ).5 These abilities are manifested in several ways: by
inferring, manipulating, and building successfully or by being able to account
for a phenomenon theoretically (see Ylikoski, this volume). Thus, understand-
ing appears as knowledge put into use, which might explain why understanding
and knowledge are often so difficult, if not impossible, to untangle.
One way of examining more closely the aforementioned entanglement of
knowledge and understanding is to ask how epistemically relevant abilities arise.
From the objectual perspective knowledge, understanding, and manipulative
abilities could arise in the very same process of interacting with external things.
This applies to modeling as well as to experimentation. While interventional
and manipulative abilities have been more readily ascribed to experimentation
than to theory (or theoretical models), our perspective on models depicts them
as objects through which we can make our theoretical ideas workable and “ex-
perimentable.” Thus, they also question the distinction between representing
and intervening (Hacking ), as they are entities constructed with the help
of different representational means for the purposes of manipulation.
 tarja knuu t til a and m artina mer z

The general principle guiding our objectual approach to scientific under-

standing is that scientific work typically revolves around the reactions of vari-
ous objects,6 including scientific instruments, experimental set-ups, and models.
We suggest that this interactive relationship between scientists and objects is
crucial to conceiving how modeling generates expertise and skills on the one
hand and scientific knowledge and understanding on the other. Conceived in
the same interactive process of modeling, our skills and knowledge are inher-
ently intertwined with each other and with the specificities of the models in
question. Moreover, even though the abilities that arise in interactions between
scientists and objects can ultimately be ascribed to the scientists as subjects,
the objects also contribute to the process: their properties have an important
role in guaranteeing understanding by influencing the outcome of the interac-
tion between scientists and their research objects.
But, one may wonder, should not the primary objects of scientific under-
standing be the real entities and processes rather than models? Surely our in-
teraction with models gives us understanding about them, but are not models
merely investigative instruments whose capability to teach us anything—their
autonomy and manipulability notwithstanding—is due to the fact that they
represent their target objects (compare Morrison and Morgan )? How-
ever, this kind of reasoning brings us back to the problem of representation.
Moreover, it seems that, as a result, a gap appears between our knowledge,
which is due to representation, and our skills, which are due to our interaction
with models. The reason for this is that if knowledge is ascribed to represen-
tation only, our interaction with models does not add any further knowledge
apart from what the representation already provides. Should this be the case, it
would indeed be difficult to say to which side scientific understanding belongs,
that of knowledge or that of skills.
One way to tackle this problem is to question the strict division between
real (natural) things and human-made intermediaries (representations, instru-
ments, and the like) that the aforementioned way of thinking implies. This
amounts to two somewhat different moves. First, from the objectual perspec-
tive there are no real objects purified from any artifice or interpretation. When
we look at the actual objects of science from a historical perspective, it seems
clear that they are entities that draw together concepts, theoretical ramifica-
tions, illustrations in different media, observations, and experimental results.
Various degrees of credibility and reality have been ascribed to these entities
(such as atoms, ether, cytoplasmic particles, unconsciousness) during their life
spans (Daston ). Thus, real things are also objects of our construction.
What is cognitively available to us is due to our conceptualizations through
representational, observational, and computational media.7
Yet, even if it were accepted that scientific objects are at least partly con-
under standing by modeling 

structed by us, it is still possible to maintain that there is a difference between

the objects that scientists aim to describe and explain and those things that
are merely means of achieving those ends. More interestingly, however, the
distinction between proper scientific objects and mere instruments can also
be questioned and relativized. In their work on scientific objects, Hans-Jörg
Rheinberger and Karin Knorr-Cetina claim that the entities being dealt with
in scientific practice can appear both as scientific objects (“epistemic things”)
and as mere tools (“technological objects”). Rheinberger (, ) zooms
in on the place of artifacts in an “experimental system” and argues that once
a scientific object (for example, a chemical reaction or a biological function)
has been thoroughly investigated, it can be turned into a tool that now assists
in the investigation of new scientific objects. Knorr-Cetina (, ), in
turn, focuses on today’s complex technologies, which are continuously being
transformed and, as a consequence, remain problematic while being used at
the same time. Thus, apart from functioning as mere tools, they are also akin
to scientific objects due to their problematic nature. For Knorr-Cetina (),
the dynamics of research are actuated by the “unfolding ontology” of objects.8
Because of their unfolding and incomplete ontology they remain provocative
for the researchers for extended periods of time, which also sheds light on the
affective component of research work. The insatiable quest of researchers to
pursue their objects in an effort to know more partly explains the intensely
pleasurable but also, at times, frustrating character of scientific endeavor.
Even though the objectual approach thus intentionally blurs the boundar-
ies between the natural and the artificial, as well as the objects of knowledge
and the means of pursuing them, it does not imply that scientific objects can be
constructed at the whim of individual researchers. Quite the contrary, the idea is
that objects themselves can be seen as “doers” (Knorr-Cetina , ). Be they
microbes, chromosomes, detectors, or mathematical objects, the various ob-
jects of knowledge have “powers, produce effects, may have their own internal
environments, mold perception, and shape the course of an experiment” ().
Unfortunately, this important role of epistemic objects is left largely unexplained
in the work of Rheinberger and Knorr-Cetina. This has led us in our earlier
work (Knuuttila , ; Merz , )—apart from targeting models as
epistemic objects—to consider their characteristic productivity and resistance
through the interplay of their material medium and their diverse uses.
In terms of productivity, the mere consideration of abstract “underlying
structures”—which constitute the kernel of the preferred semantic conception
of models—does not take us very far. This is also witnessed by the variety of
models used in science. We only need to consider the different kinds of models
used: scale models, diagrams, different symbolic formulas, model organisms,
computer models, and so on. This very wide variation suggests that the mate-
 tarja knuu t til a and m artina mer z

rial dimension of models and the diverse representational media they make
use of are crucial for promoting understanding through them. Consequently,
Knuuttila argues that, contrary to the philosophical tradition, one should take
into account the medium through which scientific models are materialized as
concrete, intersubjectively available objects (Knuuttila and Voutilainen ;
Knuuttila ). This media-specific approach focuses on their constraints and
affordances, which are partly due to the characteristics of the specific repre-
sentational means (diagrammatic, symbolic, -D, and so on) with which they
are imbued.9 There is evidence from studies in cognitive science that the rep-
resentational means used influences how well human subjects can grasp the
intended content. Zhang (), for instance, shows that different representa-
tions of the same abstract structures have different enablings as to how hu-
mans understand them. It seems that analogical and visual representations are
easier for humans to grasp than those in digital and numerical forms. This fits
well with the various pronouncements of scientists concerning the virtues of
visualization.
Apart from making use of representational media in an understanding-pro-
viding way, models typically also constrain the problem at hand, which renders
the initial problem situation more intelligible and workable. They provide us
with surrogate systems, the study of which furthers understanding by simpli-
fying or modifying the problems scientists deal with. Often the real world is
just too complex to study as such. Thus, scientific work typically proceeds by
turning the limitations built into the models into affordances; one devises the
model in such a way that one can learn and gain understanding from using or
“manipulating” it. Yet, as concrete objects, models have their own construc-
tion, which also explains how they can be doers: their construction resists some
uses and allows others. Moreover, they can lead to unexpected findings and
they also breed new problems.
We suggest that understanding through models comes typically by way of
building them, experimenting with them, and trying out their different alterna-
tive uses—which in turn explains why they are typically valued for their per-
formance and their results or output. Conceived of in this way, they may be
epistemically valuable even if they do not represent reality accurately. Purpose-
fully “wrong” models can give us knowledge that is commonplace in scientific
practice, which shows that models can be good inference generators even if
they do not represent their putative target objects accurately in any relevant re-
spect (this point is further discussed below in the parsing case; see also Suárez
). Moreover, the stress on artifactuality and materiality can accommodate
what Baird () calls “thing knowledge,” which is often tacit and bound more
to local environments and expertise than to explicit theoretical knowledge.
Furthermore, the artifactual view accommodates different kinds of epistemic
under standing by modeling 

strategies, which is testified by the development of digital technology. Whereas

it was believed earlier that understanding required us to know the structure
of the model, new simulation techniques point in a different direction, that
of pragmatic understanding oriented toward design rules and the dynamics
of the model, as well as toward predictions and output representations (Len-
hard , this volume; Humphreys ). Graphics in -D have also created
a new mode of interacting with numerical data, which allows users to gain un-
derstanding of the modeled objects through intervening kinesthetically in the
simulation process (Griesemer ; Myers ).
The productive objectual approach to models also enables us to see how the
same artifact can unfold into different objects depending on the specific func-
tion it assumes in the research process. Merz () argues that more complex
models in particular (for example, numerical and simulation models) accom-
plish different kinds of tasks in distinct contexts and for multiple actors. Spe-
cific meanings, functions, or viewpoints that are embodied in a model can be
drawn out selectively in accordance with the requirements of the local setting.
Models not only play different roles in the research process, but also oscillate
between them, which fosters dynamics that are not unidirectional. A model
does not necessarily turn from an epistemic object into an instrument or a tool
as time passes, as Rheinberger () suggests. Indeed, it may be simultaneous-
ly an epistemic object in one setting, part of the technical repertoire in another,
and yet something else in a third. In this sense, models are “multiplex objects,”
providing different kinds of understanding to various groups of scientists or
scientific fields (Merz ).
Moreover, models do not merely play a range of roles for different actors:
they also have an agency of their own. One might ask how scientists become
configured (that is defined, enabled, controlled) by specific objects—and not
the other way round.10 Salient features of models from this perspective include
how they regulate the scientists’ ways of handling them and suggest certain
kinds of solutions and forms of cooperation (Merz ). This configurative
power, which is important in terms of giving us understanding, is related on
the one hand to their material and medium-specific characteristics and on the
other hand to their multiplexity in scientific practice. It also connects with the
issue of autonomy, albeit, as we have previously argued, in a broader sense than
initially conceived of by Morrison and Morgan ().

Ild8VhZhd[JcYZghiVcY^c\WnBdYZa^c\

Since philosophical theorizing is bound to stay on the general level, we will

next turn to two empirical cases in order to exemplify the productivity of mod-
els and their role as understanding-providing interactive objects. The cases
 tarja knuu t til a and m artina mer z

concern event generators in particle physics and parsers in language technol-

ogy. Both belong to the category of computer models and simulations, which,
despite their rising importance in contemporary science, remain a rather un-
charted area philosophically. Models, with their characteristic multiplexity,
give us knowledge under their various guises as tools, objects, and inference
devices.

:kZci<ZcZgVidgh^cEVgi^XaZE]nh^Xh
Models in particle physics come in a variety of disguises, of which the mod-
el-based practice of computer simulation is particularly prevalent.11 Extensive-
ly applied, simulation contributes to the production of scientific knowledge in
various ways. We suggest that simulation work is predominantly “understand-
ing work” targeted at distinct kinds of understanding tasks. It can be produc-
tively viewed through its objects, the simulation programs (or models), which
draw together an entire web of scientific activities, different physics communi-
ties, and temporal phases.
Simulation is rendered productive for the purpose of a collider experiment.
Such experiments target new types of elementary particles (for example, the
Higgs boson) and their properties, which are traced and measured by particle
detectors with the aim of learning more about the basic constituents of mat-
ter. The novel particles and properties are mediated by collisions between two
particle beams, each collision generating a profusion of debris particles whose
properties are at stake. Simulation maps and mimics the aforementioned ex-
perimental process, that is, the production and subsequent detection of ele-
mentary particles as they are thought to occur in a “real”12 experiment. For this
purpose, the two simulation models that assume the roles of the collider (the
“event generator”) and the detector are coupled, and the output of the event
generator is fed into the detector simulation.
Event generators come in varying forms and sizes. Of particular impor-
tance are the comprehensive, multipurpose event generators that are made
publicly available throughout the particle-physics community. Event genera-
tors are not based on a unique theoretical framework: the models embodied in
them are of different theoretical origins. They account for the successive stages
of the collision process and its consequences (models of hard processes, parton
distributions, initial- and final-state radiation, beam remnants, fragmentation,
and decays, for example) and contain a combination of exact calculations, ap-
proximations, corrections, and parameterizations. Another set of ingredients
consists of the values of quantities determined in earlier experiments (particle
masses, decay widths, couplings, and cross sections, for example). The random
nature of physical processes is accounted for by Monte Carlo techniques. In a
simulation program, all of these ingredients are coupled by way of algorithms
under standing by modeling 

that turn the input (for example, type of colliding particles, beam energy) into
the output that consists of listings of the properties of all particles produced in
the event.
Due to their complex makeup and continuing adaptation to novel experi-
mental conditions and theoretical insights, event generators retain the status
of preliminary, unfinished, and unfolding objects. While their “authors” (that
is, constructors) continuously modify, improve, and update them to meet new
requirements, other physicists use them, at the same time, as tools for a range
of tasks. The object’s multiplex nature, which we describe below, goes hand in
hand with simulation as performed by members of different particle-physics
communities and thus produces different kinds of outcomes. The communities
comprise theoretical physicists (including the event-generator authors) and
experimental particle physicists (including most event-generator users), who
may be further differentiated in terms of their specific expertise (for example,
in detector systems or physics analysis). Event generators assist in coordinating
work across these communities. Below we outline three ways in which develop-
ing and working with simulations creates scientific understanding in particle
physics, in each case involving different actors, types of expertise, and simula-
tion roles.

Understanding Theoretical Models. The first understanding-providing

practice concerns event generators and their authors. A considerable propor-
tion of the models underlying an event generator constitutes original research
by its authors. These models are not always well understood before they are
implemented in the program. For this reason, using the event generator and
understanding the underlying theoretical models goes hand in hand. In order
to promote understanding of the models the authors run the simulation and
analyze the produced data, thus using the event generator interactively as a
concrete working object. Through working with it, they gain knowledge of the
consequences of models when processed through the code, of the interplay
between the different implemented models, and of the conditions under which
they can be applied with sufficient validity.
In experimenting with event generators physicists assume two distinct but
related perspectives. On the one hand, they study the interconnection between
the event generator and the implemented models, which allows them to bet-
ter understand the workings of the former. Understanding the event generator
here serves to turn it into a reliable instrument, which can be handed over to
other physicists for routine operation as a publicly available tool. On the other
hand, authors employ it as a testing ground for the underlying models, using it
explicitly to investigate distinct kinds of models: event generators make these
models workable. This is important when, for example, physicists explore the
 tarja knuu t til a and m artina mer z

physics potential of new collider features. The event generator is thus used as a
kind of processor that turns the new model features into numbers, the simulat-
ed events, which can then be scrutinized. This exemplifies what we introduced
above as the productive and interactive nature of models.

Understanding Analysis Strategies. Once a specific collider layout has been

judged appropriate for investigating novel types of physics events (for example,
a Higgs particle decaying into two gammas), the project enters into a phase
devoted to devising adequate analysis strategies. This requires the physicists
to identify smart ways of distinguishing “signal” from “background,” the lat-
ter denoting all the unwanted processes that dominate and cover the events
of interest (the signal). Strategies for “beating down” the background rely on
theoretical reasoning and are tested for their effectiveness by feeding them
into simulation studies. These studies exclusively rely on simulation and are
conducted long before any actual experiment is producing data. Data-analysis
experts conceive of the event generator as a concretization of underlying phys-
ics models, as a material medium producing concrete outcomes that can be
further processed. This interpretation implies that users can get a grip on the
underlying models by dealing with their realization and results via a computer
program. Consequently, in this instance the event generator is used as a tool
that serves an epistemic goal.
This case points to yet another interpretation of the model’s interactive and
productive capacities, as simulation thus also serves in the exploration and de-
velopment of procedures and practices. Understanding in this case concerns
epistemic strategies, and it comes about through the abilities that develop
when scientists interact with the simulation programs—as one event-genera-
tor author put it, “playing a lot with these tools gives you a very good feeling,
a very good judgment of physics itself ” (interview). Let us recall that in this
phase of an experiment, too, simulation is employed to produce knowledge for
the future: the data analysts work exclusively with simulated data in order to
understand how to deal with the (expected) real data at a later date.

Understanding the Detector. Yet another major undertaking in the prepa-

ratory phases of a collider experiment concerns the design and layout of the
particle detector that is to be built. Once more, simulation studies are crucial,
now relying on an event generator coupled with a detector simulation program.
What is at stake is the detector’s capacity to be sensitive to the physics events
of interest, like Higgs going to gamma gamma. In this enterprise the event gen-
erator is used as a mere tool, while the detector simulation program stands as
a surrogate for the (as yet unbuilt) detector. The physicists want to know how
to render the “real” detector sensitive to the events of interest, and this comes
under standing by modeling 

by way of interacting with its surrogate. In order to optimize the detector, they
vary the design features in the simulation and interactively try out which con-
figurations are the most sensitive. Simulation enables scientists to flexibly re-
consider a variety of design features with the possibility of isolating specific
aspects. Through its materialization as a concrete object the detector simula-
tion constitutes an enhanced environment that allows physicists to zoom in
on contested theoretical areas and test out a variety of potential solutions to
technical problems. In this application the simulation program thus becomes
an inference generator that assists physicists in negotiating competing physics
programs and technical design options.
Due to the brevity of our account we could only provide a glimpse into the
multifunctional nature of simulation work in particle physics, suggesting that
scientific models (in this case simulation programs) inhabit a field of tension
spanned by their epistemic productivity and their instrumental character. Thus,
both event generators and detector simulation programs are multiplex objects.
When they are coupled for the purpose of specific applications both can play
different roles as tools, scientific objects, and inference generators. This com-
plex interweaving makes it possible for simulation programs, as multiplex ob-
jects, to enhance understanding of entities as varied as conceptual objects (for
example, theoretical ideas), epistemic strategies (such as data analysis) and the
workings of technological artifacts (as in detectors).

EVghZgh^cAVc\jV\ZIZX]cdad\n
Caught between different disciplines, programming, and practical purpos-
es, language technology provides an interesting case for studying how techno-
logically oriented model-building that strives primarily for new applications
and efficient solutions can give us scientific understanding. In the following we
consider the case of the Constraint Grammar parser, which challenges many of
our established views on scientific knowledge and understanding. First, parsers
are both unrepresentational and atheoretical things in the sense that they are
not expected to be realistic depictions of human linguistic competence, nor are
they based on any autonomous formal linguistic theory. Thus, second, a large
part of the scientific understanding they provide is not theoretical but rather
inherent in the expertise that develops in the construction process. This un-
derstanding concerns language and cognition as well as the technologies used.
Third, like event generators, parsers produce many kinds of understanding, de-
pending on the purposes for which they are used.
A parser could be characterized as a computational model of a certain (as-
pect of a) language implemented as a computer program. To parse is to describe
a word or a sentence grammatically; in the context of language technology, it
refers to “the automatic assignment of morphological and syntactic structure
 tarja knuu t til a and m artina mer z

(but not semantic interpretation) to written input texts of any length and com-
plexity” (Karlsson et al. , ). There are two primary approaches to parsing,
one of which is grammar based, linguistic, and descriptive and the other proba-
bilistic, data-driven, and statistical. The methods of the data-driven approach
include corpus-based learning rules, Hidden Markov models, and machine
learning. Both approaches are grounded basically in the notion of parsing be-
ing based on the information given by the neighboring words.
The words we use in natural language are typically ambiguous: the same
word form, say “bear,” can be read differently depending on the context. The
parser works by making use of the morphological and syntactic information of
the words in the vicinity of the word in question. For instance, the word form
“bear” has both nominal and verbal readings; thus, if it is preceded by an article,
the parser assigns it a nominal reading. This is not how we humans proceed,
since unlike parsers we also have the semantic dimension of language available
and understand the intended meaning of the words and phrases.
The Constraint Grammar parser (CG) (Karlsson et al. ) is a product
of long-term research in language technology carried out at the University of
Helsinki.13 It is a grammar-based parser built on hand-crafted rules made with
the help of the Constraint Grammar, which is a language-independent parsing
formalism devised for describing the rules of any language. In contrast to tra-
ditional and formal grammars, which are licensing grammars, a CG grammar
is reductive. Instead of defining the rules for the formation of correct expres-
sions in a language, the constraint grammar of a particular language specifies
the constraints that discard as many improper alternatives as possible. The CG
parser is firmly grounded in linguistic corpora. A parser for a certain language
is created in a piecemeal fashion involving continuous rounds of rule writ-
ing, implementing the rules as a computer program, and testing the grammar
against manually disambiguated benchmark corpora. For example, the EngCG
(Voutilainen et al. ; Karlsson et al. ), which is still one of the best pars-
ers for English, took a group of linguists and a computer scientist several years
to construct.

Theory and Expertise in Parser Building. The Constraint Grammar parser

contradicts the philosophical expectations concerning models in many ways: it
cannot be understood as a clear-cut representation of human linguistic com-
petence, nor is its relationship to theory straightforward. On a more theoretical
level, parsing poses the problem of how to infer the grammatical structure of
words and phrases by automatic means. Since humans and machines cannot be
expected to accomplish this task in the same way, parsers are not conceived of
as realistic models of human linguistic competence. They are only required to
produce output that “approximates human performance sufficiently,” as one of
under standing by modeling 

the developers put it. Thus, they are primarily productive things, being rather
models for than models of.
In addition to treating models as representations, the philosophical tradi-
tion also sees them as interpretations of formal theoretical templates. In fact,
from this perspective representation and interpretation are interwoven, since
interpretation gives a theoretical template empirical content and makes it thus
represent some target object; representation, in turn, is based on the structure
of the model. What is more, if this structure is formally simple and preferably
shared by many other theories, then from the unificationist perspective at least
it imparts scientific understanding: we have succeeded in identifying an under-
lying structure that is common to many other phenomena. All this, however,
does little to account for the part that theory plays in language technology. Due
to its corpus-based nature, the CG grammar carries constraints that are quite
different from the grammar rules described by formal syntactic theories.14 In-
stead of giving one general rule, it sets out a few dozen down-to-earth con-
straints that state bits and pieces of the phenomenon by ruling out what cannot
be the case in some specific context.
Yet, it would be misleading to characterize the CG parser as a nontheoreti-
cal entity. There is a theory behind constraint grammar parsing, but rather than
being an abstract description of the general mechanisms and laws prevailing
in external reality, it is a set of instructions for building a specific artifact (see
Karlsson et al. ). These principles remain on a very general level, how-
ever, and do not in themselves really amount to any guide as to how to build
a CG parser. What this shows is that a large part of knowledge concerning
parser construction can be attributed to a special kind of expertise rather than
to overall explicit theoretical understanding. Expertise has typically been at-
tributed to local, experience-bound, and tacit dimensions of science in recent
discussions, yet it has remained rather difficult to tell what expertise consists of
and how it is gained.15 The case of parsing suggests that it is firmly tied to work-
ing with certain specific methods and artifacts, taking into account the tasks to
be accomplished. The researchers in the field highly appreciate methodological
know-how—the ability of the researcher to (fore)see what can be done with
different methods.

Parsers as Tools, Objects, and Inferential Devices. Being a multiplex epis-

temic artifact, the CG parser is a productive thing that can assume variously
the roles of a tool, a research object, or an inferential device in the scientific
endeavor. As tools, parsers are valuable for both their product, a linguistically
annotated text, and as a first stage in many natural language-processing (NLP)
systems. Models and computational methods such as parsers and parsing for-
malisms are also interesting and important scientific objects in their own right,
 tarja knuu t til a and m artina mer z

a fact that is often neglected in the philosophy of science with its strong focus
on the representational aspects of models. Interestingly, the CG parser is a dif-
ferent kind of object to linguists and to computer scientists: while linguists
see parsers primarily as linguistic descriptions and tools for linguistic analysis,
computer scientists abstract from the linguistic description and consider them
in terms of their computational characteristics.
In their capacity as inferential devices, parsers function largely like any oth-
er models, and as such they are able to impart understanding about language
and cognition despite their primarily instrumental nature. Their understand-
ing-providing features derive from the fact that they are artifacts with known
principles, mechanisms, and structures (although we need not be in command
of the structure, as the machine-learning approaches to parsing show). Yet the
parsing case indicates that the structures themselves do not necessarily provide
any understanding. If we are to understand how language actually functions,
the description has to be made operative and, as a consequence, productive.
In the case of the CG parser, this is done by implementing the language de-
scription as a computer program—a possibility that has made grammars more
interesting as scientific objects. Traditional evaluation of grammars has relied
on conventional academic discussion, but it is practically impossible for hu-
mans to consistently follow such complex rule systems as set out in extensive
grammars. This is possible for computers, which can check whether such rule
systems are consistent, for example. In terms of implementation, the linguist
also has the possibility to learn from the performance of the model by finding
out which rules cause trouble and by trying out different ones: this emphasizes
the importance of the workability of models in creating scientific understand-
ing in the sense of licensing different inferences.
Interestingly, the instrumental fitness of the parser as a tool is also impor-
tant from the epistemic perspective. Once it functions well, it provides a start-
ing point for diverse interpretations and questions. For instance, it is possible
to study what properties distinguish successful models from unsuccessful ones,
and what assumptions can be made about human language faculties. Last but
not least, the making of language-technological artifacts has given us a new
understanding of language that we could not have gained by any other means.
Here, “understanding” is a more appropriate word than “knowledge,” since over
and above any explicit knowledge it has enabled researchers to grasp how poly-
semous, context-bound, and ambiguous our language use indeed is.

8dcXajh^dch

We began this chapter with the observation that models provide an impor-
tant locus of understanding in science. This observation is backed by the fact
under standing by modeling 

that the various intelligibility criteria (for example, visualizability, familiarity,

tractability)—called “tools for understanding” by de Regt and Dieks ()—
are in fact typical objectives of model building. In an attempt to account for
model-based understanding, we suggested that it is best to take a productive
approach, according to which models are concrete, human-made objects that
can be considered multiplex epistemic artifacts. Thus, this approach, the philo-
sophical gist of which is to attribute scientific understanding to our interaction
with our objects of knowledge, accommodates many of the features that are
commonly attributed to understanding.
We illustrated our objectual approach to models and, consequently, to sci-
entific understanding, with two cases: event generators in particle physics and
parsers in language technology. What these two cases have in common is that,
even though the respective models and their places in scientific endeavor are
very different, their way of imparting scientific understanding becomes fully
apparent only when one approaches them from the productive and interac-
tive perspectives. For instance, the semantic approach does not make much
sense of them because neither event generators nor parsers start out with one
underlying theoretical framework (which seems to be typical of simulations in
general). Moreover, in both cases the implementation of the models is crucial
for their epistemic productivity. Thus, models can be seen as productive ob-
jects that bring us understanding in different ways, depending upon the uses to
which they are put.
According to the objectual approach, the understanding-providing features
of models exist due to the interplay between their materially embedded medi-
um-specific and technological features, and their multiple uses. This interplay
gives them their characteristic multiplexity: the very same thing may be a dif-
ferent kind of object for different audiences and thus afford understanding dif-
ferently. This simultaneously heterogeneous understanding provided by what
seems to be one single thing is a rather usual phenomenon in modeling, yet it
has not received much attention from philosophers so far. The multifunctional
“nature” of models also implies that the understanding that comes about in
practice does not necessarily mark any closure: models as multiplex artifacts
generate new understanding-requiring situations. Furthermore, as scientific
understanding results from scientists’ interactions with variously materialized
objects and instruments, it is hard to tell it apart from technological under-
standing (conceived of in the wide sense of also comprising different represen-
tational techniques).
The objectual approach to understanding also solves some philosophical
puzzles concerning understanding. It gives us a clue to the subjective and even
affective characteristics of modeling, thus not leaving these aspects of scientific
work at the mercy of the idiosyncrasies of individual minds. If a substantial part
 tarja knuu t til a and m artina mer z

of the understanding created is coupled with the specific uses and reactions of
the artifacts the scientists work with, then the special skills and experiences
of the scientists can be attributed to the features of the objects themselves.
In this respect, the material dimension of scientific objects is crucial: objects
are intersubjectively present to us only to the extent that they materialize as
recordings, experimental effects, or as purposefully constructed epistemic ob-
jects—in other words, models. Moreover, the objectual relations inherent in
scientific practice also explain the link between familiarity and understanding.
Working with certain kinds of objects makes them familiar, but this does not
mean that they have to be familiar in an everyday sense.
Finally, the affective and emotional component of scientific work—an
important issue we have not touched upon in this chapter—can also be ap-
proached from the objectual point of view. There exists a large body of litera-
ture in psychology on object relations, which are taken to be constitutive in the
creation of our individual selves from early childhood on. In conclusion, one
may wonder whether our scientific identities are also intricately tied to the ob-
jects we deal with in our scientific work, often over extended time spans.

CdiZh
We wish to thank the three editors of the present volume, Henk de Regt, Sabina
Leonelli, and Kai Eigner; and the members of the Philosophy of Science Group at the
University of Helsinki, Uskali Mäki, Petri Ylikoski, Jaakko Kuorikoski, Aki Lehtinen, Tomi
Kokkonen, and Jani Raerinne, for their insightful and fruitful comments on earlier drafts of
this chapter.
. This overall commitment of philosophers to the representational character of
models has gone as far as asserting that models are representations (see, for example, Frigg
; Hughes ; Teller ).
. The epistemic value of models refers here to the different ways in which they give
us knowledge. It has traditionally been ascribed to the context of justification as opposed
to the context of discovery. In our opinion, this distinction does not fit well with modeling
because part of the justification of models is typically built into them in that we tend to
trust separately in many of the ingredients they consist of (Boumans ). Moreover, the
question of representation has provided to several philosophers a way to evade the concept
of truth, which has been taken as the epistemic concept par excellence.
. We do not deny that, in many cases, scientific models do represent some target
objects, but would rather stress, in line with the pragmatists, that representation is always
also an accomplishment of representation users. On the level of actual scientific work, this
involves an extended process of building and working with models—the work often pro-
ceeding quite apart from the production of links to observable phenomena and data. The
respective practices of theoretical modelers and experimentalists often do remain rather
separate. Thus, the unilateral interest in representation tends to leave unnoticed a large
part of modeling work and its epistemic significance.
under standing by modeling 

. In this, scientific models function like boundary objects, which according to Star
and Griesemer () manage the cooperation and understanding between different
“social worlds.”
. We have intentionally left out of our discussion the so-called feeling of understand-
ing. We agree with Ylikoski (this volume) that the feeling of understanding and under-
standing as a kind of ability should be distinguished, but from the objectual perspective it
is possible to explain from which kinds of experiences the feeling of understanding arises.
For example, Myers (, ) describes how scientists in the field of protein crystal-
lography cultivate a “feeling for the possible forms and movements of the protein in vivo”
through extended periods of building and navigating through protein structures on screen.
. By “objects,” we refer to entities that are both able to resist (“object to”) our interpre-
tations and manipulations, and that are variously materialized, thus also having a corporeal
and concrete dimension—if only through representations and measurements—and experi-
mental or simulated effects and outcomes.
. This is essentially a Kantian point, though historicized and materialized.
. This characteristic is also shared by computer models and simulations (see below),
which continually acquire new features and changes in appearance, and makes them
resemble theoretical entities.
. See de Chadarevian and Hopwood () for studies on the epistemic and didactic
importance of three-dimensional scientific models made out of diverse materials such as
wood, plastic, and wax. Klein  is a beautiful study on the importance of the material
dimension of the different “paper tools” (such as diagrams, symbols, and formulas) that
scientists use. Leonelli () argues for the productivity of a multimodel approach that
considers the complementary use of material and theoretical models.
. This perspective is inspired by investigations of computer development (Woolgar
; Mackay et al. ).
. Our first case is based on the results of a long-term ethnographic study of particle
physics that one of us (M. Merz) conducted at CERN, the European Laboratory for Particle
Physics, in Geneva, Switzerland. The empirical material presented here draws on the more
detailed discussions in Merz (, , and ).
. Particle physicists label experiments, data, and so on “real” in contrast to their
“simulated” counterparts.
. This case is based on an empirical study by one of us (T. Knuuttila) conducted at
the Department of General Linguistics at the University of Helsinki and reported in several
articles in both Finnish and English (for example, Knuuttila ).
. Certain parsers are based on autonomous grammar theory, but their success in
parsing running text has been rather modest.
. See Collins and Evans () for an overview of this discussion.

GZ[ZgZcXZh
Bailer-Jones, D. M. . When scientific models represent. International Studies in the
Philosophy of Science :–.
 tarja knuu t til a and m artina mer z

Baird, D. . Thing knowledge. In The philosophy of scientific experimentation, edited by

H. Radder, –. Pittsburgh: University of Pittsburgh Press.
Black, M. . Models and metaphors. Ithaca: Cornell University Press.
Boumans, M. . Built-in justification. In Models as mediators: Perspectives on natural
and social sciences, edited by M. S. Morgan and M. Morrison, –. Cambridge:
Cambridge University Press.
Collins, H. M., and R. Evans. . The third wave of science studies: Studies of expertise
and experience. Social Studies of Science :–.
Darden, L. . Mechanisms and models. In Cambridge companion to the philosophy of
biology, edited by D. Hull and M. Ruse, –. Cambridge: Cambridge University
Press.
Daston, L. . Biographies of scientific objects. Chicago: University of Chicago Press.
de Chadarevian, S., and N. Hopwood, eds. . Models: The third dimension of science.
Stanford: Stanford University Press.
de Regt, H. W. . Discussion note: Making sense of understanding. Philosophy of Sci-
ence :–.
de Regt, H. W., and D. Dieks. . A contextual approach to scientific understanding.
Synthese :–.
French, S., and J. Ladyman. . Reinflating the semantic approach. International Studies
in the Philosophy of Science :–.
Frigg, R. . Re-presenting scientific representation. PhD diss., London School of
Economics.
Giere, R. N. . Explaining science: A cognitive approach. Chicago: University of Chicago
Press.
———. . How models are used to represent reality. Philosophy of Science :–.
Glennan, S. . Modeling mechanisms. Studies in History and Philosophy of Biological
and Biomedical Sciences :–.
Griesemer, J. . Three-dimensional models in philosophical perspective. In Models:
The third dimension of science, edited by S. de Chadarevian and N. Hopwood, –.
Stanford: Stanford University Press.
Grimm, S. R. . Is understanding a species of knowledge? British Journal for the Phi-
losophy of Science :–.
Hacking, I. . Representing and intervening: Introductory topics in the philosophy of
natural science. Cambridge: Cambridge University Press.
Hartmann, S. . Models and stories in Hadron physics. In Models as mediators:
Perspectives on natural and social science, edited by M. S. Morgan and M. Morrison,
–. Cambridge: Cambridge University Press.
Hughes, R. I. G. . Models and representation. Philosophy of Science :s–s.
Humphreys, P. W. . Extending ourselves: Computational science, empiricism, and
scientific method. New York: Oxford University Press.
Karlsson, F., A. Voutilainen, J. Heikkilä, and A. Anttila, eds. . Constraint Grammar: A
language-independent system for parsing unrestricted text. Berlin: Mouton de Gruyter.
Keller, E. F. . Models of and models for: Theory and practice in contemporary biology.
Philosophy of Science :s–s.
under standing by modeling 

Klein, U. . Experiments, models, paper tools: Cultures of organic chemistry in the nine-
teenth century. Stanford: Stanford University Press.
Knorr-Cetina, K. . Sociality with objects: Social relations in postsocial knowledge
societies. Theory, Culture and Society :–.
———. . Objectual practice. In The practice turn in contemporary theory, edited by
T. R. Schatzki, K. Knorr-Cetina, and E. von Savigny, –. London: Routledge.
———. . Epistemic cultures. In Science, technology, and society: An encyclopedia, ed-
ited by S. Restivo, –. Oxford: Oxford University Press.
Knuuttila, T. . Models, representation, and mediation. Philosophy of Science
:–.
———. . From representation to production: Parsers and parsing in language technol-
ogy. In Simulation: Pragmatic constructions of reality, edited by J. Lenhard, G. Küp-
pers, and T. Shinn, –. Dordrecht: Springer.
Knuuttila, T., and A. Voutilainen. . A parser as an epistemic artefact: A material view
on models. Philosophy of Science :–.
Kuorikoski, J. . Simulation and the sense of understanding. Paper presented at Models
and Simulations , Charlottesville, VA, March –, . Available at http://philsci_
archive.pitt.edu/archive/.
Lenhard, J. . Surprised by a nanowire: Simulation, control, and understanding. Phi-
losophy of Science :–.
Leonelli, S. . What is in a model? Using theoretical and material models to develop
intelligible theories. In Modeling biology: Structures, behavior, evolution, edited by
M. D. Laubichler and G. B. Muller, –. Cambridge, MA: Massachusetts Institute of
Technology Press.
Mackay, H., C. Carne, P. Beynon-Davies, and D. Tudhope. . Reconfiguring the user:
Using rapid application development. Social Studies of Science :–.
Mäki, U. . Models are experiments, experiments are models. Journal of Economic
Methodology :–.
Merz, M. . Multiplex and unfolding: Computer simulation in particle physics. Science
in Context :–.
———. . Kontrolle—Widerstand—Ermächtigung: Wie Simulationssoftware Physiker
konfiguriert. In Können Maschinen handeln? Soziologische Beiträge zum Verhältnis
von Mensch und Technik, edited by W. Rammert and I. Schulz-Schaeffer, –.
Frankfurt a.M.: Campus.
———. . Locating the dry lab on the lab map. In Simulation: Pragmatic constructions
of reality, edited by J. Lenhard, G. Küppers, and T. Shinn, –. Dordrecht: Springer.
Morgan, M. S. . Models, stories and the economic world. Journal of Economic Meth-
odology :–.
Morrison, M. . Models as representational structures. In Nancy Cartwright’s phi-
losophy of science, edited by S. Hartmann, L. Bovens, and C. Hoefer, –. London:
Routledge.
Morrison, M., and M. S. Morgan. . Models as mediating instruments. In Models as
mediators: Perspectives on natural and social science, edited by M. S. Morgan and M.
Morrison, –. Cambridge: Cambridge University Press.
 tarja knuu t til a and m artina mer z

Myers, N. . Animating mechanism: Animations and propagation of affect in the lively
arts of protein modelling. Science Studies :–.
Nagel, E. . The structure of science. New York: Harcourt, Brace, and World.
Rheinberger, H.-J. . Experiment, difference, and writing, part , Tracing protein syn-
thesis. Studies in History and Philosophy of Science :–.
———. . Toward a history of epistemic things: Synthesizing proteins in the test tube.
Stanford: Stanford University Press.
Star, S. L., and J. R. Griesemer. . Institutional ecology, “translations” and boundary
objects: Amateurs and professionals in Berkeley’s Museum of Vertebrate Zoology,
–. Social Studies of Science :–.
Suárez, M. . Scientific representation: Against similarity and isomorphism. Interna-
tional Studies in the Philosophy of Science :–.
———. . An inferential conception of scientific representation. Philosophy of Science
:–.
Teller, P. . Twilight of the perfect model model. Erkenntnis :–.
van Fraassen, B. C. . The scientific image. Oxford: Clarendon Press.
Voutilainen, A., J. Heikkilä, and A. Anttila. . Constraint grammar of English: A perfor-
mance-oriented introduction. Department of General Linguistics, publication no. .
Helsinki: University of Helsinki.
Woodward, J. . Making things happen: A theory of causal explanation. New York:
Oxford University Press.
Woolgar, S. . Configuring the user: The case of usability trials. In A sociology of mon-
sters: Essays on power, technology and domination, edited by J. Law, –. London:
Routledge.
Zhang, J. . The nature of external representations in problem solving. Cognitive Sci-
ence :–.
the gre at deluge

0
The Great Deluge
Simulation Modeling and Scientific Understanding

?D=6CC:HA:C=6G9

 Recently, a friend went to see her general practitioner in order to

give her thyroid gland a checkup. The doctor took a blood sample to test a few
parameters in the laboratory, and she performed an ultrasound scan on the
thyroid gland. The result was not completely clear, and the doctor referred my
friend to a specialized radiologist with the necessary instruments for a reliable
diagnosis. She said that while my friend could simply hand over the results of
the blood tests, she thought the radiologist would have little use for the ultra-
sound printout. This was not because it was irrelevant, but because the radiolo-
gist would certainly want to carry out his own test. Indeed, this is exactly what
happened.
This suggests a consensus among physicians that blood tests performed in
a laboratory provide objective and communicable information, whereas this
is not quite the case for an ultrasound printout. Anyone who has ever seen an
ultrasound printout knows how necessary it is to interpret the structures and
shadings it reveals. Nonetheless, it is not the case that the ultrasound scan is of
decreasing information value, because the radiologist did not drop it in favor of
some other procedure. Quite the opposite, he performed a second scan himself.
My friend could well have brought the radiologist an X-ray, for example, as a
source of information just like the results of the blood tests. However, what
makes the diagnostic content of an ultrasound printout so much less transport-
able? And why did the physician prefer to do his own scan?


 jo hanne s lenhar d

It is not the mere printout, the snapshot that my friend could have brought
with her, that delivers the information. The benefits of an ultrasound test are
gained through a complicated interaction: namely, the physician’s hand with its
changing position when guiding the probe and the accompanying changes on
the screen. It is only through a temporal sequence of controlled movement and
observation that the specialist can learn which imaged effects remain invariant,
which are perhaps only blurred shadings, and so forth. The full diagnostic po-
tential of the ultrasound scan can only be exploited through multidimensional
movement and through observing the relation between movement and image
over a certain period of time. The procedure draws on an essentially dynamic
aspect; hence, it is basically not a single representation, but a relational tableau
with which the physician gradually becomes familiarized through the tempo-
rally extended interaction of body movement, instrumental evaluation, and its
visualization—an orientation is gained through a sequence of images that the
physician can influence through his or her own movement. Dynamics, itera-
tion, and visualization are essential aspects of the instrumental access through
the ultrasound scan that could never be conveyed, even in principle, through
a single image.
This chapter considers which kind of understanding (a certain class of typi-
cal) computer simulations can provide, and how they do so. Discussing these is-
sues will reveal major parallels to the ultrasound scan scenario sketched above
because dynamics, feedback with visualization, and the orientation that can be
attained in a model characterize the understanding that researchers can obtain
in computer simulation as well.1
Models come in different types. Basically, we can distinguish between theo-
retical models (in mathematical language), simulation or computational mod-
els (for example, discretized versions of a theoretical model), and algorithmical
implementation (sometimes also called the computational model). Simulation
modeling involves going back and forth between all of them. The modeling
process, however, can be described with more levels of modeling (Winsberg
), with emphasis on experimental aspects (Dowling ; Winsberg ;
Lenhard ), or by analyzing whether theoretical models may be dropped
entirely in certain circumstances (Keller ). The details do not matter here
(see Lenhard et al.  for a sample of the modeling discussion specializing in
simulation).
Simulations weaken the link between theory and understanding. In this
book, Tarja Knuuttila and Martina Merz’s “Understanding by Models” points
out the affinities between studies of modeling and of understanding. However,
if the investigation of scientific understanding should be located in the context
of modeling, what distinguishes simulation models from models in general, if
such characteristics exist at all? Knuuttila and Merz claim that the “workable
the gre at deluge 

nature” of models determines their epistemic value, and part of this chapter
will spell out what this nature can be in simulation.
Simulation is leading to a pragmatic mode of scientific understanding. The
“subjective” concept of understanding is adapting to a new instrument—com-
puter simulation. Prediction and control on the one side and understanding
and insight on the other have formed the dual goal of science more or less since
its beginnings in modernity. Understanding is traditionally conceived as a mat-
ter of course in a theoretical sense. It is a deeply rooted point of view that those
who understand a phenomenon theoretically also have the appropriate means
to control it. Although one might well assign Platonic roots to this conception,
it had certainly become effective during the course of the seventeenth century
after Galileo and Newton had built the new science on the basis of mathema-
tization. Admittedly, work in the philosophy and history of technology reveals
that building and controlling devices is not entirely dependent on a foregoing
theoretical understanding, but this may well develop from engineering prac-
tice. I do not intend, however, to delve into that discussion. Instead, I want to
investigate an area that is a strong, arguably the strongest, case in favor of the
theory-based view of understanding and show that even in this area simula-
tions weaken the link between theory and understanding. Hence, this chapter
will focus on the mathematical instruments of science.
Although the computer in general and simulations in particular draw on
mathematics directly, simulation models usually reveal one particular feature
that threatens to shatter the traditional approach, namely, simulation models,
which are epistemically opaque (particularly in complex cases), so it is no lon-
ger possible to grasp in a real sense the ways in which the mathematical com-
ponents and algorithms interact. Instead, researchers have to test and study
the models in simulation experiments. Even when some parts of such models
have to be treated as a black box, simulation experiments can still render the
results of, for example, a whole field of parameters both visible and accessible
to experience. Researchers can establish a degree of orientation within a model,
a familiarity with how the model behaves.
This is the point at which the approach of the physician who is performing
an ultrasound scan meets up with that of a numerical optimizer who is using
simulation processes to explore the quality of a certain procedure, with the
methods used by a climate researcher who is varying parameters to try to find
out something about the stability of the Gulf Stream, or with the approach of
a researcher in nanotechnology who is looking for ways to control nanoscale
friction. In all these cases, an action-oriented understanding is constructed
that makes it possible to exercise control and make predictions, even when the
dynamics have not been grasped in theoretical terms. Hence, one can justifi-
ably talk about a pragmatic concept of understanding.
 jo hanne s lenhar d

Yet, at the same time, it is questionable whether it is right to use the term
“understanding” in this context. The traditional concept of understanding is
linked to theoretical insight. Even if simulations do not make any understand-
ing possible at all in this sense, one should not interpret this as a weakness of
an instrument that, so to speak, waters down the possibilities of mathematics.
One could just as well see it as a strength: the instrument allows one to do with-
out a complete understanding of the theory—control and technology become
possible in other ways as well.
My argument starts with a closer consideration of the call by de Regt and
Dieks () to place scientific understanding on the philosophical agenda. It
will analyze the work of Gunter Dueck at IBM on the so-called Great Deluge
Algorithm. This case study on how simulation is used to deal with complex
problems in a thoroughly technological context of application will reveal that
some characteristics of simulation-based understanding do not fit easily into
the framework of theory-based understanding. Some scientists argue that the
latter understanding should be replaced by a model-based one. Two options for
balancing the perceived relation between simulation and scientific understand-
ing exist: either stick to the traditional “insight” concept of understanding and
take simulation as a new instrument that can provide control and options for
intervention without understanding, or enlarge (weaken) the concept of un-
derstanding and admit a pragmatic simulation-based mode of understanding
that works partly without the intelligibility of theory, but fits to an engineering
context of application.

JcYZghiVcY^c\VcY6eea^XVi^dc

Since the seminal contributions of C. G. Hempel, understanding has had a

bad reputation in philosophy of science because of its individual and psycho-
logical connotations. The closely related concept of explanation has received a
great deal more attention, and various, though not unanimous, analyses stress
its objective features. Nonetheless, this does not mean that understanding is
irrelevant. Humphreys, for instance, points out that Hempel also maintains
that understanding, in particular theoretical understanding, was the goal of
explanatory activity (Humphreys , ). The merit of a recent article by de
Regt and Dieks () has been to put understanding effectively on the agenda
of philosophy of science.
De Regt and Dieks () summarize the current discussion in philosophy
of science and point out that even if scientific understanding is included, it
is usually treated as a mere supplement to the concept of explanation. They
contend that both leading parties in explanation, that is, the unification and
the causal-mechanism views (compare Kitcher and Salmon ), proceed in
the gre at deluge 

this way. De Regt and Dieks argue convincingly that scientific understanding
is neither a simple derivative of explanation nor restricted to the realm of in-
dividual psychology. I agree with their account, especially when they stress the
pragmatic and context-dependent nature of understanding. Moreover, they are
well aware of the difficulties confronting their project of a general criterion of
understanding: “The history of science shows great variation in what is and
what is not deemed understandable. Even at one particular moment in history
opinions about what is understandable often diverge” (, ).
Hence, every philosophical standpoint on scientific understanding has to
find the appropriate level of analysis that does not dissipate into individualistic
criteria of single researchers nor get trapped by positions that are too general
and unable to take account of the observed variability in understanding. They
formulate a Criterion for Understanding Phenomena (CUP) that makes under-
standing dependent on the existence of an intelligible theory:

CUP: A phenomenon P can be understood if a theory T of P exists that

is intelligible (and meets the usual logical, methodological and empirical
requirements). ()

Consequently, they introduce a Criterion for the Intelligibility of Theories (CIT):

CIT: A scientific theory T is intelligible for scientists (in context C) if

they can recognize qualitatively characteristic consequences of T with-
out performing exact calculations. ()

If I interpret de Regt and Dieks correctly, they are making a twofold assertion
here: First, understanding is the ability to recognize qualitatively characteristic
consequences. Second, this will be possible via intelligible theories. In my view,
their account is preoccupied with theory and underrates the role of models and
modeling.
Computer simulation is a new instrument of mathematical modeling that
can provide researchers with the ability to recognize and to interact with quali-
tatively characteristic consequences of modeling assumptions even though the
model dynamic remains partly opaque. I acknowledge the ability account of de
Regt and Dieks but do not see theory as a necessary ingredient. Furthermore,
simply replacing “theory” with “model” misses one crucial point: the ability to
draw consequences does not depend entirely on the intelligibility of the dynam-
ics (neither of model nor theory). The contribution of P. Ylikoski (this volume)
also analyzes understanding fundamentally as an ability. In my view, an attempt
to define understanding will have to take a closer look at how the mentioned
ability is brought about. This amounts to a kind of implicit definition—defining
 jo hanne s lenhar d

understanding by the ability it generates. This ability, in turn, depends on the

available means, for instance, simulation, which seems to suggest that changes
in instrumentation or modeling techniques can lead to changes in what under-
standing is.
Theoretical physics may impose a bias on the debate about simulation;
therefore, let us discuss a typical case of simulation that starts in mathematical
theory but attains its features in a distinctly application-oriented and techno-
logical context of modeling.

H^bjaVi^dcVhi]Z<gZVi9Zaj\Z

Since the invention of the electronic computer in the middle of the twen-
tieth century, calculating speed has increased enormously. Work that, at one
time, required the most powerful computers in major research institutes is
nothing out of the ordinary for machines that can be purchased throughout
the world. However, mere calculating speed does not necessarily correspond to
a capacity to solve problems. Alan Turing, who worked out a machine on pa-
per that corresponds essentially to the modern computer, already offered fun-
damental insights into computability and its limitations. The vast majority of
real numbers, for example, cannot be computed. The question regarding which
types of problem cannot be solved with the calculating speed of computers (of
the Turing machine type) led to the explication of Kolmogorov-Chaitin com-
plexity (see Chaitin ). One also uses the term “computational complexity”
to describe the sort of complexity that measures how well a computer can cope
with processing a problem.
What makes this complexity so interesting is that a great number of prob-
lems emerge in scientific and technical practice that are complex in exactly this
sense. I should like to take the well-known “traveling salesman problem” (TSP)
as a prototype. A number of cities are given with the appropriate distances be-
tween them, and the problem is to find a route that links together all the cities
while simultaneously also being as short as possible. For each such problem
with n cities, there exists a solution—so much is obvious. Several routes may
well provide the shortest traveling distance. However, how can one find such an
optimal route? This question formulates a typical optimization problem that is
in no way simple to solve even with the enormous computing power available
today. Why is this so? Eventually, it would not be all that difficult to work out
an algorithm that would try out all the different routes, and (at least) one would
have to be the shortest. This approach is correct in principle, but not feasible
in practice.
The problem is that even a moderately large n will result in so many po-
tential routes that working through them all would overwhelm even the most
the gre at deluge 

powerful computer. The computational effort as a function of n expands more

rapidly than any polynomial in n, revealing that “brute force” cannot be the key
to any practical solution of the problem (technically speaking, the TSP belongs
to the class of “np-hard” problems).
Such and similar optimization problems are widespread in technical prac-
tice. A number of scientific disciplines are involved in the practical implemen-
tation of solutions—the computer as the instrument has to be, so to speak,
fed from all sides. This includes not only numerical and combinatorial opti-
mization and scientific computing but also informational and technological
approaches in programming and implementation. On the mathematical level,
there are a host of different approaches that, in part, include very advanced
ideas (see Gutin and Punnen ). In the following, I shall present and discuss
the so-called Great Deluge Algorithm developed by Gunter Dueck that simu-
lates a flood that compels a virtual test person to optimize results (Duek ).
This approach does almost completely without theoretical considerations. This
is what makes this example seem so suitable for illustrating characteristics of
simulation modeling.
The problem is to solve the following task: One has a mathematically de-
fined space such as the space containing the possible routes of the traveling
salesman in the TSP. A point x in this space is assigned a specific value F(x)—in
the TSP, the length of Route x. A neighboring point x΄, which emerges through
switching the sequence of two cities, has another length F(x΄), and so forth.
If one conceives the generally high-dimensional spaces as being projected
onto a level, and assigns every point x the functional value F(x), one obtains
a sort of mountain range, and the task is then to find the highest peak (or the
deepest valley, which, in mathematical terms, means only switching the sign).
How can one find this peak, even though one knows nothing about the shape
of the mountain range? One can assign the height to every specific point, al-
though one does not have an overview of all points. Bearing in mind the brief
discussion on computational complexity, one can say that in many practical
tasks there are simply far too many points for even a supercomputer to check
through. A suggestion would be to start somewhere with a random choice, and
then simply go “uphill,” that is, to check only the neighboring points n(x) of x,
and then, in the next step, to maximize F(n(x)). This algorithm would certainly
lead to a peak, though without in any way guaranteeing that it would be the
highest or even one of the high ones. This would only be the case if, by chance,
the starting point were already on the slope of this high peak. Hence, although
this is a quick procedure, in general, it does not deliver a good result.
The Great Deluge Algorithm now proposes a surprisingly simple way of
modifying this approach. Go on the following simulated hike through the
mountain range of solutions: The hiker starts somewhere at random as before,
 jo hanne s lenhar d

but does not go uphill continuously; every new step sets off in a randomly cho-
sen direction. The hiker moves according to a random process that takes no
account of the form of the mountains. Parallel to this, there is a second process
that gave the algorithm its name, namely, a great deluge that is causing the wa-
ter level to rise continuously. Only one restriction is imposed on the random
hiker: not to get wet feet. Whenever the random generator proposes a step
that would lead to a point below the water level, another, new step has to be
simulated. Evidently, this simulated hike ends at a local maximum as soon as all
neighboring points lie underwater. Because the hike can also lead downhill, the
starting point does not ascertain the final outcome in advance.
An obvious objection is that the simulated hiker is bound to become ma-
rooned on one of the islands that will inevitably become isolated at some point
by the deluge. In other words, is there any argument to explain why the local
maximum attained should be equal to or almost equal to the global maximum?
Nonetheless, practice—for example, finding good placements for components
of computer chips—shows that the algorithm delivers surprisingly good re-
sults, something that its inventor also notes:

Frankly speaking, we did not seriously expect this algorithm to func-

tion properly. What if we were to start in England or Helgoland? When
the great deluge rises, even the continents will turn into islands. The
Great Deluge Algorithm cannot escape from land that is surrounded
by water and may therefore be cut off from every good solution.
But, strangely enough, we have applied the Great Deluge Algorithm
to optimize production, for great chip placements, even to the most
difficult industrial mixing problems and have calculated thousands of
TSPs. Again, only a teeny program was required and the outcome was
excellent in each instance. (Dueck et al. ; my translation)

How can one be confident that the outcome actually is “excellent”? Being
able to improve a solution that already works in industry would indicate that
the algorithm has done a good job. The most comfortable judgment situation is
when different approaches can be compared. Consider the example of locating
holes with a driller in a circuit board. The driller head has to be placed succes-
sively at all locations of holes; hence, it proceeds like a traveling salesman who
starts and finishes in a fixed town (the null position of the driller head). Finding
a good solution for this TSP is of obvious economic benefit, because it would
raise the number of items produced per time unit. Figure . shows the circuit
board and locations of all the holes to drill (towns to visit). Figure . shows the
starting point of the simulated route, that is, a random path connecting all the
holes (towns).
the gre at deluge 

;^\jgZ.#&#8^gXj^iWdVgYl^i]]daZh# ;^\jgZ.#'#6gVcYdbeVi]XdccZXi^c\
Cjaaedh^i^dcd[Yg^aaZg^ci]ZadlZgaZ[i" Vaa]daZh#9jZX`'%%)#GZegdYjXZYWn
]VcYXdgcZg#9jZX`'%%)#GZegdYjXZYWn eZgb^hh^dc[gdbHeg^c\Zg
eZgb^hh^dc[gdbHeg^c\Zg

Figure . shows the algorithm at an intermediate stage. Figure . displays the
end result of the Great Deluge Algorithm. It took about , steps, or a few
seconds, and produced a path with a length of about . inches. The optimal
result is known in this special case due to the theoretical efforts of Holland
(). It is . inches.

;^\jgZ.#(#>ciZgbZY^VgnhiViZ!XjggZci ;^\jgZ.#)#:cYgZhjaid[dcZgjcd[
aZc\i]d[eVi]^cY^XViZYVWdkZ#9jZX` i]ZVa\dg^i]b#9jZX`'%%)#GZegdYjXZY
'%%)#GZegdYjXZYWneZgb^hh^dc[gdb WneZgb^hh^dc[gdbHeg^c\Zg#
Heg^c\Zg#
 jo hanne s lenhar d

JcYZghiVcY^c\BdYZahVcYH^bjaVi^dch

At first glance, the whole approach seems to be too simple to be successful.

The Great Deluge approach neither offers great theoretical progress in the field
of optimization nor claims to do so. Its success is located in a technological
context in which effectiveness can be superior to theoretical justification. In
the philosophical context of this chapter, one could ask whether it is possible
to identify reasons that might explain this success. These reasons, if based on
theoretical considerations, could serve as a basis for explanation and under-
standing. Indeed, Dueck describes the “attempt to understand the success”
(, ) in the following way:

According to our experimental experiences, the building of islands in

solution space seems to start only when the algorithm has reached the
neighborhood of very good solutions. Additionally we have observed
that in every problem we have treated there exist a great number of
very good solutions that all look very different. Contrary to our first
intuition, it is not the case that better and better solutions become
more and more similar to the exact best solution. (, )

The researchers exploit their growing experience with the performance of

the simulation to create an image of how the situation is structured. The focus
here is not on what knowledge they gain about the high-dimensional space but
rather on how they go about this. The quote describes how the scientists make
some kind of “experimental experiences” and “observe” how the simulation be-
haves in an unknown space. In the discussion on the nature of simulations, one
main focus in philosophy of science is on the conception of theoretical model-
experiments, or computational experiments (for example, Humphreys ;
Keller ; Winsberg ). These experiences may offer surprises and are
apt to change or create “intuitions” about the structure of the space or problem
under consideration.
The following features seem to be quite typical for how simulations are used:
t
0CTFSWBUJPO
PG
TVSQSJTJOH
CFIBWJPS
JO
UIF
TJNVMBUJPO
NPEFM
t
4JNVMBUJPO
FYQFSJNFOUT
UP
FYQMPSF
PS
FYUSBQPMBUF
NPEFM
CFIBW-
ior, thereby providing more evidence, improving intuition, or making
predictions.
t
(BJOJOH
BO
PSJFOUBUJPO
XJUIJO
FQJTUFNJDBMMZ
PQBRVF
NPEFMT
CZ
QSP-
cessing simulations; that is, scientists become acquainted with the model
and how it functions.

This point is crucial: the simulation does not have to start with a clear idea about
the gre at deluge 

the structure of the solution space; on the contrary, this space is explored via
a simulated process. Hence, the simulation acts like a dynamic tentacle in our
example: a restricted random hike through unknown space. In this way, simula-
tions are “extending ourselves”—in line with the title of Paul Humphreys’s book
on simulation ()—in a very pragmatic way.
The particular case under consideration should not lead us astray regard-
ing the relevance of theory. Quite often, simulations have an elaborated theo-
retical basis. Consider nanoscience, which is devoted to the investigation of
phenomena brought about by many, but not too many, molecules. In principle,
the Schrödinger equation (quantum theory) governs their interactions; in prac-
tice, however, it is impossible to derive behavior theoretically for reasons of
complexity. Problems with complexity are endemic in the simulation context:
Knuuttila and Merz (this volume) discuss examples of parsing (language tech-
nology) and high-energy physics that would fit well here.
Lenhard () analyzes some phenomena of lubrication on the nanoscale
as observed in molecular dynamics simulations by Landman (). Landman’s
simulations make scientifically accessible—and render visible—phenomena of
increased friction in nano-sized dimensions that nanotechnology will have to
deal with. Feynman () already pointed to the plausibility of encountering
new phenomena of friction at the nanoscale. His theoretical ideas on volume
and surface, however, could not determine what these phenomena would look
like in concrete terms. Landman’s simulations do not just visualize the behavior
of molecules in thin layers; he has also experimented with small oscillations to
restore the desired and usual properties of lubricants. The simulation at least
indicates that this might be a promising approach in technology.
Comparable to the optimization case, the governing equations tell us noth-
ing about a complex situation. It is that kind of orientation gained in an ex-
perimental mode and assisted by visualization that indicated how increased
friction could perhaps be overcome. To avoid any misunderstanding: no strong
claim is made here about the representational accuracy of simulations. I agree
that validity is one of the main problems in the epistemology of simulation.
However, I hold that simulations can deliver an adequate dynamics, although
the model structure does not represent the mechanisms at work in the actual
phenomenon. Whether validity of performance is dependent on representa-
tional accuracy is a controversial issue: Winsberg (), for instance, argues
for “reliability without truth,” whereas Morgan () favors an ontological ar-
gument for the primacy of experiments over modeling approaches in general,
just because of their more direct representational force. The case of simulation
is particularly interesting, because artificial components, that is, deliberate de-
viations of the simulation model from the theoretical model, can contribute
positively to the model performance (compare Lenhard ; Winsberg ).
 jo hanne s lenhar d

Prediction and forecast are often furnished with a considerable degree of

uncertainty. Meteorology may serve here as a case in point. Attempts to im-
prove predictions are usually based on refined simulation modeling approach-
es. Let us consider briefly the case of hurricane forecast. The hurricane Opal,
which ravaged several countries around the Gulf of Mexico in , was simu-
lated at the US National Center for Supercomputing Applications. The aim of
the research project (Romine and Wilhelmson ) was to fit characteristics
of the simulated hurricane, like track, speed, or rain, to those that had been
measured and recorded extensively in Opal. The reasoning was that a fit be-
tween historical data and simulation would lead to an improved forecast: what
is good at reproducing the known past may be good for prediction too.
This simulation example is clearly theory based because the core model of
the project is a general circulation model of the atmosphere that works with the
so-called fundamental (physical) equations of atmospheric circulation. At the
same time, however, modeling a hurricane, that is, turning a general model into
a specifically adapted one, involves many additional parameterization schemes.
To fit the track, for instance, the researchers experimented with different com-
peting schemes and tried to achieve a best fit—based on extensive experiments
and visualizations—so that they could compare their performances. For one
of the track schemes, a range of parameters was tuned to make the simulated
track more similar to the observed one. Interestingly, a theory-based physi-
cal scheme was dropped in favor of a black-box one in which the parameters
had no physical interpretation. The point here is that visualization of explor-
ative simulation experiments, optimization of model performance by tuning
parameters, repeated visualization, and so forth were crucial to achieve a fit.
This means once again, just as in the cases discussed above, that researchers
recognized and interacted with qualitative consequences of the simulation
model—documenting the kind of ability that has been described as an indica-
tion of understanding.
This account is not in line with how Feynman circumscribes the common
concept of understanding: “ ‘I understand what an equations means if I have a
way of figuring out the characteristics of its solution without actually solving
it.’ So if we have a way of knowing what should happen in given circumstances
without actually solving the equations, then we ‘understand’ the equations, as
applied to these circumstances” (Feynman et al. , -). Simulation gives
this position an interesting new twist, because repeated calculations provide
the basis for gaining experiences with how the models function; that is, for
learning what happens according to the equations. In contrast to what Feyn-
man probably had in mind, it is not theoretical intelligibility but orientation
based on numerous calculations that provides “a way of knowing what should
happen in given circumstances”—as he put it in the quotation above.
the gre at deluge 

The view expressed by Feynman and many others was: “Understand the
equations, because when calculations begin, you will be lost in complexity and
complication.” The stance extracted from the simulation cases can be summa-
rized as: “Simulate and observe the outcome; repeat that procedure to explore
the models in order to eventually gain some understanding.” Hence, simulation
presents an instrument to deal with overwhelming calculations, even making
them invisible while, at the same time, the model dynamics tend to become
black-boxed—a diagnosis that has been termed “epistemic opacity” (Hum-
phreys ; compare also Lenhard ).
The long-standing epistemological promise of theory and theoretical
models is that they can be known and understood perfectly because they are
completely constructed. It has been argued that this promise is unfounded in
complex modeling situations that, however, are unavoidable in contexts of ap-
plication. The argumentation has brought up two aspects: first, the question
about the appropriate framework regarding how theory and models are related
to understanding in general; and, second, the problem of the relationship be-
tween understanding and simulation in particular. Admittedly, theory often is
the starting point, or even a main source, in the development of simulations.
Nevertheless, the crucial concepts are models and modeling, predominantly
because predictions are not derived from theory, as the previous discussion of
simulation has shown.
Hence, the general lesson is to frame the debate on scientific understand-
ing with a discussion of models as initiated by Nancy Cartwright (), with a
clear tendency against usability of theory, and exemplified by the more moder-
ate stance of Morgan and Morrison () and others. There should be con-
ceptual room for a distinctly model-based mode of scientific understanding.
This mode fits well with the aim of de Regt and Dieks () of developing a
context-dependent and pragmatic concept of scientific understanding and also
with related accounts in this volume (Knuuttila and Merz, Ylikoski).
What does the more specific lesson about simulation and understanding
consist of? Recall why theory was claimed to be essential. De Regt and Dieks
argued by opening up an alternative: a mere list of facts or predictions on the
one side and, on the other, a general theory from which one can derive pre-
dictions. Set aside the argument that the latter may be questionable. The list
would appear like an oracle that produces a correct prediction, but for reasons
one cannot understand: “In contrast to an oracle, a scientific theory should be
intelligible: we want to be able to grasp how the predictions are generated, and
to develop a feeling for the consequences the theory has in concrete situations”
(de Regt and Dieks , ).
Consequently, scientific understanding is introduced as the element that
grants usability. The investigation of simulations in this chapter has revealed,
 jo hanne s lenhar d

however, that a Heisenbergian or Feynmanian intellectual insight into the

mechanisms of how simulation models bring about their results may be lack-
ing. Instead, observation of repeated and stepwise adapted experiments and
their visualizations are used to dynamically explore how a model functions.
This approach sums up to a “black-box mode” of acquiring orientation in the
model. The talk of black (or closed) boxes has been introduced to simulation
modeling by Norbert Wiener, who transferred the term used in engineering
(see Boumans  for a gray-scale of models).
Even though simulations partly function as a black box, the results and pre-
dictions are not arbitrary. On the contrary, simulations can provide a very de-
tailed picture of model behavior that in no way resembles the pure contingency
of a list of facts—there is much connection in it. This connection is created
by the instrument (simulation), that is, by the way the results are constructed
(feedback, stepwise tuning of parameters, and so on). It is not a matter of sud-
den insight, but rather one that has to grow piecewise through iterated (quasi-
empirical) experiences. This process shows some resemblance to what Chang
() calls “epistemic iteration.”
As we have seen, simulation is good at this: developing a feeling for the
consequences, an acquaintance with how the model functions—thus providing
the ability that is thought is to be an indicator of understanding. I would like
to maintain that the instrument does not just enlarge possibilities, that is, does
not simply make intractable mathematics tractable. Simulation enforces a con-
ceptual adaptation: understanding loosens its close ties to the intelligibility of
theory. Researchers may become acquainted with a model and how it functions
by iterated simulation experiments and visualizations.

6EgV\bVi^X6XXdjcid[JcYZghiVcY^c\

If we adopt a thoroughly contextual and pragmatic concept of understand-

ing, does it not become even more elusive than before? Do simulations help
us to understand complex phenomena? Or, on the contrary, are simulations
associated with a “loss of understanding,” as Humphreys (, ) argues,
because they do not allow for theoretical understanding in the common Feyn-
manian sense? Hence, do simulations enlarge control at the cost of understand-
ing? In my view, these questions have no definite answer, because an answer
depends on the concept of understanding one adopts, and I shall discern two
options. Consider the following standpoint:

(*) Scientific understanding is characterized by providing a feel-

ing for the modeled situation and for the consequences of the
modeling assumptions.
the gre at deluge 

I have argued above that simulation fulfills this criterion. Hence, it achieves
what is commonly assumed to be achievable by understanding. In short, ac-
cording to (*), simulations provide understanding. However, a certain ambigu-
ity remains: the Great Deluge example shows how simulation can provide a
feeling for the consequences, even though Dueck and his colleagues recognized
a clear deficit in their understanding. Two options can resolve this ambiguity
by discerning a strong and a weak conception of understanding:
(i) Stick to the traditional “insight” concept of understanding; that is,
take (*) as a necessary but not sufficient criterion, and take simulation as
a new instrument that makes it possible to satisfy the criterion without
involving scientific understanding.
(ii) Stick to (*) as being sufficient, and enlarge (or weaken) the concept
of understanding. That means, adopt (*) as an implicit definition of un-
derstanding, implying that if simulation models are suitable instruments
to satisfy the criterion, they provide a mode of understanding that works
partly without the intelligibility of theory, thereby making epistemic opac-
ity and understanding partly compatible.

To vote for option (i) would mean to drop the concept of scientific understand-
ing from the agenda of investigating simulation and, instead, to analyze how
simulation permits intervention and manipulation. This would perhaps find
favor with many philosophers of science who are suspicious of understanding
because of its subjective connotations.
A vote for option (ii) would imply a more intricate standpoint. Think of a
science-studies-inspired endeavor to analyze whether and how the formation
of a simulation community takes place in science. This formation would po-
tentially include the development of a shared instrument-based way of under-
standing the world and point in the direction of option (ii). Simulation would
then come close to a paradigm in the sense of Rouse’s () interpretation of
Kuhn: paradigms form a shared way of seeing the world. The simulation-based
way would then be linked to understanding the world in the sense of (*). The
point is that for principled reasons of complexity, one cannot expect more than
this type of insight for many fields of science and technology in the context of
application. Thus, the lucidity of theory-based explanations is partly replaced
by the quasi-empirical access of simulations, constituting a genuinely novel
mode of scientific understanding.
This mode is a new one in the context of mathematical modeling in the sci-
ences. However, it is a well-known one in other contexts. Thus, expressed more
precisely, simulations expand a certain mode of understanding to mathemat-
ics-based modeling. What does this new yet old mode consist of? It is a thor-
oughly pragmatic one—it employs experimental exploration, visualization, and
 jo hanne s lenhar d

feedback, thus relying on activities of simulation modeling. It provides orienta-

tion by sounding out the territory (may it be a virtual one); humans themselves
learn in a very similar way. Actually, this mode is found in biological systems
and may well explain a tendency in some current sciences to be inspired by
such systems. The discussion on situated and embodied cognition in robotics,
such as in Brooks , can serve as one example. This mode is employed, for
instance, in elaborated visualization techniques (see Myers ). In a sense,
simulation theories of understanding as discussed in the philosophy of mind
touch upon this mode, although they are not dealing with simulation models in
the mathematical sense (see, for example, Stone and Davies  on the “men-
tal simulation” debate).
I hold that simulation is going to establish a new practice that, in turn, calls
for conceptual adaptations or transformations, and it is still unclear whether
option (i) or (ii) will become the leading opinion. This is by no means a ques-
tion of psychology; rather, it is a matter of the still developing epistemology of
simulation.

CdiZh
I would like to thank the audience of the conference Philosophical Perspectives on Sci-
entific Understanding held in Amsterdam in , and the editors of this volume for highly
valuable comments and suggestions.
. My argument is located in the growing debate about models in science. I view
simulations as a kind of model, and I will discuss simulations and simulation models
interchangeably in this chapter. My view on models has much in common with Mary
Morgan and Margaret Morrison’s account of “models as autonomous mediators” (). In
particular, Morgan and Morrison strengthen the (partially) independent role of models in
relation to theory.

GZ[ZgZcXZh
Boumans, M. . The difference between answering a “why” question and answering a
“how much” question. In Simulation: Pragmatic constructions of reality, edited by J.
Lenhard, G. Küppers, and T. Shinn, –. Dordrecht: Springer.
Brooks, R. . Intelligence without reason. MIT Artificial Intelligence Lab memorandum
.
Cartwright, N. . How the laws of physics lie. Oxford: Clarendon Press.
Chaitin, G. J. . The limits of mathematics: A course on information theory and the
limits of formal reasoning. Berlin: Springer.
Chang, H. . Inventing temperature: Measurement and scientific progress. New York:
Oxford University Press.
the gre at deluge 

de Regt, H. W., and D. Dieks. . A contextual approach to scientific understanding.

Synthese :–.
Dowling, D. . Experimenting on theories. Science in Context :–.
Dueck, G. . New optimization heuristics: The great deluge algorithm and the record-
to-record travel. Journal of Computational Physics :–.
———. . Das Sintflutprinzip. Berlin: Springer.
Dueck, G., T. Scheuer, and H.-M. Wallmeier. . Toleranzschwelle und Sintflut: Neue
Ideen zur Optimierung. Spektrum der Wissenschaft :–.
Feynman, R. P. . There is plenty of room at the bottom. Engineering and Science
:–.
Feynman, R. P., R. B. Leighton, and M. Sands. . The Feynman lectures on physics. Vol.
. Reading: Addison-Wesley.
Gutin, G., and A. P. Punnen, eds. . The Traveling Salesman Problem and its variations.
Dordrecht: Kluwer.
Holland, O. A. . Schnittebenenverfahren für Travelling-Salesman-und verwandte Prob-
leme. Ph.D. diss., University of Bonn.
Humphreys, P. W. . Computer simulations. In PSA , vol. , edited by A. Fine, M.
Forbes, and L. Wessels, –. East Lansing, MI: Philosophy of Science Association.
———. . Analytic versus synthetic understanding. In Science, explanation, and ratio-
nality: Aspects of the philosophy of Carl C. Hempel, edited by J. H. Fetzer, –. New
York: Oxford University Press.
———. . Extending ourselves: Computational science, empiricism, and scientific
method. New York: Oxford University Press.
Keller, E. F. . Models, simulation, and “computer experiments.” In The philosophy of
scientific experimentation, edited by H. Radder, –. Pittsburgh: University of
Pittsburgh Press.
Kitcher, P., and W. C. Salmon, eds. . Scientific explanation. Minneapolis: University of
Minnesota Press.
Landman, U. . Lubricating nanoscale machines: Unusual behavior of highly confined
fluids challenges conventional expectations. Georgia Tech Research News. Available at
http://gtresearchnews.gatech.edu/newsrelease/landman/landman_news.htm.
Lenhard, J. . Surprised by a nanowire: Simulation, control, and understanding. Phi-
losophy of Science :–.
———. . Simulation modeling: The cooperation between experimenting and modeling.
Philosophy of Science :–.
Lenhard, J., G. Küppers, and T. Shinn, eds. . Simulation: Pragmatic constructions of
reality. Dordrecht: Springer.
Morgan, M. S. . Experiments versus models: New phenomena, inference, and sur-
prise. Journal of Economic Methodology :–.
Morgan, M. S., and M. Morrison, eds. . Models as mediators: Perspectives on natural
and social science. Cambridge: Cambridge University Press.
Myers, N. . Molecular embodiments and the body-work of modeling in protein crys-
tallography. Social Studies of Science :–.
 jo hanne s lenhar d

Romine, G., and R. Wilhelmson. . Parameterization impacts on a simulation of Hur-

ricane Opal (). th PSU/NCAR Mesoscale Model User’s Workshop. Preprint.
Boulder, CO. –.
Rouse, J. . Kuhn’s philosophy of scientific practice. In Thomas Kuhn, edited by T. Nick-
les, –. Cambridge: Cambridge University Press.
Stone, T., and M. Davies. . The mental simulation debate: A progress report. In
Theories of theories of mind, edited by P. Carruthers and P. Smith, –. Cambridge:
Cambridge University Press.
Winsberg, E. . Sanctioning models: The epistemology of simulation. Science in Con-
text :–.
———. . Simulated experiments: Methodology for a virtual world. Philosophy of Sci-
ence :–.
———. . Models of success versus the success of models: Reliability without truth.
Synthese :–.
chap ter title ver so 

G8IK@@@
JcYZghiVcY^c\^cHX^Zci^ÒXEgVXi^XZh
 pa rt title ver so
 ('
Understanding in Biology: The Impure
Nature of Biological Knowledge

H67>C6A:DC:AA>

 This chapter offers an analysis of understanding in biology based on

characteristic biological practices: ways in which biologists think and act when
carrying out their research. De Regt and Dieks have forcefully claimed that a
philosophical study of scientific understanding should “encompass the histori-
cal variation of specific intelligibility standards employed in scientific practice”
(, ). In line with this suggestion, I discuss the conditions under which
contemporary biologists come to understand natural phenomena and I point
to a number of ways in which the performance of specific research practices
informs and shapes the quality of such understanding.
My arguments are structured in three parts. In the first section, I consider
the ways in which biologists think and act in order to produce biological knowl-
edge. I review the epistemic role played by theories and models, and I em-
phasize the importance of embodied knowledge ( “know-how”) as a necessary
complement to theoretical knowledge (“knowing that”) of phenomena. I then
argue that it is neither possible nor useful to distinguish between basic and
applied knowledge within contemporary biology. Technological expertise and
the ability to manipulate entities (or models thereof ) are not only indispensable
to the production of knowledge, but are as important a component of biologi-
cal knowledge as are theories and explanations. Contemporary biology can be
characterized as an “impure” mix of tacit and articulated knowledge.


 sabina leonelli

Having determined what I take to count as knowledge in biology, in the

next section I analyze how researchers use such knowledge to achieve an un-
derstanding of biological phenomena. My interest lies in the activities through
which researchers achieve understanding: that is, in the role of human agency
in creating, applying, and interpreting biological knowledge to comprehend the
world. I present a view of scientific understanding as a cognitive achievement of
individual biologists that results from their ability to interact with (models of )
phenomena and from their training and participation in specific research com-
munities. To obtain understanding of a given phenomenon, biologists need to
exercise specific epistemic skills and research commitments, which allow them
to materially intervene on and reason about the world. Understanding can only
be qualified as “scientific” when obtained through the skillful and consistent use
of tools, instruments, methods, theories, and/or models: these are the means
through which researchers can effectively understand a phenomenon as well as
communicate their understanding to others.
In the final section, I conclude that there are different ways in which bi-
ologists can understand a given phenomenon, depending on which skills and
commitments they are able to muster in their research. The quality of biologi-
cal understanding obtained by an individual depends on his or her capacity
to perform some of the several activities that, together, constitute biological
knowledge.

I]Z>bejgZCVijgZd[7^dad\^XVa@cdlaZY\Z

Biology is growing increasingly disunified. Biological research is fragment-

ed into countless epistemic cultures, each with its own terminologies, research
interests, practices, experimental instruments, measurement tools, styles
of reasoning, journals, and venues.1 These cultures are typically constructed
around a specific set of issues, the investigation of which requires training in a
series of techniques, software applications, and instruments that are regarded
as appropriate to this aim.2 Within this context, it is not surprising to find that
different biological fields provide not only different theories about the same
sets of phenomena, but also different types of theoretical results, ranging from
the mostly descriptive knowledge gathered by experimental or field biologists
to the largely speculative claims pursued by theoretical biologists through
mathematical reasoning.3 The knowledge produced by the epistemic cultures
engaging in biological research takes various forms, ranging from a mechanis-
tic, narrative, functional, or causal type to mathematical equations, descrip-
tions, representations, and categorizations of phenomena. Researchers are not
divided as to which type of approach is most informative or correct. More often
under standing in biol o gy 

than not, different types of theories can each have relative significance toward
understanding a given phenomenon (Beatty , s). As highlighted by
Longino () and Mitchell (), this pluralism in theoretical approaches
is more conducive to the development of biological knowledge than a uniform
landscape centered on a few unifying laws would be.
Notably, theoretical results are often expressed and used by researchers
in forms other than the propositional statements contained in an explana-
tion. Apart from symbolic representations (such as mathematical equations),
we find diagrams, photographs, maps, and various three-dimensional objects
playing important roles not only in the context of discovery, but also in the
expression and justification of biological knowledge. Biologists refer to these
representational tools as models of biological phenomena.4 The philosophical
position that most closely accounts for biologists’ eclectic views on models was
put forward in  by Morgan and Morrison, according to which a model is
defined not by its physical features, but by its epistemic function as a “media-
tor” between theories and phenomena. Models represent phenomena (through
processes including idealization, abstraction, and analogy) in ways that allow
scientists to investigate them. Their representational value depends on their
material features as well as on the interests and beliefs characterizing the re-
search context in which they are used (Morgan and Morrison , ). Model-
based reasoning, broadly defined as the use of representational tools toward
gaining epistemic access to natural processes, is widely recognized as playing a
crucial role in the production of scientific knowledge.

:meaV^c^c\Wn>ciZgkZc^c\
How precisely do biologists reason through models? The answer to this
question comes from studies of research practice in various biological disci-
plines that emphasize that model-based reasoning involves the manipulation
of models; that is, the reasoning has the possibility to modify them in order to
fit different research contexts. Models are tailored to pursuing specific research
purposes, which can vary depending on changes in the features of the phenom-
ena to be investigated as well as changes in the material, social, and institutional
settings of the researchers involved.5 Good models are ones that can be used by
the researchers to fulfill their goals in ways deemed appropriate by the relevant
research communities. If a researcher is unable to use a model because he or
she lacks the relevant skills, it will not matter that the model could in principle
be useful to the investigation: it cannot be used, and therefore cannot prove
useful toward the research. Consider the case of research on model organisms.
The worm C. elegans, the fruit fly Drosophila melanogaster, the plant Arabi-
dopsis thaliana, and the mouse Mus musculus are protagonists of experimental
 sabina leonelli

research in various fields, from molecular biology to immunology, physiology,

and ecology. Biologists are aware of the limitations imposed by the use of very
few species as representatives of the immense variety of existing organisms:
there is a high degree of arbitrariness in assuming that facts discovered through
experiments on mice (such as susceptibility to chemicals) can hold for species
as different as monkeys or humans. Yet, the preference for using a small set of
model organisms as models of the whole of organic life remains largely unques-
tioned. This is because there are substantial advantages to focusing research ef-
forts on the same organism.6 Not only do biologists come to accumulate a great
deal of data about its structure and functioning: the very quality of those data is
enhanced by the possibility of devising measuring instruments and procedures
that are finely tuned to study precisely that organism. The model organisms
thus selected become models of other life forms. Limiting the number of mod-
els makes it easier to focus on how to use them, and indeed, biologists keep
improving their ability to intervene on the selected model organisms in the
laboratory. This strategy is so productive as to trump biologists’ worries about
the extent to which these organisms actually embody biodiversity.
“Using” a model can involve straightforward manipulation, as in the case of
the handling of actual specimens or cell cultures in the lab, and/or reasoning, as
in the case of researchers thinking about a specific description, formula, or im-
age as applying to an aspect of the organism. What I wish to emphasize is that,
in all cases, biologists reach an explanation for the phenomena that interest
them through an intervention on the models used to study the phenomenon in
question. This holds whether the phenomenon in question is a concrete object,
such as an organism, or an intangible process, such as the dynamics of evolving
populations: explanation in biology is always obtained through direct interven-
tion on models of the phenomenon to be explained. By “direct intervention” I
mean the modification of a model to investigate the biological phenomenon
that that model is taken to represent. This can happen through reasoning, as
when mentally shifting the terms of an equation used to represent evolutionary
dynamics, or through physical movements, as when shifting the order of the
balls standing for nucleotides in a plastic representation of DNA’s double helix.
Whether expressed in thoughts or gestures, human agency constitutes an ines-
capable condition for producing explanations of the natural world.7
This position nicely accounts for one of the few characteristics shared by
the vast majority of contemporary life sciences. This is the ever-increasing
blurring of boundaries between “applied” and “pure” or “basic” research. With-
in theoretical physics, the term “basic research” has been used to identify ways
of producing scientific knowledge that aim primarily at explaining phenomena.
Physicists involved in campaigns for the construction of supercolliders, for in-
stance, insist that satisfying human curiosity about how the smallest compo-
under standing in biol o gy 

nents of matter look and behave is worth huge investments, no matter what
the instrumental value of that knowledge turns out to be. Research of this kind
has been contrasted with research explicitly aimed at intervening on natural
processes and shaping them at will. Cryptography was born out of the military
need to send messages without fear of being intercepted. As such, cryptogra-
phy can be portrayed as a paradigmatically applied science: it does not produce
explanations of natural phenomena, but rather provides the means for efficient
and secure communication.
When considering the practices and goals characterizing medicine, molec-
ular biology, ecology, or evolutionary biology, one finds no obvious distinctions
between research aimed at producing ways of intervening in the world and
research aimed at producing explanations of biological phenomena. These two
types of outcome are part of the same investigation processes and are deeply
intertwined with each other. Take biotechnology, which has traditionally been
referred to as an “applied” field because of its explicit focus on enhancing tech-
nical expertise and devising methods to apply theoretical knowledge to real
processes. A well-known instance of biotechnological research is the develop-
ment of genetically modified organisms (GMOs), which requires scientists to
“apply” knowledge coming from molecular biology to producing plants and an-
imals with characteristics that are socially and/or economically advantageous.
The classification of GMO research as applied does little justice, however, to
the contributions to theoretical knowledge by biologists and bioengineers en-
gaging in it. Similarly, the vast majority of supposedly “basic” molecular biolo-
gists explicitly aims to produce techniques through which organisms can be
genetically modified. This is because the possibility to intervene on an organ-
ism generates opportunities to explain some of its features (as in knock-out ex-
periments, where researchers take away specific chunks of the DNA sequence
of an organism so as to gain insight on which genes play a role in the develop-
ment of which phenotypic traits). At the same time, the ability to explain a
given phenomenon is tightly connected with the ability to intervene on it: the
production of GMOs is informed by available explanations of how the genome
is structured.
Let me emphasize again that I am not contending that biologists aim at
intervening and manipulating nature rather than at explaining it. I am stressing
that scientific efforts at explaining the biological world are necessarily inter-
twined with efforts to intervene in the world.8 At least in the case of biology,
explaining and intervening are two sides of the same coin, both in terms of the
types of knowledge employed in research (that is, theoretical and applied) and
in terms of the goals that research is expected to fulfill (such as truthful expla-
nations of natural processes and control over those same processes).
 sabina leonelli

I]Z>bedgiVcXZd[:bWdY^ZY@cdlaZY\Z
Already in the s, the intuition that theories and explanations consti-
tute but a part of scientific knowledge was put forward by philosophers Gilbert
Ryle and Michael Polanyi.9 Both men presented systematic reflections on what
I shall refer to as embodied knowledge—“know-how” in Ryle’s words () and
“personal knowledge” in Polanyi’s (/). They suggested that scientific
knowledge never involves merely a theoretical content, that is, that which is
known (as expressed in propositional or symbolic terms through theories and
explanations). It also involves the ability to apply such theoretical content to
reality and to interpret it in order to intervene effectively in the world as self-
aware human agents.
Polanyi sees such ability as an “unspecifiable art,” which can only be learned
through practice or by imitation, but which is rarely self-conscious, as it needs
to be integrated with the performer’s thoughts and actions without swaying
attention from his or her main goals (/, –). This is why Polanyi
refers to this knowledge as “tacit,” that is, as informing scientists’ performance
(actions as well as reasoning) in ways that cannot be articulated in words. Po-
lanyi distinguishes between two ways in which scientists can be self-aware: fo-
cal awareness, which is tied to the intentional pursuit of “knowing that”; and
subsidiary awareness, which concerns the scientists’ perception of the move-
ments, interactions with objects, and other actions that are involved in their
pursuit. In this view, focal and subsidiary awareness are mutually exclusive; one
cannot have one without temporarily renouncing the other. Switching from
“knowing that” to “knowing how” and back is a matter of switching between
these two ways of being self-aware. In fact, his characterization of know-how as
subsidiary awareness relegates this type of knowledge to a secondary position
with respect to “knowing that” (/, ). The acquisition of theoretical
content constitutes, for Polanyi, the central focus of scientists’ attention. Tacit
knowledge makes it possible for scientists to acquire such theoretical content,
but it is not valuable in itself: it is a means to a (theoretical) end.
While drawing inspiration from Polanyi’s emphasis on tacit expertise, I do
not think that his hierarchical ordering of types of knowledge (“knowing how”
being subsidiary to “knowing that”) does justice to the intertwining of “knowing
that” and “knowing how” that characterizes biological research. Polanyi bases
his reflections on the study of physical chemistry—and indeed, it might be that
some chemists value theoretical knowledge more highly than the practices and
skills used to construct it. Within most of biology, however, the development of
skills, procedures, and tools appropriate to researching specific issues is valued
as highly as—or, sometimes, even more highly than—the achievement of theo-
ries or data. The evolution of tools and procedures (as in the case of modeling)
under standing in biol o gy 

is not only crucial to furthering theoretical developments: it constitutes an im-

portant research goal in its own right.10
On this point, Ryle proves more accommodating. His analysis of the dis-
tinction between “knowing that” and “knowing how” does not assume either
type of knowledge to be predominant over the other. This is due to an under-
lying difference between his outlook and Polanyi’s. While the latter thinks of
the distinction as descriptive, thus presenting the two types of knowledge as
actually separate in practice, Ryle presents the distinction as a purely analytic
tool with no descriptive value. Indeed, he refers to the idea of an actual distinc-
tion between “knowing how” and “knowing that” as a false “intellectualist leg-
end”: “when we describe a performance as intelligent, this does not entail the
double operation of considering and executing” (, ). Also in contrast to
Polanyi, Ryle defines “knowing how” as involving both intentional actions and
unconsciously acquired habits. In fact, he prefers an intelligent agent to avoid
as much as possible the enacting of habits:

to be intelligent is not merely to satisfy criteria, but to apply them; to

regulate one’s actions and not merely to be well regulated. A person’s
performance is described as careful or skilful, if in his operations he
is ready to detect and correct lapses, to repeat and improve upon suc-
cess, to profit from the examples of others and so forth. He applies
criteria in performing critically, that is, in trying to get things right.
(Ryle , )

Ryle does not specifically apply his account to the case of scientific knowledge.
Nevertheless, his characterization of agents “trying to get things right” by it-
eratively improving their performance fits my characterization of biologists’
practices and interests. As I illustrated, biologists do learn to use a variety of
tools in order to pursue their intellectual interests. They need to keep improv-
ing their ability to handle models and instruments, so as to reach their goals.
In Morgan’s words,

learning from models happens at two places: in building them and in

using them. Learning from building involves finding out what will fit
together and will work to represent certain aspects of the theory or
the world or both. Modelling requires making certain choices, and it is
in making these that the learning process lies. (Morgan and Morrison
, )

The function of models is not exhausted by their usefulness as thinking tools:

they are tools for acting as well as for thinking.
 sabina leonelli

Thus, the knowledge acquired by a biologist successfully pursuing a research

project encompasses both some results (a theory, a set of data, or a taxonomi-
cal system) and the procedures needed to develop, interpret, and reproduce
those results. A biologist is not born with the ability to perform the model-
ing activities required by his research. The skills and expertise enabling him to
perform these activities are important products of his work, in the same way
as theoretical results are. In fact, as I shall argue below, understanding those
theoretical results becomes very difficult without reference to the know-how
acquired via relevant research experience.11
On the basis of these considerations, I characterize biological research as
involving two types of knowledge, which are closely intertwined and virtually
inseparable in practice, but which it is useful to distinguish analytically. The
first type includes what is typically characterized as the articulated content of
knowledge, that is, what we regard as facts, theories, explanations, and con-
cepts concerning phenomena that are available independently of specific pro-
cedures or ways of acting. Access to cell theory and relevant data about cellular
organization, for instance, can be obtained without any expertise in cell biology
or experimental research. I shall refer to this type of knowledge as theoreti-
cal knowledge. The second type is what I call embodied knowledge. This is the
awareness of how to act and reason as required to pursue scientific research.
This awareness is essential for biologists to intervene in the world, improve
their control over phenomena, and handle the representations made of those
phenomena. Embodied knowledge is expressed through the procedures and
protocols allowing scientists to intervene on the entities and processes of inter-
est; the ability to implement those procedures and modify them according to
the research context; the acquired skill of handling instruments and models;
the perception, often based on the scientists’ experience in interacting within a
specific space, of how to move and position oneself with respect to the models
and/or phenomena under study (Myers ); and the development of meth-
ods allowing replication of experimental results.12
I thus posit that biological knowledge has a dual nature, encompassing a
theoretical and an embodied component. This follows from the recognition
that, in biology, explaining goes hand in hand with intervening. Effective agen-
cy improves not only biologists’ ability to control their environment, but also
their capacity to reason, abstract, and theorize. Further, this position accounts
for the impossibility to distinguish between basic and applied research in con-
temporary biology. Against the portrayal of biology as a “pure” science devoted
to acquiring theoretical knowledge of living organisms, I propose to view biol-
ogy as an “impure” mix of tacit and abstract, local and general, basic and ap-
plied knowledge, whose aim is at once to explain and control nature.
under standing in biol o gy 

JcYZghiVcY^c\^c7^dad\n

My view on biological knowledge is very different from the one offered by

Carl Hempel, who tried to formalize scientific knowledge by emphasizing the
axiomatic formalization of its theoretical results (as Ryle and Polanyi would put
it, its “knowing that” dimension). Hempel (, ) classified understanding
as a by-product of scientific explanations, thus implying that possessing expla-
nations automatically leads to understanding the phenomena to which those
explanations apply. In other words, Hempel claimed that access to theoretical
knowledge is a sufficient condition for the understanding of that knowledge
and of its applicability to real processes: all one needs to acquire understanding
is to consult scientific explanations of the phenomena of interest.
Hempel is right to point out that access to theoretical knowledge does not,
per se, imply recourse to embodied knowledge. Scientific theories express the
analytic concepts and categories employed in the study of phenomena through
specific formulations, which can be propositional (as in a textbook explanation),
symbolic (as in mathematical equations), or even pictorial (as in the depiction
of a biological mechanism). These formulations can be independent of the use
of models to develop and study them, as well as of knowledge about how to
apply the theories. In this sense, theoretical knowledge can thus indeed be pos-
sessed without recourse to embodied knowledge. Using theoretical knowledge
to understand natural phenomena is, however, a different matter. Understand-
ing phenomena by reference to theories and explanations requires some level
of know-how about how those theories and explanations apply to reality—such
as the assumptions and models that have been employed, the goals character-
izing the context in which theories were produced, and so on. In the process of
acquiring understanding, one needs to be able to intertwine one’s material ex-
perience of the phenomenon (observations and experimental interactions with
it) with a theory-informed interpretation of that same phenomenon.

6BViiZgd[H`^aa[ja6\ZcXn
Understanding natural phenomena by means of theoretical knowledge
does require some embodied knowledge; and, vice versa, embodied knowl-
edge needs to be coupled with theory-informed interpretations to provide un-
derstanding. To capture this insight, I define understanding as the cognitive
achievement realizable by scientists through their ability to coordinate theoreti-
cal and embodied knowledge that apply to a specific phenomenon.
By the verb coordinate, I mean the strategies that a scientist can learn to
use in order to () select beliefs, thought processes, and experiences that are
relevant to the phenomenon in question, and () integrate these components
 sabina leonelli

with the goal of applying them to the phenomenon.13 Modeling is not the only
strategy used for gathering scientific understanding (reflection, observation, or
taxonomic exercises are good examples of other such strategies). Still, I empha-
size modeling here insofar as it is a prominent activity in contemporary biol-
ogy as well as a good exemplification of the means by which a coordination of
knowledge that is relevant to a specific phenomenon can be achieved.
Take again the case of model organism research. Bioinformaticians have
been collecting data concerning the most prominent model organisms (includ-
ing Arabidopsis, C. elegans, and Drosophila) into large databases, which can
be accessed by biologists across the globe. The purpose of these databases is to
facilitate the retrieval of available knowledge about these organisms, so as to
allow biologists to integrate the insights that they gain from their own research
with the understanding acquired by others. Notably, these databases make
abundant use of digital models (pictorial representations of varying degree of
abstraction, depicting entities as disparate as cells, metabolism, and gene ex-
pression) and even simulations (of processes such as flowering).
To enhance biologists’ understanding of the information that is provided,
the curators of these databases do not consider it sufficient to provide a text
describing what is known about a particular biological process together with
the relevant experimental data. Curators found that biologists can better assess
the significance of those data and descriptions when they are directly related
to the components of the organism that they are taken to provide evidence
for—for instance, by means of images.14 These images, which some researchers
call “virtual organisms,” constitute models with which biologists can work. By
tinkering with the images, modifying parameters, and comparing how differ-
ent data sets fit the models, researchers coordinate their theoretical knowledge
(for example, of the chemical composition of mRNA, the mechanisms through
which mRNA travels across the cell, and the available data sets documenting
this phenomenon) with their embodied knowledge (such as what mRNA actu-
ally looks like under the microscope and how those data sets are produced),
thus obtaining some understanding of processes such as gene expression.
Even more remarkably, the majority of experimental biologists that use da-
tabases insist on being able to interact not only with virtual representations of
organisms, but also with actual specimens. This is because physical interac-
tion with samples of the phenomena in question provides different information
from interaction with digital models: particularly, the embodied knowledge ac-
quired through the manipulation of a plant or a fruit fly is not the same as the
embodied knowledge derived through interacting with images of their meta-
bolic cycle. Laboratory specimens offer countless possibilities for exploration
and surprising results.15 Most importantly, they offer the possibility to match
the researcher’s sensory experience of a phenomenon (and of the system in
under standing in biol o gy 

which the phenomenon manifests itself ) with his or her hypotheses and back-
ground knowledge. This is another example of the process that I wish to cap-
ture through the notion of “coordination.”16
It will be clear by now that, in my view, understanding a claim or explanation
is a largely subjective matter. Understanding is not an attribute of knowledge
itself, which can be measured quantitatively (as in “how much do you under-
stand?”). Rather, it is a cognitive achievement that is acquired by individuals in
a variety of ways and that can therefore take different forms depending on the
instruments that are used to obtain it. In the case of biological understanding,
the quality of the understanding displayed by a researcher will depend on his
or her acquisition of appropriate background knowledge as well as expertise in
handling the instruments, models, and theories that make it possible to pro-
duce and apply any scientific explanation. Importantly, both types of expertise
can only be obtained through participation and training in one or more scien-
tific communities.
It is indeed important to keep in mind that scientific understanding is not
the result of mere individual introspection. Its features are shaped by the ne-
cessity of intersubjective communication. The understanding acquired by one
researcher will not be accepted as a contribution to science unless he or she
is able to communicate insight to his or her peers, so as to make it vulnerable
to public scrutiny and evaluation. Individual understanding becomes scientific
only when it is shared with others, thus contributing to the growth of scientific
knowledge and partaking in the rules, values, and goals characterizing scien-
tific research.
I do not wish to argue that individuals can disseminate their understand-
ing of a phenomenon in an unmediated, direct way—for instance, by talking to
other individuals. This position would be paradoxical vis-à-vis my definition of
understanding as the cognitive achievement of individuals, based at least in part
on knowledge that cannot be articulated. I propose that individuals possessing
a specific understanding of a phenomenon enhance other individuals’ chances
of acquiring it in an indirect manner: that is, by constructing tools (including
models, explanations, experimental set-ups, and materials) that might enable
other individuals to experience the same kind of understanding. This indirect
sharing of understanding characterizes both the acquisition and the dissemina-
tion of understanding in scientific communities. On the one hand, biologists
seeking an understanding of a phenomenon are required to learn as much as
possible from similar efforts by other scientists, so as to use the understanding
accumulated by others as a vantage point for starting their own research. On
the other hand, biologists who acquire new understanding of a phenomenon
are required to contribute to the body of theoretical and embodied knowledge
available on that phenomenon in ways that will help other scientists acquiring
 sabina leonelli

the same insight. The social processes through which biological understand-
ing is acquired and disseminated have a strong impact on the features of such
understanding.17 A biologist’s training and professional life constitute experi-
ences that are highly efficient in enabling him or her to understand a series of
phenomena in a number of specific ways.

:e^hiZb^XH`^aahVcYGZhZVgX]8dbb^ibZcih
Prima facie, epistemic skills may be broadly defined as the abilities to carry
out a number of activities in order to increase one’s understanding of reality.
They may be partly innate, such as the skill of drawing an object (which depends
to some degree on the talent of the individual attempting such action), yet they
are most often acquired through the imitation of others and/or through expe-
rience, for instance by trial and error. Skillful actions are instrumental in the
sense of being necessarily targeted toward the achievement of a goal, which in
the case of epistemic skills is an improved understanding of (some aspects of )
reality. At the same time, the notion of skill concerns also the means and man-
ner by which action is undertaken. An action can be judged as skillful even if
it does not result in the accomplishment of its intended goal. This is because
the successful achievement of an aim involves more factors than just the inten-
tions and ability of the individual acting to that aim: adverse conditions in the
environment, bad timing, social context, interference with other individuals’
goals and actions—all these factors can influence the outcome of an action, no
matter how skillfully that action is carried out.
The assessment of an action as skillful depends as much on the manner in
which the action is undertaken as it depends on its effectiveness in achieving
the intended goal. Possessing an epistemic skill implies more than the ability to
perform an action: it requires performing that action well. This means that we
need criteria to determine what it means for an action to be well performed.
However, these regulative criteria are context-dependent and arguably impos-
sible to classify in an analytic fashion. They are dictated by factors as disparate
as the nature of the goal to be achieved; the interests of the person perform-
ing the action; the social as well as the material context in which the action is
carried out; and the tools available to carry out the action. For instance, what
counts as adequate actions in designing an experiment on rats depends on the
phenomenon to be explored (for example, crowding behavior, which requires
a space where rats are kept together, or intelligence, which involves setting up
tasks for individual rats to perform); the eventual hypothesis to be tested (if
researchers want to prove that “high population density induces pathologies,”
they will test rat behavior in increasingly smaller communal cages over a long
period of time); current regulations for ethical treatment of laboratory animals
(and consequently, whether rats are given sufficient food and space to survive);
under standing in biol o gy 

and, last but not least, whether animal ecologists are likely to consider the re-
sults of the experiment as representative for the behavior of rats in the wild.
This extreme context-dependence makes it uninteresting to list regulatory
criteria without looking at specific practices. Skillful actions in biological re-
search have only two features in common: () they involve the exploitation of
tools in a way deemed appropriate to effectively pursue a proposed goal; and
() judgment on what constitutes an “appropriate” course of action depends on
standards upheld within the relevant social contexts. The tools to be exploited
range across models, theories, experimental instruments, features of the en-
vironment, as well as samples of phenomena themselves, while the relevant
social context is constituted by the community of scientific peers in charge of
examining the methods and results of any research program. These insights
are captured by the following definition of epistemic skill: the ability to act in
a way that is recognized by the relevant epistemic community as well suited to
understanding a given phenomenon.
I wish to distinguish three types of epistemic skills. Theoretical skills involve
mastering the use of concepts, theories, and abstract models (that is, being able
to manipulate various expressions of theoretical knowledge) toward the un-
derstanding of a phenomenon. These skills enable biologists to reason through
given categories and classification systems according to specific inferential
rules, while at the same time judging the validity of those categories with refer-
ence to alternative theoretical frameworks. A second type of skill encompasses
the performative skills enabling biologists to exploit material resources toward
the acquisition of biological understanding of a phenomenon. These skills can
only be acquired through direct interaction with the environment (including
laboratory equipment and specimens). For instance, a biologist can be told how
to cultivate a plant so that it develops as required by a specific experimental set-
up. The corresponding performative skills, however, are acquired only through
practice, that is, by trying over and over again to act in the desired manner, thus
gradually adapting movements and sense perception to the tools and materi-
als used in an experiment, as well as to the standards enforced by the research
community of interest. After having experienced the results of sowing seeds in
the wrong type of soil, with too little or too much humidity and under varying
lighting conditions, the biologist will hopefully have acquired the ability to as-
sess the health of a growing plant and thus to produce specimens that fit his or
her experimental needs.
The ability to conform to existing standards can also be seen as an example
of the third type of epistemic skill, that is, of the social skills denoting the ability
of researchers to behave and express their insights in ways that are recognized
by their peers and/or other participants in their social context. Social skills
such as patience, charisma, and self-assurance enable scientists to secure an
 sabina leonelli

environment that is suitable to creating tools (models, theories, and instru-

ments) that might help others to acquire and exercise the same theoretical and
performative skills as the ones exercised by the creators of those tools. Once
other researchers are able to interact with phenomena under the same condi-
tions and with the same skills used to acquire a specific type of understanding,
they will be more likely to undergo a similar cognitive experience.
Considerable time and effort is invested in learning a set of skills, certainly
in the case of most theoretical or performative skills used in contemporary bi-
ology. This means that the set of skills available to any one researcher is limited.
Biologists bother to learn and perfect skills that, they believe, will help them to
pursue their research interests. Trained researchers, who already master sev-
eral skills, have a strong tendency to develop projects where those skills can
be exercised, rather than embarking on ones where () they will have to in-
vest time and effort in acquiring new skills, and () their old skills might prove
useless.18 Through acquiring specific skills, researchers become committed to
the actions, concepts, and research directions enabled by the exercise of these
skills. Whether researchers explicitly acknowledge it or not, certain principles,
concepts, ways of acting and handling objects become entrenched in the prac-
tice of researchers engaged in the study of a specific phenomenon. They repre-
sent knowledge that is assumed, rather than hypothesized, to be relevant to the
study of the phenomenon in question. When this happens, those bits of theo-
retical and embodied knowledge start to be regarded as a necessary platform
to carry out research.19 They become what I shall call research commitments,
encompassing items as diverse as the theoretical perspective held by biologists;
their research goals and interests; their ways to perform research and interpret
protocols; and their assumptions about the representativeness of their research
materials and the applicability of their results.
Research commitments and epistemic skills are different aspects of the
same process. Acting skillfully implies committing to the activities and results
that those skills bring forth, while making a commitment to a specific tech-
nique or concept requires learning skills adequate to follow up on such com-
mitment. Skills denote the abilities to perform an action in a specific manner;
commitments consist of a tendency toward (or preference for) pursuing goals
that can be obtained through performing that action. The possession of a skill
entails a commitment to exploiting that skill (thus taking advantage of an oth-
erwise useless investment). Similarly, once a commitment is made, it implies
learning and exercising skills that are relevant to fulfilling that commitment.
This is what makes commitments different from a mere promise or a pledge:
they imply, and result from, skillful action.
To highlight the link between skills and commitments, I trace a categoriza-
tion of commitments that runs parallel to my categorization of skills. Theoreti-
under standing in biol o gy 

cal commitments are commitments to using or investigating specific concepts,

theories, and principles. Much has been written about this type of commit-
ments in terms of conceptual biases, theoretical perspectives, and background
knowledge. Performative commitments consist of commitments to specific
habits, that is, ways of thinking or moving that become entrenched to a scien-
tist’s practice. Social commitments apply to the interests and values endorsed
by scientists as a consequence of their financial, professional, and personal de-
pendence on specific social hierarchies, sponsors, or practices. For example,
peer review mechanisms engender a series of commitments toward wedding
one’s work to recognized keywords, the expectations of prospective peers, and
the content of journals in which one hopes to publish; funding agencies might
require scientists to observe specific criteria and/or values in their research; or,
scientists working in a context dominated by powerful professors might have
to adopt their views to be granted funding for research.
Together, skills and commitments represent two fundamental tools for the
achievement of understanding. Commitments guide researchers toward ac-
quiring skills relevant to the pursuit of these commitments. As a result of exer-
cising these skills, scientists become aware of how their experiences, as well as
the concepts and tools that they use, fit or challenge available explanations of
phenomena. Adherence to commitments and skillful research practices consti-
tute basic conditions under which researchers coordinate theoretical and em-
bodied knowledge toward understanding phenomena.

BVcnIneZhd[JcYZghiVcY^c\

The prominence of scientists’ skills and experiences in achieving coordina-

tion among thoughts and actions, conceptual analysis, and sense-perceptions,
effectively separates the understanding thus obtained from the mere possession
of theoretical knowledge in the form of a theory or an explanation. Theoretical
knowledge, per se, does not provide biologists with the theoretical and perfor-
mative skills needed to trace its significance for natural phenomena. In biology,
to “know” implies not only to possess theoretical knowledge, but also to be able
to use it toward understanding actual phenomena. The ability to use theoretical
knowledge to that aim is granted by embodied knowledge, and particularly by
the components of embodied knowledge that I called epistemic skills.
Researchers possess different combinations of skills and commitments de-
pending on the epistemic culture to which they belong, their goals, training,
and professional experience. Of course, all researchers, no matter what their
fields and occupations are, possess a minimal amount of theoretical knowledge
as well as theoretical commitments to specific concepts and theories guiding
their research; and they possess embodied knowledge, in the form of both the-
 sabina leonelli

oretical and performative skills, as well as performative commitments. Yet, the

balance between theoretical, performative, and social skills, as well as commit-
ments, can vary. Depending on such variation, there can be different ways to
understand the same biological phenomenon.
The first, or theoretical understanding, denotes a situation where under-
standing of biological phenomena is acquired through recourse to theoretical
commitments and skills, with performative skills and commitments playing a
subsidiary role. A good illustration of this kind of understanding is provided by
population biology, one of the main branches of biology studying evolutionary
patterns with the help of mathematical models of populations. In the case of
understanding gathered in this field, such as the understanding of the evolv-
ing interaction of two populations endowed with different traits, research is
accompanied by extensive theoretical commitments: researchers are not ques-
tioning what is meant by evolution, how populations and population interac-
tions are characterized, or whether traits are inheritable or acquired by nurture,
as such assumptions are usually a platform from which they can try to derive
results. Similarly, theoretical biologists tend to make strong performative com-
mitments toward specific ways of looking at phenomena, such as diagrammatic
or mathematical representations and computer programs allowing for specific
types of simulations (and not others). This is the extent to which such research-
ers interact with phenomena to be understood: with the help of modeling tools
that are conceptually abstracted from the phenomena themselves. Performa-
tive commitments and skills are therefore subsidiary to the development of
theoretical skills and knowledge: they are not an aim in themselves, and em-
bodied knowledge of the phenomena under scrutiny is not valued as a crucial
component of researchers’ understanding of them.
The second kind of understanding, embodied understanding, illustrates
the opposite situation: theoretical skills and commitments are used to pursue
and develop a set of performative commitments and skills, which constitute
the main interest of the researcher. Embodied understanding does not require
making substantial theoretical commitments. What is central to it is the ability
to intervene in phenomena, rather than to explain or predict them. Embod-
ied understanding is geared toward acquiring control over phenomena. The
emphasis is on mastering performative skills that will be useful to explore and
modify phenomena at will. This does not mean that theoretical skills are ab-
sent. In some cases, theoretical skills and commitments might be extensively
used to obtain embodied knowledge—for instance, in the biomedical sciences,
where an understanding of germ theory, genetics, and physiology is crucial to
being able to diagnose and treat a patient. Still, doctors rate their ability to
manipulate organisms over and above their ability to explain how organisms
function. What they aim for is an embodied understanding of human beings,
under standing in biol o gy 

which can guide their actions even when they have no idea of why or how the
processes referred to as “symptoms” take place.
The third kind of understanding is the integrated understanding deriving
from a balanced exercise and coordination of theoretical and embodied knowl-
edge. Neither theoretical nor embodied knowledge are given a privileged role
in the understanding of a phenomenon. Rather, researchers are interested both
in acquiring a theoretical interpretation of a phenomenon and in obtaining
tools and methods that will enable them to match such theoretical interpre-
tation to the actual features of the phenomenon. To this aim, the embodied
knowledge used to acquire this type of understanding includes both theoretical
and performative skills: researchers exercise their ability to reason as much as
their ability to observe and/or intervene on the phenomena under scrutiny.
Each type of understanding is suited to different research contexts, since
the quality of understanding acquired by biologists depends largely on their
research settings and goals. For instance, theoretical biologists need to con-
centrate all their skills and resources on acquiring theoretical interpretations of
phenomena. This means committing to using mathematical modeling, that is,
models with high explanatory power but little empirical content. Arguably, this
leads to gaining insights that might not have been obtained if biologists insisted
on making empirical sense of their findings throughout their research. It thus
seems entirely justified to regard theoretical understanding as the most useful
kind of understanding in theoretical biology. Another example is the field of
natural history, where researchers commit to accumulating detailed knowledge
about the morphology of as many species as possible. They thus acquire em-
bodied understanding of what organisms look like, where they can be found,
and how they should be spotted, kept, and eventually embalmed. It does not
matter to them that the skills and commitments of natural historians greatly
limit their theoretical understanding of the differences among organisms (and
the causes of such difference): for their research purposes, the best understand-
ing of organisms is embodied rather than theoretical.
Do theoretical, embodied, and integrated understanding then have the
same epistemic value in biology? I think not. From a normative viewpoint,
integrated understanding constitutes the most desirable form of understand-
ing biological phenomena, since it involves considering theoretical and em-
bodied knowledge as equally relevant to understanding a phenomenon. As I
have claimed above, biology is about the exploration as well as the analysis of
phenomena: researchers who understand a phenomenon in one of these two
ways should ideally strive to understand it also in the other, thus balancing the
amount of theoretical and embodied knowledge used to this aim and acquiring
as rich as possible an understanding. The theoretical understanding of evolving
systems acquired by theoretical biologists could actually be viewed as very par-
 sabina leonelli

tial, given that these researchers often have no idea of how their results could
apply to actual phenomena. Reliance on mathematical models often implies
losing the (embodied) sense of how the results acquired through these models
and the parameters used therein could be matched to observations, measure-
ments, statistics, and other empirical data, since what is taken to define an or-
ganism or a population in the models differs substantially from the complex
features of actual organisms and populations. Similar problems plague the field
of natural history, which has lost much of its eighteenth-century attractiveness
precisely because it does not provide ways to make theoretical sense of the ob-
served differences among species. In both examples, researchers seeking either
theoretical or embodied understanding seem to be missing out on something
important. In Hacking’s (, ) words, what biologists should strive for are
“happy families, in which theory and experiment coming from different direc-
tions meet.”
Whether biologists can actually pursue integrated understanding in prac-
tice, given the disunity and extreme specialization characterizing contempo-
rary research, is a matter of dispute. Cutting-edge fields such as system biology,
bio-ontology, and evolutionary-developmental biology are attempting to build
tools that might indeed enable biologists to better coordinate their theoretical
and embodied knowledge of the natural world (such as databases and in silico
experiments). Philosophers interested in the practice of understanding will
hopefully be in a position to contribute to the success of these efforts.

CdiZh
. Dupré (), Rosenberg (), Mitchell (), and Winther (), among oth-
ers, propose careful analyses of these elements.
. For the mechanisms through which specific tools come to be regarded as indispens-
able to an epistemic culture, see Knorr-Cetina () and Galison ().
. Mitchell  critically surveys different notions of biological theory formulated by
philosophers of biology.
. The collections of essays edited by de Chadarevian and Hopwood (), Laubichler
and Müller (), and Suárez () provide examples of the many types of models used
to gain knowledge about natural phenomena.
. Mattila () discusses the significance of tailoring models.
. On the advantages of model organism research, see also Ankeny () and Leonelli
(a, b).
. A similar point is made in Hacking’s () defense of the significance of manipulat-
ing entities in science, even if his notion of intervention, focusing on experimental interfer-
ence with phenomena, is narrower than the one I propose.
. Woodward’s () manipulationist account of causal explanation takes a view that
is more general than mine—in the sense of applying to causal explanation in all scienc-
under standing in biol o gy 

es—but also more restrictive, insofar as it is limited to causal explanations. My approach is

meant to hold for all types of explanations, including functional and nomological, while I
remain agnostic on whether this could apply to sciences other than biology. Verifying this
latter claim would require extensive comparative analyses among the sciences, which is
beyond the scope of this chapter.
. This is leaving aside the contributions to the study of agency in science made by
scholars focused on experimentation (for example, Gooding  and Radder ) and
the early pragmatists (particularly William James, Charles Peirce, and John Dewey).
. See Rheinberger () and Suárez () about the role of instruments and tech-
niques in shaping the experimental culture in molecular biology; and Gilbert () on
the paradigm shift occasioned in biology by the advent of new data-producing techniques
(such as sequencing) and bioinformatic resources (such as Internet databases).
. Acknowledgments of the tight connection between learning and the material
manipulation of objects have a long history and are increasingly accepted in cognitive sci-
ence. Indicative of this trend is the recent focus on embodied cognition—see, for instance,
Anderson  and references therein.
. Latour () argues that any site of scientific research becomes a laboratory set-
ting. Hans Radder () maintains that the development of methods to replicate experi-
mental results is a necessary step toward developing theories.
. In a similar vein, Galison () uses the expression “coordination of action and
belief ” to describe interactions between experimentalist and theoretician culture in
physics.
. For curators’ reflections on this point, see Huala et al. ().
. On the importance of surprises in exploratory modeling, see Morgan ().
. Experimentation is a very efficient way to coordinate theoretical and embodied
knowledge. It has become a trademark scientific activity at least in part because of that.
However, I do not take it to be the only way to achieve such coordination, as shown by my
first example about the manipulation of digital images.
. This important point is developed in Leonelli (b).
. Under pressure to produce as many short-term results (publications or patents) as
possible, researchers are rarely given time and funding to acquire new skills without being
able to prove their usefulness.
. Elihu Gerson (, ) remarks that: “our tentative alliances [with phenomena]
become stronger as they are used successfully as tools and materials in other projects. It
is at this point that our theories become ‘facts’, firm commitments to act in a certain way.”
Note again the similarity to Hacking’s (, ) formulation of entity realism, where the
ability to manipulate an unobservable entity in the laboratory constitutes a criterion for
believing in the existence of that entity.

GZ[ZgZcXZh
Anderson, M. L. . Embodied cognition: A field guide. Artificial Intelligence
:–.
Ankeny, R. A. . Wormy logic: Model organisms as case-based reasoning. In Science
 sabina leonelli

without laws: Model systems, cases, exemplary narratives, edited by A. N. H. Creager,

E. Lunbeck, and N. Wise. Chapel Hill, NC: Duke University Press.
Beatty, J. . Why do biologists argue like they do? Philosophy of Science :s–.
de Chadarevian, S., and N. Hopwood, eds. . Models: The third dimension of science.
Stanford: Stanford University Press.
de Regt, H. W., and D. Dieks. . A contextual approach to scientific understanding.
Synthese :–.
Dupré, J. . The disorder of things: Metaphysical foundations of the disunity of science.
Cambridge, MA: Harvard University Press.
Galison, P. . Image and logic: A material culture of microphysics. Chicago: University
of Chicago Press.
Gerson, E. M. . The American system of research: Evolutionary biology, –.
PhD diss., Department of Sociology, University of Chicago.
Gilbert, W. . Towards a paradigm shift in biology. Nature :.
Gooding, D. . Experiment and the making of meaning. Dordrecht: Kluwer.
Hacking, I. . Representing and intervening: Introductory topics in the philosophy of
natural science. Cambridge: Cambridge University Press.
Hempel, C. G. . Aspects of scientific explanation. In Aspects of scientific explanation
and other essays in the philosophy of science, –. New York: Free Press.
Huala, E., et al. . The Arabidopsis Information Resource (TAIR): A comprehensive
database and web-based information retrieval, analysis, and visualisation system for a
model plant. Nucleic Acids Research :–.
Knorr Cetina, K. . Epistemic cultures. Cambridge, MA: Harvard University Press.
Latour, B. . Science in action. Cambridge, MA: Harvard University Press.
Laubichler, M., and G. B. Müller, eds. . Modeling biology: Structures, behavior, evolu-
tion. Cambridge, MA: Massachusetts Institute of Technology Press.
Leonelli, S. a. Arabidopsis, the botanical Drosophila: From mouse cress to model
organism. Endeavour :–.
———. b. Weed for thought: Using Arabidopsis thaliana to understand plant biology.
PhD diss., Department of Philosophy, Vrije Universiteit Amsterdam.
Longino, H. . The fate of knowledge. Princeton: Princeton University Press.
Mattila, E. . Interdisciplinarity “in the making”: Modelling infectious diseases. Per-
spectives on Science :–.
Mitchell, S. . Biological complexity and integrative pluralism. Cambridge: Cambridge
University Press.
Morgan, M. S. . Experiments versus models: New phenomena, inference, and sur-
prise. Journal of Economic Methodology :–.
Morgan, M. S., and M. Morrison, eds. . Models as mediators: Perspectives on natural
and social science. Cambridge: Cambridge University Press.
Myers, N. . Molecular embodiments and the body-work of modeling in protein crys-
tallography. Social Studies of Science :–.
Polanyi, M. /. Personal knowledge: Towards a post-critical philosophy. London:
Routledge and Kegan Paul.
under standing in biol o gy 

Radder, H., ed. . The philosophy of scientific experimentation. Pittsburgh: University of
Pittsburgh Press.
———. . The world observed / the world conceived. Pittsburgh: University of Pittsburgh
Press.
Rheinberger, H.-J. . Toward a history of epistemic things: Synthesizing proteins in the
test tube. Stanford: Stanford University Press.
Rosenberg, A. . Instrumental biology, or the disunity of science. Chicago: University of
Chicago Press.
Ryle, G. . The concept of mind. Chicago: University of Chicago Press.
Suárez, E. . Satellite DNA: A case-study for the evolution of experimental techniques.
Studies in History and Philosophy of Biological and Biomedical Sciences :–.
———, ed. . Variedad sin limites: Las rapresentationes en la ciencia. Mexico City:
Universida Nacional Autónoma de México y Editorial Limusa.
Winther, R. G. . Parts and theories in compositional biology. Biology and Philosophy
:–.
Woodward, J. . Making things happen: A theory of causal explanation. New York:
Oxford University Press.
 m a r c e l b oum a n s

((
Understanding in Economics
Gray-Box Models

B6G8:A7DJB6CH

Economists use models to understand the economy. Models

are theories that summarize, often in mathematical terms, the
relationships among economic variables. Models are useful because
they help us to dispense with irrelevant details and to focus on
important economic connections more clearly.
N. G. Mankiw, Macroeconomics

 In economics, models are built to answer specific questions. Each

type of question requires its own type of model; it defines the empirical crite-
ria that a model should meet and thereby instructs how the model should be
constructed. This chapter will investigate a particular kind of question: namely,
questions that ask for understanding, and which will be labeled as how’s that-
questions. To do so, these questions will be compared with other types of scien-
tific questions, labeled as why-questions and how much-questions. The answer
to a why-question is an explanation (see Nagel , ; van Fraassen ,
). The answer to a how much-question is a measurement.
This investigation will be done by clarifying the differences between three
types of models: () models that provide an explanation (answering a why-ques-
tion); () models that provide understanding (answering a how’s that-question);
and () models that provide a measurement (answering a how much-question).
I will show that in economics a white-box model provides an answer to a why-
question. To answer how much-questions, economists can make use of black-


under standing in economic s 

box models. Many, if not most, economic phenomena cannot be investigated

in a laboratory, so any explanation of them will also have to include an account
of the environment in which the phenomena manifest themselves. As a result,
white-box models that function as explanations of economic phenomena out-
side the laboratory are generally so comprehensive and complex that they are
not intelligible. Their comprehensiveness and complexity prevent these models
from providing understanding of the phenomena under investigation.1 It will
be shown that for how’s that-questions, gray-box models are more adequate.
In other words, with respect to phenomena outside the laboratory, there is a
crucial distinction between providing understanding of these phenomena and
explaining them. For these phenomena, due to their complexity and compre-
hensiveness, white-box models do not provide understanding; therefore, econ-
omists have developed another type of model, the gray-box model. Gray-box
models are modular designed models; or to put it differently, gray-box models
are assemblies of modules; these latter items are black boxes with standard
interface.
Because models are built with the purpose of answering questions, their
assessment is closely connected to the type of question for which they are
designed. An important way to assess models is to evaluate their validity. In
Barlas’s essay on model validation in system dynamics, validity of a model is
defined as “usefulness with respect to some purpose” (Barlas , ). I will
show that, although Barlas’s account aims to discuss model validity in system
engineering, it also helps us to understand the assessment of models in eco-
nomics. Barlas notes that for an exploration of the notion of validation, it is
crucial to make a distinction between white-box models and black-box models
(in his account, gray-box models do not appear). In black-box models, what
matters is the output behavior of the model: “the model is assessed to be valid
if its output matches the ‘real’ output within some specified range of accuracy,
without any questioning of the validity of the individual relationships that ex-
ists in the model” (). White-box models, in contrast, are causal-descriptive
statements on how real systems actually operate in some aspects. Generating
accurate output behavior is not sufficient for model validity; the validity of the
internal structure of the model is crucial too. A white-box model must not only
reproduce or predict the behavior of a real system, “but also explain how the
behavior is generated” (–).
Barlas discusses three stages of model validation: direct structure tests,
structure-oriented behavior tests, and behavior pattern tests. Direct struc-
ture tests assess the validity of the model structure by direct comparison with
knowledge about real system structure. This involves taking each relationship
individually and comparing it with available knowledge about the real system.
The list of direct structure tests includes tests such as structure-confirma-
 m a r c e l b oum a n s

tion tests and parameter-confirmation tests. The structure-oriented behavior

tests assess the validity of the structure indirectly by applying certain behavior
tests on model-generated behavior patterns. These tests, such as the extreme-
condition test, the behavior sensitivity test, and the Turing test, involve simula-
tion and can be applied to the entire model, as well as to its separate submodels.
Barlas emphasizes the special importance of structure-oriented behavior tests:
“These are ‘strong’ behavior tests that can provide information on potential
structure flaws” (). Pattern and point prediction tests are examples of be-
havior pattern tests. For white-box models, all three stages are equally impor-
tant; for black-box models, only the last stage matters.
Although Barlas emphasizes that structure-oriented behavior tests are de-
signed to evaluate the validity of the model structure, his usage of the notion of
structure with respect to these kinds of tests needs some further qualification.
The way in which he describes and discusses structure-oriented behavior tests
allows for a notion of structure that is not only limited to realistic descriptions
of real systems; it also includes other kinds of arrangements, like modular orga-
nizations. Structure-oriented behavior tests are also adequate for the validation
of modular-designed models, and for these models the term “structure” refers
to the way the modules are assembled. A module is a self-contained compo-
nent (to be treated as a black box) with a standard interface to other compo-
nents within a system. Modular designed systems may be best known from
industrially produced technical artifacts, like modern personal computers with
their easily insertable and removable plug-and-play devices. These modular-
designed models—in line with the labeling of the other two types of models—
could be called “gray-box models,” and these should pass the structure-oriented
behavior tests and behavior pattern tests.

IVWaZ&&#&#I]gZZIneZhd[FjZhi^dch^c:Xdcdb^Xh

FjZhi^dc 6chlZg BdYZa KVa^YVi^dc

L]n4 :meaVcVi^dc L]^iZWdm 9^gZXihigjXijgZiZhih

   HigjXijgZ"dg^ZciZYWZ]Vk^dgiZhih
   7Z]Vk^dgeViiZgciZhih
=dlÉhi]Vi4 JcYZghiVcY^c\ <gVnWdm HigjXijgZ"dg^ZciZYWZ]Vk^dgiZhih
   7Z]Vk^dgeViiZgciZhih
=dlbjX]4 BZVhjgZbZci 7aVX`Wdm 7Z]Vk^dgeViiZgciZhih

The subsequent sections will show that gray-box models are considered to
be the most appropriate answers to how’s that–questions in economics. I will
show that a validated explanation of an economic phenomenon outside the lab-
oratory will lead unavoidably to comprehensive but incomprehensible models.
under standing in economic s 

The next section discusses Lucas’s program of general-equilibrium economics.

According to this program, which gradually became the dominant program for
macroeconomics, understanding economic phenomena requires models that
pass a Turing test, one of Barlas’s structure-oriented behavior tests, and repre-
sent a general-equilibrium system. Economic general equilibrium systems are,
however, complex systems par excellence. Influenced by modeling methodol-
ogies developed in computer science, which has a strong tradition in coping
with complexity, representations of these complex systems were proposed to
be gray-box models in order to keep them intelligible. In conclusion I will sum-
marize the main claims, critically discuss de Regt and Dieks’s () account
of scientific understanding, and present alternative criteria for understanding
in economics.

:meaVcVi^dc

The common view on explanation is that it involves subsumption under

laws, and the standard assumption about laws is that they are exceptionless
generalizations that make no reference to particular objects or spatio-temporal
locations and have a very wide scope. Woodward (), however, shows that
these criteria are not helpful for understanding the features that characterize
explanatory generalizations in, for example, economics. “In general, it is the
range of interventions and changes over which a generalization is invariant and
not the traditional criteria that are crucial both to whether or not it is a law and
to its explanatory status” ().
Woodward’s idea of invariance is this: a generalization describing a rela-
tionship between two or more variables is invariant if it would continue to
hold—that is, it would remain stable or unchanged—as various other condi-
tions change. The set or range of changes over which a relationship or gen-
eralization is invariant is its domain of invariance. So, invariance is a relative
matter: a relationship is invariant with respect to a certain domain.
Two sorts of changes can be distinguished that are relevant to the assess-
ment of invariance. First, there are changes in the background conditions to
a generalization, that is, changes that affect other variables besides those that
figure in the generalization itself. Second, there are changes in those variables
that figure explicitly in the generalization itself. In his discussion of invariance,
Woodward emphasizes that only a subclass of this latter sort of change is im-
portant: namely, changes that result from an intervention, that is, from a caus-
al process that has the right causal characteristics as described in his article.
The reason, Woodward says, is that some background conditions are causally
independent of the factors related by the generalization in question and are
therefore of no importance. However, other background conditions might be
 m a r c e l b oum a n s

causally connected to some of the factors related by the generalization, and

changes in these conditions might disrupt the relationship. A relationship that
holds under certain specific background conditions and for a restricted range
of interventions might break down outside of these.
The interesting question for economists is not only whether a relationship
is invariant under certain specific kinds of changes and interventions but also
under which changes it remains invariant; they want to know the domain of
changes for which it holds. Economists are faced with constantly changing
background conditions, and they would like to know whether the relationships
on which they base their policy advice will still hold tomorrow.
Although Woodward considers the notion of invariance under interven-
tions as the key feature that a generalization must possess if it is to play an
explanatory role, he admits implicitly that the discussion of invariance in vari-
ous sciences is broader than merely in terms of intervention. He views an inter-
vention as an idealization of an experimental manipulation—by human beings
or Nature—and probably therefore de-emphasizes the role of unstable back-
ground conditions. However, when he discusses the idea that invariance comes
with degrees, he uses the notion of Haavelmo’s autonomy (“just another name
for what we have been calling invariance” [Woodward , ]) to clarify
the relativistic characteristic of invariance. In Haavelmo’s account of autonomy,
invariance is defined with respect not only to interventions but also to changes
in background conditions, as we will see below.
In Woodward’s account of explanation in relation to invariance, the differ-
ence between laws and invariant generalizations is considered as a matter of
degree: laws are generalizations that are invariant under a large(r) and (more)
important set of changes. So, any strategy to find laws outside a laboratory has
to deal with the question: invariant with respect to what domain? In economics
(and econometrics), such an account is captured by the notion of “autonomy.”
Haavelmo () dubbed the problem of finding invariant relationships
in economics “the problem of autonomy,” which he defined as the problem of
“judging the degree of persistence over time of relations between economic
variables,” or, more generally speaking, “whether or not we might hope to find
elements of invariance in economic life, upon which to establish permanent
‘laws’ ” (). The problem we typically face in economics is that real economic
phenomena cannot be separated from other potential influences falling outside
the theoretical domain. We have to deal with passive observations, and these
are influenced by a great many factors not accounted for in theory which can-
not be eliminated by creating a ceteris paribus environment—a laboratory.
To explore the problem of autonomy, consider the following problem of
white-box modeling. Let y be an economic variable whose behavior is deter-
mined by a function F, of independent causal factors, x, x, . . . :
under standing in economic s 

y = F(x1, x2, . . . ) (1)

The way in which the factors xi might influence y can be represented by the
following equation:

Δy = ΔF(x1, x2, . . . ) = F1 Δx1 + F2 Δx2 + . . . (2)

The delta’s, Δ, indicate a change in magnitude. The terms Fi indicate how much
y will change proportionally as a result of a change in magnitude of factor xi.
Suppose we are trying to discover an invariant generalization that could
be used to explain the behavior of y. In principle, there are an infinite number
of factors x, x, . . . , that could influence the behavior of y, but we hope that it
may be possible to establish a constant and relatively simple relation between y
and a relatively small number of explaining factors, x. In a laboratory, we would
artificially isolate a selected set of factors from the other influences. In other
words, we would take care that ceteris paribus (CP) conditions are imposed:
Δxn+ = Δxn+ = . . . = 0, so that a simpler relationship can be investigated:

Δycp = F1 Δx1 + . . . + Fn Δxn (3)

Moreover, in a controlled experiment the remaining factors, xi , can be changed

in a systematic way to gain knowledge about the Fi’s and, so, establish the rela-
tionship between y and a limited number of factors x, . . . xn.
However, in economics we are not able to carry out such experiments. Hav-
ing only passive observations available, Haavelmo’s distinction between two
different notions of influence, namely potential influence and factual influence,
is fundamental to judge invariance. When Fi differs significantly from zero, fac-
tor xi has potential influence. The combination Fi·Δxi indicates the magnitude
of the factual influence of a factor xi upon y.
We usually passively observe (PO) a limited number of factors that have a
nonnegligible factual influence:

Δypo = F1 Δx1 + . . . + Fn Δxn (4)

Thus, the relationship y = F(x, …, xn) explains the actual observed values of y,
provided that the factual influence of all the unspecified factors together was
very small as compared with the factual influence of the specified factors x,
. . . , xn.
The problem, however, is that it is not possible to identify the reason for the
factual influence of a factor, say xn+, being negligible, that is, Fn+·Δxn+ ≈ 0. By
passive observations alone, we cannot distinguish whether its potential influ-
 m a r c e l b oum a n s

ence is very small, Fn+ ≈ 0, or whether the factual variation of this factor over
the period under consideration was too small, that is, Δxn+ ≈ 0. We would like
only to get rid of factors whose influence was not observed because their po-
tential influence was negligible to start with. At the same time, we want to re-
tain potential factors whose influence was not observed because they varied so
little that their potential influence was veiled.
The variation of xn+ is determined by other relationships within the sys-
tem. In some cases, a virtually dormant factor may become active because of
changes in the economic structure elsewhere. However, deciding whether a
factor should be accounted for in the relationship under investigation should
not depend on such changes, but only on its potential influence. The relation-
ship should be autonomous with respect to structural changes elsewhere. Au-
tonomous relations are those relations that should be expected to have a great
degree of invariance with respect to various changes in the economic structure.
However, this kind of invariance should not be equated with the observable
degree of constancy or persistence of a relation.
Haavelmo’s design rules for white-box modeling in econometrics were
considered as an alternative to the experimental methods of science (Mor-
gan , ). However, although researchers at the Cowles Commission2
adopted Haavelmo’s “blueprint” for econometrics (Morgan , ), they
scrapped the term “autonomy” because they believed that the theoretical rela-
tionships they were trying to measure were obviously autonomous. The reason
the Cowles Commission assumed this was because Haavelmo had pointed out
the possibility that the empirically found relationships may be simpler than
theory would suggest. This could lead researchers to discard potential influ-
ences, which could be avoided by building models as comprehensive as possi-
ble, based on a priori theoretical specifications. To explain phenomena outside
the laboratory, white-box models cannot discard the phenomena’s background
conditions; they should not only provide causal-descriptive statements about
the phenomena to be explained but also statements about the environments in
which they appear.
So, the Cowles Commission’s solution to the problem of autonomy was to
build increasingly comprehensive white-box models by building in as many po-
tential influences as possible. In the s, Lawrence Klein was commissioned
to build Cowles Commission–type models of the United States. The program’s
aim was to build increasingly comprehensive models to improve their predict-
ability so that they could be used as reliable instruments for economic policy.
Klein maintained this view throughout his life:

In contrast with the parsimonious view of natural simplicity, I believe

that economic life is enormously complicated and that the successful
under standing in economic s 

model will try to build in as much of the complicated interrelation-

ships as possible. That is why I want to work with large econometric
models and a great deal of computer power. Instead of the rule of par-
simony, I prefer the following rule: the largest possible system that can
be managed and that can explain the main economic magnitudes as
well as the parsimonious system is the better system to develop and
use. (Klein , )

One of the major efforts of the s in this respect was the Brookings
econometric model of the United States (Dusenberry et al. ). This white-
box model was a joint effort of many individuals, and at its peak it contained
nearly four hundred equations. Although much was learned from this exer-
cise, the model never achieved the success that was initially expected, and it
was laid to rest around . Such an enormous model, of which the builders
only survey the part they have contributed, and of which the overall dynamic
behavior can only be revealed by computer simulations, does not provide any
understanding of the economic system it represents. Such models can only an-
swer policy-related questions, which can be labeled as what-if–questions (see
also Lenhard in this volume).

AjXVhÉhEgd\gVbd[<ZcZgVa":fj^a^Wg^jb:Xdcdb^Xh

At the end of the s a new practice of simulations arose which has had a
lasting impact on how to assess models. Klein together with Goldberger ()
had arrived at a model of twenty-five difference equations with a correspond-
ing number of endogenous variables; it was nonlinear in character and some of
the difference equations were of the fifth order—at that time clearly the most
advanced macroeconometric model. The model had been applied to annual
projections of economic activity in the United States with some success, but its
dynamic properties were only analyzed using highly simplified assumptions.
The innovative element of the research carried out by Adelman and Adelman
() was that, rather than making simplifying assumptions, the complexity of
this white-box model was left intact and the equations were programmed for a
computer and simulated for one hundred annual periods.
Adelman and Adelman were interested in an explanation of the persistent
business fluctuations of the United States. They found that random shocks su-
perimposed on the extrapolated values of the exogenous quantities and ran-
dom shocks introduced into each nondefinitional model equation induced
cycles with three- to four-year periods and amplitudes that were “reasonably
realistic.” That the amplitudes and the periods of oscillations observed in this
model were “roughly the same as those which are found in practice” (Adelman
 m a r c e l b oum a n s

and Adelman , ) was seen by the Adelmans as “merely a necessary con-
dition for an adequate simulation of the cyclical fluctuations of a real industrial
economy” (). The question now was whether the shocked model could pro-
duce business cycles in the “technical” sense: “If a business cycle analyst were
asked whether or not the results of a shocked Klein-Goldberger computation
could reasonably represent a United States-type economy, how would he re-
spond? To answer these questions we shall apply to the data techniques devel-
oped by the National Bureau of Economic Research (NBER) for the analysis of
business cycles” ().
A comparison between the characteristics of the cycles generated by the
shocked model and the business-cycle characteristics summarized in NBER3
publications was considered to be “quite a stringent test of the validity of the
model” by the Adelmans (, ). The striking result was that when random
shocks of a “realistic order of magnitude” were superimposed on the Klein-
Goldberger model equations, the characteristics of the resulting cyclical fluc-
tuations appeared to be similar to those observed in the U.S. economy. The
Adelmans concluded, “it is not unreasonable to suggest that the gross char-
acteristics of the interactions among the real variables described in the Klein-
Goldberger equations may represent good approximations to the behavioral
relationships in a practical economy” ().
To Lucas, the Adelmans’ achievement signaled a new standard for what it
means to “understand” business cycles: “One exhibits understanding of busi-
ness cycles by constructing a model in the most literal sense: a fully articulated
artificial economy which behaves through time so as to imitate closely the time
series behavior of actual economics” (Lucas , ). To see that the Adel-
mans’ test works as a stringent test, Lucas understood that the facts that are
to be reproduced should be a list of characteristics providing as much detail as
possible, to be found in the various NBER empirical studies.
However, Lucas paraphrased the Adelmans’ question above as follows: “The
Adelmans posed, in a precise way, the question of whether an observer armed
with the methods of [the NBER] could distinguish between a collection of eco-
nomic series generated artificially by a computer programmed to follow the
Klein-Goldberger equations and the analogous series generated by an actual
economy” (). This paraphrasing is important because the characteristics test
of the Adelmans is thereby reinterpreted as a Turing test. This test was original-
ly described by Turing () as an “imitation game” to investigate the question
“Can machines think?” Today, a Turing test is generally described as follows:
Reports based on the output of the quantitative model and on measurements
of the real system are presented to a team of experts. When they are not able to
distinguish between the model output and the system output, the model is said
to be valid (see, for example, van Daalen, Thissen, and Verbraeck ). As one
under standing in economic s 

can see, the test is in principle the same as the Adelmans’ test: an observer (in-
terrogator or expert) has to decide whether a distinction can be made between
a computer output and output from the “real” world.
The enormous advantage of Turing’s approach to artificial intelligence is
that it freed scientists from building replicas of the human mind to achieve ma-
chine thinking that meets the standard of human intelligence. In the same way,
the Adelmans’ characteristics test freed macroeconometricians from having to
build white-box models: “detailed, quantitatively accurate replicas of the actual
economy” (Lucas , ). As mentioned above, economic white-box models
of macroeconomic systems are inevitably too complex and too comprehensive.
Turing testing legitimized Lucas to work with very simple (and therefore un-
realistic) models. Lucas’s approach was not to aim at models as “accurate de-
scriptive representations of reality”: “Insistence on the ‘realism’ of an economic
model subverts its potential usefulness in thinking about reality. Any model
that is well enough articulated to give clear answers to the questions we put to
it will necessarily be artificial, abstract, patently ‘unreal.’ ” (Lucas , ).
According to Lucas, the model assumptions need not be assertions about
the world:

A “theory” is not a collection of assertions about the behavior of the

actual economy but rather an explicit set of instructions for building
a parallel or analogue system—a mechanical, imitation economy. A
“good” model, from this point of view, will not be exactly more “real”
than a poor one, but will provide better imitations. Of course, what
one means by a “better imitation” will depend on the particular ques-
tions to which one wishes answers. (Lucas , )

In the “general equilibrium program” dominated by Lucas’s instructions, it be-

came standard practice to run an experiment with an artificial economy: “One
of the functions of theoretical economics is to provide fully articulated, arti-
ficial economic systems that can serve as laboratories in which policies that
would be prohibitively expensive to experiment with in actual economies can
be tested out at much lower cost” ().
In Lucas’s view, the ability of models to imitate actual behavior in the way
tested by the Adelmans is a necessary, but not sufficient, condition to use these
kinds of macroeconometric models for policy evaluation. Policy evaluation re-
quires “invariance of the structure of the model under policy variations” (Lucas
, ). The underlying idea, known as the Lucas Critique, is that estimated
parameters that were previously regarded as “structural” (or “autonomous” in
Haavelmo’s terminology) in econometric analysis of economic policy actually
depend on the economic policy pursued during the estimation period. Hence,
 m a r c e l b oum a n s

the parameters may change with shifts in the policy regime (Lucas ). Lu-
cas’s  essay is perhaps the most influential and most cited paper in mac-
roeconomics (Hoover ), and it contributed to the decline in popularity of
the Cowles Commission approach. The Lucas Critique was an implicit call for
a new research program that amounted to the aim and justification of develop-
ing black-box and gray-box models. This alternative to the Cowles Commission
program involved formulating and estimating macroeconometric models with
parameters that are invariant under policy variations and can thus be used to
evaluate alternative policies. The only parameters Lucas assumed to be invari-
ant under policy changes are the parameters describing “tastes and technology”
(Lucas , ). By “tastes” Lucas referred to microeconomic characteristics
like individual preferences, and by “technology” he referred to stable institu-
tional characteristics.
Lucas’s “general equilibrium program” was most successfully carried out
by Kydland and Prescott (). In a special symposium, Computation Experi-
ments in Macroeconomics, they explained their general equilibrium program
methodology. Their “experiment” was an implementation of Lucas’s program
that ran a simulation experiment on an artificial economy. According to Kydland
and Prescott (, –), any economic computational experiment involves
the following methodology: The purpose of a computational experiment is to
derive a quantitative answer to some well-posed question. Therefore, what is
needed is a general-equilibrium theory. A theory is not a set of assertions about
the actual economy; rather, following Lucas (), it is defined as an explicit set
of instructions for building a mechanical imitation system to answer a question.
A model can be judged only relative to some given question. The features of a
given model may be appropriate for some specific question (or specific class of
questions, like how’s that–, how much– or what if–questions), but not for oth-
ers. Generally, some economic questions have a known answer, and the model
should give an approximately correct answer to them if we are to have any con-
fidence in the answer given to the question with an unknown answer. Kydland
and Prescott called this way of model assessment “calibration.” Thus, data are
used to calibrate the model economy so that it mimics the world as closely as
possible along a limited but clearly specified number of dimensions.
Kydland and Prescott’s specific kind of assessment is similar to Lucas’s idea
of testing, although he did not call it “calibration.” I argued above that Lucas’s
idea of testing is similar to a Turing test. It should be noted that a Turing test
does not make sense if you raise questions to which you yourself do not know
the answers. To have confidence that a computer is intelligent, it should give
approximately correct answers to questions to which you know the (approxi-
mately correct) answer in advance. To test models as “useful imitations of real-
ity,” we should subject them to shocks to “which we are fairly certain how actual
under standing in economic s 

economies, or parts of economies, would react. The more dimensions on which

the model mimics the answer actual economies give to simple questions, the
more we trust its answer to harder questions” (Lucas , –). This kind
of testing is similar to calibration as defined by Franklin (, ): “the use of
a surrogate signal to standardize an instrument. If an apparatus reproduces
known phenomena, then we legitimately strengthen our belief that the appa-
ratus is working properly and that the experimental results produced with that
apparatus are reliable.”
Kydland and Prescott called this method of testing “calibration” instead of
a Turing test because they were only interested in how much–questions, that
is, in measurement. The research of Prescott and coauthors explicitly aims at
answering how much–questions and certainly not why-questions: “In our busi-
ness cycle studies, we do not try to fit or explain anything . . . theory is a set
of instructions for constructing a model to measure something” (Kydland and
Prescott , –). They considered models as measuring instruments.
One of the “harder” questions Kydland and Prescott wanted one of their mod-
els to answer was “What is the quantitative nature of fluctuations induced by
technology shocks?” (Kydland and Prescott , ). The answer to this ques-
tion was “the model economy displays business cycle fluctuations  percent as
large as did the U.S. economy” (). In other words, the answer is supposed to
be a measurement result carried out with a calibrated instrument.
It should be noted, however, that though they claim only to be interested
in answering how much–questions, their modeling methodology agrees with
Lucas’s more general research program, which also includes how’s that–ques-
tions. Black-box models suffice as measuring instruments, but their modeling
strategies lead to gray-box models, which are also appropriate for answering
how’s that–questions.

<gVn"7dmBdYZah

Gray-box model building in the general-equilibrium program can be char-

acterized by what Orcutt () in one of the first accounts of simulations in
economics called the “building-block approach.” This approach makes com-
puter simulation of large-scale complex systems feasible.

Extensive testing of individual pieces must be carried out before the

pieces are assembled, and even after they have been assembled, it fre-
quently may be necessary to modify some pieces. Also in finding and
in eliminating the errors in a large and complex computer program
it is important to be able to do it piece by piece. And even after it is
assembled, it frequently may be desirable to alter a particular operat-
 m a r c e l b oum a n s

ing characteristic, or parameter, or the initial composition of compo-

nents and their status variables. For these reasons it is highly useful
to take the individual components of a model as building blocks and
construct them and the over-all model so that they are like the fully
plugable components of a modern piece of electronic equipment. (Or-
cutt , –)

Given a building-block approach, another important method of facilitating the

design of a computer simulation is, according to Orcutt, to replicate the com-
ponents. While large numbers of components may be used in a model, the de-
signer must attempt to reduce these components “to a relatively small number
of major types” (). Components of the same major type have identical oper-
ating characteristics. Replication of this kind not only saves labor in developing
and programming a model, but also has advantages that relate to the estimation
and testing of the operating characteristics of the basic components.
In current systems engineering, this gray-box approach is better known as
“modular design”:

Modular design simplifies final assembly because there are fewer

modules than subcomponents and because standard interfaces typi-
cally are designed for ease of fit. Each module can be tested prior to
assembly and, in the field, repairs can be made by replacing defective
modules. Custom systems can be realized by different combinations
of standard components; existing systems can be upgraded with im-
proved modules; and new systems can be realized by new combina-
tions of existing and improved modules. (White , )

In the general equilibrium literature (see, for example, the two standard survey
volumes, Cooley  and Stokey and Lucas with Prescott ), two modules
are always part of the models: a Cobb-Douglas production function, which has
the form: F(kt, ht) = ktθht1-θ, where kt denotes capital and ht labor; and a house-
hold utility function defined by u(ct, –ht) = (–α)log ct + α log (–ht), where
ct denotes consumption. They are parameterized, that is, incorporate a list of
required stylized economic growth facts, and need only be slightly adapted to
meet the wishes of the customer.
In systems engineering, an important step in the design process is param-
eterization: “Parameterization consists of identifying the numerical quantities
that specify each element in the system and their permissible ranges. A sys-
tem specification consists of a set for the system parameters together with the
system configuration. Alternatively, we could view this as the specification of
a set of parts and the instructions to assemble and put these to use” (White
under standing in economic s 

, ). In the general equilibrium literature, this way of parameterization

is similar to how the parametric classes are specified by incorporating the styl-
ized facts of growth into the modules of “technology” (production function)
and “tastes” (utility function).
The development of models for understanding complex systems has a long
tradition in computer science. Von Neumann () is paradigmatic to this
approach. According to von Neumann, the problem of complexity consists of
two parts. The first part is partitioning into elements: “The natural systems are
of enormous complexity, and it is clearly necessary to subdivide the problem
that they represent into several parts. One method of subdivision, which is par-
ticularly significant in the present context, is this: The organisms can be viewed
as made up of parts which to a certain extent are independent, elementary
units” (von Neumann , ). The second part consists of understanding
how these elements are organized into a whole, and how the functioning of the
whole is expressed in terms of these elements. This second part is something
that can be solved by mathematicians and logicians. The first part of the prob-
lem could be removed by the “process of axiomatization”:

The Axiomatic Procedure. Axiomatizing the behavior of the elements

means this: We assume that the elements have certain well-defined,
outside, functional characteristics; that is, they are to be treated as
“black boxes.” They are viewed as automatisms, the inner structure of
which need not be disclosed, but which are assumed to react to certain
unambiguously defined stimuli, by certain unambiguously defined re-
sponses. ()

The general approach is of partitioning into elementary units, which can be

treated as black boxes, but it is remarkable that these elements are labeled as
“axioms,” indicating that they have some fundamental status.
Simon  is an elaboration of von Neumann’s axiomatic procedure. The
central thesis of Simon’s article is that complex systems frequently take the
form of hierarchic systems. A hierarchic system is one composed of interre-
lated subsystems, each of the latter being, in turn, hierarchic in structure until
we reach some lowest level of elementary subsystems. Each subsystem can be
treated as a black box whose inputs and outputs, but not its internal structure,
are of interest.
A system is decomposable when the subsystems are independent of one
another, that is, when there is no interaction between them. Decomposable sys-
tems are much easier to investigate, but complex systems are by definition not
decomposable. It was Simon’s solution for the description and comprehension
of complex systems to show that hierarchic systems have the property of near-
 m a r c e l b oum a n s

decomposability. Nearly decomposable systems are those in which the interac-

tions between the subsystems are weak but not negligible. As a result of this
feature, two propositions can be stated: “(a) in a nearly decomposable system,
the short-run behavior of each of the component subsystems is approximately
independent of the short-run behavior of the other components; (b) in the long
run, the behavior of any one of the components depends in only an aggregate
way on the behavior of the other components” (Simon , ).
An important consequence of the fact that subsystems only interact in an
aggregative fashion is that, in studying the interaction of two different subsys-
tems, the detail of their interaction can be ignored. As a result, complex sys-
tems, if they have a hierarchical structure, can be intelligible:

The fact, then, that many complex systems have a nearly decompos-
able, hierarchic structure is a major facilitating factor enabling us to
understand, to describe, and even to “see” such systems and their
parts. Or perhaps the proposition should be put the other way round.
If there are important systems in the world that are complex with-
out being hierarchic, they may to a considerable extent escape our
observation and our understanding. Analysis of their behavior would
involve such detailed knowledge and calculation of the interactions
of their elementary parts that it would be beyond our capacities of
memory or computation. (Simon , )

Reviewing Simon’s contributions to economics, Ando () formulates

quite accurately a theme that runs consistently throughout Simon’s writings:
“to construct a comprehensive framework for modeling and analyzing the be-
havior of man and his organizations faced with a complex environment, rec-
ognizing the limitation of his ability to comprehend, describe, analyze and to
act, while allowing for his ability to learn and adopt” (Ando , ). One of
Simon’s research areas was therefore to find a description of that complex envi-
ronment so that it is both comprehensible and manageable for decisionmakers.
Early in his career, Simon had already found that the description of a very com-
plex system could be simplified by partitioning it hierarchically into a number
of weak interacting subsystems in such a way that each subsystem could be
analyzed separately and, moreover, in an aggregate (that is, a nondetailed) way.
As a result, these simplified representations of complex systems do provide un-
derstanding. This strategy can be found in several of Simon’s articles that deal
with complexity.4 Simon’s work on hierarchical systems became the starting
point for most of the later modularity literature in systems engineering, sys-
tems biology (see, for example, Krohs and Callebaut ), and management
science, and provides the basic tools for gray-box modeling.
under standing in economic s 

8dcXajh^dch

Morrison and Morgan () argue that models function as “instruments

of investigation.” We can learn about the world from them because they involve
some form of representation. The relevant question about instruments is not
“How true are they?” but “How useful are these instruments to answer specific
questions?” The usefulness of an instrument, nevertheless, depends on the em-
pirical adequacy of the representation it involves. The empirical adequacy of a
representation depends on the complexity of the system under investigation
and our ability to construct representations of this complex system that cap-
ture the essential (for example, invariant) characteristics. This is, in principle,
a technical problem, labeled by Hoover () as the “Cournot problem,” after
the man who formulated it explicitly:

The economic system is a whole of which all the parts are connected
and react on each other. . . . It seems, therefore, as if, for a complete
and rigorous solution of the problems relative to some parts of the
economic system, it were indispensable to take the entire system into
consideration. But this would surpass the powers of mathematical
analysis and of our practical methods of calculation, even if the values
of all the constants could be assigned to them numerically. (Antoine
Augustin Cournot, quoted in Hoover , –)

The general equilibrium program attempts to provide solutions to this Cour-

not problem by setting up a manageable interdependent system to represent a
whole economy. A mid-twentieth-century version of this program is the Cowles
Commission approach. This was a combination of the Walrasian method,
which attempts to construct a mathematical skeleton of a system, and econo-
metrics, to put empirical flesh on the bones of the system. Although the white-
box models that resulted from this approach provided explanations, they were
not, however, intelligible because of their complexity and comprehensiveness.
Lucas’s general equilibrium program of constructing “analogue economies”
introduced an alternative approach for dealing with the Cournot problem. His
approach was admittedly influenced by developments in computer science and
systems engineering, particularly with respect to simulations. In this chapter, I
have shown that the gray-box modeling methodology that resulted from these
disciplines led to tools for understanding.
The assessment of a model that functions as an answer to a question de-
pends on the specific type of question. The premises of an explanation have
to include invariant relationships (Woodward ), and thus the reliability of
such an answer depends on whether the domain of invariance of the relevant re-
 m a r c e l b oum a n s

lationships covers the domain of the question. A measurement is reliable when

it is an output of a calibrated measuring instrument. Calibration is defined as
a set of operations that establish, under specified conditions, the relationship
between values of quantities indicated by a measuring instrument and the cor-
responding values realized by standards. The idea of a standard is that it is often
based upon naturally occurring phenomena when these possess the required
degree of stability. These stable phenomena need not necessarily be invariant
relationships but can also be stable facts. In economics, what are known as the
stylized facts of economic growth fulfill this role of calibrating facts.
Models built to provide understanding have to meet validation criteria oth-
er than those that provide explanations or measurements. While each module
(black box with standard interface) of a gray-box model has to be calibrated
individually, the overall test for the assemblage of these modules is a kind of
Turing test.
These results about gray-box understanding in economics imply a twofold
revision of de Regt and Dieks’s () account of understanding, which consists
of two criteria:
. Criterion for Understanding Phenomena (CUP): A phenomenon P
can be understood if a theory T of P exists that is intelligible (and meets
the usual logical, methodological, and empirical requirements) ().
. Criterion for the Intelligibility of Theories (CIT): A scientific theory
T is intelligible for scientists (in context C) if they can recognize qualita-
tively characteristic consequences of T without performing exact calcula-
tions ().

The first revision is that economists use models, instead of theories, to un-
derstand economic phenomena.5 Theories are considered to act only as guides
to find the main causal factors or, more generally, as a set of instructions to
built models. The second revision is that for phenomena outside the labora-
tory, the methodological requirements imply that these models are modular
designed gray-box models (the logical requirements), which are validated by
structure-oriented behavior tests and behavior pattern tests (the empirical re-
quirements). Simon’s work on nearly decomposable systems shows that econo-
mists can recognize qualitatively characteristic consequences of these kinds
of models. These two revisions lead to the following two adjusted criteria for
economics:
. Criterion for Understanding Economic Phenomena (CUEP): An
economic phenomenon P can be understood if a model M of P exists
that is intelligible.
. Criterion for the Intelligibility of Models (CIM): A scientific
model M is intelligible for economists (in context C) if it is a gray-box
under standing in economic s 

model (modular-designed and validated by structure-oriented behav-

ior tests and behavior pattern tests).

In other words, surveying how understanding is achieved in economics, de

Regt and Dieks’s criteria for understanding phenomena have to be revised: their
focus must be shifted from theory to model, and the methodological require-
ments have to be specified for the design and assessment of a gray-box model.

CdiZh
. I will assume here that for phenomena investigated in a laboratory, how’s that–ques-
tions can be answered by white-box models. In other words, for laboratory phenomena,
explanations can also provide understanding.
. The Cowles Commission for Research in Economics was set up in  and funded
by Alfred Cowles specifically to undertake econometric research. The journal Economet-
rica, in which Haavelmo’s essay appeared, was run from the commission. The Cowles
Commission’s econometric approach, developed in the s and s, became the
standard approach as presented in econometric textbooks.
. Founded in , the National Bureau of Economic Research (NBER) is a research
organization focused on empirical investigations of how the U.S. economy works.
. Boumans () discusses several economic examples for which Simon applied his
framework of hierarchical systems.
. See also the quotation from Mankiw () at the beginning of this chapter, which
is from a standard textbook in macroeconomics used worldwide. A similar suggestion for
a model-based understanding is also proposed by Lenhard, and Knuuttila and Merz in this
volume.

GZ[ZgZcXZh
Adelman, I., and F. L. Adelman. . The dynamic properties of the Klein-Goldberger
model. Econometrica :–.
Ando, A. . On the contribution of Herbert A. Simon to economics. Scandinavian
Journal of Economics :–.
Barlas, Y. . Formal aspects of model validity and validation in system dynamics. Sys-
tem Dynamics Review :–.
Boumans, M. . A macroeconomic approach to complexity. In Simplicity, inference and
modelling, edited by A. Zellner, H. A. Keuzenkamp, and M. McAleer, –. Cam-
bridge: Cambridge University Press.
Cooley, T. F., ed. . Frontiers of business cycle research. Princeton: Princeton University
Press.
de Regt, H. W., and D. Dieks. . A contextual approach to scientific understanding.
Synthese :–.
 m a r c e l b oum a n s

Dusenberry, J. S., G. Fromm, L. R. Klein, and E. Kuh, eds. . The Brookings Quarterly
econometric model of the United States. Chicago: Rand McNally.
Franklin, A. . Calibration. Perspectives on Science :–.
Haavelmo, T. . The probability approach in econometrics. Supplement to Economet-
rica .
Hoover, K. D. . The new classical macroeconomics. Oxford: Basil Blackwell.
———. . The problem of macroeconometrics. In Macroeconometrics, developments,
tensions, and prospects, edited by K. D. Hoover, –. Boston: Kluwer.
Klein, L. R. . My professional life philosophy. In Eminent economists: Their life philoso-
phies, edited by M. Szenberg, –. Cambridge: Cambridge University Press.
Klein, L. R., and A. S. Goldberger. . An econometric model of the United States,
–. Amsterdam: North-Holland.
Krohs, U., and W. Callebaut. . Data without models merging with models without
data. In Systems biology: Philosophical foundations, edited by F. C. Boogerd, F. J. Brug-
geman, J.-H. S. Hofmeyr, and H. V. Westerhoff, –. Amsterdam: Elsevier.
Kydland, F. E., and E. C. Prescott. . The computational experiment: An econometric
tool. Journal of Economic Perspectives :–.
———. . A response from Finn E. Kydland and Edward C. Prescott. Journal of Eco-
nomic Perspectives :–.
Lucas, R. E. . Econometric policy evaluation: A critique. In The Phillips Curve
and labor markets, edited by K. Brunner and A. H. Meltzer, –. Amsterdam:
North-Holland.
———. . Understanding business cycles. In Stabilization of the domestic and in-
ternational economy, edited by K. Brunner and A. H. Meltzer, –. Amsterdam:
North-Holland.
———. . Methods and problems in business cycle theory. Journal of Money, Credit,
and Banking :–.
Mankiw, N. G. . Macroeconomics. rd edition. New York: Worth.
Morgan, M. S. . The history of econometric ideas. Cambridge: Cambridge University
Press.
Morrison, M., and M. S. Morgan. . Models as mediating instruments. In Models as
mediators: Perspectives on natural and social science, edited by M. S. Morgan and M.
Morrison, –. Cambridge: Cambridge University Press.
Nagel, E. . The structure of science: Problems in the logic of scientific explanation. Lon-
don: Routledge and Kegan Paul.
Orcutt, G. H. . Simulation of economic systems. American Economic Review
:–.
Simon, H. A. . The architecture of complexity. Proceedings of the American Philosophi-
cal Society :–.
Stokey, N. L., and R. E. Lucas, with E. C. Prescott. . Recursive methods in economic
dynamics. Cambridge, MA: Harvard University Press.
Turing, A. M. . Computing machinery and intelligence. Mind :–.
van Daalen, C. E., W. A. H. Thissen, and A. Verbraeck. . Methods for the modeling
under standing in economic s 

and analysis of alternatives. In Handbook of systems engineering and management,

edited by A. P. Sage and W. B. Rouse, –. New York: Wiley.
van Fraassen, B. C. . The pragmatic theory of explanation. In Theories of explanation,
edited by J. C. Pitt, –. Oxford: Oxford University Press.
von Neumann, J. . The general and logical theory of automata. In John von Neumann:
Collected works, vol. , edited by A. H. Taub, –. Oxford: Pergamon.
White, K. Preston. . Systems design. In Handbook of systems engineering and manage-
ment, edited by A. P. Sage and W. B. Rouse, –. New York: Wiley and Sons.
Woodward, J. . Explanation and invariance in the special sciences. British Journal for
the Philosophy of Science :–.
 pa rt title ver so

()
Understanding in Physics
Bottom-Up versus Top-Down

9:CC>H9>:@H

 Physics is the paradigmatic example of a successful science. One of its

great successes is its impressive track record of giving explanations of natural
phenomena, by which these phenomena are made understandable. This much
is generally granted, but things become less clear when one asks what these
physical explanations exactly consist in. Philosophers of science have proposed
a variety of analyses of explanation (nomological-deductive, causal, unification,
to mention but a few), and it is not immediately obvious which of these propos-
als best captures physical practice.
My position is that there is no unique answer to this question, that the ques-
tion is even ill-posed. Explanations and ways of achieving understanding are
contextual in physics, no less than in other disciplines. As a consequence, there
exists no uniquely best explanatory scheme. Instead, there is a plurality of pos-
sible physical explanations and ways of understanding physical processes, and
it depends on the type of question that is asked and on the aim and interests of
the scientist that poses the question which one is the most appropriate. In other
words, what is the best explanation and the best strategy for achieving under-
standing depends on contextual, pragmatic factors. In the following I shall il-
lustrate this general point by focusing on a specific instance of the plurality that
is involved: bottom-up versus top-down approaches in fundamental physics.
Let me start with some words about the historical context in which this distinc-


under standing in ph ysic s 

tion was first explicitly introduced—this will, quite fittingly, help to make clear
what its role and status are.
Shortly after the First World War, in , Albert Einstein unexpectedly
rose to worldwide public fame. In  the final version of his General Theory
of Relativity had appeared, securing his reputation in the academic world. Even
during the war years preparations had started in English university circles to
test one of the most significant predictions of this new theory, namely the bend-
ing of light by massive bodies. It is true that Newton’s theory of gravitation also
predicts such an effect, if light is conceived as a stream of particles attracted
by gravity, but the numerical value of the deflection predicted by Einstein was
significantly different. In the case of star light bent by the sun general relativity
yields a deflection about twice as big as the Newtonian value, which suggests
the possibility of a crucial experiment.
So it happened that in  two sun eclipse expeditions were sent off from
England in order to measure the actual magnitude of the light deflection. The
outcomes were presented at a special joint meeting of the Royal Astronomi-
cal Society and the Royal Society of London on November  of the same year,
where the majority of those present considered the results to favor Einstein’s
theory most, although there was no unanimity. The next day, November , ,
however, the London Times carried an extensive article about the meeting, with
the headline “Revolution in Science: New Theory of the Universe.” This marked
the beginning of Einstein’s role as a public hero and genius. The editors of the
London Times invited Einstein to write a popular piece about his new theory,
a request he gladly agreed to, and on November  his now-famous article My
Theory appeared (Einstein /, ).
In addition to its role as one of the milestones in the public perception of
Einstein, this  article has become known for the methodological consider-
ations with which Einstein prefaced his explanation of relativity theory. Here,
Einstein made a distinction between two ways of constructing theories and
giving explanations in physics, one “bottom-up” and one “top-down.” As he
put it,

We can distinguish between various kinds of theories in physics. Most

of them are constructive. They attempt to build up a picture of the
more complex phenomena out of the materials of a relatively simple
formal scheme from which they start out. Thus the kinetic theory of
gases seeks to reduce mechanical, thermal, and diffusional processes
to movements of molecules—i.e., to build them up out of the hypoth-
esis of molecular motion. When we say that we have succeeded in
understanding a group of natural processes, we invariably mean that
 dennis diek s

a constructive theory has been found which covers the processes in

question.
Along with this most important class of theories there exists a sec-
ond, which I will call “principle theories.” These employ the analytic,
not the synthetic, method. The elements which form their basis and
starting point are not hypothetically constructed but empirically dis-
covered ones, general characteristics of natural processes, principles
that give rise to mathematically formulated criteria which the separate
processes or the theoretical representations of them have to satisfy.
Thus the science of thermodynamics seeks by analytical means to de-
duce necessary conditions, which separate events have to satisfy, from
the universally experienced fact that perpetual motion is impossible.
The advantages of the constructive theory are completeness,
adaptability, and clearness, those of the principle theory are logical
perfection and security of the foundations.
The theory of relativity belongs to the latter class. In order to grasp
its nature, one needs first of all to become acquainted with the prin-
ciples on which it is based. (Einstein /, )

Thus, constructive theories and explanations start from the basic con-
stituents and elementary processes that build up a phenomenon, like when
we explain the pressure exerted by a gas from the collisions of the gas mol-
ecules against the walls of the container. By contrast, a principle theory begins
with postulating some general principle, suggested by experience. Properties
of individual processes are subsequently deduced from the requirement that
these processes should behave in accordance with what the general principle
stipulates.
Einstein’s introduction of this methodological dichotomy evidently served
an immediate purpose: he was facing the important task of explaining his own
theory of relativity, which he indeed had presented and developed starting from
general principles. The latter is true both for the special theory of relativity,
published in , and the general theory, published in . In particular the
special theory, on which I shall focus here, closely follows the axiomatic-deduc-
tive model. It starts with two clearly stated principles: () the relativity principle
(the physical laws have the same form in all inertial frames of reference), and
() the light principle (the speed of light is independent of the velocity of the
emitting source) (Einstein /, –). As it turns out, these two simple
and general starting points suffice for the prediction of many concrete physical
phenomena, like time dilation and length contraction. As we shall shortly see, a
striking and important point is that these predictions are possible without go-
ing into any detail about how the clocks and rods exhibiting the dilations and
under standing in ph ysic s 

contractions are built up from atoms and molecules, and without any specific
information about how these elementary constituents interact. Everything fol-
lows from very general considerations: the type of explanation is top-down.
In his  London Times article, Einstein is not very outspoken about the
relation between the two types of physical explanations that he had distin-
guished: he cites advantages on both sides. If anything, he seems to favor the
constructive approach, given his statement that “when we say that we have suc-
ceeded in understanding a group of natural processes, we invariably mean that
a constructive theory has been found which covers the processes in question.”
But given the context, namely the purpose of making the theory of relativity
understood and acceptable to a general audience, and given that Einstein him-
self had developed his theory from general principles, it is only natural that the
constructive approach is mentioned in the beginning of the  London Times
article but does not receive much further attention.
The theory of relativity soon acquired enormous fame and came to be con-
sidered as the paradigm of modern physical thinking, on a par with or even
surpassing Newton’s achievements. Given this course of events, it is under-
standable that both in circles of physicists and of philosophers of science the
idea has not infrequently taken root that at least as far as relativity theory is
concerned, top-down, principle explanations possess a privileged status. In-
deed, it is not difficult to find claims in the literature that bottom-up explana-
tions of typically relativistic effects like time dilation and length contraction are
inappropriate or even impossible (see below for a sample of such claims).
However, quite generally both types of account—bottom-up and top-
down—are viable. The typically relativistic dilations and contractions can
surely also be understood in a constructive, bottom-up fashion. This does not
deny that the top-down derivations that have become standard in the literature
possess explanatory value and lead one to understand why these effects must
exist according to relativity theory. Indeed, the choice between the different
explanatory strategies has a pragmatic character and depends on contextual
factors. There is no clear-cut and general difference between the two types of
explanation with regard to their power to generate understanding because the
notion of understanding is contextual in the same way explanation is.
As already mentioned, for concreteness I shall focus on the choice between
bottom-up and top-down explanations in special relativity. However, I think it
will become clear from the general line of argument that with respect to these
possible explanatory options there is no important relevant difference between
relativity theory and other physical theories. Moreover, bottom-up and top-
down explanations are just two examples from a wider gamut of possible expla-
nation forms (compare de Regt ). Quite generally, physics is pluralistic as
far as explanations and ways of obtaining understanding are concerned.
 dennis diek s

HeZX^VaGZaVi^k^inVhVI]Zdgnd[Eg^cX^eaZ

In the very first sentence of his “On the Electrodynamics of Moving Bod-
ies,” the  essay in which he first formulates the special theory of relativity,
Einstein (/, ; Stachel et al. , ) sets the tone for both his own
article and for most of the work on relativity that was to follow later. He draws
attention to the fact that the hitherto usual theoretical treatments of electrody-
namic phenomena contain distinctions and asymmetries that are absent from
these phenomena themselves. This rather abstract observation, about possible
“excess baggage” in the form of theories, then motivates Einstein’s first pos-
tulate, the principle of relativity. As still formulated in the introduction of his
essay, this principle tells us that the laws of electrodynamics and optics, like the
laws of mechanics, hold good in their same standard form in all inertial frames
of reference—a little bit later on in the article it becomes clear, however, that
this is taken to be valid for all laws, even for yet to be discovered nonelectrody-
namical and nonmechanical ones. To this postulate Einstein immediately adds
a second that stipulates that light is always propagated in empty space with a
definite velocity c, independent of the state of motion of the emitting body. As
Einstein promises us, still in the same introductory remarks, these two pos-
tulates will suffice for the attainment of a simple and consistent theory of the
electrodynamics of moving bodies.
Here we see the principle-theory approach in full swing: the starting point
is the formulation of very general requirements, suggested by empirical data.1
Einstein’s two starting points are very general in character, even to the point
that one wonders how they could lead to such concrete and specific predictions
as changes in the lengths of moving rods and retardations in the pace of mov-
ing clocks. But Einstein is able to honor his promise and indeed deduces these
effects from his axioms.
This remarkable feat is achieved by Einstein through his insistence that
lengths and times are not merely abstract concepts, but correspond to what is
indicated by concrete measuring instruments, rigid rods and clocks. At the end
of the introductory section of his  essay, Einstein famously declared, “The
theory to be developed is based . . . on the kinematics of the rigid body, since the
assertions of any such theory have to do with the relationships between rigid
bodies (systems of co-ordinates), clocks, and electromagnetic processes.” Thus,
coordinates are identified with notches in rigid material axes, distance is what
is measured by rigid measuring rods, and time corresponds to what is indicated
by the hands of synchronized clocks. It certainly appears at first sight that a
strong operationalist flavor emanates from these statements, a theme that was
soon picked up by logical positivists and operationalists (see Reichenbach 
and Bridgman )—although it is not so clear whether Einstein himself was
under standing in ph ysic s 

here committed to an outspoken philosophical position (Dieks ). For our

theme it is only important, however, to note that the two general postulates are
supplemented by Einstein with this concrete interpretation of spatial and tem-
poral concepts in terms of rods and clocks, and that this makes a connection
possible with concrete physical phenomena. The latter becomes evident when
we realize that the relativity postulate now says that phenomena measured in
one inertial system (also referred to as inertial frame, a system of Cartesian axes
in inertial motion) with the help of rods and clocks resting in that system, can
have an exact counterpart in any other inertial system, in which phenomena
are measured with rods and clocks co-moving with that other system.
The regularities found in the phenomena as judged from within all these dif-
ferent frames must be the same. Likewise, the light postulate comes to stipulate
that in any inertial system the velocity of light assumes the value c if the distance
covered is measured with rods resting in the frame and time is determined with
clocks resting in it. These requirements can only be fulfilled if rods and clocks
behave in a way that is different from what is assumed in prerelativistic physics.
According to Newton’s theory, good (in the sense of not deformed) measuring
rods do not change their length when they are given a velocity, and good clocks
similarly keep ticking at the same pace. But it is easy to see that this behavior
cannot be in accordance with Einstein’s two postulates. Indeed, according to
Newtonian theory, the speed of light must be c-v if measured with rods and
clocks that rest in a frame that is moving with speed v in the same direction as
the light (assuming that the speed of light is c in the original frame, with respect
to which the new system with its rods and clocks moves); and this would be in
conflict with the light postulate.
As Einstein goes on to show in his article, the requirement that the two
relativistic postulates be always exactly fulfilled uniquely fixes what has to hap-
pen to moving rods and clocks: the rods have to shrink, undergo Lorentz con-
traction, and the clocks must slow down, be affected by time dilation. This is
precisely the principle-theories procedure described by Einstein in his 
London Times article: the postulates “give rise to mathematically formulated
criteria which the separate processes or the theoretical representations of them
have to satisfy.” Whatever the physical and chemical constitution of the rods
and the clocks, if the two postulates of special relativity are to be fulfilled, mov-
ing rods have to be shorter and moving clocks have to go slower.

I]ZE]nh^XVaGZVa^ind[GZaVi^k^hi^X:[[ZXihVcYi]ZHiVijhd[
GdYhVcY8adX`h

The axiomatic-deductive way in which the theory of relativity was intro-

duced, together with the very general nature of the axioms, has from the start
 dennis diek s

given rise to questions about the nature of the relativistic effects. Even if it is
clear that the existence of contractions and dilations can be logically derived
from the postulates, does this also mean that they represent physically real
phenomena, real material changes in the rods and clocks? It has repeatedly
been suggested that this is not the case: that the changes have to do more with
the theoretical description given to the rods and clocks, or with the conditions
of observation than with things happening in the bodies themselves; that in
this sense the effects are more apparent than real. Indeed, no connection with
bottom-up accounts in terms of molecules, and forces between them, is pro-
vided so that a “mechanical” understanding of the effects is lacking in the deri-
vation from the postulates. Some authors have been misled by this, even to
the extent of claiming that the relativistic effects are more psychological than
physical in character. Among early examples are the American physicists Gil-
bert Lewis and Richard Tolman, who wrote in  that in Einstein’s theory
“the distortion of a moving body is not a physical change in the body itself,
but is a scientific fiction”; the changes in the units of space and time are “in a
certain sense psychological.” This they contrasted with older proposals by H.
A. Lorentz according to which there is “a real distortion of the body” caused
by deforming forces (Lewis and Tolman ). In the same vein, Laub ()
commented on Lorentz’s theory that in it, “observed changes in a moving body
are of an objective nature,” implying that in Einstein’s theory the situation was
different. Similarly, von Ignatowski () wrote, “measurements on a moving
body yield only an apparent value.” And the mathematician Vladimir Varičak
() commented that the relativistic “contraction is, so to speak, only a psy-
chological and not a physical fact.”
These allegations did not remain unchallenged. Already in  Paul Eh-
renfest intervened with a thought experiment that showed how the relativistic
contractions can give rise to undoubtedly physical effects, like the explosion
of a solid body. Ehrenfest’s () article invites us to consider a solid material
cylinder that is set into rotating motion around its axis. As a result of the rota-
tion, strains will arise in the material: the elements of the cylinder will tend to
undergo relativistic contraction into the direction of their rotational motion
but will remain undeformed in the directions perpendicular to their rotational
velocity. For the cylinder as a whole this leads to conflicting tendencies that
may result in the destruction of the cylinder if the velocity of rotation becomes
great enough. There can therefore be no doubt that causal processes, having to
do with the forces between the atoms and molecules, are involved in the con-
tractions and dilations.
Also Einstein himself participated in the discussion with a reply to Varičak’s
paper (Einstein ). Opposing the idea that the Lorentz contraction is subjec-
tive or an artefact of conventions of measurement, he described a method for
under standing in ph ysic s 

determining quite objectively that moving ideal rods are shorter, even without
the use of synchronized clocks. Moreover, in a letter of  to the philosopher
Joseph Petzoldt, Einstein commented on Ehrenfests’s thought experiment: “It is
well to remark that a rigid circular disk at rest must break up if it is set into rota-
tion, on account of the Lorentz contraction of the tangential fibers and the non-
contraction of the radial ones. Similarly, a rigid disk in rotation (produced by
casting) must explode as a consequence of the inverse changes in length, if one
attempts to put it at rest” (Stachel , ). In a preserved draft of a letter from
, Einstein likewise pointed out that to obtain a rotating rigid disk without
relativistic strains and tensions, one would have to first melt a disk at rest, then
put the molten disk into rotation and finally solidify it while rotating ().
Clearly then, the principle-theory approach should not lure us into the mis-
taken belief that no bottom-up, causal stories about the relativistic effects are
possible. In fact, at several points in his career Einstein in retrospect gave vent
to related reservations about his original introduction of special relativity. An
early example is in his lecture “Geometry and Experience” (Einstein /,
), delivered not long after the just-mentioned discussions about the physical
reality of the contraction effects. In “Geometry and Experience,” Einstein states,
“The idea of the measuring rod and the idea of the clock in the theory of relativ-
ity do not find their exact correspondence in the real world. It is also clear that
the solid body and the clock do not in the conceptual edifice of physics play the
part of irreducible elements, but that of composite structures, which must not
play any independent part in theoretical physics” (/, ).
This is very relevant to our topic: as we have seen, Einstein’s original deri-
vation of the contraction and dilation effects proceeded through the identifi-
cation of distances and periods with the indications given by ideal rods and
clocks, without mentioning anything about the causal processes going on in
the interior of these devices. Nothing at all was said about their atomic or mo-
lecular constitution or about the forces that keep them together, and as we
have seen this could easily create the false impression that no ordinary causal
processes are involved at all in the contractions and dilations. But in the quoted
passage, Einstein emphasizes that rods and clocks are ordinary bodies with a
microscopic structure and therefore determined in their macroscopic features
by what occurs at the microscopic level.
In the same  lecture Einstein continues: “It is my conviction that in
the present stage of development of theoretical physics these concepts [that is,
rods and clocks] must still be employed as independent concepts; for we are
still far from possessing such certain knowledge of the theoretical principles
of atomic structure as to be able to construct solid bodies and clocks theoreti-
cally from elementary concepts” (/, ). Note how the last part of
the quote resonates with Einstein’s description of the constructive method in
 dennis diek s

his  newspaper article. It is pretty clear that Einstein never thought that
general principles about rods and clocks should replace considerations about
atomic constitution and causal processes. The axiomatic approach was intro-
duced as an alternative that was particularly appropriate, given the context of
the situation Einstein was facing. This situation is that we are accustomed to
using rods and clocks for making space and time coordinates physically con-
crete, but are not able to directly describe these devices in terms of their atomic
and molecular constitution; this makes it expedient to introduce them as inde-
pendent concepts. Moreover, we are interested in questions of a very general
character, not pertaining to particular bodies or particular causal processes.
The axiomatic approach is fitting here: it is able to explain why the contractions
and dilations must be there in a quite general way, whatever the details of the
underlying causal processes.
Einstein later characterized the top-down approach as a practical decision,
made for the time being. As soon as a direct causal characterization via funda-
mental physical theory becomes available, Einstein tells us, this treatment will
have to replace the axioms-about-rods-and-clocks account. We find this state-
ment at various places in Einstein’s writings. Thus, at the end of his career he
writes in his autobiographical notes,

One is struck by the fact that the theory introduces two kinds of physi-
cal things, i.e., () measuring rods and clocks, () all other things, e.g.,
the electromagnetic field, the material point, etc. This, in a certain
sense, is inconsistent; strictly speaking measuring rods and clocks
would have to be represented as solutions of the basic equations (as
objects consisting of moving atomic configurations), not, as it were, as
theoretically self-sufficient entities. (Schilpp , )

He then immediately goes on to explain that the appeal to measuring rods and
clocks should be seen as a makeshift procedure for the time being, as long as
no complete fundamental bottom-up treatment of rods and clocks is available,
“with the obligation, however, of eliminating it at a later stage of the theory.”
In fact, Einstein seems overcautious here. He could have resorted to Her-
mann Minkowski’s work: in his famous  address Minkowski already
showed how special relativity can be built up in terms of fundamental physical
laws alone, without any appeal to rods and clocks as independent entities.

B^c`dlh`^VcYi]ZEg^cX^eaZ6eegdVX]EZg[ZXiZY

On September ,  Hermann Minkowski (/, ) delivered

his lecture Raum und Zeit.2 The lecture has become famous, and the sentence
under standing in ph ysic s 

from the introductory statement, “Henceforth space by itself and time by it-
self are doomed to fade away into mere shadows, and only a kind of union of
the two will preserve an independent reality,” has acquired proverbial status.
Still, there are aspects to Minkowski’s ideas that so far have not received the
attention they deserve (Dieks ). I here shall focus on how Minkowski
is able to avoid an appeal to rods and clocks in his presentation of special
relativity.
The impression created by Minkowski’s just-quoted winged words, namely
that space-time is an entity existing independently of matter, should be treat-
ed with caution. Reading Minkowski’s essay, rather, suggests that his position
was sympathetic to Leibnizean relationism, and that he was close to Einstein’s
sympathies—although both authors were not explicit about their philosophy
of space and time, writing as they did on problems in physics rather than phi-
losophy. How Minkowski viewed the ontological status of space-time becomes
clear if we look at his discussion of how we can lay down space-time coordi-
nates. Unlike Einstein, Minkowski did not engage in a discussion of rods and
clocks in order to define such coordinates; he suggested a very different proce-
dure. This alternative procedure has not been the subject of serious study in the
philosophy of physics literature—rather surprisingly, for Minkowski’s proposal
implements a way of constructing the edifice of special relativity in which only
general features of fundamental physical laws play a role, without the need of
invoking macroscopic measuring instruments. In a nutshell, Minkowski’s sug-
gestion is to do without any pre-given interpretation of coordinates in terms of
distances and clock settings, and to start with a theoretical account of elemen-
tary physical phenomena in terms of completely arbitrarily chosen variables.
The form of the regularities formulated in this way will most likely be extremely
complicated, but by going to new independent variables, that is, by performing
mathematical coordinate transformations, one may simplify. As the final result
of this process of successive simplifications, a preferred set of space-time co-
ordinates will be found as that “system of reference x, y, z, t, space and time, by
means of which these phenomena then present themselves in agreement with
definite laws” (Minkowski /, ).
As can be expected, the “definite laws” Minkowski explicitly refers to are
Maxwell’s equations of electrodynamics in their standard form. But later in his
article he assumes that all fundamental laws of nature, including those yet to be
discovered and responsible for the stability of matter, should exhibit the same
symmetry properties as the equations of electrodynamics. This boils down to
executing Einstein’s program of doing away with rods and clocks; in spite of
the absence of complete knowledge of all laws and in spite of a lack of de-
tailed insight into how macroscopic measuring devices should be described by
fundamental theory—as Minkowski’s elegant mathematical treatment shows,
 dennis diek s

such knowledge is not necessary to fulfill, albeit in abstract form, Einstein’s

desideratum.
More in detail, Minkowski starts by introducing coordinates x, y, z, t as
arbitrary address labels of “world-points”; the manifold of all world-points is
the world. These world-points are not meant as abstract geometrical entities:
Minkowski makes it immediately clear that the coordinates should be seen as
attributes of physical, even “observable” (wahrnehmbare) things. As he puts it,
“Not to leave a yawning void anywhere, we will imagine that everywhere and
everywhen there is something perceptible. To avoid saying ‘matter’ or ‘electric-
ity’ I will use for this something the word ‘substance’ ” (/, ). The
use of the term “observable” here may seem strange: Minkowski can hardly
be supposed to require that miniscule portions of his substance, whatever its
nature, should be accessible to the unaided senses. But we should bear in mind
that this is a science article, not an exercise in philosophy—let alone an appli-
cation of logical positivist ideas avant la lettre. Physicists are wont to use the
term “observable” as a way to denote properly physical things, with which it is
possible to enter into causal interaction. Minkowski’s wahrnehmbar should be
interpreted accordingly, as simply denoting “physical” or “material.”
After this introduction of substantial points, Minkowski focuses on the
career of one such point, for which he coins the term “worldline.” He contin-
ues, “The whole universe is seen to resolve itself into such worldlines, and I
would like to state immediately that in my opinion physical laws might find
their most perfect expression as reciprocal relations between these worldlines”
(/, ). Minkowski’s idea that the laws of physics represent relations
between physical worldlines plays a central role in his subsequent argument.
Starting from the laws of physics, which thus express relations between world-
lines and, consequently, regularities in the behavior of material systems,
Minkowski starts his space-time analysis: as already explained, the usual iner-
tial coordinates x, y, z, t are determined as those coordinates in terms of which
the laws take on their standard forms.
Thus, we have found a way to present special relativity that does not de-
pend on the existence of rigid rods and clocks. The theory is now given a more
abstract rendering, in which only the existence of fundamental physical laws is
assumed. Einstein’s relativity postulate now becomes the requirement that the
form of these laws be invariant under a certain group of transformations (the
Lorentz transformations). The operations in this group can be interpreted as
transitions from one inertial frame of reference to another—the spirit of the
invariance requirement is therefore not different from that of Einstein’s origi-
nal relativity postulate. With Minkowski’s approach, however, special relativity
attains a more fundamental form than in Einstein’s own presentation, because
we no longer need rods, clocks, and materially realized inertial systems; in
under standing in ph ysic s 

Minkowski’s presentation special relativity deserves the predicate “principle

theory” more than ever. The overarching principle now states the invariance
of all laws of nature under Lorentz transformations, and Einstein’s derivations
can be repeated with greater generality in this new context. Again, the typically
relativistic effects can be derived without going into the specific details of the
material processes that are involved—the invariance properties of the laws suf-
fice for reaching the desired conclusions. However, now that everything derives
from the fundamental laws that govern the behavior of the elementary build-
ing blocks of physical objects and processes, the possibility of thinking about
bottom-up, causal accounts of these relativistic phenomena also starts to force
itself upon us, and the question of the relation between bottom-up and top-
down becomes more urgent.

HeZX^VaGZaVi^k^inVhV8dchigjXi^kZI]Zdgn

Typically relativistic effects like length contraction and time dilation can
give rise to phenomena that one at least is inclined to describe in terms of mi-
croscopic causal processes. In the case of the rotating and eventually exploding
cylinder, for example, it is natural to ask about the forces that are responsible,
the amounts of energy that are involved, and so on. Likewise, one may won-
der about the detailed mechanisms that make rods shrink when a velocity is
imparted to them. Also in this latter case effects may arise like the breaking of
objects, which seems to cry out for causal explanation. A notorious example is
furnished by John Bell’s (, ) two rockets problem in which two rockets
undergo equal accelerations in such a way that their mutual distance remains
the same. A rope connecting the rockets will tend to Lorentz contract during
the process, because it acquires a velocity. As a consequence, the rope will be-
come tighter and may eventually even break when the velocity becomes high
enough.
As Bell recounts, he experienced that many physicists are amazed when
confronted with this outcome of the thought experiment. Bell attributes their
confusion to the top-down manner in which the theory of relativity is usu-
ally taught, and pleads for a different approach in which the focus is shifted
from top-down to bottom-up, constructive and causal accounts of relativistic
phenomena. I will comment on the relation between these two approaches in
the next section; here, I merely want to emphasize that such a constructive ap-
proach is indeed possible.
I mentioned above that Einstein already had warned against the idea that
macroscopic bodies like rods and clocks are not susceptible to bottom-up anal-
ysis: these bodies evidently consist of atoms and molecules, so their behavior
should be determined by the laws that govern these submicroscopic constitu-
 dennis diek s

ents and their interactions. I mentioned in passing that before the advent of
relativity theory, H. A. Lorentz had already developed a theoretical scheme
where he tried to explain that moving bodies contract.3
Lorentz’s (/, ) approach focused on the forces that hold mac-
roscopic bodies, like measuring rods, together. These forces change when the
bodies are set in motion, and this has the effect of changing the mutual dis-
tances between the atoms and molecules. Lorentz proved that this leads to the
right macroscopic contraction in the direction of motion (the same contraction
we encounter in relativity theory) if it is assumed that under motion all possible
forces change in the same way as the electromagnetic ones. This latter assump-
tion anticipates Minkowski’s requirement that all laws should be invariant un-
der Lorentz transformations, but still, the type of explanation here is clearly
causal and bottom-up. The account works by determining the equilibrium
positions of the atoms and molecules in the interatomic and intermolecular
force fields, and by finding out how these equilibria are altered when the forces
change due to motion.
Exactly the same explanations can be given in relativity theory. If we know
what the fundamental building blocks of matter are, and know how the forces
between them vary when a system is set into motion, it becomes possible in
principle to calculate how the macroscopic features of these bodies will change
when they start moving. Actually, it seems pretty obvious that bottom-up ex-
planations of this type are possible for everything that happens according to
relativity theory.

I]ZGZaVi^dcWZilZZc7diidb"JeVcYIde"9dlc

Minkowski’s approach makes it easy to confront and compare our two ex-
planatory strategies. In Minkowski’s presentation of special relativity the basic
ingredients are matter and fundamental laws governing matter, and although
Minkowski’s own approach is top-down, these ingredients also set the stage for
a Lorentz-like bottom-up account in terms of atoms, molecules, and the forces
between them. Minkowski’s scheme further contains the general principle that
all laws are invariant under Lorentz transformations. The way Minkowski in-
troduces his scheme—starting from concrete phenomena, investigating their
regularities and thus arriving at laws, and finally inquiring about the invari-
ance properties of these laws—might create the impression that this principle
cannot be an independent component of the theory: once the laws have been
found, whether or not they are Lorentz invariant is a matter of inspection and
does not need independent stipulation. However, the idea obviously is that it is
part of relativity theory to suppose that new laws that may be found as a result
of future research will possess these same invariance properties. Interpreted
under standing in ph ysic s 

in this way the principle can be considered an independent part of relativity

theory, on whose authority we are entitled to assume certain general proper-
ties of all physical processes, even those whose specific features we do not yet
know. Now, for the logic of the relation between top-down and bottom-up it is
a central point that these general features, shared by all physical laws, are suf-
ficient to show that all accounts, either bottom-up or top-down, will lead to the
same essential aspects of the typically relativistic effects like length contraction
and time dilation.
The way this works may be sketched as follows. Take the example of length
contraction. Suppose we have a macroscopic body at rest in some inertial
frame of reference. Whatever its microscopic constitution and whatever the
forces that hold it together, it must be the case that the microscopic processes
that are responsible for the macroscopic features of the body are governed by
Lorentz invariant laws. There will exist some kind of equilibrium configuration
(from a macroscopic viewpoint) of the atoms and molecules, and this configu-
ration determines the length of the body. From the Lorentz invariance of the
laws responsible for the equilibrium we now deduce that a body with the same
composition but resting in another inertial frame, therefore one that moves
with velocity v with respect to our original frame, possesses a corresponding
equilibrium state. This second equilibrium state has a description in terms of
the internal Cartesian coordinates of the second inertial frame that is exactly
the same as the description given of the first equilibrium in terms of the coor-
dinates of the first frame. But we know the relations between the coordinates
of the two frames: these are given by the Lorentz transformations. This implies
that the equilibrium length of the second, moving, body is related to the length
of the first body by the Lorentz contraction. For example, if the first body ex-
actly fills the space between the coordinates x =  and x = , so that it has length
, its moving mirror image will exactly fit between xψ =  and xψ = , in which
the primed coordinates belong to the moving frame. The Lorentz transforma-
tion formulas now tell us that this coordinate difference  between the primed
coordinates corresponds to a coordinate difference

 (1–v²
c²)

between the unprimed coordinates (assuming that the velocity v is in the x di-
rection). So as judged from the first frame the length of the moving body is less
than : it has the Lorentz contracted value.
As will be clear from this sketch, the relativity principle does all the work
in securing that the right contraction and dilation factors will result. The mi-
 dennis diek s

croscopic details can be left out: by omitting the references to submicroscopic

constituents and forces, and only mentioning the macroscopic body itself, we
would recover Einstein’s original approach, the archetypical top-down type of
account of a relativistic effect. Still, the just-sketched derivation itself is essen-
tially identical to the one proposed by Lorentz. Lorentz started from the mi-
croscopic building blocks of bodies, analyzed their equilibrium positions, and
went on to prove that if there is one such equilibrium in a resting body, there
must be corresponding equilibria in moving bodies such that the resulting mac-
roscopic lengths of these moving bodies relate to the rest length via the Lorentz
contraction. In this proof he had to assume that all laws change in the same way
in the transition from rest to motion as the laws of electrodynamics—which is
equivalent to assuming that all laws are Lorentz invariant. So it turns out that
both Lorentz’s constructive approach and Einstein’s/Minkowski’s principle ap-
proach use exactly the same theoretical ingredients. To some extent, this puts
the distinction between top-down and bottom-up into perspective: there is no
difference in the relevant theoretical machinery.
Is there any remaining difference, then, between the two approaches? There
certainly is, but the relation between the two ways of explaining relativistic
effects is not one of mutual exclusion or of right versus wrong (compare de
Regt ). Both derivations use the same physical theory, and both are clearly
equally valid as far as logical deduction is concerned. The difference between
the two arguments is in the emphasis on one premise instead of another, and
in where the deduction starts. These differences correspond to differences in
exactly what one wishes to know, and from which point of view; in other words,
the difference is pragmatic: it relates to our desires and interests and not to an
objective context-independent superiority of one approach over the other.
Indeed, suppose we are considering a problem like that of Ehrenfest’s para-
dox or Bell’s rockets, and we ask ourselves at which point during the accelera-
tion process, and where, something will break. Or, more generally, think of the
question of whether and when a rod that is set in motion by an accelerating
force will shrink. These are questions about the dynamics of the contraction
process, and in this context of dynamical questions a dynamical bottom-up
approach is obviously the right one. For example, to know whether an acceler-
ated rod will become Lorentz-contracted we have to investigate whether it will
be able to reach a state of internal equilibrium, and this depends on the in-
terplay of internal and external forces and on thermodynamic considerations.
Knowledge of these relevant causal factors will make it possible to determine
the details of the deformation process; moreover, the resulting causal bottom-
up picture provides us with a clear insight into what happens and in this way
generates understanding. It was especially the latter point that motivated John
Bell () to call for a drastic revision of the way relativity is taught: namely, to
under standing in ph ysic s 

abandon the usual top-down approach initiated by Einstein and to revert to a

Lorentz-like procedure in which one starts with atoms and forces. Bell thought
that this bottom-up method was the way par excellence to obtain understand-
ing of the relativistic effects (actually, Einstein’s own  remarks about the
constructive method go in the same direction).
However, we have already seen that the top-down approach uses exactly
the same theoretical elements, although with different emphasis. This by itself
already casts doubt on the notion that no explanation and understanding could
be achieved by the top-down method. Indeed, the more than one hundred years
of history of special relativity have abundantly demonstrated that in certain
contexts physicists accept the original Einstein-like derivations of the relativ-
istic effects as real explanations and ways of acquiring genuine insight and un-
derstanding. The contexts in question are different from the dynamical ones we
mentioned a moment ago: they pertain to the comparison of different inertial
frames of reference, and bodies resting in them, without asking anything about
what happens when a body is transferred from a state of rest in one frame to a
state of rest in another. So the comparison is between bodies of which it is given
that they are moving uniformly with respect to each other in completely similar
internal states—a “stationary” situation, to be contrasted with the earlier men-
tioned “dynamic” cases in which accelerations played a role. In the stationary
case, it is illuminating (as we have already seen in some detail) to concentrate
on the general invariance properties of the laws of physics. Lorentz invariance
explains why rods that are each other’s exact counterparts in different frames
possess lengths that are related via a Lorentz transformation. The picture that
arises here has a simplicity and compactness that makes things understand-
able.4 The Lorentz contraction is here represented as the natural relation be-
tween moving bodies, and as reflecting general space-time symmetries. This
is what has become accepted as the “orthodox” understanding of relativistic
effects, against which Bell fulminated but which physical practice has demon-
strated to be perfectly defensible in itself.

8dcXajh^dc

Since its introduction, the special theory of relativity has usually been re-
garded as a typical principle theory in which the direction of explanation is
top-down. In accordance with this, it is not uncommon to find statements in
the literature to the effect that attempts at bottom-up, causal explanations are
out of place in relativity. However, there have certainly been many physicists,
among them the old masters Einstein, Minkowski, Ehrenfest, and their like,
who never fell prey to this misunderstanding. The possibility of bottom-up
strategies in relativity, which I have stressed here, is therefore not something
 dennis diek s

new, but it is still worth stressing in view of the continuing dominance of the
top-down approach both in the teaching of relativity and in philosophical ac-
counts of the theory. Bell () was right in resisting this dominant attitude by
drawing attention to the possibility of bottom-up accounts. From my point of
view, however, “he struck a spark but threw no light”: he overshot the mark by
suggesting that the bottom-up approach is the understanding-providing one.
The issue has been receiving growing attention in the philosophy of physics
literature, especially after the publication of Harvey Brown’s () book Physi-
cal Relativity, which defends a position very similar to that of Bell. In the wake
of Bell’s and Brown’s publications, a serious dispute has arisen among philoso-
phers of physics about the correct way of explaining relativistic effects. On the
one hand there are those who defend the orthodox viewpoint: these argue that
the top-down approach is appropriate since it alone fully captures and reflects
the general space-time structure of the world according to special relativity
(see, for example, Janssen , ; Balashov and Janssen ). On the other
hand, there are those who maintain that this top-down approach ignores the
physical mechanisms that are at work (for example, Brown and Pooley ,
; Brown ), and therefore is unable to provide us with genuine explana-
tions and understanding.
What I have argued here is that this dispute derives from a misconception.
There is no uniquely best way of explaining the relativistic effects. The differ-
ences between the explanations are differences in the use we make of one and
the same theoretical scheme. They relate to decisions about where to put the
emphasis and where to begin and end the analysis; these decisions in turn are
conditioned by what we are interested in. In other words, the difference is one
of pragmatics. Explanation and understanding are relative to questions we ask
and interests we have. Explanation and understanding, not only in relativity
theory but also in theoretical physics and even in theoretical science in general,
are inherently pluralistic.

CdiZh
. That no difference between inertial frames can be detected as far as the form of
applicable theoretical principles is concerned, and that the velocity of light is always the
same, had been suggested by several nineteenth-century experiments performed precisely
with the purpose of finding empirical counterparts to the theoretical asymmetries men-
tioned by Einstein in the first sentence of his article. It was the failure of these experiments
that provided positive empirical input for Einstein’s postulates.
. Many of the ideas can already be found in a lecture Minkowski gave a year earlier
(Minkowski ).
. Even before special relativity, the contraction idea had gained currency as a way of
accommodating the famous null experiments, such as the Michelson-Morley experiment.
under standing in ph ysic s 

. See de Regt and Dieks () for more on the relevance for understanding of the
intuitive accessibility of such theoretical pictures.

GZ[ZgZcXZh
Balashov, Y., and M. Janssen. . Presentism and relativity. British Journal for the Phi-
losophy of Science :–.
Bell, J. S. . How to teach special relativity. Speakable and unspeakable in quantum
mechanics, –. Cambridge: Cambridge University Press.
Bridgman, P. W. . Einstein’s theories and the operational point of view. In Albert Ein-
stein: Philosopher-scientist, edited by P. A. Schilpp, –. LaSalle: Open Court.
Brown, H. R. . Physical relativity: Space-time structure from a dynamical perspective.
Oxford: Oxford University Press.
Brown, H. R., and O. Pooley. . The origins of the spacetime metric: Bell’s Lorentzian
pedagogy and its significance in general relativity. In Physics meets philosophy at the
Planck scale, edited by C. Callender and N. Huggett, –. Cambridge: Cambridge
University Press.
———. . Minkowski space-time: A glorious nonentity. In The ontology of spacetime,
edited by D. Dieks, –. Amsterdam: Elsevier.
de Regt, H. W. . Wesley Salmon’s complementarity thesis: Causalism and unification-
ism reconciled? International Studies in the Philosophy of Science :–.
de Regt, H. W., and D. Dieks. . A contextual approach to scientific understanding.
Synthese :–.
Dieks, D. . The adolescence of relativity: Einstein, Minkowski, and the philosophy of
space and time. Forthcoming in Minkowski spacetime: A hundred years later, edited by
V. Petkov. New York: Springer.
Ehrenfest, P. . Gleichförmige Rotation starrer Körper und Relativitätstheorie. Physika-
lische Zeitschrift :.
Einstein, A. /. On the electrodynamics of moving bodies. In The principle of
relativity, by H. A. Lorentz, A. Einstein, H. Minkowski, and H. Weyl, –. London:
Methuen.
———. . Zum Ehrenfestschen Paradoxon. Physikalische Zeitschrift :–.
———. /. My theory. In Ideas and opinions, –. New York: Crown.
———. /. Geometry and experience. In Ideas and opinions, –. New York:
Crown.
Janssen, M. . Reconsidering a scientific revolution: The case of Lorentz versus Ein-
stein. Physics in perspective :–.
———. . Drawing the line between kinematics and dynamics in special relativity. Stud-
ies in History and Philosophy of Modern Physics :–.
Laub, J. . Über die experimentellen Grundlagen des Relativitätsprinzips. Jahrbuch der
Radioaktivität und Elektronik :–.
Lewis, G. N., and R. C. Tolman. . The principle of relativity and non-Newtonian me-
chanics. Philosophical Magazine :–.
Lorentz, H. A. /. Electromagnetic phenomena in a system moving with any
 dennis diek s

velocity smaller than that of light. In The principle of relativity, by H. A. Lorentz , A.

Einstein, H. Minkowski, and H. Weyl, –. London: Methuen.
Minkowski, H. /. Space and time. In The principle of relativity, by H. A. Lorentz,
A. Einstein, H. Minkowski, and H. Weyl, –. London: Methuen.
———. . Das Relativitätsprinzip, Vortrag gehalten  November . Jahresbericht der
Deutschen Mathematiker-Vereinigung :–.
Reichenbach, H. . The philosophical significance of the theory of relativity. In Albert
Einstein: Philosopher-scientist, edited by P. A. Schilpp, –. LaSalle: Open Court.
Schilpp, P. A., ed. . Albert Einstein: Philosopher-scientist. LaSalle: Open Court.
Stachel, J. . Einstein and the rigidly rotating disk. In General relativity and gravitation:
One hundred years after the birth of Albert Einstein, vol. , edited by A. Held, –.
New York: Plenum.
Stachel, J., D. C. Cassidy, J. Renn, and R. Schulmann, eds. . The collected papers of
Albert Einstein. Vol. . Princeton: Princeton University Press.
Varičak, V. . Zum Ehrenfestschen Paradoxon. Physikalische Zeitschrift :–.
von Ignatowski, W. . Der starre Körper und das Relativitätsprinzip. Annalen der
Physik :–.
chap ter title ver so

(*
Understanding in the Engineering Sciences
Interpretative Structures

B>:@:7DDC

 My account of scientific understanding focuses on scientific prac-

tices, especially the intellectual activities and abilities of scientists. I will use
engineering sciences—which I consider laboratory sciences (compare Hacking
)—as a case for illustrating how scientists gain scientific understanding of
phenomena, and how they exercise their understanding of scientific theories.
Although my account was developed through the perspective of the engineer-
ing sciences, it is appropriate to the “basic” laboratory sciences as well.

I]Z:c\^cZZg^c\HX^ZcXZh

Engineering science is scientific research in the context of technology.

The engineering sciences strive to explain, predict, or optimize the behavior
of devices or the properties of diverse materials, whether actual or possible.
Devices can perform different tasks, for instance, manufacture relevant physi-
cal phenomena (such as X-rays, superconductivity, catalytic reactions, and
sono-luminescence); measure relevant physical variables (such as tempera-
ture, morphology, mass, electrical conductivity, chemical concentration, and
wave-length); and produce materials or energy (see Boon ). Likewise,
technologically relevant materials exhibit specific properties. Examples are su-
perconductive materials, piezoelectric materials, catalysts, membranes, dyes,
medicines, and fertilizers.


 mieke b o o n

Scientific research involves conceiving of how the device functions in terms

of particular physical phenomena that affect a proper or improper functioning.
Similarly, scientists conceive of different materials in terms of specific material
properties that produce the proper or improper material functions. The engi-
neering sciences aim thus at developing scientific understanding of physical
phenomena and material properties in terms of relevant physical mechanisms
and/or mathematical representations that allow scientists and engineers alike
to reason how to create, produce, improve, control, or design various devices
and materials.
Theories and models published in scientific articles represent scientific
understanding of phenomena that determine the functioning of devices and
materials. Hence, engineering science and “basic” laboratory sciences are sim-
ilar in their concern with understanding phenomena. In this chapter, the term
phenomena will stand for physical phenomena and material properties, and
the term theories will denote theories and models (see Boon and Knuuttila
).

JcYZghiVcY^c\E]ZcdbZcV

De Regt and Dieks () assume that scientific understanding is one of

the key objectives of science. This idea seems trivial, but it is not. In most phi-
losophy of science, “explaining a phenomenon” is taken to be synonymous with
“understanding a phenomenon.” Insofar as the notion of understanding aims to
add something extra to explaining, most authors reject it as a notion of philo-
sophical interest because they assume that it is a mere psychological surplus of
explaining. J. D. Trout (), for instance, argues that the sense of understand-
ing often leads scientists astray when it acts as a cue for adopting an expla-
nation as a good or correct one. Understanding has no epistemological value
because what makes an explanation good (that is, epistemically reliable) is not
determined by this sense, but instead “concerns a property that the explanation
has independent of the psychology of the explainers; it concerns features of
external objects, independent of particular minds” ().
The tradition within which Trout rejects the importance of scientific under-
standing presupposes that a philosophical account of science can be explicated
in terms of a two-placed relation between world and knowledge. In this view,
true theories represent the world that is responsible for the occurrence of phe-
nomena, and are independent of scientists’ intellectual activities and abilities in
a scientific practice. Trout’s argument also stands in a tradition that takes true
theories as the ultimate aim of science. An important drawback of the latter
presupposition is that it does not account for the results of current laboratory
sciences. Scientific articles in the laboratory sciences present representations
u nd e r stan d i n g i n th e e ngi ne e ri ng s c ience s 

of new phenomena, and/or theories that represent scientific explanations of

phenomena rather than scientific theories, and how a theory was justified as
suggested by traditional philosophical accounts of science. This implies that
phenomena are at the center of interest in the laboratory sciences, instead of
being mere tools for justifying theories. Thus, since laboratory sciences cer-
tainly are forms of science, we must reconsider the nature of science in light
of this wider range of activities and purposes. Laboratory science is an ongo-
ing practice in which scientists aim at discerning, creating, and understanding
phenomena; scientists use and produce scientific knowledge and understand-
ing, which is represented in theories.1
Consequently, scientific practice must be explicated in terms of a tripartite
relation that takes into account the intellectual activities and abilities of scien-
tists, who use scientific knowledge when () discerning phenomena that are
relevant (such as to a technological application), () predicting how phenom-
ena can be created or manipulated—for instance, by predicting physical condi-
tions at which a desired phenomenon will be produced, and () developing a
scientific understanding of these phenomena, thus producing new scientific
theories. On this account, theories do not represent the world. Instead, theories
represent scientists’ scientific understanding of the world. Additionally, the in-
tellectual activities of scientists who are doing steps one through three involve
the ability to use theories, which means that they understand theories.2 Hence,
scientific understanding is not a mere psychological surplus of a scientific ex-
planation; it is essential to how science gets done. In explicating the triadic
relation between scientists-world-knowledge, this notion is crucial to account
for the nature of theories and the intellectual activities and abilities of scientists
in developing these theories.
My two proposals—that theories represent scientists’ understanding of
phenomena, and that scientists understand the theories and models they use
in developing other theories and models—asks for further explication of the
notion of understanding. This involves questions such as: What is scientific
understanding of a phenomenon? What is it about understanding a theory that
allows for using it?
De Regt and Dieks () have proposed criteria for understanding phe-
nomena and theories. I will analyze their account and demonstrate where it
leads to difficulties. Then I will explain my claim that theories of phenomenon
P represent scientists’ scientific understanding of P. My basic idea is that theo-
ries result from scientists’ intellectual activity of structuring and interpreting
phenomena, thereby producing interpretative structures. I propose interpre-
tative structure of P as an alternative to a theory of P. I will spell out why and
how understanding a theory allows for using it in structuring and interpreting
phenomena. Finally, I will argue that different types of structuring and inter-
 mieke b o o n

preting must be distinguished that cannot be reduced to one another, and play
complementary roles in the development of science.

8g^iZg^V[dgJcYZghiVcY^c\

De Regt and Dieks () have proposed a method for explicating the role
of understanding by articulating criteria for understanding—thereby avoiding
a merely psychological interpretation. Their basic assumption is that under-
standing is acquired by scientific explanation, and their two criteria, CUP and
CIT, aim to provide a general account of what that means. CUP is their crite-
rion for understanding phenomena: “A phenomenon P can be understood if a
theory T of P exists that is intelligible” (, ) which is related to their cri-
terion for the intelligibility of a theory (CIT): “A scientific theory is intelligible
for scientists (in context C) if they can recognize qualitatively characteristic
consequences of T without performing exact calculations” (, ). In these
formulations, de Regt and Dieks are not explicit about what they mean by “the-
ory.” I will assume that their use of “a theory of P” means “a scientific explana-
tion of P,” which by my account is synonymous with “a scientific model of P.”
De Regt and Dieks have several objectives. First, by articulating CUP and
CIT, they aim to provide a more precise interpretation of the nature of skills in-
volved in achieving scientific understanding. In further explicating these skills,
they put emphasis on the qualitative and intuitive dimensions of understand-
ing. For instance, “theories may become intelligible when we get accustomed to
them and develop the ability to use them in an intuitive way” (, ). From
the perspective of the engineering sciences, their emphasis on intuitive aspects
is problematic for, in many cases, scientists do understand theories in terms
of how and where to use them, whereas they often do not have an idea of their
qualitative consequences, that is, the outcomes of explanations or predictions
that make use of the theory. This is particularly the case when scientists, in de-
veloping an explanation of a complex phenomenon, PC, integrate their scientific
understanding of physical phenomena P, . . . , Pi for which respective scientific
explanations T, . . . , Ti are available.
One example stems from the way scientists have developed an explana-
tion of the phenomenon of sono-luminescence, which is the emission of a light
pulse from imploding bubbles in a liquid when excited by sound. Brenner and
colleagues (, ) state that “an enormous variety of physical processes
is taking place inside this simple experiment [that produces the phenomenon
of sono-luminescence], ranging from fluid dynamics, to acoustics, to heat and
mass transfer, to chemical reactions, and finally to the light emission itself.”
What scientists usually do in generating predictions and explanations is to in-
terpret a complex phenomenon PC (for example, the emission of a light pulse
u nd e r stan d i n g i n th e e ngi ne e ri ng s c ience s 

from an imploding gas bubble), in terms of mutually interacting, physical phe-

nomena P, . . . , Pi. CUP and CIT are too limited to account for how an explana-
tion of a phenomenon PC is developed, since these formulations only express
that scientific understanding of a phenomenon P allows grasping how (in con-
crete circumstances) P follows from the explanation of P. That account of un-
derstanding P, . . . , Pi does not explicate how scientists use this understanding
in developing an explanation of PC; nor does it explicate how the intelligibility
of scientific explanations of P, . . . , Pi does the work in a scientist’s understand-
ing of PC.
Therefore, when suggesting that understanding the world in terms of scien-
tific theories is important for generating predictions and explanations, de Regt
and Dieks (, ) seem to have something different in mind than I have.
Their focus is on criteria that determine whether scientists understand phe-
nomena, whereas I wish to put emphasis on the ability to use scientific explana-
tions T, . . . , Ti of phenomena P, . . . , Pi in developing scientific explanations
and predictions of other phenomena PC. In brief, our difference seems to exist
in our point of departure: whereas de Regt and Dieks assume that explanations
provide understanding, my assumption is that using an explanation requires
understanding of that explanation.
A second objective of de Regt and Dieks () is to accommodate the his-
torical diversity of conceptions of understanding in actual scientific practices,
and hence to do justice to the fact that, in the history of science, scientists have
had different standards for understanding. They argue that different types of
scientific explanation have understanding as an aim, which implies that un-
derstanding can be achieved through different types of scientific explanation.
For instance, causal models and mechanisms lead to scientific understanding
because they allow “scientists to grasp how the predictions of a theory come
about, and to develop a feeling for the consequences the theory has in a par-
ticular situation” (de Regt , ). Which theories conform to the general
criterion for the intelligibility of scientific theories (CIT) depends on the his-
torical context (). In this way, they also reconcile conflicting views of ex-
planatory understanding, such as the causal-mechanical and the unificationist
conception, which is their third objective.
However, the assumption that the intelligibility of a certain type of explana-
tion is a matter of preference determined by the historical context generates a
problem. As I just argued, constructing an explanation of a complex phenom-
enon often involves different types of theories simultaneously. Constructing
an explanation of sono-luminescence, for instance, involves thermodynamics,
which is a theory that typically fits in the unification view, as well as theories
on chemical reactions, which usually have the character of causal-mechanistic
explanations. Accordingly, developing an explanation of sono-luminescence
 mieke b o o n

involves understanding at least two different types of theories. This also results
in two types of explanatory models of PC, which are a causal-mechanistic and
a nomo-mathematical model of PC (compare Boon ). The first model-type
allows for causal-mechanistic reasoning, whereas the second allows for nomo-
mathematical reasoning about the phenomenon PC. Although the two types
of models are related and explain different aspects of the phenomenon, they
cannot be reduced to each other. Therefore, in current scientific practices, dif-
ferent types of theories are combined in order to develop explanations and pre-
dictions of phenomena. Usually this is not a matter of preference, but related
to the intended uses of explanations and predictions of the phenomenon—for
example, models that predict how to create the phenomenon, models for cal-
culating optimal conditions for the phenomenon to occur, models used in the
design of processes that manufacture materials, and models that represent how
new properties of materials are generated.
Hence, de Regt and Dieks’s () two criteria of understanding cannot suf-
ficiently account for the engineering sciences. In particular, their proposals that
understanding is acquired by scientific explanation, and that understanding a
theory consists in recognizing qualitatively characteristic consequences of T,
are unsatisfactory as an answer to the question, “What is it about understand-
ing a theory that allows for using it?” One problem with their account is that
the proposed criteria do not explicate why and how an explanation provides
understanding, for example, how it is possible that a scientist recognizes the
qualitative consequences of a scientific explanation in concrete conditions, and
why does understanding of a scientific explanation allow for generating new
explanations and predictions. In brief, it does not sufficiently explicate how
scientific understanding does the work it is supposed to do in the intellectual
activities and abilities of scientists.
My account of scientific understanding expands on de Regt and Dieks
() and aims to overcome the mentioned difficulties. My leading questions
are: “What is it about an explanation that allows for understanding it?” and
“What is it about understanding an explanation that allows for using it?”

>ciZgegZiVi^kZHigjXijgZh

I propose that a scientific explanation of Pi is intelligible for scientists in

context C if it allows for reasoning about questions and problems relevant to C.
This is a refinement of CIT. Reasoning about questions and problems involves
using (parts of ) causal-mechanistic models and/or (parts of ) nomo-math-
ematical models in C. Such reasoning takes place while developing theoretical
interpretations of observed phenomena, when making calculations of experi-
mental conditions for determining values of relevant physical parameters, or
u nd e r stan d i n g i n th e e ngi ne e ri ng s c ience s 

when making theoretical predictions about possible refinements of explana-

tory models and/or experimental set-ups for achieving a better fit between
models and measurements. As a result, the intelligibility of a theory does not
primarily depend on intellectual preferences, or on recognizing qualitatively
characteristic consequences of T; it sooner depends on whether scientists can
use it in their reasoning with the theories of C in hand about questions and
problems relevant to C.
How is it possible that scientists can use scientific explanations (such as
theories and models) in their reasoning about questions and problems? My
core claim is that producing intelligible explanations consists of structuring and
interpreting phenomena in terms of relations between objects, thus producing
interpretative structures.3 These interpretative structures must be empirically
adequate (compare van Fraassen ), internally consistent, and coherent
with other relevant interpretative structures.
The point of replacing the notion of scientific explanations with the notion
of interpretative structures incorporates an epistemological presupposition,
which is that our knowledge does not represent pictures of the world (that is,
“state of affairs” or “matters of fact”), but rather how we have structured and
interpreted the world, that is, how we have drawn relations. We draw relations
when we carve the world up into phenomena, and when we “observe” relations
between them. This holds for knowledge of the “observable” world represented
in judgments, and for knowledge of the “unobservable” world represented in
interpretative structures.4 The epistemological issue at stake is that we cannot
“observe” structures and relations. In cognitive activities, we can draw differ-
ent kinds of relations, such as logical, mathematical, statistical, and causal. Our
choices depend on our cognitive aims.
Yet we can articulate how it is possible that scientists use interpretative
structures (that is, theories or models) in their scientific reasoning. Qualita-
tive recognition of characteristic consequences of T (see CIT) is too weak an
account. Instead, using interpretative structures is possible because scientists
make use of the relations represented by these structures for drawing infer-
ences. This leads me to a rephrasing of de Regt and Dieks’s criteria, yielding
CUPa: “A phenomenon P can be understood if an interpretative structure IS
of P exists that is intelligible (and meets the usual logical, methodological and
empirical requirements)”; and CITa “An interpretative structure is intelligible
for scientists if they can use IS in developing explanations and predictions at
new circumstances.”
CITa must be adapted somewhat further, because developing scientific ex-
planations and predictions of concrete phenomena involves the ability to use
the fundamental theories, interpretative structures, and theoretical methods
of C—which I will call the scientific field. This implies that scientists must have
 mieke b o o n

the ability to know where and how to apply the theories of a scientific field, and
how to infer consequences from them. I have illustrated this with the explana-
tion of sono-luminescence above, which involves the ability to recognize that
certain general theories (such as thermodynamics, fluid dynamics, acoustics,
theories of heat and mass transfer, and theories of chemical reactions) may
apply to this phenomenon and how to apply these theories to the case at hand.
Hence, criterion CITa must account for the role of scientists’s understanding
of a scientific field within which the phenomenon is relevant. This results in
a more general criterion for understanding scientific fields (CUSF): scientists
understand a scientific field if they have acquired the ability to use scientific
knowledge of the field (for example, fundamental and general theories, and
interpretative structures of relevant phenomena) in developing explanations
and predictions of other phenomena that are relevant to the field (including the
ability to meet the methodological criteria of the field).
These new criteria CUPa and CUSF, as alternatives to CUP and CIT, still
do not explicate why scientists understand interpretative structures, and what
it is about an interpretative structure that allows for using it in generating ex-
planations and predictions in new situations. I propose that scientists are able
to understand an explanation of P if this explanation presents an interpretative
structure of P. This implies that scientists understand theories and explanations
because theories and explanations are interpretative structures. But why is that?
My brief answer is, because interpretative structures represent how scientists
conceive of objects and relations among these objects. In general, conceiving of
relations among objects allows for making inferences. Hence, the specific char-
acter of explanations is that explanations are interpretative structures that rep-
resent how scientists conceive of relations between objects, which allows for
using explanations, for example, in making inferences under new conditions.
The idea that understanding T consists of scientists’ ability to use T in their
reasoning—where such reasoning is possible because interpretative structures
represent relations from which scientists can draw inferences—also avoids the
rather intuitive aspect of CIT, which is that “scientists recognize qualitative
consequences of T.”

I]Z6eegZci^XZ8VW^cZibV`Zg/68VhZd[<ZdbZig^XVaHigjXijg^c\

HigjXijg^c\VcY>ciZgegZi^c\Å>ciZgegZiVi^kZHigjXijgZhÅ
>ciZgegZiVi^kZ;gVbZldg`h
Structuring and interpreting can be done in terms of different interpretative
frameworks. These frameworks determine the type of phenomena that scien-
u nd e r stan d i n g i n th e e ngi ne e ri ng s c ience s 

tists discern and the type of relations they “see,” for instance, in terms of logical,
geometrical, nomo-mathematical, causal-mechanistic, or statistical relations.
I will present a case that aims to show how the development of an interpre-
tative structure of a phenomenon works, and why an interpretative structure
allows for understanding a phenomenon. I also will clarify how understanding
the field (for example, Euclidean geometry)—that is, the ability to use the fun-
damental theories of the field articulated in CUSF—is involved when develop-
ing an interpretative structure that explains the phenomenon. Additionally, the
case of the apprentice cabinetmaker aims to illustrate that mathematics can
do the explanatory work, that is, that a mathematical explanation can provide
understanding. This goes against those who believe that only causal or causal-
mechanistic explanations are satisfactory in that sense.
An apprentice cabinetmaker has done his best to make a door that perfectly
fits into the opening of a cupboard, but when he hangs it on its hinges, he finds
that the door does not fit. He does not understand what has gone wrong since
he is absolutely certain that he has made the correct measurements. The master
cabinetmaker explains the observed phenomenon of this nonfitting door, PO,
as follows. First, the master cabinetmaker structures and interprets PO with-
in the framework of geometry. She abstracts from properties such as weight,
color, or properties of the wood, and focuses on proportions and dimensions
of PO, which she projects onto a two-dimensional geometrical space. She also
abstracts from the concrete values of the dimensions and proposes to assume
that the width of the opening of the door is W, the width of the door is D, and
the thickness of the door is d. She represents the structure thus obtained in fig-
ure .. When seeing figure ., most of us will immediately recognize how
the explanation works. But that is not the point of this case. What I aim to show
are several, more general points relevant to my argument.

W
Door
Door

D
;^\jgZ&(#&#AZ[i^aajhigVi^dcgZegZhZcihVc^ciZgegZiV"
i^kZhigjXijgZ>Hl^i]^ci]Z[gVbZldg`d[:jXa^Y^Vc
\ZdbZignd[i]Ze]ZcdbZcdcEi]ViVYddgZmVXian
Òihi]ZdeZc^c\d[VXjeWdVgY#G^\]i^aajhigVi^dc
gZegZhZcih>HD"<d[ED#
 mieke b o o n

How should we analyze the first step in the development of an explana-

tion of the observed phenomenon, PO? It is important to recognize that this
first step consists of structuring and interpreting PO within the interpretative
framework of Euclidean geometry, thus producing an interpretative structure
of PO, which I will call ISO-G. The formula ISO-G means: an interpretative struc-
ture IS, of an observed phenomenon (O of observed) in terms of an interpreta-
tive framework (G of geometry).5
A second and also important point is that the interpretative structure rep-
resented in figure . (that is, ISO-G), represents geometrical relations between
objects in terms of which PO (the observed concrete, material, nonfitting door
hanging on the hinges of the cabinet) has been structured and interpreted.
Hence, ISO-G represents PO in terms of geometrical relations between geometri-
cal objects, which are rectangles (the geometrical objects) and the rotation of
one of the rectangles relative to the others (a geometrical relation). This concep-
tion rejects the idea that ISO-G in figure . is a true or essential representation
of PO in the realist sense. Thus, it rejects a conception of the semantic relation
between PO and ISO-G as some kind of correspondence, or (partial) isomorphic,
or similarity, or analogy relation. Moreover, my conception rejects a classical
Lockean conception, which assumes that ISO-G represents the primary proper-
ties of PO, for this geometrical structure does not represent primary properties
either. Instead, it represents how PO is interpreted and structured in terms of a
specific interpretative framework (geometry).
A third point to make is that figure . simultaneously presents an inter-
pretative structure ISO-G and a phenomenon. For the interpretative structure of
PO, ISO-G must also be understood as a phenomenon PG that occurs in a geo-
metrical space, whereas the observed phenomenon PO occurs in physical space.
The geometrical phenomenon PG is represented by rectangles, with one rect-
angle connected to another at one point around which it can rotate. PG, then, is
the phenomenon that the rectangle “door” does not fit between the other two
rectangles when rotating it around one point (illustration on right of fig. .),
whereas it does fit when simply sliding it between the rectangles (illustration
on left). My former point also holds for the semantic relation between PO and
PG, which is that we must not conceive of PG as an objectively true represen-
tation of PO. Instead, PG represents how the cabinetmaker conceives of PO in
terms of geometrical relations between geometrical objects.
A fourth point is that the interpretative structure of PO, ISO-G, frames the
type of questions that can be asked (such as why the rotating rectangle does not
fit between the other two), together with the type of explanations that can be
developed (for example, because the diagonal of the rotating rectangle is longer
than W). These questions and answers are about phenomenon PG. Therefore,
asking the right questions (such as why the concrete material door cannot be
closed) involves going back and forth between PO and PG, which requires un-
u nd e r stan d i n g i n th e e ngi ne e ri ng s c ience s 

derstanding how PG has been constructed, that is, understanding how to struc-
ture and interpret the observed phenomenon PO within a geometrical space.
Accordingly, structuring and interpreting PO in the framework of geometry
has produced PG. In the second step of developing an explanation of PO, PG can
be structured and interpreted in terms of the axioms and laws of Euclidian ge-
ometry, producing ISG-G. Figure . represents how this works.

r=W
Dmax W–a
Door
W W

Dmax a W–a a
d d
W W
Dmax = √(W2–d2) (W–a) = √(W2–d2)

;^\jgZ&(#'#AZ[i^aajhigVi^dcgZegZhZcih>H<"<&d[
E<#G^\]i^aajhigVi^dcgZegZhZcih>H<"<'d[E<#

In a traditional language we would say that PG “obeys” the axioms and laws
of Euclidean geometry; in the modern language of the semantic conception of
theories, we would say that PG “satisfies” these laws (see, for example, Suppe
). The master has different possibilities in developing interpretative struc-
tures ISG-G to explain PG. I will call them ISG-G() and ISG-G(). For instance, by
applying Pythagoras’ theorem, she shows that the width of the door, D, should
not exceed √(W 2-d 2). If the apprentice understands this explanation ISG-G() of
PG, he is able to infer two consequences from it (see fig. ., left). First, as the
thickness, d, of the rectangle “door” exceeds zero, the width of this rectangle, D,
must be less than the width between the fixed rectangles, W. Second, the maxi-
mum width of “door” can be calculated from the formula: Dmax = √(W 2-d 2). Ac-
cordingly, understanding ISG-G() as an explanation of PG within the framework
of Euclidean geometry allows for inferring consequences from ISG-G(), which
in this case consists of making predictions about the effects of interventions,
such as making the rectangle “door” smaller by removing a slice with thickness
a.
However, the cabinetmaker has a specialty: she prefers to make her doors
skewed, since removing slice a would cause a gap between rectangle “door”
and the rectangle to the right. Therefore, she expounds on her explanation
 mieke b o o n

ISG-G() by telling the apprentice that although he has done a good job, he is not
yet finished. He should cut off a triangle from the edge of the door by drawing
a line between the outside corner to a point, a, on the inside width-line of the
door. Since the apprentice understands this explanation, and because he also
understands geometry, he can infer how to calculate point a. He infers that
point a must satisfy the formula (W-a) = √(W 2-d 2). Also this second formula
has been derived from structuring PG in terms of Pythagoras’ theorem, which
involves a further refinement of ISG-G().
On the basis of his understanding of geometry, the apprentice also under-
stands that he can easily construct point a by applying the law to which all
circles obey, which is that the distance is equal from the center of a circle to
any point on the circle. This is how a second interpretative structure ISG-G() of
PG is developed within the framework of geometry. He constructs point a by
drawing a circle with the dot in its center and taking W as the length of radius
r. A rounded edge is constructed by means of a circle with a radius of r=W. In
order to make the “door” fit, he must then draw a straight line from point a to
the outside corner of the door in order to mark the triangle that he should cut
away. In this way, PG has been structured and interpreted in terms of the law
that defines a circle. Inspired by these arcs, the apprentice decides that he will
develop his own specialty once he becomes a fully qualified cabinetmaker. This
specialty will be rounded edges, since, on the basis of the axioms of geometry,
he understands that the door can only be closed if the distance from the hinge
to the edge of the door at all points along d is exactly equal to W. He finds this
solution (right-hand side in fig. .) the most elegant.

8g^iZg^dc[dgJcYZghiVcY^c\HX^Zci^ÒX;^ZaYh8JH;
By further analyzing this case, I will illustrate the abilities involved in under-
standing a scientific field. How does the criterion for understanding scientific
fields (CUSF) apply to the case at hand? Understanding Euclidean geometry
involves the ability of the cabinetmaker to structure and interpret the phenom-
enon of the nonclosing door, PO, within the framework of Euclidean geometry,
producing PG. Understanding the interpretative structure of PG, ISG-G, guides
the apprentice and the cabinetmaker in exploring the effects of possible inter-
ventions with the door’s geometrical structure. The cognitive abilities of the
apprentice and his master thus comprise their being able () to recognize that
the observed phenomenon PO can be structured within the framework of Eu-
clidean geometry; () to abstract from properties of the material door that are
irrelevant within that framework; () to construct a geometrical structure of
PO, ISO-G, that represents how the cabinetmaker conceives of the phenomenon
in terms of geometrical relations between rectangles; () to conceive of ISO-G as
a phenomenon in a geometrical space, PG; () to know which theorems are fit-
u nd e r stan d i n g i n th e e ngi ne e ri ng s c ience s 

ting to PG; () to construct an interpretative structure ISG-G of PG by structuring

PG in terms of axioms and theorems of Euclidean geometry; and finally, () to
infer consequences of interventions to PG that aim at solving the problem by
means of a further development of ISG-G. A cabinetmaker with these cognitive
abilities has an understanding of Euclidian geometry, which means that he or
she knows where and how to use it.
Hence, a crucial point about scientific understanding is that it guides sci-
entists in using theories for structuring and interpreting phenomena within
an interpretative framework. This cognitive ability concerns questions such as
within which interpretative framework the phenomenon should be framed, or
which theories should be used, and how to use the theories in constructing the
interpretative structure. In exactly this manner, scientific understanding is cru-
cial for developing scientific knowledge.

8VjhVa"BZX]Vc^hi^XVcYCdbd"BVi]ZbVi^XVaHigjXijg^c\

HigjXijg^c\VcY>ciZgegZi^c\
As illustrated in the case above, the first step in an explanation involves
structuring and interpreting the observed phenomenon, PO, in terms of a cer-
tain type of interpretative framework. A simplified example may illustrate how
simultaneous structuring and interpreting works within a causal-mechanistic
(CM) and a nomo-mathematical (NM) framework works.
Assume a Boyle type of experiment, in which a correlation is measured
between the pressure and the volume of a gas in an air-tight cylinder. The ob-
served phenomenon PO is that “decreasing the volume of the cylinder requires
someone to increase the force on the piston.” This phenomenon is framed in
terms of a CM framework, that is, “decrease of the volume of the gas causes
an increase of the pressure of the gas,” which describes a causal-mechanistic
phenomenon PCM. The measured values of the pressure, P, and the volume, V,
are framed in terms of an NM framework, where nomo-mathematical means
“mathematical relations between physical variables.” For instance, the set of
data points plotted in a P-V diagram represents the nomo-mathematical phe-
nomenon PNM. Accordingly, within the CM framework, the relation between
pressure and volume observed in this experiment is conceived as a causal-
mechanistic relation, whereas within the NM framework the relation between
the measured values of pressure and volume is conceived as a mathematical
relation between data points.6
Analogous to the example of the cabinetmaker, the next step consists of
developing interpretative structures of PCM in terms of CM relations, and of
PNM in terms of NM relations. A CM interpretation employs (or introduces
new) theoretical entities, such as gas molecules, and properties of these enti-
 mieke b o o n

ties, such as velocity and kinetic energy. An NM interpretation is developed in

the same way as the second step taken by the cabinetmaker. A rather simplistic
approach is to structure and interpret PNM in terms of a mathematical equation
that merely fits the data points (for example, a polynomial function), which pro-
duces ISNM-M.7 Examples are Hooke’s law, Ohm’s law, and Balmer’s law, which
are usually called phenomenological laws. Also in this case, the mathematical
equation thus obtained is both an interpretative structure ISNM-M of PNM and a
phenomenon, which I will call PNM-M.
In a next step, scientists will aim at structuring and interpreting phenom-
enon PNM-M in terms of the axioms of a fundamental theory, which means struc-
turing and interpreting PNM-M within the fundamental laws of an NM framework
in order to develop a more general interpretative structure, ISNM-NM.8 An ex-
ample of ISNM-NM is the Maxwell-Boltzmann equation, which is central to the
kinetic theory of gases. The equations of this ISNM-NM represent, for instance,
the (most probable) speed distribution of gas molecules, which is derived from
structuring and interpreting the behavior of gas molecules in terms of theoreti-
cal principles such as Kolmogorov’s probability axioms, and conservation of
energy and momentum.9
However, the idea that the kinetic theory of gases can be mathematically
derived from fundamental axioms clearly is a misrepresentation of what actu-
ally happens in the development of an ISNM-NM. I claim that in the development
of interpretative structures, different interpretative frameworks are comple-
mentary.10 What I mean is that scientists integrate structuring and interpret-
ing in different frameworks, producing ISNM-NM and ISCM-CM simultaneously;
these interpretative structures cannot be reduced to each other. For example,
in developing ISNM-NM, scientists will make use of PCM. They first need to struc-
ture and interpret PCM in causal-mechanistic terms, thus producing ISCM-CM.
For instance, PCM is structured and interpreted in terms of properties and in-
teractions of gas molecules, such as, “gas molecules have constant random
motion,” “collisions between gas molecules are elastic,” and “gas molecules do
not interact.” These sentences are examples of what is meant by structuring in
terms of how scientists conceive of things and relations between these objects.
Next, structuring and interpreting within the NM framework makes use of the
ISCM-CM thus produced. In the next step, ISCM-CM is considered as a phenom-
enon PCM-CM. In that way, PCM-CM will be structured and interpreted in a three-
dimensional geometrical space, producing a phenomenon, PG, which consists
of randomly moving balls with velocity vectors attached to them.11 Only then
can PG be structured and interpreted in terms of fundamental principles such
as the Kolmogorov axioms, within an NM framework, producing an ISNM-NM,
(for example, the Maxwell-Boltzmann equations).
Conclusions inferred from ISNM-NM are true for PG, whereas the empirical
u nd e r stan d i n g i n th e e ngi ne e ri ng s c ience s 

adequacy of ISNM-NM must be tested for observed phenomena of real gases. The
ideal gas law, for instance, is only empirically adequate for ideal gases.12 A next
step in scientific research is to explain deviations between conclusions inferred
from ISCM-CM, on the one hand, and measurements, on the other. Such devia-
tions are structured and interpreted in terms of ISCM-CM, for instance, in terms
of van der Waals’s forces between gas molecules. This is how science develops:
structuring and interpreting proceeds ever further, which involves developing
interpretative structures ISNM-NM of NM phenomena within NM frameworks,
and interpretative structures ISCM-CM of CM phenomena in CM frameworks, in
an ongoing mutual interaction.

HX^Zci^ÒXJcYZghiVcY^c\
Scientists understand interpretative structures because these represent
how they conceive of objects and relations between objects. This understand-
ing allows for inferring conclusions from these structures. For instance, un-
derstanding Hooke’s law as an interpretative structure for the behavior of a
spring allows for predicting the stretch of a spring that results from exerting
a force. Developing interpretative structures involves an understanding of rel-
evant scientific fields, which in turn implies the involvement of specific cog-
nitive abilities. First, an observation or a measurement must be represented
adequately within an interpretative framework, such as a causal-mechanistic
or a nomo-mathematical framework, in order to produce a PCM or PNM, respec-
tively. Second, scientists must decide on the right fundamental principles for
structuring and interpreting a PNM. Third, when structuring and interpreting
within a causal-mechanistic framework, scientists must envisage physical enti-
ties and properties relevant to PCM. Fourth, deductions from fundamental laws
must be mathematically correct. Often, approximations and simplifications are
needed in order to make the mathematical equations manageable. An example
is neglecting viscosity terms in the Navier-Stokes equation. Doing this correct-
ly usually involves scientific understanding in terms of a causal-mechanistic
framework. These are the kinds of abilities that determine someone’s scientific
understanding of a scientific field.

:e^hiZbdad\^XVa>hhjZh/DW_ZXihVcYGZaVi^dchWZilZZcDW_ZXih
The core aspect of interpretative structures is that they represent objects
and relations between objects. These structures represent relations in a similar
way as to how judgments represent relations between disparate objects and/or
events. For instance, “the cat is on the mat” represents a relation in space, or an
individuation of objects, that is, the cat and the mat; “the cat was on the mat”
represents as a relation in time between two events; “if the cat sits on the mat,
then the dog is out” represents a logical relation; “the cat is skinny” represent a
 mieke b o o n

relation between an object and a property; “the cat has dirtied the mat” repre-
sents a causal relation between two events; “the cat sits one meter beside the
mat” represents a mathematical relation—in this case, the length of a straight
line between two points; and “this is the path that the cat always takes” repre-
sents a relation between locations on a geographical map.
The basic idea of these examples is that in observing the world, humans
must structure and interpret their observations in order to arrive at even the
most elementary judgments, for instance, PG, PCM, or PNM. In doing so, they
relate what they observe in terms of interpretative structures. As may be clear
from the examples, different kinds of interpretative structures are involved,
which result in different types of relations, such as relations in space and/or
time, relations that individuate (discern) objects or events, relations between
objects and properties, causal relations or interactions, mathematical, and logi-
cal relations. Instead of conceiving our “observations” of relations as something
that exists independent of us, we must conceive of “observed” relations as ways
in which we have structured and interpreted our observations.13
Knowledge represented in theories and scientific explanations must be
understood as similar in form to knowledge represented in judgments. Ac-
cordingly, theories and scientific explanations also represent objects and rela-
tions between objects, such as logical, geometrical, causal-mechanistic, and/or
nomo-mathematical objects and relations. These are not objects and relations
that scientists could somehow observe if they had the proper means, but rela-
tions in terms of which scientists structure and interpret what they observe.
Why do we have a basic interest in structuring and interpreting the world?
Several reasons can be given. First, structuring and interpreting the world is a
basic cognitive need, because without this activity the world would appear cha-
otic and arbitrary to us. Second, interpretative structures often have an aesthet-
ic value. Third, structuring and interpreting the world according to ever-higher
epistemic standards is an intellectual challenge. Fourth, we need interpretative
structures in order to think rationally about our acting in, and intervening with,
the world. Again, knowledge of these relations is not gained by observing them,
but by means of “structuring and interpreting” our observations in terms of
certain frameworks. These frameworks guide how to construct or invent rela-
tions between objects or events.

8dcXajh^dch

Why are the engineering sciences such an interesting case with regard to
perspectives on scientific understanding? I propose that considering the engi-
neering sciences as science presents us with consequences on how to conceive
of the laboratory sciences in general.
u nd e r stan d i n g i n th e e ngi ne e ri ng s c ience s 

My argument started from the assumption that there is a strong, tradition-

al, normative idea that says that the ultimate aim of science is true theories. This
idea implies that there is a fundamental distinction between basic sciences,
which aim at truth, and applied sciences (such as the engineering sciences),
which aim at use. However, from the perspective of the engineering scienc-
es, an alternative normative idea can be formulated, which holds that science
(as an intellectual enterprise) aims at discerning, creating, and understanding
phenomena. This idea accords with actual scientific practices, in which “basic”
laboratory sciences and the engineering sciences are highly similar in their sci-
entific approaches, which aim at creating and intervening with phenomena,
and at developing an understanding of phenomena in terms of causal-mecha-
nistic and/or nomo-mathematical structures.
In view of this alternative aim of science, I have developed an alternative ac-
count of understanding phenomena and theories. Central to my account is an
epistemological issue, which is that neither objects nor relations between objects
are simply observed. I have thus proposed a philosophical conception which im-
plies that scientists structure and interpret a phenomenon P in terms of differ-
ent types of interpretative frameworks, such as logical, statistical, geometrical,
causal-mechanistic, or nomo-mathematical. These interpretative frameworks
determine the types of objects and the types of relations between these objects
in terms of which scientists structure and interpret the world. For instance, ()
formal objects with properties, and logical relations between them, in a logi-
cal framework; () mathematical objects (such as points, lines, shapes, vectors),
and mathematical relations between them, in a mathematical framework; ()
physical objects with physical properties, and causal relations between them, in
a causal-mechanistic framework; () nomo-mathematical objects (such as data
points that result from measurements), and nomo-mathematical relations be-
tween them, in a nomo-mathematical framework. I have explicated this point
somewhat further by relating it to the character of judgments (as opposed to
a propositional conception of language). Judgments are central to our reason-
ing in common language and serve this function because they represent how
we conceive of objects and relations between objects that can be of different
types.
Accordingly, scientists gain understanding by structuring and interpret-
ing within particular interpretative frameworks, thus producing interpretative
structures IS. They understand a phenomenon P if they have structured and
interpreted P in terms of an interpretative structure IS, whereas they under-
stand an explanation IS if they can draw inferences from it, and, more generally
speaking, if they can use IS in structuring and interpreting other phenomena
(such as complex phenomena). However, these criteria do not explain why sci-
entists understand an interpretative structure, and why they gain understanding
 mieke b o o n

by structuring and interpreting. I have proposed that structuring and inter-

preting produces understanding of a phenomenon because scientists develop a
conception of P in terms of certain types of objects and relations between these
objects. Conceiving of a phenomenon in terms of an interpretative structure
that represents objects and relations between objects that—metaphorically
speaking—underlie the phenomenon, provides understanding, not because an
IS states what the world is like, but because it allows for reasoning about the
phenomenon, for example, how the phenomenon results from particular con-
ditions, or how interventions will affect the phenomenon.
More advanced structuring and interpreting in science makes use of fun-
damental theories of different types. Examples of fundamental theories are
first-order proposition logic within the logical framework, Euclidean geom-
etry within a geometrical framework, chemistry within the causal-mechanistic
framework, and Maxwell’s equations within the nomo-mathematical frame-
work. Scientists understand a fundamental theory, T, if they can use it for
structuring and interpreting a phenomenon.
An important issue with regard to the engineering sciences is how scien-
tists gain understanding of complex phenomena. I have proposed that this is
possible because scientists understand relevant scientific fields. The criterion
of understanding a scientific field, CUSF, aims to account for this possibility:
scientists understand a scientific field if they have acquired the ability to use
scientific knowledge of the field (for example, fundamental and general theo-
ries, as well as interpretative structures of phenomena) in developing explana-
tions and predictions of phenomena that are relevant to the field (including the
ability to meet the methodological criteria of the field). This criterion expands
on CIT, as proposed by de Regt and Dieks (). Developing interpretative
structures (that is, gaining understanding) of (complex) phenomena involves
combining different types of structuring and interpreting (for example, logical,
geometrical, causal-mechanistic, and nomo-mathematical). This is particularly
obvious in the engineering sciences, but a closer look reveals that this is also
the case in “basic” laboratory sciences. Therefore, I suggest that the account of
understanding proposed in this chapter is relevant to the laboratory sciences
in general.

CdiZh
I would like to thank Henk Procee, Bas van Fraassen, Tarja Knuuttila, Isabelle Pe-
schard, Henk de Regt, Sabina Leonelli, and Kai Eigner for their constructive contributions.
This research is supported by a grant from the Netherlands Organisation for Scientific
u nd e r stan d i n g i n th e e ngi ne e ri ng s c ience s 

Research (NWO). I also wish to thank the ZIF Centre for Interdisciplinary Research, Biele-
feld University, for their financial support, and the members of the ZIF research group
(–), Science in the Context of Application, for their intellectual support.
. My view on science stands in the tradition of the New Experimentalists. Authors
who have worked at the line between philosophy of science and history of science, such as
Ian Hacking (), Alan Franklin (), Peter Galison (), Robert Ackerman (),
Margaret Morrison (), and Hasok Chang (), give reasons to believe that the tradi-
tional aim of science is also too narrow for most of the present-day “basic” laboratory sci-
ences. A severe critique of this traditional picture shared by these authors is that it neglects
the crucial role of experiments and instruments in science. In the laboratory sciences,
experiments and instruments are not mere spectacles on the world. They are more than
just tools for observing the world; they also have a life of their own. For instance, instru-
ments produce phenomena that do not exist outside the laboratory, and those phenomena
become topics of scientific and technological interest. An example is the phenomenon of
superconductivity, which can only exist in artificially produced superconductor materials
at extremely low temperatures.
. Petri Ylikoski (this volume) defends a similar thesis. He refers to Wittgenstein (Phil-
osophical Investigations, ), who claims that understanding is an ability, since, when a
person understands something, he or she is able to do certain things. Michael Dummett
(/, –), to whom I am indebted, stands in the same tradition. Against the
generally accepted idea that understanding the meaning of a word or a sentence consists in
its method of verification, or in its truth conditions, Dummett suggests that understanding
a word or sentence involves knowing how to use it. Similarly, I propose that understand-
ing a theory does not depend on its truth, which would mean that a scientist understands
the theory because it provides a literally true story of what the world is like (compare van
Fraassen ). Rather, understanding a theory involves the ability to use it.
. My notion of interpretative structures is much broader than the common use of
structures in philosophy of science, which indicates logico-mathematical structures that
represent logical and mathematical relations.
. The distinction between the “observable” and “unobservable” world refers to an
intuitive distinction that we usually make between the world that is directly accessible for
sense experiences, and the world behind the phenomena. Van Fraassen () used this
distinction for explicating his antirealism, and his work has initiated a lot of philosophi-
cal discussion on this subject. Nevertheless, it is still a clarifying distinction in our daily
language, which is how I aim to use it here. My point in this chapter is that observability is
problematic for another reason, which is that we cannot make sense of “directly” observing
the relations presented in our judgments of the “observable” world. This is why I have put
these terms in quotes.
. Radder () analyzes in great depth the claim that observation presupposes
conceptual interpretation. Radder’s emphasis is on the idea that concepts structure our
observations of the world. My emphasis is on the idea that cognitive activities structure our
observations of the world; these activities consist of structuring and interpreting in terms
of specific types of objects and relations between objects.
 mieke b o o n

. The semantic view of theories gives a refined account of how data points are
converted into a data model, (see, for example, Mayo , chapter ). I agree with this
account, but will not use “data-model” in this chapter.
. The formula ISNM-M means an interpretative structure IS of measured data points
in a nomo-mathematical framework NM. Data points are related in terms of an arbitrary
mathematical equation M (for example, a polynomial function).
. My proposal that the axioms of fundamental theories present a theoretical frame-
work within which scientists structure and interpret phenomena agrees with the semantic
view of theories of van Fraassen (, –), and Giere (, ). From this per-
spective, Giere (, ) argues against Cartwright (). According to Giere, funda-
mental “laws cannot tell lies about the world, because they are not really statements about
the world. . . . They are, as Cartwright herself sometimes suggests, part of the characteriza-
tion of theoretical models, which in turn may represent various real systems.” Opposite
to this view stands structural realism. An overview of the constructive empiricism versus
structural realism debate can be found in Psillos ().
. The engineering sciences involve at least five topics that are considered as general
basic theory, necessary for taking a scientific approach to the development of technology,
which are typical examples of how nomo-mathematical structuring works. These topics
are mechanics, flow phenomena, electricity and magnetism, thermodynamics and statisti-
cal mechanics, and quantum mechanics. My basic assumption is that, like the axioms
of Euclidean geometry, the axioms of fundamental theories are not true or empirically
adequate, but fitting for structuring and interpreting certain types of phenomena. For in-
stance, Newton’s three laws of motion and the law of gravity play a central role in mechan-
ics. Conservation laws, such as conservation of mass, energy, momentum and chemical
substance play a central role in flow phenomena. Electricity and magnetism draw on
Maxwell’s laws; thermodynamics on the basic laws of thermodynamics; statistical mechan-
ics on the Boltzmann equation and the laws of probability theory, and quantum mechanics
on the Schrödinger equation.
. With the complementarity of interpretative frameworks, I mean something differ-
ent than Salmon’s () complementarity thesis, which claims that explanations in terms
of causal-mechanisms, and explanations in terms of derivations from fundamental laws are
complementary. De Regt’s () analysis of this claim shows that Salmon interprets his
causal conception of scientific explanation as an “ontic” claim: “in contrast to ‘epistemic’
conceptions which regard explanations as arguments, ontic conceptions assert that ‘events
are explained by showing how they fit into the physical patterns found in the world’ ” (de
Regt , ). My conception of science as developing interpretive structures within
interpretative frameworks aims at avoiding ontic or metaphysical connotations.
. More correctly, the phenomenon is interpreted and structured in analytic geom-
etry. Scientific explanations of phenomena are often structured within a framework that
integrates Euclidian geometry and analytic geometry that uses Cartesian coordinates for
defining the space. As a result, the IS can be represented in mathematical spaces that are
no longer strictly geometric, but constructed from combinations of x, y, z, t, and physical
variables such as the velocity, v. Introduction of principles from analytic geometry allows
for instance for the construction of differential equations. Another aspect that I will not
u nd e r stan d i n g i n th e e ngi ne e ri ng s c ience s 

explain in more detail is that the deduction of mathematical structures does not usually
start from the fundamental principles, but from equations already deduced from them,
such as the Maxwell-Boltzman equation, or Navier-Stokes equation.
. Clearly, ideal gases are defined in this way: a gas is ideal if it obeys the ideal gas
law. Therefore, the belief that the ideal gas law is empirically adequate for ideal gases is
problematic.
. On this matter I am deeply indebted to Kant’s criticism of Hume’s empiricism;
compare Kant (/) and Henry Allison’s () interpretation of Kant’s work.
Rather than using Kant’s notion of categories, I propose the notion of interpretative
frameworks.

GZ[ZgZcXZh
Ackerman, R. . The new experimentalism. British Journal for the Philosophy of Science
:–.
Allison, H. E. . Kant’s transcendental idealism: An interpretation and defense. New
Haven: Yale University Press.
Boon, M. . Technological instruments in scientific experimentation. International
Studies in the Philosophy of Science :–.
———. . How science is applied in technology. International Studies in the Philosophy
of Science :–.
Boon, M., and T. Knuuttila. . Models as epistemic tools in engineering sciences. In
Handbook of the philosophy of science, vol. , Philosophy of technology and engineering
sciences, edited by Anthonie Meijers, –. London: Elsevier Science.
Cartwright, N. . How the laws of physics lie. Oxford: Clarendon Press.
———. . Nature’s capacities and their measurement. Oxford: Clarendon Press.
Chang, H. . Inventing temperature: Measurement and scientific progress. New York:
Oxford University Press.
de Regt, H. W. . Wesley Salmon’s complementarity thesis: Causalism and unification-
ism reconciled? International Studies in the Philosophy of Science :–.
de Regt, H. W., and D. Dieks. . A contextual approach to scientific understanding.
Synthese :–.
Dummett, M. /. The philosophical basis of intuitionist logic. In Truth and other
enigmas, –. London: Duckworth.
Franklin, A. . The neglect of experiment. New York: Cambridge University Press.
Galison, P. . How experiments end. Chicago: University of Chicago Press.
Giere, R. N. . Explaining science: A cognitive approach. Chicago: University of Chicago
Press.
———. . Science without laws. Chicago: University of Chicago Press.
Hacking, I. . Representing and intervening: Introductory topics in the philosophy of
natural science. Cambridge: Cambridge University Press.
———. . The self-vindication of the laboratory sciences. In Science as practice and
culture, edited by A. Pickering, –. Chicago: University of Chicago Press.
Kant, I. /. Critique of pure reason. Cambridge: Cambridge University Press.
 mieke b o o n

Mayo, D. G. . Error and the growth of experimental knowledge. Chicago: University of
Chicago Press.
Morrison, M. . Unifying scientific theories: Physical concepts and mathematical struc-
tures. Cambridge: Cambridge University Press.
Psillos, S. . Is structural realism possible? Philosophy of Science :s–s.
Radder, H. . The world observed / The world conceived. Pittsburgh: University of Pitts-
burgh Press.
Salmon, W. C. . Scientific explanation: Causation and unification. In Philosophy of
science: Contemporary readings, edited by Y. Balashow and A. Rosenberg, –.
London: Routledge.
Suppe, F. . The semantic conception of theories and scientific realism. Urbana: Univer-
sity of Illinois Press.
Trout, J. D. . Scientific explanation and the sense of understanding. Philosophy of Sci-
ence :–.
van Fraassen, B. C. . The scientific image. Oxford: Clarendon Press.
———. . Laws and symmetry. New York: Oxford University Press.
chap ter title ver so

(+
Understanding in Psychology
Is Understanding a Surplus?

@6>:><C:G

 Since Thomas Kuhn’s characterization of science by means of a list

of epistemic values that provide “the shared basis for theory choice,” there is a
debate in philosophy of science about what the epistemic values of science are
(for example, Kuhn ; McMullin ; Longino ; Lacey ). Kuhn’s
list comprised such values as empirical accuracy, consistency, scope, simplic-
ity, and fruitfulness (Kuhn , –), which, as he argued, can be seen as
constitutive of science. An enterprise could have different criteria for judging
theories, but then it would not be science ().
In this chapter, I argue that the lists of epistemic values that are constitutive
of science as proposed by philosophers, such as Kuhn, Ernan McMullin, Helen
Longino, and Hugh Lacey, lack a highly important element, namely intelligi-
bility. By means of an empirical case study I will show that scientific models,
which enable the connection between theory and the world, can only give us
scientific knowledge about phenomena if these models are intelligible.
My argument is based on a case study of neobehaviorism. This case is in-
teresting for the study of philosophy of science not only because of the huge
influence of neobehaviorism on psychology in particular and on science in gen-
eral, but also on account of its extraordinary positivistic inclination. Between
 and , when psychology as academic enterprise grew enormously
and neobehaviorism was the leading discipline in American psychology, the
neobehaviorists advocated a view of science that was heavily influenced by


 k ai eigner

logical positivism, which included, among other things, the requirement that
all subjective elements should be banished from science. Because of their posi-
tivistic attitude, neobehaviorists were called “black-box psychologists”: they
considered the organism as a black box with observable stimuli as input and
observable responses as output, and they rejected any allegedly metaphysical
speculation about the mechanisms inside this black box. Metaphorically speak-
ing, opening the box would not yield “insight.” Instead, they declared as the aim
of psychology to find functional relationships that describe the correlations be-
tween stimuli and responses. This positivistic tendency makes neobehaviorism
highly relevant to a study on intelligibility and understanding. Insofar as un-
derstanding and intelligibility, with their subjective connotations, were allowed
to play a role at all by the neobehaviorist doctrine, this was only in the context
of discovery and not in the context of justification. By focusing on two leading
figures of neobehaviorism, Edward C. Tolman (–) and Clark L. Hull
(–), I will show that, despite their positivistic ideas, in practice even
the neobehaviorists aimed at insight and implicitly embraced intelligibility as
an epistemic value. What might seem to be one of the strongest counterex-
amples of my claim appears in fact a corroboration of it, and as such provides a
strong case in favor of my claim concerning intelligibility and understanding.
In this chapter, I will offer a philosophical account of intelligibility and un-
derstanding based on the work of Henk de Regt, but with a shift in focus from
theories to models. I will then present the case study of neobehaviorism and,
by means of an analysis of the work of Tolman and Hull, in which I will use my
philosophical account of intelligibility and understanding, I will demonstrate
my claim about the constitutive value of intelligibility.

JcYZghiVcY^c\E]ZcdbZcVWnBZVchd[>ciZaa^\^WaZBdYZah

A sketch of de Regt’s account of scientific understanding can be found in

an article that was published in  (see also chapter  of this volume). In
that article de Regt draws on the intuition that scientifically understanding a
phenomenon is not merely knowing the relevant scientific theories, laws, and
background conditions, but in addition “being able to use” them in the case
at hand (de Regt , ). Furthermore, he suggests that understanding a
phenomenon requires a theory that is intelligible to the scientists who use it—a
suggestion that we find again in de Regt and Dieks () as the criterion for
understanding phenomena (CUP). In line with this suggestion, de Regt defines
the intelligibility of a theory as a positive value that scientists attribute to the
cluster of virtues of this theory that help them in using the theory (de Regt
, ).
According to de Regt, intelligibility is not an intrinsic property of a theory;
u nd e r stan d i n g i n p s yc h ol o gy 

instead, it is a value that scientists project onto representations of theories.

Which theories are deemed intelligible depends on certain properties of the
theory as well as certain contextual factors, such as the skills of the scientists
(de Regt , ). Even though intelligibility is a context-dependent value, de
Regt and Dieks argue that it is still possible to formulate general criteria for the
intelligibility of theories. An example is the following criterion (CIT): a theory
is intelligible for scientists in a specific context if they can recognize qualitative-
ly characteristic consequences of the theory without performing exact calcula-
tions (de Regt and Dieks , –). As de Regt and Dieks argue, to fulfill
this criterion, merely knowing the statements of a theory is not enough. In ad-
dition, scientists need skills to exploit (pragmatic) virtues of the theory as tools
to apply this theory successfully to concrete situations. Examples of such skills
are causal reasoning and visualization; examples of the virtues of a theory that
can be used as tools are its causal nature, its visualizability, or its simplicity.
But is this account relevant for a discussion about understanding in psy-
chology? Because of the central position of quantitative theories and exact cal-
culations in de Regt’s account, it seems that its domain is restricted to that
of the natural sciences, such as physics. Furthermore, it can be argued that
the theory-dominated view of science in the work of de Regt—we understand
phenomena if we have intelligible theories about them—has the disadvantage
that it does not sufficiently take into consideration an intuition that is shared by
many philosophers of science, namely that understanding is usually acquired by
means of scientific models (for example, Hartmann , Bailer-Jones ).
However, it is my opinion that these possible points of critique can be cir-
cumvented and that the work of de Regt can be of use for the formulation of an
account of understanding that is more general. The limitations of de Regt’s ac-
count originate from the approach he uses to formulate this account, in which
an important role is played by the criterion for understanding phenomena
(CUP) and by the criterion for the intelligibility of theories (CIT). The first
criterion focuses on theories, the second on exact calculations. However, these
criteria should not be seen as the fundamental basis of de Regt’s account, and
accordingly the account should not be seen as restricted by them. In my view,
de Regt introduces these criteria in order to explicate intuitions about under-
standing, for instance in the case of the physicist Werner Heisenberg, which
are used to bring in the more fundamental part of his account of understand-
ing, namely the part about skills that are required to use relevant theoretical
knowledge.
In order to formulate a more general account, I will elaborate on the idea
mentioned above that understanding is related to the ability to use the relevant
theoretical knowledge. I will focus on two aspects of this ability, basing my
analysis on the article of de Regt and Dieks (). The first aspect is the ability
 k ai eigner

to apply theoretical principles to particular phenomena (compare de Regt and

Dieks , ). I will discuss this aspect using the work of Ronald Giere (
and ). As we will see, this aspect of applying principles is closely connected
to the use of models in science—which may be one of the reasons that many
philosophers of science share the intuition that understanding and modeling
are related. The second aspect is the ability to develop insight in the conse-
quences of theoretical principles in concrete situations (compare de Regt and
Dieks , ). Although these two aspects can be conceptually separated, in
practice they are intertwined. In a general account of understanding it should
be made clear how these abilities are related to each other and what their role
is in the generation of scientific understanding. What I regard as an important
contribution to the development of such a general account is the idea in de
Regt’s work that scientific understanding is brought about by the interplay be-
tween skills and virtues; although in my view these virtues are not principally
of theories but of models (compare Boumans in this volume, who also pleads
for a shift in focus in de Regt and Dieks’s account of understanding from theory
to model). I will give a sketch of such an account in which I concentrate on the
role of models for understanding and employ de Regt’s ideas about skills and
virtues to achieve understanding.

6eean^c\I]ZdgZi^XVaEg^cX^eaZhidEVgi^XjaVgE]ZcdbZcV
The first aspect of the ability to use theoretical knowledge is the ability to
make a connection between theory and the world. The relation between scien-
tific theories and the world—which is epistemologically relevant for the empir-
ical justification of theoretical knowledge—is a topic in philosophy of science
that has received much attention. Since the s we have seen a gradual shift
in the philosophical ideas concerning this relation. The traditional theory-
dominated view was superseded by a view in which models play an important
role in representing aspects of the world. For instance, Cartwright, Shomar, and
Suárez () hold the view that the connection between scientific theory and
the world is accomplished by constructing models that form representations of
phenomena. Constructing these models is a skillful activity that involves evalu-
ations and decisions. Along the same line, de Regt and Dieks (, ) write
that the ability of scientists to apply a theory to a particular phenomenon de-
pends both on their skills and on the pragmatic virtues of the theory, such as
its visualizability or simplicity. According to de Regt and Dieks, this activity of
connecting theory and the world, which has a pragmatic nature, has epistemic
relevance because it is a necessary ingredient for achieving the epistemic aims
of science. Although I agree with this, the argumentation for it needs more
sophistication in order to avoid the objection that the activity of constructing
models—and thus forming the connection between theory and the world—
u nd e r stan d i n g i n p s yc h ol o gy 

belongs to the context of discovery, and that the pragmatic dimension of it does
not play a role anymore once the connection between theory and phenomena
has been established. An important step in circumventing this objection is to
formulate more precisely what we mean by concepts like theory, model, and
representation. For this I will make use of the representational view of models
of Giere (, ), which, similar to the view of Cartwright and colleagues,
shares the contemporary idea mentioned above that models play an essential
role in representing aspects of the world.
Giere (, ) prefers to speak of “theoretical principles” instead of
“theories” or “laws” because the latter terms are used quite broadly and even
ambiguously, both in scientific practice as well as in philosophical discussions
about the sciences. Examples of theoretical principles are Newton’s laws of me-
chanics, the evolutionary principle of natural selection, and, as we will see in
this chapter, Hull’s principles of behavior. Although such principles have often
been interpreted as empirical statements, Giere sees them merely as defini-
tions of the abstract terms that compose theoretical models. For instance, in
Newton’s three laws of motion, quantities called force and mass are related
to well-understood quantities such as position, velocity, and acceleration. The
laws, however, do not spell out in more specific terms what might count as a
force or mass (Giere , ). According to Giere, the function of these theo-
retical principles is to act as general templates for the construction of models.
Giere (, ) calls his view on models representational because it takes
models as tools for representing the world. He is aware that the models used in
scientific practice seem to form a heterogeneous class, including physical mod-
els, scale models, analogue models, and mathematical models, but he argues
that with his account it is possible to understand models in a way that usefully
encompasses this heterogeneity to a large extent. “What is special about mod-
els is that they are designed so that elements of the model can be identified
with features of the real world. This is what makes it possible to use models
to represent aspects of the world” (Giere , ). In short, models can be
used to represent phenomena, and theoretical principles constitute definitions
of “abstract objects” that can be used to construct these models. As we will
see, this view of the relation between theory and phenomena fits nicely with
the practice of the neobehaviorists, who dealt with principles of behavior (for
example, Hull a). These principles functioned exactly as Giere described
in his representational view, as definitions of the abstract objects that compose
theoretical models.
In order to use a model to represent some aspect of the world, one should
be able to pick out relevant features of the model that are similar in certain
respects and degrees to what is modeled. Giere notices that there are no objec-
tive rules for picking out some specific features of the models. Moreover, he
 k ai eigner

argues that the concept of similarity is context dependent (Giere , ) and
that there is no objective measure of similarity between the model and the real
system.

Anything is similar to anything else in countless respects, but not any-

thing represents anything else. It is not the model that is doing the
representing; it is the scientist using the model who is doing the rep-
resenting. One way scientists do this is by picking out some specific
features of the model that are then claimed to be similar to features of
the designated real system to some (perhaps fairly loosely indicated)
degree of fit. It is the existence of the specified similarities that makes
possible the use of the model to represent the real system in this way.
(Giere , –)

Giere’s description of what it means to use models to represent phenom-

ena, and especially the remark that the model does not represent by itself but
that it is the scientist who is “doing the representing,” clarifies the idea that the
pragmatic dimension of connecting theory and phenomena by means of mod-
els has epistemic relevance. For, it is not the case that once the model is con-
structed there is automatically a connection between theory and phenomena.
As described in recent literature on models (for example, Bailer-Jones ),
such a connection, without which the theory has no empirical content and thus
no empirical relevance, requires an active involvement of the model users. In
Giere’s terminology, this connection is only established if someone is “doing
the representing.” Furthermore, Giere’s representational view makes clear that
the skills required for this are those that enable the selection of relevant fea-
tures and the judgment of similarities. The active and skillful process of using
a model to represent a phenomenon will in practice be intertwined with the
process of constructing models, and the picture of the whole skillful process of
modeling forms a good description of the ability to apply scientific theory to a
particular phenomenon. This gives a specific content to the first aspect of the
intuition about the use of theoretical knowledge as found in de Regt ().

9ZkZade^c\>ch^\]i^c8dchZfjZcXZhd[I]ZdgZi^XVaEg^cX^eaZh
The second aspect of the ability to use theoretical knowledge, which I will
discuss in relation to the intuitions about scientific understanding as found in
de Regt and Dieks (), is the ability to develop qualitative insight into the
consequences of theoretical principles in concrete situations. According to de
Regt and Dieks (, ), this ability requires conceptual tools that make it
possible to circumvent the calculating stage and make an intuitive jump to the
conclusion. In their account, the virtues of the theory can be used as such tools
u nd e r stan d i n g i n p s yc h ol o gy 

if they realize “the right combination” with the skills of the scientist (de Regt
, ). As mentioned above, I regard this as a valuable insight for an ac-
count of understanding, although in my view these virtues are better to be seen
as virtues of models instead of theories.
An illustration provided by de Regt and Dieks is the way Ludwig Boltzmann
introduced the kinetic theory of gases. Boltzmann () devoted the introduc-
tory section to a purely qualitative analysis of gases in which he concluded that
a gas can be pictured as a collection of freely moving molecules in a container.
This model of the gas provides qualitative insight in the behavior of gases. For
instance, it is straightforward to explain in a qualitative way that a gas exerts a
pressure on the walls of the container. If a gas molecule collides with a wall of
the container, it gives a little push, and the total effect of the pushing produces
the pressure. This picture also makes clear that a decrease of volume results
in an increase of pressure. A decrease of volume will cause an increase of the
number of molecules per unit of volume, and this causes an increase in the
number of impacts per unit of time on the wall surface, and thus an increase
in pressure. As de Regt and Dieks argue, in this way qualitative understanding
about gases is obtained which is not based on calculations:

It is important to note that the above reasoning does not involve any
calculations. It is based on general characteristics of the theoretical
description of the gas. Its purpose is to give us understanding of the
phenomena, before we embark on detailed calculations. Such calcu-
lations are subsequently motivated, and given direction, through the
understanding we already possess. (de Regt and Dieks , –)

In this example, the conceptual tool used to recognize qualitative consequenc-

es without performing exact calculations is causal reasoning. This tool is used
to achieve the goal of qualitative prediction. According to de Regt and Dieks, it
is an instrument that helps to get a feeling for the situation. We are able to use
this tool because we possess the skill of making causal inferences and because
Boltzmann’s model of the gas has the virtue of providing a causal mechanical
picture.
Taking a closer look at the example of the ideal gas law reveals that the right
combination of virtues and skills not only facilitates developing qualitative in-
sight into the consequences of theoretical principles in concrete situations, but
also facilitates applying theoretical principles to particular phenomena. Both
aspects of the ability to use relevant theoretical knowledge are intertwined.
The causal story about the pushing molecules enables us make the connection
between, on the one hand, Boltzmann’s kinetic model and, on the other hand,
the concrete pressure on the wall. We are able to “do the representing”—which
 k ai eigner

means picking out the relevant features of the model and judging the similari-
ties between the behavior of the model and the real gases—due to the causal
nature of Boltzmann’s model, which in light of our skill of making causal infer-
ences is a virtue that renders the model intelligible.
We now have the ingredients of the account of understanding used in
this chapter. In order to understand a phenomenon, an intelligible model is
needed, which is a model that has virtues skillful scientists can use for two
purposes, namely for “doing the representing” of the phenomenon and for
seeing intuitively the characteristic consequences of the model. I will use this
account to give an analysis of the role of intelligibility and understanding in
the case study about neobehaviorism, where I will mainly look at the work of
Tolman and Hull, who are considered to be the main representatives of the
methodology of neobehaviorism. I will use the work of Tolman to get an idea
of the character of the skills and virtues that play a role in neobehaviorism and
I will use the work of Hull in order to demonstrate the epistemic relevance of
these virtues and skills, and thus of the epistemic relevance of intelligibility
and understanding.

7Z]Vk^dg^hiBZi]dYdad\n

Classical behaviorism, the precursor of neobehaviorism, was founded al-

most a century ago as a reaction against subjectivism and “introspective” psy-
chology. Usually, the American psychologist John B. Watson (–) is
given the credit for founding this field in  with the publication of a very
influential manifest called “Psychology as the Behaviorist Views It,” which, as
already can be seen in the first lines of his manifest, was an appeal to construct
an objective psychology:

Psychology as the behaviorist views it is a purely objective experimen-

tal branch of natural science. Its theoretical goal is the prediction and
control of behavior. Introspection forms no essential part of its meth-
ods, nor is the scientific value of its data dependent upon the readiness
with which they lend themselves to interpretation in terms of con-
sciousness. The behaviorist, in his efforts to get a unitary scheme of
animal response, recognizes no dividing line between man and brute.
The behavior of man, with all of its refinement and complexity, forms
only a part of the behaviorist’s total scheme of investigation. (Watson
, )

The two basic ideas of classical behaviorism, which can already be found in
this manuscript of Watson, were that the (almost exclusive) subject matter of
u nd e r stan d i n g i n p s yc h ol o gy 

psychology should be overt behavior, and that the major goals of psychology
should be the prediction and control of behavior instead of the description and
explanation of states of consciousness (Hergenhahn , ).
When logical positivism arose in the s it heavily influenced behavior-
ism, and since then we speak of neobehaviorism. Although logical positivism
offered the behaviorists a philosophical foundation for their objectivist views,
it was not evident how this body of thought could be translated into recom-
mendations for scientific practice. Therefore, leading figures of neobehavior-
ism, such as Tolman and Hull, initiated a behaviorist methodology that was
meant to give shape to an objective psychology in which there was no room for
subjective elements. A major ingredient of this methodology was operational-
ization, a procedure by which the theoretical terms used in psychology were
given objective definitions and that was meant to eliminate possible subjective
connotations, originating, for instance, from prescientific insight.
We will see, however, that, in spite of their efforts, the meaning of the the-
oretical terms used by Tolman and Hull surpassed their objective definition:
apparently the terms had indispensable “surplus meaning.” An analysis of this
surplus meaning, in which I make use of my account of understanding, will
show that its function is to render intelligible the scientific models in which
these terms appear. From the indispensable character of this surplus meaning I
will argue that actually, despite their aversion to subjective elements in science,
the neobehaviorists aimed at constructing models that are intelligible. My con-
clusion from this will be that even in the case of neobehaviorism intelligibility
functioned as a significant epistemic value.

:YlVgY8#IdabVc
In the s, Tolman developed a methodology for neobehaviorism that
he called “operational behaviorism” (Tolman /). A motive for Tolman
to develop this methodology was his interest in the logical positivism of the
Vienna Circle. After a sabbatical in Vienna, Tolman expressed the intention
to apply the ideas of logical positivism to psychology. “The ‘logical positivists,’
that is such men as Wittgenstein, Schlick, Carnap in Europe and Bridgman,
C. I. Lewis, Feigl and Blumberg in this country, have already done the task for
physics. But no one as yet, so it seems to me, has satisfactory done it for psy-
chology” (Tolman /, ). In a logical positivistic spirit he defined the
aim of psychology to find functional relationships between the behavior of an
organism and its environmental “determiners,” labeled by him respectively as
“dependent” and “independent variables.” He schematized this in the following
manner (Tolman /, ):

Independent variables ——— f ——— Dependent variables

 k ai eigner

For complex behavior, finding this function f is a difficult task, and therefore
Tolman proposed to break up this complicated function into more manageable
component functions, f and f, which led to the introduction of theoretical
terms, or “intervening variables.” Tolman schematized the new situation as fol-
lows (Tolman /, ):

Independent variables ——— f ——— Intervening variables ——— f ——— Dependent variables

The classical example of an intervening variable is a theoretical term labeled

“demand” or “hunger,” which we introduce to relate an independent experimen-
tal variable (for example, the time since the organism last received food) with
a certain behavior (for example, the amount of food consumed). The introduc-
tion of intervening variables was not in conflict with the positivistic ideals of
neobehaviorism as long as the meaning of these terms was captured in an ob-
jective way by operational definitions. From that point of view, an intervening
variable is nothing more than a label used in an observed functional relation-
ship between behavior and environment. It does not stand for an entity, event,
or process occurring in an unobserved region in an organism’s body or mind,
and it should not to be reified, or assigned a causal role (Zuriff , ).
In accordance with the second scheme, Tolman’s methodology can be di-
vided in three steps, namely () picking out the appropriate independent, de-
pendent, and intervening variables; () formulating the f functions between
the independent variables and the intervening variables by operationally defin-
ing the intervening variables; and () formulating the f functions between the
intervening variables and the dependent variables. I will exemplify Tolman’s
methodology by looking at his concrete example about rats in mazes. By means
of this example I will examine the three steps in his methodology and find that,
in practice, Tolman required more than only his methodological ideas. In each
step, Tolman relied on his prescientific understanding of rats, which he used as
a heuristic guide in carrying out his methodology.
As I stated above, Tolman (/) analyzed the behavior of rats in T-
mazes as an illustration of his methodology. A T-maze consists of a very simple
maze with only one choice-point where the rats in sequence choose between
an alley to the left and an alley to the right. Only if a rat enters the alley to the
right will it find food (see fig. .).
Let us now, as first step, look more carefully at how Tolman picked out the
appropriate independent, dependent, and intervening variables. It appears that
Tolman simply proposed to look at certain variables. As a dependent variable,
Tolman proposed to investigate the “left-turning tendency” of the rats, which
he defined as the percentage of the rats that enter the left alley. In this example,
Tolman gave no reasons why he considered especially this aspect of behavior
u nd e r stan d i n g i n p s yc h ol o gy 

“WRONG” “CORRECT”

OGL OL OR OGR
BL OC BR

;^\jgZ&)#&#6I"bVoZl^i]h^c\aZX]d^XZed^ci#D82i]Zed^cid[
X]d^XZ0DA$G2i]ZXdbeaZmd[hi^bjajhdW_ZXihbZi\d^c\Ydlci]ZaZ[i$
g^\]iVaaZn0D<A$G2i]Z\dVaVii]ZaZ[i$g^\]i07A$G2i]ZaZ[i$g^\]i"ijgc^c\
WZ]Vk^dgIdabVc&.(-$&.++!&*%#

worthy of note, and not other aspects, such as whether the rats moved quickly
or slowly. One may wonder why he decided to measure the tendency of the rats
to enter the left alley, that is, the alley where there is no food. Maybe his inten-
tion for choosing this as a dependent variable was to give the impression of
being unprejudiced and therefore objective. For independent variables, Tolman
proposed to look at the number of previous trials in the maze made by the rats
and the period of food deprivation. He gave no reasons for this choice either;
apparently, he considered this choice of variables to be obvious. That it was
based on a prescientific understanding of the behavior of rats becomes visible
when we look at Tolman’s choice of intervening variables.
In Tolman’s methodology, intervening variables are introduced in the pro-
cess of breaking down the function between the independent and dependent
variables into component functions. In his example, Tolman introduced the
variable “demand” that intervenes between the period of food deprivation and
the percentage of rats that enters the left alley, and also the variable “hypoth-
eses” that intervenes between the number of trials and the percentage of rats
that enters the left alley. Although intervening variables were not supposed
to be seen as representing entities or processes, it is obvious that this way of
breaking down the function is based on the ideas that the rats are getting hun-
gry if they are deprived of food for a long time, that during the trials in the maze
the rats form hypotheses about where food can be found, and that hungry rats
will use these hypotheses in order to find food. Apparently Tolman used his
prescientific understanding of the behavior of rats as a heuristic guide in his
choice of dependent, independent, and intervening variables.
As second step, let us look at the formulation of the f functions between
 k ai eigner

the independent variables and the intervening variables. These functions form
the operational definitions of the intervening variables. Formulating these
functions was a central element in Tolman’s methodology of operational be-
haviorism. In this procedure of operationalization, Tolman made use of what
he calls “standard experiments” in which he measured the functional relation
between a specific independent and a specific dependent variable while keep-
ing the other independent variables at a constant (standard) value. The results
of these experiments formed the basis for the definition of the intervening vari-
ables. The rationale behind this second step was that fixing the definitions of
the intervening variables using an objective methodology removes the surplus
meaning. However, we will see that the definitions Tolman provided were not
entirely based on his methodological rules. This follows, for instance, already
from their dependency on “standard values,” which are not specified in Tolman’s
methodology. In most experimental situations this is not a problem because it
is obvious what the standard values are. For instance, the rats used for train-
ing experiments in a maze should not be completely underfed or overstuffed.
Determining the right (standard) level of nourishment is a judgment that is not
based on general methodological rules.
In his example, the operational definition of the intervening variable “de-
mand” was based on measurements of the relation between the period of food
deprivation and the percentage of rats that entered the left alley, where the
amount of training of all the rats was set to a standard value. The result of this
standard experiment is visible in figure ..

BL
BL + BR

%
50

M- (hours
0 12 24 36 48 60 72 since
feeding)

;^\jgZ&)#'#EZgXZciV\Zd[gVihZciZg^c\i]ZaZ[iVaaZn
Vh[jcXi^dcd[i]ZeZg^dYd[[ddYYZeg^kVi^dc#IdabVc
&.(-$&.++!&*-#
u nd e r stan d i n g i n p s yc h ol o gy 

With the increase of the number of hours passed since feeding, the number
of rats that entered the left alley decreased (and later increased). On the basis
of figure ., Tolman formulated a definition of the intervening variable “de-
mand.” To turn this figure into a definition of “demand,” Tolman argued, one
modification was necessary, namely that “demand” should really be defined as
the inverse of this ratio (Tolman /, ). This led to the following op-
erational definition of “demand” (see fig. .):

DEMAND
%
100

M- (hours
0 12 24 36 48 60 72 since
feeding)

;^\jgZ&)#(#9ZbVcYVh[jcXi^dcd[i]ZeZg^dYd[[ddY
YZeg^kVi^dc#IdabVc&.(-$&.++!&*.#

The inversion (and translation) of the graph, which is not based on meth-
odological rules, indicates that these rules are not sufficient for formulating
the operational definitions. Apparently Tolman regarded it to be essential that
demand increases with the time since feeding—at least for the first fifty hours.
That it does not increase any more and even decreases after that amount of
time is presumably an unwanted effect of this operationalization. Although
Tolman made no remark about it, he did make the judgment to stop the experi-
ment after about eighty hours, which suggests that he regarded this last period
to be irrelevant or uninteresting. This is again an example of a decision that is
not based on methodological rules.
The notable act of transforming graphs without methodological motiva-
tions is not exceptional. Similar transformations can be found in the case of
other intervening variables. For the development of the operational definition
of the intervening variable “hypotheses,” Tolman measured the function be-
tween the number of previous trials made by hungry rats and the percentage
of the rats that entered the left alley, and it appeared that with the increase of
 k ai eigner

the number of previous trials, the number of rats that entered the left alley
decreased. He transformed this function into the definition of “hypotheses” by
taking the inverse of this function. The result was, as Tolman argued, “no more
than our old friend, the learning curve” (Tolman /, ), which is a
frequently used function in neobehaviorism (see, for example, fig. . below).
In this way he provided “hypotheses” with an objective, operational definition
in terms of the number of previous trials. Tolman did not explain why both “de-
mand” and “hypotheses” should be defined as an inverse of the graphs on which
they were based. This raises the question about why Tolman had to apply these
transformations. As will be clear when we look at the next step regarding the
formulation of the f function, Tolman applied these transformations, which
were based on his prescientific understanding of the behavior of rats, because it
enabled him to attach surplus meaning to the terms in order to make his model
intelligible. Instead of ensuring that the meaning of the theoretical terms in his
behavioral model did not exceed their operational definition (as he ought to
have done, according to the philosophical doctrine he preached), Tolman in
fact ensured that these terms had surplus meaning.
As the third step, I will examine the formulation of the f function, about
which Tolman wrote, “It is by means of this f function (if we but knew what it
was) that we would be able to predict the final outcome for all possible values
of the intervening variables” (Tolman /, ). Finding the f function
is easier said than done. Tolman confessed that this is the part of his doc-
trine about which he was “haziest” (). He proposed to follow a strategy of
anthropomorphism:

I am at present being openly and consciously just as anthropomorphic

about it as I please. For, to be anthropomorphic is, as I see it, merely
to cast one’s concepts into a mold such that one can derive useful pre-
liminary hunches from one’s own human, everyday experience. These
hunches may then, later, be translated into objective terms . . . I . . .
intend to go ahead imagining how, if I were a rat, I would behave as
a result of such and such a demand combined with such and such . . .
[hypotheses] and so on. And then on the basis of such imaginings,
I shall try to figure out some sort of f rules or equations. And then
eventually I shall try to state these latter in some kind of objective and
respectable sounding terms. (–)

Although Tolman called it the haziest part of his illustration of operational

behaviorism, it actually is the part that gives us the clearest indication that in
his scientific work Tolman aimed at formulating intelligible models. In this ex-
ample it is the collection of intervening variables, like demand and hypothesis,
u nd e r stan d i n g i n p s yc h ol o gy 

that together constitute the model of the behavior of rats in mazes. Tolman
used his understanding of rats to construct this model. This understanding was
based on certain skills, of which imagining was the most important. Tolman
imagined how, if he were a rat, he would behave. This is imagining in the sense
of empathizing. By means of this skill, Tolman was able to exploit a fruitful an-
thropomorphic conceptualization. He was able to cast concepts like “demand”
and “hypotheses” into a mold such that he could derive useful hunches from
everyday experience.
Obviously the transformation of the graphs mentioned above was already
part of this framing process. The objective definitions of the theoretical terms
used in Tolman’s behavioral model of the rat in the maze were shaped in such a
way that Tolman was able to give an anthropomorphic interpretation of them.
The meaning of the theoretical terms therefore exceeded their objective defini-
tion; these terms had acquired surplus meaning. By using these terms in his
model, Tolman ensured that the model had the virtue of being interpretable in
the above-mentioned way, which made it intelligible (provided that one pos-
sessed the skill of imagining).
For Tolman this model was intelligible. First, he was able to apply the model
to the behavior of the rats, or, in other words, he was able to “do the repre-
senting.” By putting himself in the position of a rat, he was able to judge the
similarities between the model and the phenomenon of rats in mazes and pick
out the relevant features of the model (for example, from everyday life he knew
that demand is a relevant feature). Second, Tolman was able to develop quali-
tative insight into the consequences of the model in concrete situations. His
skill of imagining enabled him to derive useful preliminary hunches that he
could use to see intuitively how, in concrete situations (such and such demands
combined with such and such hypotheses), the rats were expected to behave.
The model is intelligible for Tolman, and thus provides understanding of the
phenomenon of rats in mazes.
However, one might question whether this understanding can be labeled
as scientific understanding. Tolman’s example seems to be rather unscientific,
without reference to theoretical principles. The model does not seem to be
composed of abstract objects defined by theoretical principles, as in Giere’s
() representational view of scientific models. Therefore, one could argue
that this model does not enable Tolman to use theoretical principles in the case
at hand. However, Tolman’s remark—that the graph that formed the basis for
the definition of “hypotheses” was actually the learning curve—reveals, in fact,
that Tolman’s model does incorporate theoretical principles. According to the
learning rule, which is an important principle in neobehaviorism, the learning
curve describes the function between the strength of a stimulus-response con-
nection and the number of “reinforcements” (rewards and punishments). By
 k ai eigner

identifying the hypotheses graph with the learning curve, Tolman apparently
considered the hypotheses to be a stimulus-response connection of which the
strength is defined by the principle of the learning rule.
The above analysis of Tolman’s example corresponds with the claim that in-
telligibility is an epistemic value of science. However, Tolman’s assertion—that
he would eventually try to translate the outcome of his methodology into ob-
jective and respectable-sounding terms—indicates that his striving for intel-
ligible models is merely an intermediate step in the development of objective
scientific claims. The final step would consist of a translation of his claims into
objective and respectable-sounding terms without surplus meaning. If Tolman
would succeed in removing the surplus meaning of the terms, the virtues of
the models as described above would also be gone, and accordingly the “right
combination” between the virtues of the model and the skills of the scientists
would be lost.
In actual scientific practice, do neobehaviorists like Tolman rigorously
perform this last step? And, if not, is there a special reason why the surplus
meanings should not be annihilated? In the article “On a Distinction between
Hypothetical Constructs and Intervening Variables,” which caused ardent dis-
cussion among neobehaviorists and logical positivists, MacCorquodale and
Meehl () dealt with these questions. They ascribed to Tolman the position
that theoretical terms should not have surplus meaning, whereas they ascribed
to Hull the opposite position. Tolman () contributed to this discussion
with an article in which he advocated—quite remarkably according to some of
his colleagues—that theoretical terms should have surplus meaning: “I am . . .
convinced that ‘intervening variables’ to which we attempt to give merely op-
erational meaning . . . really can give us no help unless we can also imbed them
in a model from whose attributed properties we can deduce new relationships
to be looked for” (). As he admitted, even Tolman himself was surprised by
his own statement, because for many years he had objected to it. Nevertheless,
he came to realize that it is “inevitable and to be desired” to provide the theo-
retical terms with surplus meaning by framing them in a model that provides a
conceptual substrate—“a substrate which is endowed by its author with certain
intrinsic properties of its own.” (–)
The first part of my case study, about Tolman, highlights the important role
of the surplus meaning of theoretical terms used in models for enabling the
right combination between the virtues of the model and the skills of the scien-
tist, which brings about understanding of the phenomena in the case at hand.
The second part of the case study, about Clark L. Hull, will show that this sur-
plus meaning cannot be removed; it is essential for scientific knowledge, thus
revealing the epistemic relevance of intelligibility.
u nd e r stan d i n g i n p s yc h ol o gy 

8aVg`A#=jaa
Hull’s methodology can be seen as an elaboration of Tolman’s objective be-
haviorism. Hull used Tolman’s idea to operationally define theoretical terms
and, like Tolman, he constructed ingenious theoretical models of the adaptive
behavior (or learning behavior) of organisms. Hull’s methodology was very in-
fluential. At the peak of his career Hull was the most widely known behaviorist,
and his methodology dominated psychology (Smith , ). In , Hull’s
magnum opus Principles of Behavior appeared, which was not only “the begin-
ning of an attempt to sketch a systematic objective theory of the behavior of
higher organisms” (Hull a, ) but also a clarification of Hull’s view of sci-
ence and methodology.
Hull’s ideas about science were in many respects closely similar to those of
the logical positivists. He shared the interest in the use of formal logic and was
attracted to British empiricism. Hume’s Treatise of Human Nature especially
fascinated him (Smith , ). Hull regarded behaviorism as a direct de-
scendant from British associationism, of which Hume’s work was exemplary,
and he shared Hume’s admiration for Newton’s scientific method. In a lecture
for the international congress of the Unity of Science Movement of  in
Paris, Hull described his scientific procedure as “logical empiricism.” In this
lecture, and also in the lecture for the international congress of the Unity of
Science Movement of  in Chicago, where he was an invited speaker, Hull
emphasized the kinship between logical positivism and behaviorism.

There is a striking and significant similarity between the physicalism

doctrine of the logical positivists (Vienna Circle) and the approach
characteristic of the American behaviorism originating in the work
of J. B. Watson. Intimately related to both of the above movements
are the pragmatism of Peirce, James, Dewey on the one hand, and the
operationism of Bridgman, Boring and Stevens, on the other. These
several methodological movements, together with the pioneering ex-
perimental work of Pavlov and the other Russian reflexologists, are, I
believe, uniting to produce in America a behavioral discipline which
will be a full-blown natural science; this means it may be expected
to possess not only the basic empirical component of natural science,
but a genuinely scientific theoretical component as well. (Hull b,
)

Hull’s work can be seen as an attempt to apply the methodology of logical em-
piricism in psychology. In his view, by adopting the logical positivistic ideal
of objectivity, psychology could become a natural science. To put it simply, as
Hull did in his  lecture, this “methodology of logical empiricism” has three
 k ai eigner

phases. “The methodology begins with an empirical determination of its pos-

tulates, and ends with an empirical check on the objective validity of its theo-
rems; between these two there lies the integrating symbolic structure of logic
and mathematics. Thus arises the kinship to ‘logical empiricism’ ” (Hull ,
–).
Although Hull was an advocate of a positivistic psychology, by which he
meant, in particular, a psychology that was in agreement with the logical empiri-
cist norm of objectivity, it was disputed whether in his own scientific work Hull
lived up to this requirement. For instance, MacCorquodale and Meehl ()
argued that important theoretical terms in Hull’s work had surplus meaning. A
few years earlier, Kenneth W. Spence, one of Hull’s former pupils, had already
expressed similar worries about Hull’s use of theoretical terms. In his writings,
Hull often supplemented the formal definitions of hypothetical constructs with
informal formulations in which the constructs were depicted as mechanisms
mediating between stimuli and responses. Spence discarded this, and in 
and  he wrote several letters in which he repeatedly warned Hull against
his habitual practice: “I have always been very unhappy about the fact that you
have been inclined to throw in hypotheses as to the mediational mechanisms
underlying the abstract mathematical concepts” (Spence to Hull, quoted in
Smith , ). Illustrative is the comparison Spence made between Hull
and the physicist Maxwell, who considered it to be necessary for a scientific
explanation to involve mechanisms: “What you are doing would be analogous
to Maxwell insisting on the mediational mechanism of an ether-like medium
to explain electromagnetic phenomena rather than depending upon the purely
mathematical aspects of his theory” (Spence to Hull, quoted in Smith ,
).
This habit of giving an informal, mechanical interpretation of theoreti-
cal terms formed one of the sources of surplus meaning. Another source was
formed by naming the theoretical terms. Like Tolman, Hull used names that
were already meaningful, such as “drive,” “demand,” “habit strength,” and “re-
inforcement.” Although he formulated his scientific claims in objective terms
with operational definitions, this caused the meaning of these terms to exceed
these definitions. We will see that Hull provided theoretical terms with surplus
meaning because, as for Tolman, this enabled him to connect the theoretical
principles of behavior to concrete behavior of higher organisms through the
construction of intelligible theoretical models.
For an illustration of Hull’s approach, I will focus on his use of the theoreti-
cal terms “reinforcement” and “habit strength,” which were important terms
in the learning principle as described in his Principles of Behavior (a). In
his ideas about adaptation and learning of organisms, Hull was strongly influ-
enced by Pavlov’s theory of conditioned learning. Hull saw the reflex between
u nd e r stan d i n g i n p s yc h ol o gy 

receptors and effectors as a causal mechanism. He even built a mechanical ap-

paratus to simulate the conditioned reflex. Hull wrote that receptor-effector
connections of an organism roughly correspond to what are commonly known
as habits (Hull a, ). The strength of such a connection he called “habit
strength.” Hull’s theoretical term “habit strength” is comparable to Tolman’s
theoretical term “hypotheses.” As mentioned above, Tolman argued that the
graph he used to give a definition of “hypotheses” as a function of the number
of previous trials was no more than the learning curve. The shape of the graph
was similar to that of the learning curve, and the parameters were also similar:
the trials of the rats in the maze, which could result in food (reward), were a
form of reinforcement. Just as Tolman did with “hypotheses,” Hull used the
learning curve to give an operational definition of “habit strength” as a function
of successive reinforcements (see fig. .).

100
UNITS OF HABIT STRENGTH ( S H R )

80

60

40

20

0 5 10 15 20 25 30
SUCCESSIVE REINFORCEMENTS (N)

;^\jgZ&)#)#I]ZaZVgc^c\XjgkZ#=jaa&.)(V!&&,#

Hull’s “habit strength” is more general that Tolman’s “hypotheses.” In his

attempt to be as general as possible, Hull did not give a specification of the
meaning of “reinforcement.” It is simply a condition that increases the habit
strength as indicated by the learning curve (Hull a, ). This definition of
habit strength exemplifies Giere’s view that theoretical principles (the learn-
ing curve) function as definitions of the abstract objects (habit strength) that
compose theoretical models. For concrete situations, one has to specify what
counts as reinforcement. An example of such a concrete situation is Tolman’s
T-maze, where reinforcement is indicated by the number of previous trials of
the rats in the maze. By giving a “local” (see Radder ) interpretation of
the meaning of the theoretical terms in his behavioral models, Hull was able
 k ai eigner

to connect these models to concrete instances of learning behavior of higher

organisms.
The connections between theoretical principles and phenomena have em-
pirical relevance in the sense that they are required for the empirical justifica-
tion of the knowledge contained in the theoretical principles. I will investigate
how Hull applied theoretical principles to concrete phenomena and analyze
the role played by the surplus meaning of the theoretical terms in these prin-
ciples. We will see that there are similarities between the way in which Hull
constructed his models to connect principles to phenomena and the way Tol-
man constructed his model as described in the first part of this case study. The
learning models that Hull constructed were composed of theoretical terms,
such as “habit strength” and “reinforcement,” that were defined by the theo-
retical principles of behavior. Like Tolman, Hull put surplus meaning into the
theoretical terms so that the models became intelligible. They could be used
to represent the phenomena and to develop qualitative insight into the conse-
quences of the model in concrete situations.
In Principles of Behavior (a), Hull discussed a number of experiments,
which he claimed were all measurements of habit strength in an “indirect way”
(). These measurements were indirect because Hull’s experiments only mea-
sured functional dependencies between independent and dependent variables,
and not between these variables and habit strength. In that sense, these indirect
measurements are similar to Tolman’s standard experiments, in which also only
functional dependencies between independent and dependent variables were
measured, and not between these variables and the intervening variables. It is
remarkable that these experiments—which Hull argued were measurements
of the same theoretical term “habit strength”—were totally different from each
other. They were executed with different organisms as subjects of investigation
and with totally different techniques. How was Hull able to argue that all these
experiments were about habit strength? Or more precisely, how did Hull make
the connection between his learning principles, which were used to define hab-
it strength, and the different experiments?
In one of the experiments, a human subject repeatedly received a mild elec-
tric shock after hearing a noise. After this training period, only the noise was
presented and the strength of the galvanic skin reaction of the subject was mea-
sured as a function of the number of “reinforcement repetitions” received in
the training period. In the graph in which the result was depicted (see fig. .),
Hull recognized the learning curve and argued that this meant that the result of
the experiment was a manifestation of habit strength (Hull a, ). Such a
claim can only be made if it is reasonable to suppose that receiving a mild elec-
tric shock after hearing a noise can be seen as (negative) reinforcement, and
that it is reasonable to see the galvanic skin reaction, after receiving a shock,
u nd e r stan d i n g i n p s yc h ol o gy 

as a habit of which the amplitude corresponds to the habit strength. Appar-

ently Hull considered this to be reasonable, and the similarities with Tolman’s
example make clear that he had this opinion because of the surplus meaning of
theoretical terms, such as “habit strength” and “reinforcement.” By being em-
pathic (comparable with Tolman’s empathy mentioned above), Hull is able to
understand that receiving a mild electric shock is a form of reinforcement. The
role played by the surplus meaning of these terms is similar to that played in
Tolman’s case, where the surplus meaning of “demand” made it possible to view
this theoretical term in the light of everyday experiences, which enabled mak-
ing the connection with concrete situations.
In another experiment, albino rats were trained to depress a bar in order to
obtain food. After this training period, bar pressing was no longer followed by
a reward and the number of unreinforced reactions of the rats was measured as
a function of the number of “reinforcement repetitions” in the training period.
Again Hull recognized the learning curve in the graph that depicted the result
(see fig. .), and again he argued that this meant that the result of the experi-
ment was a manifestation of habit strength (Hull a, ).
Apparently the meaning of the theoretical terms transcends the “local”
meaning they have in concrete situations. To put it in the words of Hans Rad-
der (, esp. chap. ), these terms have “nonlocal meaning”: their scope is
not constrained to a specific domain but can be enlarged by successfully apply-
ing these terms to a materially different domain (Radder , ). In Hull’s

16
AMPLITUDE OF GALVANIC SKIN REACTION (A)

14

12

10

0
0 8 16 24 32 40 48
NUMBER OF REINFORCEMENT REPETITIONS (N)

;^\jgZ&)#*#BVc^[ZhiVi^dcd[]VW^ihigZc\i]^cVcZmeZg^bZcil^i]
]jbVchjW_ZXih#=jaa&.)(V!&%(#
 k ai eigner

NUMBER OF REACTIONS TO PRODUCE EXTINCTION (n)

16

14

12

10

0
0 10 20 30 40 50 60 70 80 90
NUMBER OF REINFORCEMENT REPETITIONS (N)

;^\jgZ&)#+#BVc^[ZhiVi^dcd[]VW^ihigZc\i]^cVcZmeZg^bZcil^i]
VaW^cdgVih#=jaa&.)(V!&%+#

view, the theoretical term “habit strength” can successfully be applied to dif-
ferent domains: he regards his experiment with the rats as a replication of the
experiment with the electric shock—they are both experiments that measure
the functional relationship between habit strength and the number of rein-
forcement repetitions, even though the material realization of the experiments
differs completely. Indeed, the experiments produce similar graphs, but this is
not enough to call them replications of each other. This also requires certain
judgments, for instance, that it is reasonable to suppose that obtaining food
after depressing a bar can be seen as reinforcement, and that it is reasonable to
see the number of unreinforced reactions made by the rats as corresponding to
habit strength. In my view, Hull was able to make these judgments, by means
of which he could apply his behavioral model in different domains, by virtue of
the surplus meaning of the terms in this model.
This goes beyond the neobehaviorist analysis, for in that analysis the con-
nection between model and learning phenomenon is not allowed to depend on
judgments of the user of the model, but instead should be stated in objective
terms by means of operational definitions. For every situation, the neobehav-
iorist ought to provide operational definitions for “stimuli,” “responses,” and
“reinforcement.” In a specific experimental situation concerning humans hear-
ing a noise, “reinforcement” is to be defined as receiving a mild electric shock
after hearing the noise. In another experimental situation concerning rats de-
u nd e r stan d i n g i n p s yc h ol o gy 

pressing bars, “reinforcement” is to be defined as obtaining food after depress-

ing the bar.
It was a point of discussion among neobehaviorists whether this maneuver
of giving operational definitions to the theoretical terms implied a restriction
to the domain of application, which was the view of strict operationalism (Zu-
riff , –). For Hull’s theoretical principles this would mean that their
domain of application would only cover Hull’s examples about habit strength in
his Principles of Behavior (a). The principles would only denote relation-
ships among specific variables in a very limited domain, and accordingly have
no epistemological relevance in other domains. They would, in fact, not even
be applicable in situations very similar to one of the examples, because that
would require slightly different operational definitions of the theoretical terms.
But this is not in accordance with Hull’s intention, which was to formulate a
general theory of behavior that is epistemologically relevant in a much larger
domain than merely the domain covered by his examples. Therefore, more ap-
propriate to scientific practice is the competing view among neobehaviorists,
namely that operational definitions should be seen as partial or “open” and not
as complete definitions (Zuriff , ). This view leaves open the possibility
of extending the domain of application. The meaning of theoretical terms like
“reinforcement” and “habit strength” is not entirely captured by their opera-
tional definitions. Therefore, neobehaviorists who share this view would agree
that theoretical terms bear nonlocal meaning in the sense discussed above.
Hull never specified how the domain of application of the theoretical prin-
ciples might be extended. It seems that this was obvious to him. In his method-
ology he never discussed how the terms in his models should relate to concrete
situations. He simply attached them to concrete situations. Apart from the
two experiments discussed above, Hull gave numerous other examples of in-
direct measurements of habit strength in experiments with animals, such as
rats, dogs, chickens, and also with human subjects, and in all of these different
situations Hull recognized the learning curve and was able to relate the experi-
ment to his learning model. He simply knew what to call “stimuli,” what to call
“responses,” and what to call “reinforcement.” In my view, the reason for this
is that the models were intelligible to him due to the surplus meaning of the
theoretical terms.
For, in the first place, Hull had no difficulty applying his behavioral model to
the concrete behavior of higher organisms; in other words, he was able to “do
the representing.” In a way similar to Tolman, Hull was able to do this because
of his skill in imagining, in the sense of empathizing. Because Hull knew the
surplus meaning of the theoretical term “reinforcement,” and because he was
able to imagine himself in the position of the human subjects who received
a mild electric shock after hearing a noise, or in the position of the rats who
 k ai eigner

obtained food after pressing a bar, Hull was able to identify the shocks and the
food as being cases of “reinforcement.” By means of these kinds of identifica-
tions Hull could apply his model to concrete situations.
Second, Hull was able to use the model for developing qualitative insight
into the consequences of the model in concrete situations. Undoubtedly, like
Tolman, Hull could derive useful hunches about this due to the above-men-
tioned skill of imagining. In addition, he could use the causal-mechanical sur-
plus meaning he had allocated to the theoretical terms in the behavioral model.
Because he interpreted “habit strength” as the outcome of a causal mechanism,
of which he had even built a simulation in the form of a mechanical apparatus,
Hull could use his skill of causal reasoning to develop qualitative insight in the
consequences of the behavioral model.
The intelligibility of the models induced by the surplus meaning of the theo-
retical terms is epistemologically relevant. Making the model unintelligible by
removing the surplus meaning (if that is at all possible) would have the conse-
quence that scientists lack the occasion of using their skills to apply the model
to concrete behavior. Except for the small domain in which operational defini-
tions of the theoretical terms are in effect (Hull’s examples), the connection
between model and phenomenon would be broken. Without this connection,
the theoretical principles in the model have no empirical content, which would
make them epistemologically insignificant. The case study about Hull there-
fore shows us that there are epistemological reasons for not removing surplus
meaning. Scientific models need to be intelligible.

8dcXajh^dc

The aim of this chapter is to demonstrate, from scientific practice, that in-
telligibility is one of the epistemic values that are constitutive of science; or to
put it in other words, that understanding is required for scientific knowledge.
The account of understanding used in this chapter is based on the idea that
merely possessing relevant theoretical knowledge is not enough for the scien-
tific understanding of phenomena. In addition, one should be able to use this
knowledge, which implies applying it to concrete cases by means of intelligible
models. A model is intelligible if it can be used to represent the phenomenon
and if its characteristic consequences can be seen intuitively. Whether a model
is intelligible for a user depends not only on the characteristics of the model but
also on the user’s skills.
The point of departure of the case study discussed in this chapter is Tolman’s
concrete example of rats in mazes, by which he exemplified his ideas of an ob-
jective psychological methodology. Because Tolman used his prescientific un-
derstanding of rats as a heuristic guide in formulating an abstract model about
u nd e r stan d i n g i n p s yc h ol o gy 

their behavior, the meaning of the terms in the model was initially based on
this prescientific understanding. By applying his methodology, Tolman aimed
at fixing the meaning of these terms by objective definitions. But, as discussion
of his example shows, Tolman did not reach this goal. The terms in the model
could not be divested of all their surplus meaning originating from prescien-
tific understanding. In fact, a close look at Tolman’s example reveals that this
surplus meaning rendered the model intelligible to him. The right combina-
tion of Tolman’s skills (for example, his ability to imagine in the sense of to
empathize) and the virtues of the model originating from the surplus meaning
of the theoretical terms used in the model (such as its anthropomorphic inter-
pretability), enabled him to use the model for representing the phenomena of
rats in mazes and for developing qualitative insight into its consequences in
concrete situations.
The analysis of Tolman’s example illustrates the notion of the intelligibility
of a model. That this intelligibility is essential for scientific knowledge of phe-
nomena is demonstrated by an analysis of the models of behavior in the work of
Hull, who developed a very influential methodology for neobehaviorism, which
was an elaboration on Tolman’s work. The models of behavior in Hull’s work
contain theoretical terms that are defined by theoretical principles of behavior
(for example, by means of the learning curve). The epistemological value of
these models, and thereby of the principles of behavior, depends on the pos-
sibility of applying these models to concrete behavioral phenomena. Without
this possibility the theoretical principles lack empirical content. A close look at
Hull’s work reveals that establishing the connection with concrete phenomena
necessarily involves judgments that are not based on objective methodological
rules, for instance about the similarities of features of the models with features
of the phenomena and about the relevance of these similarities. Hull was able
to make these judgments (seemingly with ease) due to the intelligibility of the
models that, in a way similar to the intelligibility of Tolman’s model of the rat,
accrued through the surplus meaning of the theoretical terms in the models.
Only because the theoretical terms possessed surplus meaning was Hull able
to make the connection between model and phenomenon. This reveals that the
epistemological value of the behavioral models, and thereby of the theoreti-
cal principles of behavior, depends on the intelligibility of these models, which
supports the claim that intelligibility is one of the epistemic values that are
constitutive of science.
One cannot confirm, in a straightforward way, a philosophical claim about
the epistemic role of intelligibility and understanding in science. To elucidate
such a claim by means of empirical evidence from scientific practice requires
well-chosen case studies. The motivation for the choice of neobehaviorism as
a case study lies in the extraordinary positivistic inclination of the neobehav-
 k ai eigner

iorists that originated from logical positivistic influences. The pursuit of the
neobehaviorists to set up an objective science of psychology, released from all
subjective elements—which in the end would also leave no room for intelligi-
bility and understanding—can be seen as a test case against the philosophical
idea that understanding plays a fundamental role in science. An analysis of the
scientific practice of the neobehaviorists, who did not reach the goal of an ob-
jective science, reveals that they instead came to realize that the use of intelli-
gible models is inevitable and even to be desired. Their embrace of intelligibility
as an epistemic value forms a corroboration of the claim that understanding is
more than a surplus epiphenomenon. It demonstrates that understanding is a
requirement for scientific knowledge.

CdiZh
I would like to thank Henk de Regt, Hans Radder, and Sabina Leonelli for their helpful
criticism and suggestions, and especially Hans Radder for his constructive ideas about the
nonlocal meaning of theoretical terms.

GZ[ZgZcXZh
Bailer-Jones, D. M. . Tracing the development of models in the philosophy of science.
In Model-based reasoning in scientific discovery, edited by L. Magnani, N. J. Nerses-
sian, and P. Thagard, –. New York: Kluwer Academic/Plenum.
———. . When scientific models represent. International Studies in the Philosophy of
Science :–.
Cartwright, N., T. Shomar, and M. Suárez. . The tool box of science: Tools for the
building of models with a superconductivity example. Poznań Studies in the Philoso-
phy of the Sciences and the Humanities :–.
de Regt, H. W. . Discussion note: Making sense of understanding. Philosophy of Sci-
ence :–.
de Regt, H. W., and D. Dieks. . A contextual approach to scientific understanding.
Synthese :–.
Giere, R. N. . Using models to represent reality. In Model-based reasoning in scientific
discovery, edited by L. Magnani, N. J. Nersessian, and P. Thagard, –. New York:
Kluwer Academic/Plenum.
———. . How models are used to represent reality. Philosophy of Science :–.
Hartmann, S. . Models and stories in Hadron physics. In Models as mediators:
Perspectives on natural and social science, edited by M. S. Morgan and M. Morrison,
–. Cambridge: Cambridge University Press.
Hergenhahn, B. R. . An introduction to the history of psychology. rd ed. Pacific Grove,
CA: Brooks/Cole.
Hull, C. L. . Logical positivism as a constructive methodology in the social sciences.
Einheitswissenschaft :–.
u nd e r stan d i n g i n p s yc h ol o gy 

———. a. Principles of behavior. New York: Appleton-Century.

———. b. The problem of intervening variables in molar behavior theory. Psychological
Review :–.
Kuhn, T. S. . Objectivity, value judgment, and theory choice. In The essential tension:
Selected studies in scientific tradition and change, –. Chicago: University of
Chicago Press.
Lacey, H. . Is science value free? Values and scientific understanding. nd ed. London:
Routledge.
Longino, H. . Science as social knowledge. Princeton: Princeton University Press.
MacCorquodale, K., and P. E. Meehl. . On a distinction between hypothetical con-
structs and intervening variables. Psychological Review :–.
McMullin, E. . Values in science. In PSA , vol. , edited by P. D. Asquith and T.
Nickles, –. East Lansing, MI: Philosophy of Science Association.
Radder, H. . The world observed / The world conceived. Pittsburgh: University of Pitts-
burgh Press.
Smith, L. D. . Behaviorism and logical positivism: A reassessment of the alliance. Stan-
ford: Stanford University Press.
Tolman, E. C. /. Psychology versus immediate experience. In Behavior and psy-
chological man, –. Berkeley and Los Angeles: University of California Press.
———. /. Behavior at a choice point. In Behavior and psychological man, –.
Berkeley and Los Angeles: University of California Press.
———. . Discussion from “Interrelationships between perception and personality: A
symposium.” Journal of Personality :–.
Watson, J. B. . Psychology as the behaviorist views it. Psychological Review :–.
Zuriff, G. E. . Behaviorism: A conceptual reconstruction. New York: Columbia Univer-
sity Press.
 pa rt title ver so

(,
Understanding in Political Science
The Plurality of Epistemic Interests

? : G D : C  K6 C  7 D J L : A

 Upon a first encounter with the field of International Relations (IR)

studies, we stumble into a plurality of theoretical perspectives some of which,
such as realism and liberalism, have already been around for decades, while
others, such as constructivism, are more recent. A recent survey among IR
scholars working in the United States gives us a rough idea of the weight at-
tached to these different perspectives. Answering the question “What paradigm
in International Relations are you primarily committed to in your research?”
 percent chose realism/neorealism,  percent liberalism/neoliberalism, 
percent constructivism,  percent Marxism/globalism, and  percent other,
among which the most significant were rational choice theory and feminism.1
An analysis of the distribution of scholars among these six most important
theoretical perspectives in IR makes it clear that the field is neither converging
toward a broad consensus nor completely scattered. In what follows, I analyze
the dynamics of the field, and in particular two debates that concern the plural-
ity of theories and the search for unity. The first debate regards the possibility
and desirability of substituting the existing theoretical perspectives with one
perspective, that is, rational choice theory. The second debate explores the pos-
sibility of a synthesis of (some of ) these different perspectives. Through this
analysis, I will show what the participants’ idea of (increased) understanding
might be—that is, where the dynamics of the field should lead us—and, subse-
quently, I will try to make this idea of understanding philosophically explicit.


u nd e r stan d i n g i n p ol i ti c a l s ci e nce 

Finally, this philosophical elaboration is illustrated through a discussion of the

relation between feminism and rational choice theory within the practice of
social science.

Jc^ÒXVi^dcVcYJcYZghiVcY^c\
In a traditional view, scientific understanding is linked to unification, or
reducing the number of theories (see Kitcher ). From this point of view,
the question of whether the unification of the different theoretical perspectives
in IR would increase understanding has to be answered positively: “Science
advances our understanding of nature by showing us how to derive descrip-
tions of many phenomena, using the same patterns of derivation again and
again, and, in demonstrating this, it teaches us how to reduce the number of
types of facts we have to accept as ultimate (or brute)” (Kitcher , ). In
contrast, I intend to show that an analysis of social scientific practice and the
reactions of scholars to unificationist attempts indicate the need for a more
nuanced answer to this question. This analysis can provide us with a deeper
insight into scientific understanding in social science, which is different from
the traditional view.
Several projects in the social sciences have attempted to unify the field. One
of the projects I discuss here—and the one that is considered the most success-
ful nowadays—is the incorporation of the social sciences within the framework
of rational choice theory and game theory. In order to test whether Kitcher’s
() view holds, then, we have to ask: do we achieve increased understand-
ing through such an incorporation within a single overarching framework? The
second project I discuss concerns the quest to reduce the plurality of perspec-
tives through a synthesis. Does such a unifying synthesis increase understand-
ing? Let us start with the promises of rational choice theory.

I]ZEgdb^hZhd[GVi^dcVa8]d^XZI]Zdgn

In the social sciences, one of the most successful unificationist projects

is the so-called economics takeover, or economics imperialism.2 After a short
introduction to this project, I will focus on the crucial question of whether
this advocacy of rational choice theory increases understanding in political
science.
The central idea of the economics takeover is that economists (and oth-
er social scientists) improve knowledge in the social sciences by applying the
dominant or orthodox theory and method in economics far beyond its original
home. The aim is to increase our understanding of the social world, of social
action, by advocating a unifying theory for the individual in the social sciences,
 jer o en van b o u wel

one that is closely linked to economics and the idea of the homo economicus.
“Economics imperialism is a matter of persistent pursuit to increase the de-
gree of unification provided by rational choice theory by way of applying it to
new types of explanandum phenomena that are located in territories that are
occupied by disciplines other than economics” (Mäki , ). As such, it
seems to be the best candidate in the social sciences with which to test Kitcher’s
() view.
Characteristic features of the economics takeover are the development of
formal models for the social sciences, whatever the subject matter; the inclusion
of the notion of maximizing agents, in line with rational choice theory (RCT);
and a general downgrading of the significance of history, culture, and dynamics
(in space and time). Current advocates of the unificationist project include, for
example, Herbert Gintis (see his  work, in which he argues that—finally—
the conditions for unity in the behavioral sciences have been created, referring
to the theoretical tools of rational actor model and game theory); Robert Bates,
Rui de Figueiredo, and Barry Weingast, who “explore the possibilities for theo-
retical integration by suggesting how some of the fundamental concepts used
by interpretivists can be incorporated into rational choice theory” (Bates et al.
, ); Margaret Levi, Elinor Ostrom, and James Alt, who “expect the next
century to witness a major flowering of scientific achievement across the social
sciences similar to the neo-Darwinian synthesis of this past century in biology”
(Levi et al. , ).
Advocates of the economics takeover and the application of RCT in po-
litical science might claim that they provide explanatory mechanisms that are
transparent and coherent, and that their explanatory theory is eminently plau-
sible, and so should be confidently accepted, as a means of engendering under-
standing. However, it is questionable whether this unified explanatory theory,
and the class of understanding it purports to convey, are generally accepted as
providing (increased) understanding. Even if this unification induces a kind of
understanding (given some epistemic interests), certain critiques formulated
within political science point out that it does not provide the kind of under-
standing desired by some groups and communities in society.
Let us look at some of these critiques resisting unification. Some critics
start by noting that the economics takeover (that is, increasing the degree of
unification by applying rational choice theory in territories outside economics,
encapsulating political science, sociology, anthropology, history, and so on) has
indeed, especially in political science, led to a growing percentage of journal
articles using the RCT perspective. This has been documented for political sci-
ence in general by Donald Green and Ian Shapiro () and for international
relations studies by Stephen Walt (). These critics then go on to argue that
much of this growing body of formal RCT literature is irrelevant. According to
u nd e r stan d i n g i n p ol i ti c a l s ci e nce 

them, it does not answer important societal questions and “real-world prob-
lems.” Concerning the discipline of international relations studies, Stephen
Walt (, ) has drawn the following conclusion:

In this sense, much of the recent formal work in security studies re-
flects the “cult of irrelevance” that pervades much of contemporary
social science. Instead of using their expertise to address important
real-world problems, academics often focus on narrow and trivial
problems that may impress their colleagues but are of little practical
value. If formal theory were to dominate security studies as it has oth-
er areas of political science, much of the scholarship in the field would
likely be produced by people with impressive technical skills but little
or no substantive knowledge of history, politics, or strategy.

It is as a result of this tendency that the relevance of academic political science

for (parts of ) the public is decreasing.3
The unificationist project of the economics takeover has been labeled
theory- and method-driven, as opposed to problem- or question-driven. This
method-driven research leads, according to Shapiro (, ), to the “self-
serving construction of problems, misuse of data in various ways, and related
pathologies summed up in the old adage that if the only tool you have is a ham-
mer everything around you starts to look like a nail.” For example, “making a
fetish of prediction can undermine problem-driven research via wag-the-dog
scenarios in which we elect to study phenomena because they seem to admit
the possibility of prediction rather than because we have independent reasons
for thinking it worthwhile to study them” () This is the economics takeover
perceived as eager to advance the dominant economic theory and method by
the self-serving construction of political problems and by substituting existing
theories and/or methods in political science with its own, but not really inter-
ested in finding answers to unaddressed political questions.4
Reactions such as these show that some people do not get satisfactory an-
swers when they apply RCT in a (new) context related to their political ques-
tions and epistemic interests. Similar reactions have arisen recently within the
economics discipline itself, criticizing the dominance of neoclassical or main-
stream economics (related to RCT). Some students and researchers are left
with questions. A notable example is the recent petition signed by French stu-
dents (which appeared in the French newspaper Le Monde, June , ). One
quote from this is telling:

Most of us have chosen to study economics so as to acquire a deep

understanding of the economic phenomena with which the citizens of
 jer o en van b o u wel

today are confronted. But the teaching that is offered, that is to say for
the most part neoclassical theory or approaches derived from it, does
not generally answer this expectation. Indeed, even when the theory
legitimately detaches itself from contingencies in the first instance, it
rarely carries out the necessary return to the facts. . . . Furthermore,
this gap in the teaching, this disregard for concrete realities, poses an
enormous problem for those who would like to render themselves use-
ful to economic and social actors.5

In this discipline too, then, a number of questions are unsatisfactorily addressed

by the unifying theory. Instead, the French students want “a pluralism of ap-
proaches, adapted to the complexity of the objects and to the uncertainty sur-
rounding most of the big questions in economics (unemployment, inequalities,
the place of financial markets, the advantages and disadvantages of free-trade,
globalization, economic development, etc.).”6
Taking into account the above complaints and critiques from scientific
practitioners and students of the economics takeover, one tends to conclude
that the unificationist project has not necessarily increased understanding.
Several phenomena and real-world problems are being neglected by it (due,
critics claim, to unificationist attempts to apply theory- or method-driven ap-
proaches to address them). And yet it seems crucial for understanding that the
problems or questions that people (not as a unity but rather as a plurality) face
and want to answer and understand are taken into account.

I]Z>YZVd[VHnci]Zh^h

One important unificationist attempt is, as discussed above, to substitute

or replace the main theoretical perspectives in the international relations dis-
cipline, that is, realism, liberalism, constructivism, Marxism, feminism, and ra-
tional choice theory, with the last of these. A more general unificationist way
of dealing with the variety of theoretical perspectives is to plea for a synthesis
(rather than a substitution).
In his contribution to a forum discussing the question “Are dialogue and
synthesis possible in international relations?” Moravcsik () presents such
a plea. This is a view that is very common among social scientists, that scien-
tific progress is made when a set of competing theories is replaced with one
synthesizing theory: “Theory synthesis is not only possible and desirable but
it is constitutive of any coherent understanding of international relations as a
progressive and empirical social science” (Moravcsik , ). There is “the
need to combine theories to explain complex real-world events” ().
According to Moravcsik, “We should think more about the ways in which
u nd e r stan d i n g i n p ol i ti c a l s ci e nce 

theoretical syntheses might help us to understand concrete events in world

politics” (, ).7 In light of what follows, it is interesting to note that
Moravcsik uses an unqualified (“unitary,” one might say) “us” when talking
about understanding. As such, his plea not only endorses the synthesis, but
also rejects the pluralism advocated by other contributions to the same forum.
In contrast, some defend pluralism because they are skeptical about the
possibility of a synthesis:

The thrust of the questions is that dialogue and synthesis are all of
one cloth, that everything can be debated out, and that some integral
new whole is likely to emerge to command our assent if we all do our
homework. But considering how seldom debates establish such con-
sensus, it might be useful to take this experiential datum as a start-
ing point and to inquire into the reasons why communication across,
and often even within, different theoretical perspectives is so difficult.
(Kratochwil , )

Others endorse pluralism as a positive value, arguing that a synthesis theory

as defended by Moravcsik, though it may be “disguised” as a complete view,
is actually only a partial view. An example of this viewpoint is found in Smith
(): “Dialogue is not going to be easy, or even possible, in international rela-
tions until the discipline becomes less dominated by a narrow orthodoxy re-
flecting historically and culturally specific interests. Such a change will take time
but until it is achieved the discipline will continue to reflect one limited, partial
view about the structures and processes of the one world of international poli-
tics” (; italics mine). In this view, the surplus value of plurality and dialogue
should be emphasized and any acceptance of a synthesis theory as the unifying
theory (making all others obsolete) should be considered a deterioration of the
discipline: “In contrast [with synthesis], dialogue certainly implies a willingness
to concede that a particular theory might only be of limited or partial relevance
and that other viewpoints are worthy of consideration” ().
The convenor of the forum on dialogue and synthesis in the international
relations discipline, Gunther Hellmann, draws the following conclusion: “even-
tually the argument boils down to the question of how useful we find looking at
the world from the perspective we choose” (Hellmann , ). Hellmann’s
emphasis on usefulness leads me to the explication of understanding.

I]ZGdaZd[:e^hiZb^X>ciZgZhihVcY6YZfjVXn^cJcYZghiVcY^c\

In order to explicate understanding and to grasp the role of usefulness in

it let me first refer to my earlier work on explanations in social science (Van
 jer o en van b o u wel

Bouwel and Weber ; Weber and Van Bouwel ). There I emphasized
the importance of taking epistemic interests into account when making a
choice among different forms of explanation (and the plurality of theories
from which these explanations derive) and regarded explanations as answers
to explanation-seeking why-questions, the formulation of which helps to make
the explananda as explicit as possible and to draw attention to the underlying
epistemic interests of the explainee. From this viewpoint, a good explanation
should not be understood as merely the accurate (in relation to reality) ex-
planation of a social phenomenon, but also as an adequate (in relation to the
epistemic interests of the explainee) answer of an explanation-seeking ques-
tion concerning a social phenomenon.8 Usefulness depends on adequacy as
well as accuracy.
Analogous considerations have to be made when talking about understand-
ing. If we want to get a grip on what achieving understanding requires in social
science, attention should be paid to the (variety of ) interests of the understand-
ees.9 Understanding does not depend merely on the accuracy of the explana-
tions provided by a theory but also on its adequacy in relation to the interests
of the understandee.
In the objectivist view of understanding of Trout () this distinction
between accuracy and adequacy is missing. For Trout the accuracy suffices
to provide understanding. De Regt (, ) rightly criticizes Trout’s view
and pleads for the inclusion of pragmatic elements: “Understanding is not only
knowing the formula, but in addition being able to use the formula in the case
at hand.” Accuracy is not the only criterion: “Scientists prefer a more intelligible
theory over a less intelligible one, sometimes even at the cost of some accuracy,
not because it gives them a ‘feels right’ sense of understanding but rather be-
cause they have to be able to use the theory” ().
Following this approach, scientific theories provide understanding of the
social world if they offer the understandees an adequate answer to their prob-
lems, that is, one that that gives an idea of what to do, how to act to solve a
problem. To do this, they must fit with the interests of the understandees and
have a certain familiarity within their life-worlds and personal experiences.
This familiarity ensures that the adequate theory is accessible to the under-
standee, and usable given the questions to be answered or the problems to be
solved. It enables individuals and social groups to intervene in the social world
using the theory in order to reach their goals (empowerment, emancipation,
social benefits, and so on). The theories that help people to understand social
phenomena will thus provide more than merely theoretical knowledge.10
The overview of debates above provided us with several examples of how ac-
curate theories might be irrelevant for some groups, and how attention should
u nd e r stan d i n g i n p ol i ti c a l s ci e nce 

be paid to adequacy when talking about understanding. As was seen, rational

choice theory just does not raise some questions about social phenomena or
at least not in a way that coincides with the interests of the knower or under-
standee. For instance, neoclassical, mainstream economics does not provide
the French students with the understanding they are looking for concerning
questions of unemployment, inequalities, the place of financial markets, the
advantages and disadvantages of free trade, and so on. Their interests (constitu-
tive of their questions) have not been addressed adequately. Here the pragmatic
elements matter: we should not only ask whether a theory is accurate (warrant-
ed by sufficient evidence, and so on) but also whether it is cast in a form that
is adequate for the understandee (that is, cognitively accessible to the situated
knower who wants to use the theory) and whether it is useful to and in line with
his or her interests so that it will help to solve his or her problems. A theory can
be accurate (in relation to reality), yet fail these pragmatic tests.
The debate concerning the idea of a synthesizing theory in IR illustrates
the importance of epistemic interests too. Moravcsik, an advocate of synthe-
sis, either supposes that the accuracy of a synthesizing theory suffices (with
adequacy not being taken into account), or neglects the possibility of there be-
ing epistemic interests different from his own (such as the unitary us when he
discusses understanding, as noted above), or presupposes that the synthesizing
theory will always provide the most adequate answer (notwithstanding the va-
riety of possible epistemic interests). Smith is more sensitive to the differences
qua interests, as we also saw above.
When Moravcsik suggests that “theoretical syntheses might help us to un-
derstand concrete events in world politics” (, ; italics mine), one should
raise the question of what “us” or “our” understanding means—whose under-
standing? Which epistemic interests are being addressed? In social science,
this kind of question is often neglected, especially by the so-called orthodoxy
(where Smith  refers to, compare supra), by the dominant perspective. (As
a parenthesis, I would suggest that neglecting the variety of interests to be ad-
dressed opens up the way for unification and synthesis.) Smith emphasizes that
this orthodox perspective on the social world, though not necessarily false, is
only partial, addressing dominant interests. Nondominant or disadvantaged
groups might be interested in different aspects of social reality, looking for
knowledge that is useful for understanding and overcoming their systematic
disadvantages, framing questions, devising theoretical classifications, and so
forth, with this aim in mind.
Anderson (), discussing feminist epistemology and philosophy of sci-
ence, summarizes the characteristics of these nonorthodox theories in social
science:
 jer o en van b o u wel

Critical theories aim to empower the oppressed to improve their situ-

ation. They therefore incorporate pragmatic constraints on theories
of the social world. To serve their critical aim, social theories must (a)
represent the social world in relation to the interests of the oppressed,
i.e., those who are the subjects of study; (b) supply an account of that
world which is accessible to the subjects of study, which enables them
to understand their problems, and (c) supply an account of the world
which is usable by the subjects to study to improve their condition.
(Anderson , ; italics mine)

To achieve increased understanding (addressing the plurality of possible in-

terests), then, developing a plurality of theoretical perspectives seems to be a
better option than unification. The unificationist attempts seem to lead to the
neglect of some interests, obstructing a deep understanding (as in the French
students’ petition, discussed above); similarly, the elaboration of one synthesiz-
ing theory only provides a partial understanding (see Smith ).

;Zb^c^hbVcYGVi^dcVa8]d^XZI]Zdgn

To illustrate that the plurality of epistemic interests is better served by a

plurality of theoretical perspectives than by a unified one, I will discuss the
different standpoints feminists have adopted in the discussion concerning the
benefits of rational choice theory and the possibilities of a synthesis between
RCT and feminism. This also gives me the opportunity to put some critiques of
orthodoxy into perspective.
Browsing through the different evaluations feminists have made of RCT,
we can distinguish at least three positions. The first rejects RCT, disagreeing
with its ontological commitments, and articulates a different image of the actor
as the basis for feminist theory. A second position accepts RCT as the domi-
nant theory and wants to enrich it with feminist research interests. The third
position develops a balanced view by specifying the cases in which RCT is ad-
equate, and those in which it definitely is not.
Some feminist voices reject the notion of homo economicus, rational man,
outright, because of its ontological commitments (accepting these would pre-
clude them from considering the theory as useful for providing any understand-
ing). The decontextualized individualism and the privileging of reason over
other capacities trouble many feminists (for example, Nedelsky ; Meyers
). According to these scholars, the homo economicus as an autonomous,
self-transparent, opportunistic, calculating, self-reliant, self-confident, (con-
tinuously) healthy actor, ignores the significance of unchosen circumstances
and interpersonal relationships, eclipses family, friendship, passionate love,
u nd e r stan d i n g i n p ol i ti c a l s ci e nce 

and community, and dismisses dependency as a defective form of selfhood,

while caregiving responsibilities vanish along with children, the disabled, and
the frail elderly. On the basis of this kind of critique, RCT is rejected entirely.
These feminist social scientists want, on the contrary, to emphasize the con-
nectedness and embeddedness of the self.11
A lot of these feminist critiques of RCT reflect the skepticism of modernist,
unitary accounts of the self, and, as such, many feminists are joining the post-
structuralists in declaring the death of the autonomous, self-reflective individ-
ual. This might, however, be detrimental to feminist aims; giving up the idea
of a core self and coherent identity, and the capacity to describe and reflect on
one’s experience might undermine feminist emancipatory objectives for which
a view of selfhood and reason seems indispensable.
Some feminists do not reject RCT or neoclassical theory in economics, but
want instead to see its research agenda adjusted or enlarged to include, for
example, the gender division of labor (paid versus unpaid work, unequal pay
for equal work, and so on), the consequences of policy for women, and model-
ing the household production and distribution using bargaining theory. In this
way, they hope to add to RCT typically feminist research interests that have
not been addressed before because the dominant social group has not been
interested.
This viewpoint, then, does not—as the first group of feminists does—radi-
cally reject RCT and neoclassical theory. Rather, their idea is to accommodate
feminist research priorities within neoclassical economics and RCT, and, as
such, to extend them with new insights without undermining them. Indeed, ac-
cording to this position, by bringing in the feminist interests, one would achieve
increased understanding and further the unificationist attempt of RCT.
Before discussing the third position feminists have adopted in relation to
RCT, which will substantiate my pluralistic view of understanding, I want to
compare the first two positions and point to some links with the debates on
synthesis and economics imperialism. In debates on synthesis and economics
imperialism there exists an orthodoxy/heterodoxy dichotomy; critics point to
the orthodoxy’s attempt to impose its views via a synthesis or economics im-
perialism on minority or heterodox theories in the field, while the orthodoxy
sees no surplus value in the heterodox theories because its theory could explain
the same phenomena in a better and more unified way. A similar dichotomy
can be found between the two first positions in the discussion concerning ra-
tional choice theory and feminism. The second position—enriching rational
choice theory—neither questions the orthodoxy nor criticizes its unificationist
agenda, while the first position radically rejects the unificationist attempt to
impose rational choice theory and replaces the ontological view of the actor
with its own alternative.
 jer o en van b o u wel

In my earlier work, I have shown that these two positions—orthodox, uni-

ficationist theory and the radically heterodox theory—share a winner-takes-all
approach to social science and that they are monistic contenders for unifica-
tion (Van Bouwel ). This approach often leads to facile conclusions, for
instance, that showing that a theory does not work in some situations means a
falsification of the usefulness in other situations, and showing it works in some
situations is a guarantee it will work in other or all situations. Both leave out
the pragmatic aspects of explanation and understanding and neglect to take the
plurality of epistemic interests into account.
The heterodox position, in its criticism of the unificationist attempts of
orthodoxy, often claims to be pluralist. Whether they are really pluralist or
whether this claim is only a strategy to gain some (professional) space remains
to be seen. The latter would be an example of strategic pluralism:

primarily just a strategic move in the game of trying to dominate a

field or profession. Those in the minority proclaim the virtues of plu-
ralism in an effort to legitimate their opposition to a dominant point of
view. But one can be pretty sure that, if the insurgent group were itself
ever to become dominant, talk of pluralism would subside and they
would become every bit as monistic as those whom they had replaced.
(Giere , )

These remarks are important in order to understand how a pluralist view of un-
derstanding has to be conceived so as to leave unificationist attempts behind.
A third group of feminists has developed a balanced view, which does not
reject RCT outright, but limits its usefulness to specific cases, arguing that in
other cases, alternative theories will be more adequate. The selection of the
most adequate theory is made by comparing theories and taking interests and
usefulness into account. A good example can be found in Anderson ().
Analyzing the feminist angle on health care policies taken by Kristin Luker’s
Taking Chances, Anderson reaches the conclusion that the answer to her title,
Should Feminists Reject Rational Choice Theory?, “depends on our purposes,
and on the aspect of RCT being used” (Anderson , ). She believes RCT
offers both resources for and obstacles to understanding the problems women
face and helping them overcome them.
Rational choice theory can be used to answer questions addressing feminist
interests in the following way:

One feminist aim is to ensure that the health-care system effectively

delivers contraception to the women who need it. Luker’s use of the
deliberative theory of rational choice effectively illuminates the moti-
u nd e r stan d i n g i n p ol i ti c a l s ci e nce 

vations that deter many sexually active women from using contracep-
tion, even when they have an interest in doing so. Here the function of
rational choice theory is to make women’s choices intelligible in terms
of their motivations, so the health-care system responds to women’s
needs as they see them. (Anderson , ; italics mine)

An example of the benefits of using RCT would be: “Luker suggests that once
health-care practitioners recognize that women have a stake in seeing their
sexual activity as spontaneous, and tend to stop contraception between rela-
tionships, they can respond by providing drop-in contraceptive services, rather
than requiring women to make appointments weeks in advance” (–).
On the other hand, rational choice theory also presents obstacles to under-
standing the problems women face. Only using RCT precludes the identifica-
tion of obstructions to women’s rationality and autonomy. Anderson does not
agree with Luker’s use of RCT to vindicate the rationality of women:

Merely showing that women’s choices can be fit into the accounting

framework of a cost-benefit analysis is not enough to vindicate wom-
en’s rationality. Nor should feminists want to rush to judgment on this
score. We should be interested in identifying obstacles to the achieve-
ments of women’s autonomy and rationality, rather than assuming that
women face no problems in this area. (Anderson , )

Hence, social scientists should also consider alternative theories in order to

identify the social circumstances that undermine women’s rationality and
autonomy:

With respect to the feminist goal of identifying realization of, or im-

pediments to, women’s rationality, the application of the formal theory
of rational choice to women’s choices poses an obstacle. It effaces the
distinction between action on one’s own autonomous preferences,
and action governed by oppressive social norms. . . . [It] fails to dis-
tinguish between autonomous and heteronomous preferences, and
thus fails to mark distinctions of vital interest to feminists. . . . [RCT]
thus represent[s] people as acting on their own preferences, when
they may just be yielding to the demands and expectations of others,
or governed by oppressive or incoherent social norms. . . . Feminists
can target internalised sexist social norms for critical attention only if
we represent them as factors influencing women’s choices that are in-
dependent of—and indeed, contrary to—their autonomous personal
preferences. Luker’s representation of women’s sexual agency as ratio-
 jer o en van b o u wel

nal ignores the numerous social conditions that undermine women’s

rational autonomy with respect to their sexual decisions. (Anderson
, )

Anderson’s analysis shows us how, depending on the specific interests one has
(and the research questions these generate), different theories can be the most
adequate. Understanding that the existing plurality of theories in social sci-
ence can help to answer questions generated by a plurality of interests should
make us cherish theoretical pluralism rather than attempt unification. (This
does not imply that every existing theory is indispensable; comparing theories
when addressing specific questions will enable us to distinguish good from bad
theories. It does, however, entail that reducing the plurality of theories to one
unifying theory results in a loss of opportunities to achieve understanding.)
Anderson’s balanced analysis illustrates how the winner-takes-all approach can
be left behind; the proposed unification under RCT does not provide under-
standing for all at all times (independent of interests and context), but it does
serve some epistemic interests. Moreover, not rejecting RCT (and its ontologi-
cal commitments) outright solves the problem raised in relation to the first
feminist position, where questioning the rationality of the individual risks un-
dermining the emancipatory objectives of feminism. And it is precisely this
that feminism pursues: “Feminist theory . . . tries to understand the social world
so as to enable certain kinds of liberating changes in it” (Anderson , ;
italics mine).

6EajgVa^hi^XK^Zld[JcYZghiVcY^c\^ci]ZHdX^VaHX^ZcXZh

Starting from the plurality of theoretical perspectives in international rela-

tions studies, I have revisited debates on synthesis and unification to conclude
that the plurality of social theories can answer research questions in relation to
a plurality of epistemic interests. Accounts of the social world are selected on
the basis of one’s interests and the questions they generate, and of the extent to
which the answers these provide are usable and facilitate understanding in the
context of the problems to be solved.
Hence, when trying to grasp the relation between theory and understand-
ing, it is not only important to consider the accuracy of theories but also to
take their adequacy into account. This latter depends on the epistemic inter-
ests of the understandee. Thus, an adequate theory should be cast in a form
that is accessible to the understandee, and usable given the questions to be an-
swered or the problems to be solved. These interests can facilitate or obstruct
understanding; some accounts of the world are more accessible—in line with
the interests of the understandee—and enable individuals or groups to under-
u nd e r stan d i n g i n p ol i ti c a l s ci e nce 

stand their problems. For instance, feminism theorizes in ways that women can
use to improve their lives; it implies that women should be able to recognize
themselves and their lives in feminist accounts of women’s predicaments (see
Anderson , ).
Feminism gives us a good illustration of the impact of interests on the se-
lection of theories to achieve understanding. Its relation to rational choice
theory—the best candidate in the social sciences to refute Kitcher ()—was
discussed, and I concluded that if we want to maximize understanding, the
ideal of unification (be it by RCT or an alternative theory) is not desirable be-
cause of the plurality of epistemic interests to be addressed. Even if unification
conveys understanding for some (relative to their interests at a given time), it
will always obstruct understanding for others (with different interests). Recog-
nizing the plurality of possible epistemic interests will help us to move on from
the winners-takes-all approach to social science (as was shown by Anderson’s
evaluation of RCT), and to stop regarding the plurality of theoretical perspec-
tives—present in international relations studies—as a problem. On the con-
trary, theoretical pluralism is a strength in achieving understanding.

CdiZh
. These data and the methodology used to obtain them can be found in Peterson,
Tierney, and Maliniak ().
. The term “imperialism” might sound too negative, but actually, one of the pioneers
of the approach, Gary Becker (), whose work is the locus classicus, agrees with this
label: “This definition of ‘economic imperialism’ is probably a good description of what I
do” (Swedberg , ).
. Robert D. Putnam warns us—in relation to irrelevance—of the dangers of policy re-
search migrating toward schools of public administration, just as happened with practical
economic studies, which changed from the economics department to business schools: “If
one compares the size of economics departments and business schools in today’s academy,
the cost of reducing a social science to sterile theoretical endeavors is obvious” (quoted on
www.paecon.net, consulted February ).
. The increasing “colonization” of political science by economists has led to so much
discontent that an anti-imperialist movement, named “Mr. Perestroika,” has been created.
The movement was initiated in the year , with a mass e-mailing by Mr. Perestroika.
Schram (, ) discusses the Mr. Perestroika movement in political science: “the ways
in which contemporary social science all too often fails to produce the kind of knowledge
that can meaningfully inform social life.” The Perestroikans’ main focus is that major
journals in the field have been preoccupied with publishing research that conforms to the
economics takeover features.
. Italics mine; this English version of the French petition can be found on
www.paecon.net, consulted February . Out of this petition arose a broad movement
called “post-autistic economics.” It has a lot in common with the Mr. Perestroika move-
 jer o en van b o u wel

ment, although it is mainly aligned against the dominance of neoclassical economics

within economics and in support of heterodox economics. A history of the movement can
be found on www.paecon.net.
. Le Monde, June , ; English version on www.paecon.net; consulted February
.
. The reference to “concrete events” can be understood as an implicit critique of
the constructivists and reflectivists and their metatheoretical travails, which leads them
away from dealing with the explanation and understanding of world politics, according to
Moravcsik.
. For more details on this view of explanation, see Van Bouwel and Weber () and
Weber and Van Bouwel ().
. This analogy between explaining and understanding does not, however, imply that
explaining and understanding are the same. Contrary to Hempel (), I do not consider
scientific understanding a by-product of scientific explanations, so that acquiring explana-
tions of a phenomenon automatically leads to its understanding. In what follows, it will
become clear that more is needed in order to talk of understanding.
. Notice the similarities to Leonelli (this volume), albeit that the focus here is not on
the embodied knowledge and epistemic skills of the individual scientific researcher, but on
the usefulness of social scientific theories for social groups and individuals (both scientists
and nonscientists) and their ability to use these theories. Here, understanding goes public.
A similar Deweyan move can be found in Bohman ().
. The criticism of the lack of embeddedness of the actor in RCT and neoclassical
economics is not exclusively feminist. See, for instance, Davis ().

GZ[ZgZcXZh
Anderson, E. . Should feminists reject rational choice theory? In A mind of one’s
own: Feminist essays on reason and objectivity, edited by L. M. Antony and C. E. Witt,
–. Boulder: Westview Press.
———. . Feminist epistemology and philosophy of science. Stanford encyclopedia of
philosophy. Available at http://plato.stanford.edu/entries/feminism-epistemology.
Bates, R., R. de Figueiredo, and B. Weingast. . The politics of interpretation: Rational-
ity, culture, and translation. Politics and Society :–.
Becker, G. . The economic approach to human behavior. Chicago: University of Chi-
cago Press.
Bohman, J. . Theories, practices, and pluralism: A pragmatic interpretation of critical
social science. Philosophy of the Social Sciences :–.
Davis, J. . The theory of the individual in economics: Identity and value. London:
Routledge.
de Regt, H. W. . Discussion note: Making sense of understanding. Philosophy of Sci-
ence :–.
Giere, R. N. . Perspectival pluralism. In Scientific pluralism, edited by S. Kellert, H.
Longino, and K. Waters, –. Minneapolis: University of Minnesota Press.
u nd e r stan d i n g i n p ol i ti c a l s ci e nce 

Gintis, H. . Towards the unity of human behavioral sciences. Politics, Philosophy and
Economics :–.
Green, D., and I. Shapiro. . Pathologies of rational choice theory. New Haven: Yale
University Press.
Hellmann, G. . In conclusion: Dialogue and synthesis in individual scholarship and
collective inquiry. International Studies Review :–.
Hempel, C. G. . Aspects of scientific explanation. In Aspects of scientific explanation
and other essays in the philosophy of science, –. New York: Free Press.
Kitcher, P. . Explanatory unification and the causal structure of the world. In Scientific
explanation, edited by P. Kitcher and W. C. Salmon, –. Minneapolis: University
of Minnesota Press.
Kratochwil, F. . The monologue of “science.” International Studies Review :–.
Levi, M., E. Ostrom, and J. Alt. . Conclusion. In Competition and cooperation: Con-
versations with Nobelists about economics and political sciences, edited by J. E. Alt, M.
Levi, and E. Ostrom. New York: Russell Sage.
Mäki, U. . Symposium on explanations and social ontology : Explanatory ecumenism
and economics imperialism. Economics and Philosophy :–.
Meyers, D. . Self, society, and personal choice. New York: Columbia University Press.
Moravcsik, A. . Theory synthesis in international relations: Real not metaphysical.
International Studies Review :–.
Nedelsky, J. . Reconceiving autonomy: Sources, thoughts, and possibilities. Yale Jour-
nal of Law and Feminism :–.
Peterson, S., M. J. Tierney, and D. Maliniak. . Teaching and research practices, views
on the discipline, and policy attitudes of international relations faculty at U.S. colleges
and universities. Available at http://mjtier.people.wm.edu/TRIP summary Aug .pdf.
Schram, S. . Return to politics: Perestroika and postparadigmatic political science.
Political Theory :–.
Shapiro, I. . Problems, methods, and theories in the study of politics, or what’s wrong
with political science and what to do about it. Political Theory :–.
Smith, S. . Dialogue and the reinforcement of orthodoxy in international relations.
International Studies Review :–.
Swedberg, R. . Economics and sociology: Redefining their boundaries: Conversations
with economists and sociologists. Princeton: Princeton University Press.
Trout, J. D. . Scientific explanation and the sense of understanding. Philosophy of Sci-
ence :–.
Van Bouwel, J. . Explanatory pluralism in economics: Against the mainstream? Philo-
sophical Explorations :–.
Van Bouwel, J., and E. Weber. . Remote causes, bad explanations? Journal for the
Theory of Social Behaviour :–.
Walt, S. . Rigor or rigor mortis? Rational choice and security studies. International
Security :–.
Weber, E., and J. Van Bouwel. . Can we dispense with structural explanations of social
facts? Economics and Philosophy :–.
 pa rt title ver so

(-
Understanding in Historical Science
Intelligibility and Judgment

:9L>C@DHI:G

 Ideas about the role of understanding in different academic disci-

plines are sometimes incompatible. Regarding the empirical sciences, J. D.
Trout (), for instance, launches a vigorous attack on “the sense of under-
standing” as a subjective and valuable concept in the evaluation of scientific
explanations:

The sense of understanding would be epistemically idle phenomenol-

ogy were it not so poisonous a combination of seduction and unreli-
ability. It actually does harm, sometimes making us squeamish about
accepting true claims that we don’t personally understand, and more
often operating in the opposite direction, causing us to overconfident-
ly accept false claims because they have a kind of anecdotal or theo-
retical charm. ()

Concerning history, F. R. Ankersmit () comes to an—at first sight—op-

posite conclusion:

How historians relate to their own time, what are their innermost feelings
and experiences, what have been the decisive facts of their own lives—
these are all things that should not be distrusted and feared as threats to
so-called historical subjectivity but cherished as historians’ most crucial
asset in their effort to penetrate the mysteries of the past. ()


u n d e r sta n d i n g i n h i st o r i c a l s c i e n c e 

Can the tension between these views be eliminated by stressing that the
first one is about the acceptance of knowledge in empirical sciences such as
physics and the second about the role of understanding in the study of his-
tory? Or is it better to stress that these different views are not typical of physics
and history, and that the choice of other representatives of the philosophy of
these disciplines would have given other positions? To answer these questions,
insight into the phenomenon of understanding in history in particular and in
science in general is necessary.
My aim in this chapter is twofold: I first want to explore how understanding
is obtained in the study of history and then, second, to specify the concept of
scientific understanding. I will begin by summarizing the conception of sci-
entific understanding in physics defended by de Regt and Dieks. After noting
some problems regarding de Regt and Dieks’s conception and their criteria for
achieving understanding, I will introduce the concept of historical understand-
ing. It will become clear that the tension between objectivity and subjectivity
is a general feature in the concepts of scientific and historical understanding.
By focusing on the notion of intelligibility, I will show that intentional explana-
tions, colligatory concepts, synoptic judgments, and historical narratives are
modes of understanding in history—modes in which the “nonobjective crite-
ria” (the “subjective dimension”) cannot be eliminated. I will explore the nature
of these nonobjective criteria at great length. These theoretical considerations
will be illustrated by examples taken from actual historiography of the First
World War. On the basis of discussions in this field, I will try to clarify the
notions of historical understanding and scientific understanding through the
concept of judgment.

JcYZghiVcY^c\^cE]nh^Xh

In order to develop a concept of scientific understanding, I will use the con-

tribution made by de Regt and Dieks as my starting point (de Regt ; de
Regt and Dieks ). Focusing on physics, de Regt and Dieks give the initial
impetus to the formulation of such a concept by working out how various types
of scientific explanation provide understanding. De Regt and Dieks (, )
claim that they “formulate a generally applicable, non-trivial criterion for the
attainment of understanding” and that the notion of understanding “is epis-
temically relevant and transcends the domain of individual psychology.” The
last part of the quote is directed against the views of people such as Hempel
() and Trout (), who state that understanding is merely a psychologi-
cal by-product of scientific practice.
How is understanding epistemically relevant in the view advanced by de
Regt and Dieks? First of all, although they agree with the objectivist view that
 edwin koster

understanding a phenomenon requires knowledge of the relevant theories,

laws, and background conditions, they argue that this is not a sufficient condi-
tion for scientific understanding. Someone may be able to list the principles of
quantum mechanics and may know all the background conditions available;
but if he cannot apply this theoretical knowledge to different concrete situa-
tions, we would not say that he understands quantum physics. Possessing the
relevant formula is insufficient for determining the notion of scientific under-
standing. One must also be able to use this formula to deduce a prediction of
a concrete phenomenon or to give an adequate description. Following Cart-
wright (), de Regt and Dieks argue that the application of a theory to a real
system cannot be achieved by following formal principles. One also needs rules
of thumb and “good sense.” And this means the introduction of a pragmatic
element that is epistemically relevant: being able to use a theory refers to par-
ticular (not necessarily purely subjective) qualities of the scientist involved (de
Regt , –; de Regt and Dieks , –).
According to de Regt and Dieks, the ability to use a theory requires the
appropriate combination of two elements: the skills of the scientist and the
virtues ascribed to the theory. How these elements are combined depends on
the context. Skills are acquired by training and experience and make it possible
to apply theoretical knowledge. They are dependent on the context in which
they are learned and applied. Skills are thus not always part of the explicit and
conscious tools of a scientist. They belong to the presupposed disposition of
a scientist, and de Regt (, ) therefore calls them “archetypes of tacit
knowledge.” Skills are thus related to the notion of familiarity, a notion that de
Regt and Dieks associate with understanding. The skills scientists possess de-
termine their preferences for certain theoretical virtues such as causality, uni-
fying power, and visualization. The use of “field lines,” for example, visualizes
the way electrostatic systems behave and this provides understanding (de Regt
and Dieks , –, ).
De Regt and Dieks make two remarks here. First, scientific understand-
ing in this sense is indeed a nonobjective notion, but that does not make it a
purely subjective concept. Skills and theoretical virtues are, after all, relative
to scientific communities rather than to individuals (de Regt and Dieks ,
). Second, scientific understanding defined in this way is closely related to
the notion of intelligibility: “Intelligibility is a context-dependent value that sci-
entists may ascribe to a theory, on the basis of theoretical virtues (e.g., visualiz-
ability, simplicity or causality) which harmonize with the skills of the scientists
in question” (de Regt , ).
Like the notion of “understanding,” the notion of intelligibility is thus re-
lated to skills and to theoretical virtues. De Regt and Dieks relate the notions
of “understanding” and “intelligibility” to each other explicitly when they claim
u n d e r sta n d i n g i n h i st o r i c a l s c i e n c e 

that a “phenomenon P can be understood if a theory T of P exists that is in-

telligible (and meets the usual logical, methodological and empirical require-
ments)” (de Regt and Dieks , ). Next, they formulate a criterion for the
intelligibility of a theory: “A scientific theory T is intelligible for scientists (in
context C) if they can recognize qualitatively characteristic consequences of T
without performing exact calculations” (de Regt and Dieks , ; compare
de Regt , ). In my view, this formulation makes clear that understand-
ing is indeed related to familiarity: to be able to use a theory, one has to see in-
tuitively (“to develop a feeling for”) the consequences of this theory or, in other
words, one must know (“be able to grasp”) how this theory enables predictions
(de Regt and Dieks , ; de Regt , ).

JcYZghiVcY^c\VhVc>cYZÒc^iZ8dcXZeiWZilZZcDW_ZXi^k^in
VcYHjW_ZXi^k^in

According to de Regt and Dieks, the concept of scientific understanding

is thus related to the intelligibility of theories. It is epistemically relevant and
introduces a pragmatic element—the right combination of skills and theoreti-
cal virtues that makes it possible to use a theory. The objectivist conception
of scientific understanding is thus too narrow. Due to the pragmatic element,
understanding contains a subjective dimension relative to the level of the sci-
entific community and not to that of the individual scientist. It is striking that
de Regt and Dieks use notions such as “good sense,” “familiarity,” “intuition,”
and “to recognize without making calculations” to describe what they mean
by scientific understanding. These notions are not very precise. This is a con-
sequence, perhaps, of their subject: the concept of understanding—compared
to the precision of physics—is related to an elusive and vague state of our
consciousness.
In this chapter I will look at the question of understanding in the field of
the study of history. In contrast to physics, the concept of understanding has
a long history in this academic discipline, and the tension between objectivity
and subjectivity in the concept of understanding as well as its intangibility are
also present. Modern historical science begins with the historicism of Leopold
von Ranke (–) and Johann Gustav Droysen (–). Over against
idealism’s emphasis on an all-encompassing totality, Ranke stresses the neces-
sity to understand the irreducible individuality of historical events, that is, the
past should not be assessed from the anachronistic perspective of the historian
but is to be understood on its own merits. According to Ranke, the task of
the historian is to describe the past wie es eigentlich gewesen ist (“as it actu-
ally was”). Ranke thus believes that historians are capable of giving objective
descriptions of the past. This ideal of objectivity is reflected in the emphasis on
 edwin koster

empirical source investigations and critical-philological examinations. In Ran-

ke’s concept of historical understanding, one can find a strong commitment to
the notion of epistemological objectivity (de Mul , –).
It is important to distinguish Ranke’s ideal of objectivity from that of his-
torians who were inspired by the positivistic program of Auguste Comte and
John Stuart Mill. The latter wanted to follow the model of natural science and
they—in sharp contrast to Ranke’s emphasis on unique individual historical
phenomena—tried to formulate regularities by subsuming historical events
under historical laws. A late echo of this positivistic approach in history can be
found in the work of Carl Gustav Hempel (, –) and in some social
scientific approaches to history.
Droysen rejects both Ranke’s ideal of historical understanding and the posi-
tivistic approach to history. As opposed to Ranke’s objectivity, Droysen claims
that the writing of history is characterized by a subjective construction of the
past. First of all, the historian is bound by time and place, and this means, for
example, that his or her necessary selection from often unorganized sources
cannot be called objective. What one historian regards as important is different
from what another regards as important. Second, only by way of a subjective
perspective is it possible to (re)construct meaningful relationships among the
isolated facts found in the sources and thus to understand part of the past real-
ity (de Mul , –). In contrast to positivism, Droysen writes that his-
torical science is not concerned with explanation by means of general laws but
with the understanding of nonrecurring historical phenomena. The difference
between Droysen and the positivistic approach becomes clear when we turn to
Droysen’s characterization of understanding: “it is an immediate, sudden pro-
cess that takes place without our being conscious of the logical mechanism
operative in it. Thus the act of understanding [Verstehen] is like an immediate
intuition, like a creative act, like a spark of light between two electrophoric
bodies, like a receptive act” (cited and translated in Plantinga , –).
The tension between Ranke’s ideal of objectivity in historical understanding
and Droysen’s claim of the subjectivity of all history writing comes explicitly to
the fore in the work of Wilhelm Dilthey (–). On the one hand, Dilthey
was looking for objective knowledge in the study of history, but, on the other,
he was aware of the fundamental finitude and thus subjectivity of the historian.
Traces of this (apparent) antithesis can be found in several places in his work.
Thus, he claims that the possibility of understanding is dependent on the trans-
fer of the historian’s internal experience to the external reality of the historical
agents, without losing his or her desire for objective knowledge. On the one
hand, Dilthey emphasizes that this transfer has to do with “immediacy” and
“artistic intuitions”: through his or her imagination, the historian can climb
into the skin of the historical agents by which the historian has direct access
u n d e r sta n d i n g i n h i st o r i c a l s c i e n c e 

to their motives and ideas. On the other hand, Dilthey makes clear that the
understanding of historical agents also has to do with the verifiable inquiry of
documents in which their intentions are expressed (Plantinga , –; de
Mul , –, –, –, –).
Various scholars have proposed a way out of this ambiguity. Although they
are not univocal in their interpretations, they agree that Dilthey did not use
the term “objectivity” as synonymous with “universal validity.” According to
them, such a reading is impossible because of Dilthey’s emphasis on human
finitude and historicity. They further suggest that it would have been better if
Dilthey—speaking of the human sciences—would have used the concept of “in-
tersubjectivity” instead of “objectivity” because the former concept recognizes
the subjective nature of historical understanding without characterizing it as
arbitrary (Plantinga , –; de Mul , –).
With this reflection on the first phase of modern historiography, I have
gathered presumptive evidence for the claim that the tension between objec-
tivity and subjectivity is a general feature of the concept of understanding. Per-
haps the subjective dimension is to blame for the vagueness of this concept—a
vagueness to which de Regt and Dieks do not pay much attention. The plausi-
bility of this claim and the tenability of this suggestion will become clear in the
remainder of this chapter. But before gaining better insight into the concept
of understanding—much of the meaning of this concept is still implicit—we
first have to remove a source of misunderstanding. Understanding (in history)
is sometimes exclusively put on a par with our capacity to empathize with the
experiences of historical agents.
A classic expression of this identification is R. G. Collingwood’s (–)
term “reenactment,” one of the best-known aspects of his hermeneutics of his-
tory. According to Collingwood, to make history intelligible, historians reen-
act the experiences of the historical agents, that is, they empathize with their
thoughts and think through their actions (D’Oro , sec. .). Collingwood
(, ) even claims that in the act of reenactment the historian does not
think thoughts similar to those of the historical agent but the same thoughts:
“the process of argument which I go through is not a process resembling Plato’s,
it is actually Plato’s so far as I understand him correctly.” Sometimes reenact-
ment is incorrectly identified with (historical) understanding. We have already
noted that Dilthey has a broader concept of historical understanding because
he does not confine it to immediacy.
Hans-Georg Gadamer’s (, –) criticism of Collingwood’s theory
of reenactment is explicitly clear on this point. His ideas regarding the well-
known fusion of horizons and the Wirkungsgeschichte demonstrate that the gap
between the time of the historical agent and the time of the historian cannot be
bridged. This does not mean that, according to Gadamer, an understanding of
 edwin koster

the past cannot be achieved: by virtue of the historian’s awareness of his or her
own situatedness, it is possible to have access to the individuality of the past.
However, Gadamer does not offer a specific method for achieving an under-
standing either of history or of (historical) texts. Truth and Method (Wahrheit
und Methode), Gadamer’s classic in the field of twentieth-century philosophi-
cal hermeneutics, can be considered a severe critique of the idea that truth and
knowledge are connected exclusively to an unambiguous method (Gadamer
, xi–xii). From the perspective of the adherents of such a method one
could even assert that Gadamer claims that the concept of understanding is an
intrinsically loose term. The plausibility of this view and, with this, the claim
that it is impossible to specify the slightly vague notions used by de Regt and
Dieks, will be tested below.

I]Z>ciZaa^\^W^a^ind[=^hidgn

Thus far, I have shown that there are similarities between the concept of un-
derstanding in physics and history: () the tension between objectivity and sub-
jectivity and () the vagueness of the concept. This supports de Regt and Dieks
because they claim to formulate a “generally applicable criterion for the attain-
ment of understanding.” However, one may doubt whether their conception
of scientific understanding can be applied to the practice of historians. They
describe the notion of intelligibility, for example, in connection with theories,
and they formulate a criterion for the intelligibility of theories in terms of the
absence of precise calculations. How are these ideas to be applied to the field of
history? Many scholars will agree that even if one assumes a strong connection
between history and the social sciences, the development of theories is rare in
the study of history. The same obtains for making exact calculations.
However, historians do refer to the notion of intelligibility. One of the ar-
ticles by the well-known philosopher of history W. H. Walsh is called “The In-
telligibility of History.” It is concerned with the question, among other things,
of whether it is possible to find intelligible connections between historical phe-
nomena, such as actions and events (Walsh , ). Other than in physics,
the intelligible connections in history are not meant to be extrinsic: according
to Walsh and many others, these connections are not causal. Walsh wants to
show that some sort of intrinsic relation exists between different historical phe-
nomena in a certain period of the human past. According to Walsh, this con-
nection is most obviously present in the case of actions: “two actions may be
connected as parts of the realization of a single consistent policy, and the ideas
behind them . . . may reciprocally determine each other. . . . the earlier actions
themselves . . . will take the form they do because the agent intended to carry
out certain other actions later” ().
u n d e r sta n d i n g i n h i st o r i c a l s c i e n c e 

Actions are thus intrinsically related, rather than causally. Insight into
these connections makes history intelligible, writes Walsh, and can, I would
add, clarify the concept of historical understanding. Next to the reconstruc-
tion of a web of related actions, the historian tries to make a coherent whole of
the events of a selected period by putting them together under certain leading
ideas. To describe this activity of the historian, Walsh borrows a term made
famous by William Whewell and Mill: the historian “colligates” different events
according to “appropriate conceptions” such as “the Industrial Revolution” and
“the Enlightenment” (Walsh , ). An analysis of the role colligatory con-
cepts play in history throws light on the notion of historical understanding as
well. Walsh’s insights lead to at least two possible modes of understanding in
history: intentional understanding to gain insight into the actions of the histor-
ical agents and understanding by way of forming colligatory concepts to gain
knowledge of the way events are connected.

=^hidg^XVaJcYZghiVcY^c\VcY>ciZci^dcVa:meaVcVi^dc

Many people will agree that to understand, for instance, the events that
together constitute the First World War, it is necessary to acquire knowledge
about the political developments that gave rise to the outbreak of the Great
War, the interests of the superpowers at the time, the military strategies of the
armies from  to , and the consequences of the Versailles Treaty in
. These macro-level developments form, of course, the indispensable in-
gredients of every historical attempt to understand this crucial period in Eu-
ropean history.
Although knowledge at this level is considered important, historians have
shown great interest in the documents that express the emotions, feelings, and
thoughts of the soldiers fighting in the trenches or recovering in field hospitals.
The insights offered by documents recalling the war, such as war poetry, novels,
letters, long-lost memoirs, diaries, and regimental histories written by partici-
pants—the micro-level of the First World War, so to speak—resulted in the last
decades of the twentieth century in a number of movies, the foundation of the
Great War Society (whose members try to reenact the war), and new histo-
ries of particular battles based on the reports of eyewitnesses (Stummer ).
An example of this last instance is the work of the historian Lyn Macdonald
on the British Tommies and German foot soldiers—who called themselves
Frontschweine—in the trenches of Flanders. By interviewing the survivors and
then making use of memoirs and studies of the terrain of the Western front,
she succeeds in sketching a lively picture of the war at trench level (Macdon-
ald , ). All this reveals the widespread belief that only the stories of
the historical agents can give insight into this traumatic event. In other words,
 edwin koster

these documents emphasize the explanatory function of intentions in the un-

derstanding of history.
For the historian who thinks that the expression of the soldiers’ experiences
and the rulers’ motives is, indeed, the war story, “intentional explanations” are
central in writing history. From this perspective, Walsh’s intrinsic connections
come close to Collingwood’s notion of reenacting the thoughts of the histori-
cal agents who are involved in the chain of events. That this implies forms of
subjective understanding is demonstrated clearly by James Joll (), who fo-
cuses on the intentions and motives of the ruling classes and political leaders in
Europe during the crisis of July . According to Joll, in situations where the
outcome is entirely uncertain, rulers fall back on instinctive reactions, modes of
behavior, beliefs, rules, and objectives that are taken for granted and were never
written down. In other words, in moments of crisis their motives are deter-
mined by unspoken assumptions, and the historian has to find out what “goes
without saying.” Although historians sometimes know (thanks to the sources)
the words historical agents used, they have to reconstruct the assumptions
themselves, and this reconstruction “may vary according to our assessment of
quite different evidence about their characters” (Joll , ).
The result of this process of reconstruction is thus in some way dependent
on the historian. He or she decides whether certain conditions are the deter-
mining factors in the explanation of, for example, the July crisis: the technical
sequence of mobilization plans, general economic or sociological circumstanc-
es, the educational background of British Foreign Secretary Sir Edward Grey,
the mental state of the people in reference to such factors as the fear of revolu-
tion and national loyalties, or the influence of Darwinian ideas on the ideol-
ogy of imperialism—to mention just a few possibilities (Joll , –, ,
). According to Joll, all these factors played a role, but one cannot point out
the decisive one. However, he makes clear that it is impossible to apply general
features of a period mechanically to the particular actions of the politicians and
generals who were in charge during the drama of  (, ).
The emphasis on intentions in the understanding of history can be associ-
ated with the well-known Verstehen/Erklären dichotomy. Indeed, it is hard to
deny that phenomena such as empathy and so-called “double hermeneutics”
play a role in the study of history but not in, for example, physics. But “dichoto-
my” is too strong a term to use here. Joll’s discussion makes clear, for instance,
that understanding is not opposed to explanation: by making the unspoken
assumptions explicit the intentions of the historical agents are understood and
the why-question is answered. Understanding and explanation—applied to the
activity of the historian—are two sides of the same coin (Ricoeur , –,
–, –, –; compare Ricoeur , –). In line with this
claim, Stueber (, ) argues that, although the understanding of intentions
u n d e r sta n d i n g i n h i st o r i c a l s c i e n c e 

is indispensable for the explanation of individual agency, this does not entail
commitment to the claim that reenactment or empathy is the only method
historians should use. Stueber’s approach gives insight into the necessarily sub-
jective character of historical understanding.
It is sometimes suggested that our capacity to understand one’s own and
somebody else’s behavior is based on psychological theories. Therefore, the
concept of empathy would not be of any importance: to understand the mo-
tives of a historical agent, it would be enough to apply these general theories
to particular situations. However, how can one decide which theory is to be
applied in a certain situation? For that purpose, a general theory stating which
aspects of a specific situation are epistemically relevant seems to be necessary.
The existence of such a theory, Stueber writes, is not very likely. And even if
we had such a theory at our disposal, the right application of the theory in a
particular situation could not be guaranteed: there are no formal principles
that state how one is to use such a general theory. Therefore, one may look for
other theories that do state how the general theory is to be applied correctly
in different circumstances. But then we are threatened with an infinite regress
of theories and this, according to Stueber (, ), “can only be avoided by
appealing to a non-theoretical capacity.” This nontheoretical capacity is the hu-
man capacity for empathy: we can understand the intrinsic relation between
someone’s intentions and motives on the one hand and his or her actions on the
other only if we understand how these intentions and motives would lead us to
performs the same actions ().
It should be noted that our access to, for instance, complaints about trench-
feet or the short-arm inspection of soldiers serving during the First World War
is almost never immediate. This is not only due to the implications of herme-
neutical notions such as Wirkungsgeschichte—insofar as we believe that they
are tenable—but also because we know that the posthumous writings of the
men in the trenches are to a certain extent constructed: it has been shown
that popular war novels written by eyewitnesses functioned as templates for
soldiers and civilians to shape their own memories. The deliberate adaptation
of their own memories to the versions presented in these war fictions was so
natural that “the military doctor/novelist Georges Duhamel noted in  that
‘if his former patients were to read today their own stories, they would rarely
recognize them’ ” (Shapiro, , –)—an indirect proof of Gadamer’s the-
sis, so to speak. Therefore, it is an illusion to think that empathy, important as
it is, takes us directly to the soldiers’ experiences.
To conclude, the understanding of intentions requires the act of empathy,
which cannot be eliminated in favor of the “objective application” of a set of
theories. However, the historian has to be aware that the experience of seeing
through the eyes of a historical agent might be illusionary. Here the historian
 edwin koster

is faced with a definite limitation: he or she has to be critically faithful to the

remains of the past without ever being sure about the truth of history (compare
Ricoeur , ).
Although Stueber thinks that the capacity of empathy is indispensable for
historical practice, he shows that the historian also needs to be able to form
practical judgments, because in applying this capacity we have to make deci-
sions about which theory to apply and which aspects of a situation have to be
judged as relevant and which not (Stueber , ). In this process, general
(and sometimes tacit) standards such as psychological generalizations are to
be applied in particular situations without applying strict rules. And this is, I
conclude, only possible if the historian “understands.”
Finally, Stueber holds to the idea that the usefulness of empathy in account-
ing for individual agency has its limits. The explanation and understanding of
individual agency is the result of “the creative interplay of re-enactment and
various theoretical considerations” (Stueber , ). An example of this “cre-
ative interplay” can be found in Strachan’s The First World War: A New Illus-
trated History (). Strachan makes clear why the Great War is rightly called
a “world war” by focusing not only on the Western front but also on the Middle
East and the Russian front, on the fighting in Asia and Africa, and by looking
at the many colonial soldiers fighting in Europe. He writes, for instance, that an
important part of Germany’s strategy consisted in weakening the entente pow-
ers France and England by attacking them indirectly through their empires.
However, this strategy could not be realized in November  because the
German army was fighting in Europe and, even if there were some German sol-
diers and weapons left, it would still have been almost impossible to transport
them to “British” Turkey, for instance, because of the British naval supremacy.
For this reason, the Ottoman Empire could be of great help by providing troops
to the German army. However, Strachan explains that, for political and finan-
cial reasons, Helmuth von Moltke (the chief of the general staff ) did not con-
sider an alliance with Turkey (Strachan , –).
Although Strachan does not project himself into the role of Moltke by way
of empathy—he actually refers to written documents exemplifying the inten-
tions of the German military leaders—he still makes use of psychological gen-
eralizations to imagine and evaluate the deliberations of the people involved.
In addition, he evaluates their motives against the background of theoretical
considerations: the economic rules and political habits. These theoretical con-
siderations are not just reconstructed for the benefit of a better understanding
of individual agency. Since the course of history is not fully determined by the
consequences of human actions, the job of the historian does not end with
the reconstruction of intentions. He or she must also connect events that are
either the results of unintended actions or the consequences of powers that are
u n d e r sta n d i n g i n h i st o r i c a l s c i e n c e 

beyond the control of the human agents involved. One way to accomplish this
task is through the invention and application of colligatory concepts.

8daa^\Vidgn8dcXZeih!Hncdei^X?jY\bZcih!VcYCVggVi^kZ:medhjgZ

With colligatory concepts we make the transition from the first mode of
understanding—intentional understanding related to the actions of historical
agents—to the second mode, which is about the relation between historical
events. When historians consult their sources, they are confronted with a mass
of largely unrelated materials. It is their task to organize the data by showing
that they reveal certain patterns (themes or developments) that they judge to
be significant for the selected period. These patterns are seen as phases in a
continuous process and together they can be expressed by colligatory concepts
such as “the Renaissance” and “the First World War.”
According to Walsh (, –), colligation is primarily a stage in his-
torical interpretation because this operation is used principally to characterize
and analyze a certain period in history and, in doing so, colligatory concepts
produce a sense of understanding. However, as Walsh admits, the use of col-
ligatory concepts also has an explanatory function: by displaying how an event
fits into a development, the historian shows why it happened as it did (Walsh
, ; compare Von Wright , ). This again makes it clear that there
is no opposition between understanding and explanation in history.
Which conditions should govern the choice for certain colligatory con-
cepts? According to Walsh, two conditions must be fulfilled. In the first place,
a colligatory concept has to fit the facts: the statements that embody the con-
cept must be supported by evidence. To call a certain period in history “the
First World War,” for instance, it is thus necessary to show that our knowledge
of this period, the facts, can be organized into a picture of a unique and all-
embracing war. Second, it must illuminate the facts so that the use of this con-
cept increases our understanding of the selected period (Walsh , –).
Mink () explains what this last condition means. Mink makes a distinc-
tion between knowledge on the one hand and comprehension or understand-
ing on the other. We know many things as unrelated facts, such as the way
Franz Ferdinand died, the date of birth of the author of The Guns of August,
or the name of the capital of Germany. To know these facts is not to com-
prehend them. Comprehension, Mink writes, “is an individual act of seeing-
things-together, and only that. It . . . [is] the human activity by which elements
of knowledge are converted into understanding” (). This act of “seeing events
together” that are temporally and spatially separate is essentially the same as
the process of colligation. Mink calls it a “synoptic judgment” and, according
to him, this form of judgment is typical for the phenomenon of historical un-
 edwin koster

derstanding and cannot be substituted by any analytic technique. The historian

tries to understand the separate events of the past as a whole by “seeing things
together,” in the same way we recognize the individual pieces of a jigsaw puzzle
through the representation of the complete image. In a total and synoptic judg-
ment, various events are thus seen as elements in a single and complex web of
relationships (Mink , –). Like Walsh, who states that historians “know
in practice how to handle colligatory concepts” but find it difficult to give an
explicit description of the criteria for applying these concepts (Walsh ,
), Mink emphasizes the necessity of practical experience to construct and
to understand synoptic judgments. Here he uses the notion of phronèsis: “a way
of grasping a practical rule or principle which cannot be explicitly exhibited as
the basis for a mechanical method” (Mink , ).
Mink thus claims that a synoptic judgment, in which the weight of the rel-
ative importance of the factors involved is determined, cannot be exercised
through an objective procedure such as an algorithm. At least two questions
can be raised regarding this claim: how can one determine the most relevant
factors in a historical development, and why is it impossible to call this pro-
cess “objective”? Perhaps the best way to answer these questions is to refer to
the use of thought experiments in the study of history. As Ricoeur (, ,
–) has shown, important contributions to this subject have been made
by Max Weber (“In order to penetrate the real causal interrelationships, we
construct unreal ones”), Raymond Aron (“Every historian, to explain what did
happen, asks himself what might have happened”), and William Dray (“The
historian thinks away the suggested cause in order to judge what difference its
non occurrence would have made in the light of what else he knows about the
situation studied”).
Recent interest in the phenomenon of thought experiments in history shows
it is still seen as an important instrument for achieving intelligibility in history.
De Mey and Weber (, ) assert that contrastive reasoning is required to
arrive at historical explanations and that thought experiments offer such rea-
soning. They provide some examples of thought experiments and claim that
“without contrastive reasoning, causal relevance cannot be established.” One of
their examples concerns the existence of unofficial truces during World War I.
Some historians have argued that, given the circumstances of trench warfare,
the strategy of conditional cooperation is sufficient to produce truces. Thought
experiments, such as answering the question “What would happen if the units
that often have truces were involved in a different kind of warfare (for example,
Blitzkrieg or guerilla)?” are needed to sustain this argumentation (De Mey and
Weber , ). So far, their ideas are in agreement with the views of Weber,
Aron, and Dray, quoted above: thought experiments are particularly relevant
for the identification of factors that play a part in the development of events.
u n d e r sta n d i n g i n h i st o r i c a l s c i e n c e 

But unlike these earlier contributions, De Mey and Weber emphasize that
thought experiments lose their explanatory power if they are not supported
by empirical data. They claim that premises such as “trench warfare is posi-
tively causally relevant for truces,” can be empirically supported (–). And,
indeed, they show by means of several examples that contrastive reasoning is
more than plain fantasy.
However, in my view, they overemphasize the role of generalizations that
can be tested empirically (but not experimentally). De Mey and Weber (,
–) present the determination of factors that are “more important” than
other factors as an almost entirely objective process. Only incidentally do they
acknowledge that contrastive reasoning is dependent on the knowledge of the
specificity of particular situations and on the possibility of accounting for rele-
vant values that are not always quantifiable (). I agree that the notion of “more
important than” is accessible to analysis and that empirical support has a vital
part to play here. But to determine the factors that, for instance, bring about
the truces, practical experience is needed as well. Thus, contra the arguments
of De Mey and Weber, and the suggestions of Ricoeur (, ), I maintain
that this is not an objective process and that subjective deliberations—though
not irrational or a matter of taste—are therefore inevitable. Thought experi-
ments are the tools (subjective, but not arbitrary) for determining the most
relevant factors.
It thus follows that a synoptic judgment, in which knowledge is converted
into understanding, must partly be characterized by subjective—but not arbi-
trary—considerations. Perhaps the best way to illustrate the transition from
historical knowledge to historical understanding—to borrow Mink’s terminol-
ogy once again—is to look at the process of constructing a historical narrative.
From the documentary phase, in which the historical archives serve as sources
for a set of unrelated facts, through the phase in which actions and events are
explained and understood by way of synoptic judgments, the historian arrives
at the representative phase: the writing of history through the production of a
narrative. Without claiming that these phases can be identified with distinct
chronological stages, the subdivision of the work of the historian in these inter-
woven methodological moments helps in recognizing the special features of the
concept of historical understanding (Ricoeur , –). Why, one can ask,
is the representative phase confined to narrative discourse? The answer is that
synoptic judgments have their written complement in the plot of a narrative.
As Ricoeur (, x) writes, “The plot of a narrative . . . ‘grasps together’ and
integrates into one whole and complete story multiple and scattered events.”
Therefore, the attainment of a synoptic judgment and the act of constructing a
plot are similar creative processes.
A beautiful example of the narrative exposure of a synoptic judgment is the
 edwin koster

first chapter of Barbara Tuchman’s book on the beginning of the First World
War, The Guns of August. In this chapter she connects various developments
taking place at the royal courts during the first decade of the twentieth cen-
tury by describing in a grand style the funeral of Edward VII of England. She
makes a reasonable case for the important influence of the personal diplomacy
of Edward, made possible by his great gifts as a sociable king and by his fam-
ily ties: he was related to many members of Europe’s ruling houses, including
Kaiser Wilhelm of Germany and the Dowager Empress Marie of Russia. Tuch-
man writes that, when the procession of kings was passing by, the crowd “could
not keep back gasps of admiration” but also “felt a finality in the slow tread
of the marchers and in the solemn music played by the Royal Horse Guards’
band” (Tuchman , –). In this way, she renders a lively depiction of a
culture that would change enormously after August  but was also largely
responsible for the outbreak of the First World War. That is why her historical
narrative can be seen as a synoptic judgment.

=^hidg^XVaJcYZghiVcY^c\VcYi]Z8dcXZeid[?jY\bZci

In the last sections I explored two modes of historical understanding: one

in relation to intentional explanation and the other concerning colligatory con-
cepts and narrative representation. In this section I will show that both modes
are related to the notion of “judgment” and that this notion is of vital impor-
tance for the concept of historical understanding. According to the first mode,
a historian understands the actions of the historical agents (among others) in
terms of their goals, beliefs, and the circumstances they were aware of. The
intentions and motives of, for instance, politicians are partly determined by
unspoken assumptions such as beliefs, rules and objectives that go without
saying (Joll ). A reconstruction of these assumptions is not possible with-
out the historian’s judgment of the weight of different personal and cultural
factors.
This does not mean that a historian simply projects his or her own norma-
tive criteria onto the past. Arguments can be given for the choice of the selected
general features that are applied to the particular situation of a historical agent,
and the validity of the generalizations can be empirically supported. However,
there is no algorithm or specific theory that tells us how general features and
particular actions should be connected. The historian has to judge every specif-
ic action on its own merits, and in these judgments personal factors play a part.
Although psychological generalizations and theoretical regularities play a role
in these processes, an act of empathy is unavoidable. To understand whether
the action of a historical agent is appropriate becomes a matter of practical
u n d e r sta n d i n g i n h i st o r i c a l s c i e n c e 

judgment (Stueber). The historian’s argumentation resembles the reasoning of

judges who weigh opposing arguments in order to come to a decision.
Obviously, the concept of judgment returns in the second mode of his-
torical understanding. To realize and to understand the process of colligation,
historians use tacit standards (Walsh) and, to apply these standards correctly,
need to make judgments. In addition to colligatory concepts, I have discussed
synoptic judgments (Mink) in which the application of practical rules or prin-
ciples requires practical experience and not the mechanical implementation of
the subsequent steps of an algorithm. The concept phronèsis is used to refer to
this capacity. Phronèsis is a form of reasoning used to determine the right way
to act in an unknown situation. It is developed through experience and is con-
cerned with choice and involves deliberation. This type of reasoning is, above
all, similar to the formation of synoptic judgments because of the comparable
role of rules of thumb: to relate certain events to one another and to deter-
mine the most relevant factors in historical developments, one has to mediate
between general principles and concrete particular situations (see Bernstein
, ; Nussbaum , –). Although empirically supported claims play
a role in the determination of the most important factors involved in the con-
struction of a colligatory concept or the exercise of a synoptic judgment (De
Mey and Weber), subjective deliberations cannot be avoided. Finally, we saw
that it is plausible that synoptic judgments are naturally expressed in historical
narratives (Ricoeur).
Thus, the first thesis of this chapter is that judgment plays a central role in
the concept of historical understanding. Judgment in this sense has to do first
with the application of general features to particular situations and, second,
with the determination of the weight of several factors involved. We have seen
that both capacities require practical experience, characterized by a subjective
moment that cannot be eliminated. This does not mean that historians project
their own opinions onto the past they try to describe. They are, however, aware
of this threat to the results in every stage of their historical research. Therefore,
they give arguments for the way they treat their sources and try to support
their judgments empirically. This makes it possible to criticize historical rep-
resentations. Thus, for instance, when the important question “Who started
the Great War in ” is posed, historians such as Tuchman (), Strachan
(), and Fromkin () can produce different, substantiated answers that
are formed partly by subjective preferences but also are open for discussion.
In other words, their views “arise out of informed judgments” and should be
“submitted to the community of competent individuals for evaluation and criti-
cism” (Brown , viii). Thus, historical results that are distorted by subjective
preferences are criticized and can be eliminated.
 edwin koster

GZXdch^YZg^c\i]Z8dcXZeid[HX^Zci^ÒXJcYZghiVcY^c\

Central to the notion of historical understanding is the concept of judg-

ment: nonobjective deliberations that are necessary to applying general features
to particular situations and to determine which factors are more important
than other ones. I want to investigate how the concept of judgment can be used
outside the field of history and whether the use of the concept of judgment—in
relation to the view of de Regt and Dieks—allows one to describe the concept
of scientific understanding more precisely. I will make clear, and that is the
second thesis of this chapter, that the notion of judgment clarifies the concept
of scientific understanding.
One of the first philosophers to use the term “judgment” in natural science
was Thomas Kuhn. He stresses the fact that “every individual choice between
competing theories depends on a mixture of objective and subjective factors,
or of shared and individual criteria” (Kuhn , ). The objective criteria are
accuracy, consistency, scope, simplicity, and fruitfulness. These criteria play a
vital role when a scientist has to choose between competing theories. However,
as Kuhn shows by discussing some examples from the history of science, it is
impossible to make a choice based on these criteria, which is why these “objec-
tive criteria” do not function as unambiguous rules and why they must be sup-
plemented by subjective considerations (–, , ). Kuhn emphasizes
that these subjective considerations are not equal to “bias and personal likes or
dislikes.” They have a “judgmental” character: it is possible to explain a choice
and to clarify on which grounds the subjective considerations are based. In
other words, “no standards of factuality or actuality are being set aside” ().
So, according to Kuhn, the concept of judgment is also present in the natural
sciences.
This view is elaborated by Harold Brown, who claims that the concept of
judgment plays a central role in the notion of scientific rationality. He defines
judgment as “the ability to evaluate a situation, to assess evidence, and come to
a reasonable decision without following rules” (Brown , ). Such judg-
ments are fallible and can be made only if an appropriate body of informa-
tion is available (Brown , ; Brown , ). Peer review and other
mechanisms for filtering incorrectly formed results (Bauer , –) allow
judgments that are distorted by false subjective preferences to be criticized
and eliminated. Whether this concept of judgment in science can be related to
Aristotle’s notion of phronèsis is disputed (Bernstein , –; Brown ,
–).
Thus, it is not uncommon to use the concept of judgment in the case of sci-
entific understanding, even though the exact meaning of this concept may vary
a bit from the one used in the notion of historical understanding. If we assume
u n d e r sta n d i n g i n h i st o r i c a l s c i e n c e 

that the differences are negligible between the way in which the concept of
judgment is used in history and how it is used in the natural sciences, then we
can also indicate points of resemblance with the concept of understanding as
presented by de Regt and Dieks ()—for instance, the absence of algorithms
and the comparable role played by rules of thumb and skills. Regarding skills,
Brown () states, for example, “When we develop the ability to exercise
judgment in a particular field, we are developing a skill” (). And, as we have
seen, skills are also important to the approach of de Regt and Dieks.
The conditions to exercise judgment, such as being acquainted with com-
parable (for instance, historical or experimental) circumstances, come close
to their use of the notion of “familiarity.” However, it seems to me that a re-
description of the concept of scientific understanding in terms of “judgment”
is less vague than the definition of this concept in terms of “good sense” and
“intuition,” as de Regt and Dieks propose (compare de Regt, this volume, for an
attempt along these lines). I claim that, unlike the use of the latter terms, the
application of the notion of judgment gives insight into the way a scientist puts
together empirical evidence, (more or less) objective criteria, and subjective
considerations (even though based on sound reasons).
Of course, de Regt and Dieks have a great deal more to say about the con-
cept of understanding, so I am a bit unfair in my treatment of their position.
Moreover, the concept of understanding seems to possess a sense of elusiveness
that cannot be avoided. So, why does the vagueness of their concept constitute
a problem? Further, the imprecise character seems to be due to the subjective
dimension of the concept of understanding, and the presence of this dimension
in historiography has been sufficiently demonstrated in this chapter. Again, is it
not in de Regt and Dieks’s favor that they speak of nonobjective criteria and fo-
cus on the “not purely subjective” character of these criteria? It is and it is not.
It is in their favor because they rightly emphasize all this; but it is not enough,
for they do not provide sufficient insight into the subjective dimension of un-
derstanding and they do not discuss the possibility of eliminating distorted as-
pects of subjective preferences.
Another shortcoming in their approach is that they seem to confine the
concept of scientific understanding to the application of theories. They do not
say much—as far as I can see—about the possible role understanding plays in
the construction of theories and explanations. This may be due to the context
in which they raise their questions; that is, they start with the fact that we have
different kinds of explanations at our disposal and then ask how these expla-
nations provide understanding. They do not pose the question of how these
explanations are formed. However, de Regt and Dieks do not seem to be con-
cerned with the role of understanding in the formation of theories. In contrast,
in historiography the concept of understanding is central to its practice: it is
 edwin koster

understood in the first place as important at the level of constructing “theories”

(historical narratives).
To conclude, I think the discussion on historical understanding not only
has some weight of its own, but it also has clarified the slightly vague notions
(in my opinion) used by de Regt and Dieks. By looking at the nonmechanical
way in which judgments are formed in history, fresh light is thrown on the non-
objective criteria for the elusive concept of scientific understanding.

GZ[ZgZcXZh
Ankersmit, F. R. . Sublime historical experience. Stanford: Stanford University Press.
Bauer, H. H. . Scientific literacy and the myth of the scientific method. Urbana: Univer-
sity of Illinois Press.
Bernstein, R. J. . Beyond objectivism and relativism. Oxford: Blackwell.
Brown, H. I. . Rationality. London: Routledge.
———. . Judgment, role in science. In A companion to the philosophy of science, edited
by W. H. Newton-Smith, –. Oxford: Blackwell.
Cartwright, N. . How the laws of physics lie. Oxford: Clarendon Press.
Collingwood, R. G. . The idea of history. Edited by J. Van der Dussen. Oxford: Oxford
University Press.
De Mey, T., and E. Weber. . Explanation and thought experiments in history. History
and Theory :–.
de Mul, J. . The tragedy of finitude: Dilthey’s hermeneutics of life. New Haven: Yale
University Press.
de Regt, H. W. . Discussion note: Making sense of understanding. Philosophy of Sci-
ence :–.
de Regt, H. W., and D. Dieks. . A contextual approach to scientific understanding.
Synthese :–.
D’Oro, G. . Robin George Collingwood. Stanford encyclopedia of philosophy. Available
at http://plato.stanford.edu/entries/Collingwood.
Fromkin, D. . Europe’s last summer: Who started the Great War in ? New York:
Knopf.
Gadamer, H.-G. . Truth and method. New York: Seabury Press.
Hempel, C. G. . The philosophy of Carl G. Hempel: Studies in science, explanation, and
rationality. Edited by J. H. Fetzer. Oxford: Oxford University Press.
Joll, J. . The unspoken assumptions. In The origins of the First World War: Great power
rivalry and German war aims, edited by H. W. Koch, –. London: Macmillan.
Kuhn, T. S. . The essential tension: Selected studies in scientific tradition and change.
Chicago: University of Chicago Press.
Macdonald, L. . They called it Passchendaele: The story of the battle of Ypres and of the
men who fought in it. London: Penguin Books.
———. . To the last man: Spring . New York: Carroll and Graf.
u n d e r sta n d i n g i n h i st o r i c a l s c i e n c e 

Mink, L. O. . Historical understanding. Edited by B. Fay, E. O. Golob, and R. T. Vann.

Ithaca: Cornell University Press.
Nussbaum, M. C. . Love’s knowledge: Essays on philosophy and literature. New York:
Oxford University Press.
Plantinga, T. . Historical understanding in the thought of Wilhelm Dilthey. Toronto:
University of Toronto Press.
Ricoeur, P. . Time and narrative. Vol. . Chicago: University of Chicago Press.
———. . Memory, history, forgetting. Chicago: University of Chicago Press.
Shapiro, A.-L. . The fog of war: Writing the war story then and now. History and
Theory :–.
Strachan, H. . The First World War: A new illustrated history. London: Penguin Books.
Stueber, K. R. . The psychological basis of historical explanation: Reenactment, simu-
lation, and the fusion of horizons. History and Theory :–.
Stummer, R. . The war we can’t let go. The Guardian. November .
Trout, J. D. . Scientific explanation and the sense of understanding. Philosophy of Sci-
ence :–.
Tuchman, B. W. . The guns of August. London: Constable.
Von Wright, G. H. . Explanation and understanding. London: Routledge and Kegan
Paul.
Walsh, W. H. . The intelligibility of history. Philosophy :–.
———. . Colligatory concepts in history. In The philosophy of history, edited by P.
Gardiner, –. Oxford: Oxford University Press.
 contribu tor s
contribu tor s

:FEKI@9LKFIJ

B^Z`Z 7ddc studied chemical engineering and philosophy. Since ,

she has been associate professor in philosophy at Twente University of Tech-
nology. Her main research interest is philosophy of science for the engineering
sciences, and includes three perspectives: philosophy of science, epistemology,
and ethics. In  she founded the Society for Philosophy of Science in Prac-
tice (http://www.philosophy-science-practice.org).
BVgXZa7djbVch is associate professor of history and methodology of
economics at the University of Amsterdam. His research interests are in mod-
els, measurement, and mathematics. He is the author of How Economists Model
the World into Numbers (), and the editor of Measurement in Economics:
A Handbook (). His current project concerns measurement outside the
laboratory.
=Vhd`8]Vc\ is professor of philosophy of science at University College
London. He is the author of Inventing Temperature (), and a cofounder
of the Society for Philosophy of Science in Practice (SPSP). His most central
research activities are in the philosophical analysis of historical episodes in the
physical sciences from the eighteenth century onward. He is also engaged in
more abstract research in philosophy of science with a focus on realism, evi-
dence, pluralism, and pragmatism.
=Zc` L# YZ GZ\i is a lecturer at the Faculty of Philosophy of the VU
University Amsterdam. He holds a master’s of science in foundations of phys-
ics (Utrecht University) and a doctorate in philosophy of science (VU Univer-
sity). He has published on topics in the history and philosophy of science, in
particular on scientific explanation and understanding. He is cofounder of the
European Philosophy of Science Association (EPSA) and of the Society for Phi-
losophy of Science in Practice (SPSP).
9Zcc^h9^Z`h is professor of the philosophy and foundations of the natural
sciences at Utrecht University and director of the Utrecht Institute for the His-
tory and Foundations of Science. He is also a member of the Royal Netherlands
Academy of Arts and Sciences, editor of the journal Studies in the History and
Philosophy of Modern Physics, and associate editor of Foundations of Physics. He
has published widely on subjects relating to space and time, the foundations and
philosophy of quantum theory, and the philosophy of physics in general.


 contribu tor s

@V^:^\cZg studied physics and philosophy of science at Utrecht Univer-

sity, specializing in the history of the natural sciences and in the philosophy of
the cognitive sciences. He is presently affiliated with the faculty of philosophy
of the VU University Amsterdam, where he teaches philosophy of science and
is finishing his doctoral dissertation on understanding in psychology.
HiZe]Zc G# <g^bb is an assistant professor of philosophy at Fordham
University. He works primarily in epistemology and the philosophy of science,
and his articles have appeared in the British Journal for the Philosophy of Sci-
ence, Philosophy and Phenomenological Research, Philosophical Studies, and
the European Journal of Philosophy.
IVg_V@cjjii^aV is a research doctor in the Department of Philosophy,
University of Helsinki. Her work focuses on modeling and scientific represen-
tation. Recently, she has been working on the theme of fiction in science, espe-
cially in the context of economics. Her areas of expertise include science studies
and semiotics, both of which she has been teaching, in addition to philosophy,
at the University of Helsinki. She has published in Philosophy of Science, Semi-
otica, Science Studies, and Forum: Qualitative Social Research. Presently, she is
the editor-in-chief of Science Studies.
:Yl^c @dhiZg studied mathematics and philosophy of religion at the
VU University Amsterdam, the Complutense University of Madrid, and Prince-
ton Theological Seminary. He is assistant professor of philosophy at the VU
University of Amsterdam. His research interests include philosophical ques-
tions regarding human evolution, anthropology, history, and religion.
?d]VccZh AZc]VgY studied philosophy and mathematics at the Uni-
versities of Heidelberg and Frankfurt. In , he finished his dissertation in
mathematics and has been affiliated with Bielefeld University since . Cur-
rently, he is fellow and assistant of the research group Science in the Context
of Application at the university’s Center for Interdisciplinary Research (ZiF). A
main focus of his research interest is on mathematical modeling, most recently
on computer simulation.
HVW^cV AZdcZaa^ is a research fellow of the ESRC Centre for Geno-
mics in Society (Egenis) at the University of Exeter. She was trained in the phi-
losophy, history, and social studies of science in London and Amsterdam. She
investigates philosophical issues connected to the use of bioinformatics in bio-
medical research, particularly the epistemic role of data, model systems, and
classifications across diverse epistemic cultures.
EZiZg A^eidc was the Hans Rausing Professor and head of the Depart-
ment of History and Philosophy of Science at Cambridge University, and a fel-
low of King’s College, until his untimely death in November . He is the
author of Inference to the Best Explanation (, nd ed. ) and is recog-
contribu tor s 

nized as one of the leading philosophers of science and epistemologists in the

world.
BVgi^cVBZgois a professor at the Institute of Sociology, University of
Lucerne. She is also a guest professor at the Technology and Society Labora-
tory, Empa St. Gall. In her research she is particularly interested in issues per-
taining to conceptual scientific practice, to the reconfigurations involved with
computer-based research, and to the constitution of novel research areas (such
as the nanosciences).
BVg\VgZi Bdgg^hdc is professor of philosophy at the University of
Toronto. She has written on a variety of topics in the philosophy of science, in-
cluding modeling in physics and biology, unity and reduction in science, Kant’s
philosophy of science, and, more recently, on emergence in physics.
?ZgdZcKVc7djlZais a postdoctoral fellow of the Research Founda-
tion Flanders (Fonds Wetenschappelijk Onderzoek Vlaanderen) and a member
of the Center for Logic and Philosophy of Science at Ghent University. He is
interested in scientific pluralism, social epistemology (for example, the democ-
ratization of science), and philosophy of the social sciences. His articles have
appeared in journals such as Economics and Philosophy, Philosophy of the Social
Sciences, History and Theory, and Journal for the Theory of Social Behaviour. He
is currently editing a volume on the social sciences and democracy.
EZig^Na^`dh`^is professor of philosophy of science at the Department
of History and Philosophy, University of Tampere. He is interested in the as-
sessment of explanations and evidence by scientists and in the relationships
between explanations provided by different scientific fields. In particular, he
focuses on the interfaces of the social sciences with neuroscience, cognitive
science, and evolutionary biology.
 pa rt title ver so
@E;<O

Note: page numbers followed by t refer to Aron, Raymond, 

tables; those followed by n refer to note, with artifact, , –, , –, 
note number. Atkinson, P., 
autonomy: of economic variables, –; of
abilities. See skill(s) models, , n, –, n, , ,
abnormal events, call for explanation inherent , n
in, – axiomatic-deductive model. See principle
Abrahamsen, A.,  theories
abstraction(s), , , –, , ; vs. axiomatization, , , , 
idealization, , –, –, n
abstract mathematical models: in biology, background beliefs, and feeling of
–; detailed knowledge provided by, understanding, –, 
–; mediating role of models and, – background conditions, and invariance of
, n; necessity of, in physics, –; economic variables, –
relationship to concrete system, , ; Baehr, J., n
and understanding, , , , – Bailer-Jones, D. M., , 
Achinstein, Peter, , – Baird, D., 
Ackerman, Robert, n Baker, G. P., 
Adelman, F. L., – Balashov, Y., 
Adelman, I., – Barlas, Y., –
agency: and action, ; models and, –, Bates, Robert, 
; role of, , , , –, n, Batterman, R., n
–; voluntary action and,  Bauer, H. H., 
“aha” experience. See feeling of understanding Beatty, J., 
aim(s) of science, , , , , , , , , , Bechtel, W., 
, , , , ,  Becker, Gary, n
algorithm(s), –, –, , –, behavioral concept, understanding as, –
–, , –. See also Great Deluge behaviorism: classical, –; intervening
Algorithm variables in, –, ; operational, .
Allison, Henry, n See also neobehaviorism
Alston, W., n behavior pattern tests, , t
Alt, James,  Bell, John, , –, 
Anderson, E., –, – Bergson, Henri, 
Anderson, M. L., n Bernstein, R. J., , 
Ando, A.,  Big Bang theory, unintelligibility of, 
Ankeny, R. A., n biology: abstract mathematical models
Ankersmit, F. R.,  in, –; epistemic skills in, –;
anticipation, demonstration of understanding and manipulation, , , ; model
through, – organisms, –, –; pluralism of
antirealism, , n theoretical approaches in, –; research
applied science, , –, ,  commitments in, –; tacit (embodied)
approximation: explanation and, –; in knowledge in, –, –, ,
mathematical reasoning, , ; models –. See also model-based reasoning;
and, –, , , , , , –, understanding in biology
,  biotechnology, 
argument patterns, , n Bishop, M., n



Black, M., 
black-box models: assessment of, –;
Lucas Critique and, 
black-box psychologists, 
Blumberg, A. E., 
Bohman, J., n
Bohr, Niels, 
Boltzmann, Ludwig, –, n, –
Boon, Mieke, , , , 
bottom-up theories. See constructive theories
Boumans, Marcel, , , n, , n,

Bridgman, P. W., , 
Brookings Institute, model of U.S. economy,
–
Brooks, R., 
Brown, Harold I., –, , , , 
Brown, Harvey R., 
Brown, J. R., , 
Buchwald, Jen, n
calibration, ; of economic models, –,
–; of instruments, ; of sense of
understanding,  (see also illusion of
depth of understanding)
Callebaut, W., 
Carlson, T., n
Carnap, Rudolf, 
Cartwright, Nancy, –, , n, ,
n, , , 
causality, , , ; and intervention, 
causal maps (Gopnik), 
causal-mechanistic approach to
understanding, , 
causal-mechanistic models, , , , –,
–, –, n, ; vs. nomo-
mathematical models, 
causation: vs. explanation, ; as hardwired
human capacity, –; and understanding,
, ; without explanation, , –
cause and effect, Hume on, 
Chaitin, G. J., 
Chang, Hasok, , n, , , , , nn–
, –nn–, –, n
circular reasoning: detection of, ; in
scientific explanation, 
CIT. See Criterion for the Intelligibility of
Theories
Clark, A., , 
Clausius, Rudolf, 
Cleeremans, A., 
cognition: embodied, , , n;
everyday, , ; explanatory, ; and
understanding, , , , , , 
cognitive abilities or skills, , , , –,
. See also epistemic skills
cognitive achievement, understanding as, ,
, , –
inde x
cognitive aims, , , 
cognitive benefits of explanation, , –,
–
cognitive science, , 
coherence: and feeling of understanding, –
, n, ; understanding as knowledge
about, 
colligatory concepts: in historical
understanding, , –; tacit standards
for, 
Collingwood, R. G., , 
Collins, H. M., n
communication of understanding, , ,
, –, , , –, . See
also explanation
community. See scientific communities
complex systems: criteria for understanding,
–, ; understanding in, 
computational complexity, –
computational models, , –, 
computer models: and computability
limitations, ; and computational
complexity, –; as productive objectual
models, –; and understanding of
complex systems, . See also general
equilibrium models; simulation-based
understanding; simulations
Comte, Auguste, 
Constraint Grammar parser, –
constructive theories: defined, –;
relation to principle theories, –; of
special theory of relativity, , –,
–, –, –
context: explanation and, , , ,
–; intelligibility and, –, , –,
n, –, , –; models and,
, , , , , –; skills and,
, –, ; understanding and, –,
–, , , , , , , , ,
, , , 
contrastive explanation, t, , , 
contrastive reasoning, –
contrastive thesis, 
control groups, and reliability in feeling of
understanding, 
Cooley, T. F., 
Copenhagen-Göttingen circle: Einstein’s
critique of, –; and pluralism of
understanding, , n
counterfactual explanation, 
counterfactual inference, –, 
counting, and discreteness, –, t
Cournet, Anton Augustin, 
Cournot problem, 
Cowles Commission, , , , n
Craver, C., 
Criterion for the Intelligibility of Theories
(CIT), –, , , –, , –
ind e x
, –, 
Criterion for Understanding Phenomena
(CUP), , , , , –, –,
, 
CUP. See Criterion for Understanding
Phenomena
Darden, L., , , 
Darwinism, synthesis with Mendelism, –
, n
Davis, D. A., 
Davis, J., n
Davy, Humphry, 
de Chadarevian, S., n, n
decomposable systems, 
deductive chauvinism, –
deductive inference: as conditionally
reliable process, n; and principle of
noncontradiction, 
deductive-nomological (D-N) model, –, ,
–, , –, 
de Figueiredo, Rui, 
degrees of freedom of molecule, and kinetic
theory of gases, –
De Mey, T., –, 
de Mul, J., , 
de Regt, Henk, , n, n, n, n,
n, , n, , n, n, n,
n, , n, , , , , ,
n, , , , n, , , ;
on application of theory to phenomena,
, –; critique of, –; on
feeling of understanding, , –, ;
on intelligibility, , –, –; on
judgment, ; on skills, –, –, ; on
understanding, –, , , , –,
, , , , , –, –,
–, 
Descartes, René, nn–; laws of
collision, Leibniz’s critique of, –;
and understanding through potential
explanations, –
description: vs. explanation, , , –,
, n, , , –, , –,
, ; in historical science, –, ;
models and, –, –, , –,
, , , ,  (see also model(s));
skills and, , 
determinism, as commitment to engage in
rational prediction, 
Dewey, John, n
Dieks, Dennis, , , , n, n, ,
n, , , , , , , , ,
, n, , , , , , , ,
; on application of theory to phenomena,
; critique of, –; on feeling of
understanding, ; on intelligibility, –;
relevance of theories to historical studies,
; on scientific understanding, –; on

skills, ; on understanding, , , ,
–, –, –, –
Dilthey, Wilhelm, –
direct structure tests, –, t
discreteness, and counting, –, t
D-N model. See deductive-nomological (D-N)
model
Donald, M., , 
double hermeneutics, 
Dowling, D., 
Dray, William, 
Droysen, Johann Gustav, , 
Dueck, Gunter, , –, , , 
Duhamel, Georges, 
Duhem, P., , n
dumbbell model, –, n
Dummett, Michael, n
Dupré, J., n
Dusenberry, J. S., 
Earman, John, –
economics: models in, , –, –,
–; neoclassical, critique of, –, ;
post-autistic, n; validity of models in,
–, –, –. See also models in
economics; understanding in economics
economics takeover (economics imperialism),
–, n
economic variables, invariance of, –
Edwards, A. W. F., n
egocentric bias, and theory selection, n
Ehrenfest, Paul, , 
Eigner, Kai, , 
Einstein, Albert, , –, n; on
constructive theories of relativity, –,
–, –; on constructive vs.
principle theories, –; introduction of
theory of relativity, , , , ; on
physical reality of relativistic effects, ,
–; on quantum mechanics, –, 
Elgin, Catherine, , n
embodied cognition. See cognition, embodied
embodied knowledge. See tacit knowledge
embodied understanding, –
empathy: in behaviorist methodology, ,
–, ; and other minds, ; and
understanding in historical science,
–, – (see also reenactment and
understanding, in historical science)
empiricism, , , , , –, , ,
, , , , n, n. See also
logical empiricism
engineering sciences, –, , , –
, n; explanation in, ; intervention
in, –, ; models in, –.
See also models in engineering science;
understanding in engineering sciences
epistemic activities: complex, ontological
principles underlying, –; hierarchy of,

; pairing with ontological principles, –
, t, –; and scientific understanding,
–, , , –, 
epistemic humanism, 
epistemic interests, and adequacy of
understanding, , –
epistemic iteration, simulation-based
understanding and, 
epistemic skills: acquisition of, ; assessment
of, –; in biology, –; defined, ,
; types of, –
epistemic value of feeling of understanding.
See feeling of understanding, reliability of
epistemic value of models, –, –,
, n
epistemic values constitutive of science,
intelligibility and, , , –
epistemology: feminist, ; and ontology,
; and philosophy of science, -, ;
social, 
Evans, R., n
event generators, –
evidence vs. explanatory power, –
evolution: and psychology, , ;
understanding of, -, -, n,

Ewens, W., n
exemplars, , –, , 
expectation, and understanding, , , –,
–
experience: cognitive, from performing
research, -, -; everyday, -
, ; personal, of historical agents, ,
, -; practical, -, ; from
performing simulations, , , , .
See also feeling of understanding
explanandum: ambiguity in, ; contrastive
explanations of, ; in explanation, , –,
, , , , , –, 
explanans: ambiguity in, ; in explanation, ,
-, , , , , 
explanation: ambiguity of concept, –;
cognitive benefits of, , –, –;
defining of, , ; evaluating quality of,
–; history of philosophical treatment
of concept, ; imprecision in explanandum
or explanans, ; indeterminate end states
of, , ; interpretive structures and,
–, –; mechanism of providing
understanding, –, , ; models and,
–, ; need for, abnormal events as
signal of, –; objectivism on, –;
and objectivity, ; as pragmatic concept,
, –; role of deduction in, –;
simulacrum account of, –; as type of
knowledge, –; types of knowledge
provided by, , ; understanding as
ability to give, ; usefulness of, –;
why-regress, 
inde x
explanation, contrastive, and principle of
sufficient reason, t, , , 
explanation, mathematical, , –, , ,
–, , –. See also abstract
mathematical models
explanation vs. description. See description, vs.
explanation
explanation in economics, –
explanation effect, 
explanation in engineering sciences, 
explanation in historical science, as intrinsic to
understanding, –, 
explanation in physics: as pragmatic
(functional), , , –, . See also
constructive theories; principle theories
explanation vs. understanding, –; in
economic models, ; history of views on,
–, ; presence of subject and, 
explanatory claims, making explicit, 
explanatory power: ambiguity of concept, ;
vs. evidence, confusion of, –
explanatory practices: improvement of, –
; lack of attention regarding, in sciences,
–, , –
explanatory understanding, , , , ,
, –, , 
externality/objectivity, and observation, 
familiarity, , , , , , –,
; feeling of understanding and, , –,

feeling of knowing, 
feeling of understanding, –, , –, ,
n; as ability or faculty, ; coherence
and, –, n, ; compared to
other feelings, –; as motivating cause
of understanding, –, , –,
–; necessity of relying on, ; as tacit
guide to theory selection, –, ; vs.
understanding, , ; understanding
without, , n
feeling of understanding, reliability of, –,
, ; anecdotal evidence, –; clinical
research on, –, –; conditional
nature of, , –, , n; contextual
factors, –, ; manipulation and,
; methods of improving, ; and
overconfidence bias, –; sources of
unreliability, –. See also illusion of
depth of understanding
Feigl, H., 
feminism, views on rational choice theory,
–
Feynman, R. P., , –
Fisher, R. A., –, , –, ,
n
Fox Keller, Evelyn, –, , 
Frank, Philipp, 
Franklin, A., , n
ind e x
French, S., 
Friedman, Michael, , , –, , , , 
Frigg, R., , , n
Fromkin, D., 
functional explanation, , , –, 
functional relations, dependencies, , –
, , , 
fundamental theories, as interpretive
structure, 
Gadamer, Hans-Georg, –, 
Galen, , 
Galileo, , –, , , n, 
Galison, P., n, n, n
gases. See kinetic theory of gases
general equilibrium models: of Adelman and
Adelman, –; axiomatizing of system
elements, ; complexity of, –,
; and Cournot problem, ; gray-box
models, –; of Lucas, –; modular
design in, –; parameterization in,
–
general theory of relativity, Einstein’s
introduction of, 
genetically modified organisms (GMOs), 
genetics, . See also population genetics
Gerson, Elihu, n
Giere, Ronald N., , , , , n,
, –, 
Gilbert, W., n
Gintis, Herbert, 
Glennan, S., 
Glymour, C., , , n
GMOs. See genetically modified organisms
Goldman, Alvin, n
Gooding, David, , n
Gopnik, Alison, –, , nn–, –,
, , n, nn–, n, 
gray-box models, –; Lucas Critique
and, ; modular design in, , ,
–; nearly-decomposable systems,
–; parameterization in, –; and
understanding, , –
Great Deluge Algorithm, –, , 
Greco, J., n
Green, Donald, 
Griesemer, J. R., , , n
Grimm, Stephen R., nn–, , , , 
Gutin, G., 
Haavelmo, T., –, n
habit strength, Hull on, –
Hacker, P. M. S., 
Hacking, Ian, , , n, n, ,
n
Haldane, J. B. S., n
Hardy-Weinberg law, , –, 
Harris, W. S., n
Hartmann, S., , 

Heisenberg, Werner, , , 
Hellmann, Gunther, 
Hempel, Carl Gustav, n, n, n,
, , , n; on causation, ;
and D-N model, –; objectivism of,
–; positivism of, ; on subjectivity
of understanding, , –, ; on
understanding, 
Hergenhahn, B. R., 
hermeneutics, –, 
hierarchic systems, 
hindsight bias, and feeling of understanding,

historical narratives, construction of, 
historical science: description in, –,
; explanation in, –, ; ideal of
objectivity in, –; intelligibility in, –
, ; judgment in, , n, , ,
–; narrative in, , –, , ;
vs. natural sciences, , –; objectivity
in, , –; positivism in, ;
standards of evidence in, ; subjective
elements in, –, , , –;
tension between objectivity and subjectivity
in, –. See also understanding in
historical science
history of science, , –, –, n,
, , n, 
Hitchcock, Christopher, n
Holland, O. A., 
Hookway, C., n
Hoover, K. D., , 
Hopwood, N., n, n
how much-question: Lucas general
equilibrium model and, ; model type
appropriate to, t; type of answer to, ,
t
how’s that-questions: defined, ; Lucas
general equilibrium model and, ; model
type appropriate to, , , t; type of
answer to, t
Huala, E., n
Hughes, R. I. G., n
Hull, Clark L., , , –, 
human agency, -; in model-based
reasoning in biology, ; in productive
objectual models, , –, , ; in
understanding, , . See also agency; skill(s)
Hume, David, , , n, 
Humphreys, P. W., , , , , , ,
, 
Hutchins, E., , 
hypotheses: construction of, –, , ,
–; evaluation of, –, –,
–
hypothetico-deductive model, 
IBE. See Inference to the Best Explanation
idealization: vs. abstraction, , –, –

, n; in model-building, –, , 
identification, and principle of identity of
indiscernibles, 
IDU. See illusion of depth of understanding
illusion of depth of understanding (IDU):
avoiding, –; causes of, ;
experimental demonstration of, –; in
scientific understanding, –, 
images: as source of tacit causal information,
–, ; and understanding, , ,
n, 
induction, and prediction, –
inference(s): counterfactual, understanding
and, , ; tacit dimension of, –
Inference to the Best Explanation (IBE), ,
–, 
integrated understanding, 
intellectual vices, and feeling of understanding,
–, , 
intellectual virtue, as scientific method, n
intelligibility, –; as activity-dependent,
; as basic assumption of scientific world
view, ; case study (kinetic theory of
gases), –, n; criteria for, –; as
criteria for acceptance of scientific theory,
–; defined, , ; de Regt and Dieks
on, –, –; as epistemic value
constitutive of science, , , –;
frontier of, scientific advances and, –;
in historical studies, –, ; history
of concept, ; interpretive structures and,
–; neobehaviorism and, , –;
ontological principle underlying epistemic
activity, –, t; as performability of
epistemic activity, ; as pragmatic concept,
–, –, ; scientists’ preference for,
; and understanding, , , 
intentional explanation, –
interaction. See manipulation
international relations studies: theoretical
orientations in, . See also unification in
international relations studies
interpretation: causal-mechanistic vs. nomo-
mathematical, –; historical, ;
knowledge and, , , , , ;
models and, –; necessity of, ;
observation and, n; principles of, ;
and surplus meaning, –; theory and,
, –
interpretive structures: case study, –;
explanation and, , –, –; and
human cognition, –; integration of
different frameworks in, –; necessity
of, for generating theory, –, ; and
scientific understanding, , –
intervening variables, in behaviorism, –,

intervention: in economics, –; in
inde x
engineering, –, ; feeling of
understanding and, ; in life sciences, , ,
; models and, –; and principle of
causality, ; and understanding, , –,
, , 
intuition: into consequences of theory, as
knowledge, –; and intelligibility, ; in
judging of competing theories, ; origins
of, . See also empathy; familiarity; feeling
of understanding; skill(s)
isomorphism of models, and representation,
–
James, William, n, n
Janssen, M., 
Joll, James, , 
judgment: about reliability, ; defined, ;
errors in, n, ; and explanation effect,
; feeling of understanding and, –;
in historical science, , n, , ,
–; and interpretation, –; models
and, , , , , ; refinement of,
by community, –, ; and scientific
explanation, , , , –, –, ;
and scientific understanding, , , , ,
, , , –, , , –;
synoptic, –, ; and understanding in
historical science, –. See also illusion
of depth of understanding; skill(s)
Juslin, P. A., 
Kant, Immanuel, n, n, n; critique
of universal categories in, , , , –; on
intelligibility, ; and subjective certainty, 
Kargon, R., 
Karlsson, F., , 
Keil, Frank, C. , , –, 
Kelvin, Lord, , 
kinetic theory of gases, development of, –,
n
Kitcher, Philip, , , , n, , , ,

Klein, Lawrence, –
Klein, U., n
Knorr-Cetina, Karin, , n
knowing, feeling of. See feeling of knowing
knowledge, , ; ability to use, ,  (see also
skill(s)); about coherence, as understanding,
; from abstract mathematical models,
–; explanation and, –, –,
; and human agency, , ; intuition
into consequences of theory as, –;
from models, , ; objectivity of,
; rationalist philsophers on, ; vs.
understanding, –, ; understanding
and, , ; understanding required for,
–. See also tacit knowledge; theoretical
knowledge
ind e x
Knuuttila, Tarja, , , , , n, , ,
n, –, , , n, 
Kochelmans, J. J., 
Kolmogorov-Chaitin complexity, –
Koriat, A., 
Koster, Edwin, , , , n
Kratochwil, F., 
Krohs, U., 
Kuhn, Thomas S., –, n, –, ,
, –
Kuorikoski, J., , , 
Kvanvig, Jonathan, –, n
Kydland, F. E., –
laboratory sciences, , –, –,
n
Lacey, H., n, 
Ladyman, J., 
Lambert, K., 
Landman, U., 
language technology, –, 
Latour, Bruno, n, n
Laub, J., 
Laubichler, M., n
law(s): causal, , ; covering, , ;
empirical, ; fundamental, , ,
, –, –, , ; vs.
generalizations, –, –;
general or universal, ; historical,
, ; mathematical, –, ;
phenomenological, , , ; physical,
, , , , –, , –, , ,
, –; subsumption under, , ;
theoretical, –, , –, , ,
, . See also theoretical principles
Leibniz, Gottfried Wilhelm: critique of
Descartes’ physics, –; principle of
continuity, –, , ; and principle of
identity of indiscernibles, 
Lenhard, Johannes, , , , , , ,
n
Leonelli, Sabina, , , , , n, n,
n, n, n, n
Levi, Margaret, 
Lewis, C. I., n, 
Lewis, D., 
Lewis, Gilbert, 
light principle, in theory of relativity, , ,

Lin, L.-M., , 
Lipton, Peter, , , , –, , n, ,
n, , n, , , , 
local knowledge, –, , 
local vs. nonlocal meaning, –, 
logical empiricism: on intelligibility, ; and
neobehaviorism, –; on scientific
understanding, , –, 
logical positivism: on intelligibility, ; and
neobehaviorism, –, , –

Longino, H., n, , 
Lorentz, H. A., , , 
Lucas, R. E., –, , 
Lucas Critique, –
Luker, Kristin, –
MacCorquodale, K., , 
Macdonald, Lyn, 
Mach, E, n
Machamer, P., , 
Mackay, H., n
Mäki, U., , 
Maliniak, D., n
manipulation, n; biological science and, ,
, ; demonstration of understanding
through, ; and reliability of feeling of
understanding, ; simulations and, ;
as source of tacit causal information,
, –, , ; understanding and, ,
, –, . See also understanding
through models
Mankiw, N. G., , n
Mattila, E., n
Maxwell, James Clerk, , n, n, n
Mayo, D. G., n
Mayr, Ernst, , 
McAllister, J. W., n
McMullin, Ernan, n, 
mechanical explanation, ontological principles
underlying, 
mechanical models, –, , , –, ,
, –, , 
mechanism(s): causal, , , , –, ,
; exemplar, –; and extraction of
understanding, –, , , , –, ,
, –, , ; models as, , ,
, , 
mechanistic approach to understanding, 
Meehl, P. E., , 
Mendelism, synthesis with Darwinism, –
, n
mental models, understanding as possession
of, –
Merz, Martina, , , , , , ,
n, –, , , n
metaphysical belief(s), commitments(s),
principle(s), truth(s), , , , , , ,

metaphysics as guide in science, and
ontological opinion vs. ontological principle,

meteorological models, accuracy vs.
representativeness in, 
Meyers, D., 
Mill, John Stuart, , 
Mills, C. M., 
Mink, L. O., –, 
Minkowski, Hermann, –, , n
Mitchell, S., , n, n

model(s), –; as concrete, constructed
objects, , –; generation of
knowledge through, ; intelligibility of,
; as interpretation of theory, –,
; mechanical, –, , , –, ,
, –, , ; media-specific
approach to, ; as mediators, , –,
n, –; mental, – (see also
visualization); as multiplex epistemic
objects, , , –, , ; vs.
phenomenon, –; productive account
of, –; semantic conception of, –;
skill required to use, ; as source of
tacit causal information, ; vs. theory,
n; types, , . See also abstract
mathematical models; black-box models;
computer models; general equilibrium
models; gray-box models; productive
objectual models; representativeness of
models; simulation-based understanding;
simulations; understanding through models;
white-box models
model-based explanation, , –, –,
–, n, , –, –, –
. See also explanation, mathematical
model-based reasoning: abstract mathematical
models and, , –, –; defined,
; human role in, , –, , ,
; pragmatic (functional) nature of, ,
–; productive objectual models, –
; representational value, ; types of, 
model-based understanding: with abstract
mathematical models, , , , –;
in biological science, –, –;
of complex systems, ; with economic
models, –, –; with gray-box
models, , –; by model building,
, –, , –, 
model building: intelligibility and, –;
knowledge generated by, , ; as
nondeductive process, –; subjective
aspects of, –; understanding generated
by, , –, , –, 
model organisms, –, –
models in economics, , –, –,
–. See also general equilibrium
models; gray-box models
models in engineering science, –
modular design models: assessment of, ;
gray-box models as, , , –
molecular biology, productive objectual
models in, –
Moltke, Helmuth von, 
Moravcsik, A., n, –, 
Morgan, M. S., , , , n, ,
n, ; on autonomy of models, ;
on experimentation vs. modeling, ; on
learning through models, , , ;
inde x
on models as connection between theory
and phenomena, , ; on pragmatic
(functional) nature of models, 
Morrison, Margaret, , , , –nn–
, n, n, , , n, ,
n; on autonomy of models, ; on
learning through models, , ; on
models as connection between theory
and phenomena, , ; on pragmatic
(functional) nature of models, 
Mr. Perestroika movement, n
Müller, G. B., n
Myers, N., , , 
Nagel, E., , 
nanoscience, simulation-based understanding
in, 
narration, and principle of subsistence, t, 
narratives: and feeling of understanding,
; historical, , –, , ; as
knowledge, 
natural science(s): vs. historical science, ,
–; judgment in, ; psychology as, ,
, ; vs. social sciences, 
nearly-decomposable systems, –
necessity, as concept rooted in human action,
, –, 
Nedelsky, J., 
neobehaviorism: Hull’s work in, –;
and intelligibility, ; tenets of, –;
Tolman’s work in, –
Neurath, Otto, 
New Experimentalists, n
Newton, Isaac, , , , , n, ,

nomo-mathematical models, , n; and
interpretive structures, –
noncausal explanations, , n
normative frameworks: on aim of science,
; for evaluation of understanding, , ;
and historical science, ; and integrated
understanding, 
Norton, B. J., n
Nozick, R., 
Nussbaum, M. C., 
object, defined, n
objectivism: de Regt on, –, ; on
explanation and understanding, –, ,
–, –
objectivity: explanation and, ; feeling of
understanding and, ; in historical science,
, –; limitations of, , , –;
models and, –; and observation,
–; psychology and, –
observable(s), , , , n, , ,
n
observation, and externality/objectivity, 
ind e x
Olsson, H., 
ontological opinion, vs. ontological principle,
, 
ontological principles: of complex epistemic
activities, –; definition of, –;
necessity of, derived from epistemic
activities, ; non-universality of, ; vs.
ontological opinion, , ; pairing with
epistemic activities, –, t, –
ontology: definition of, ; of objects, ;
pragmatic roots of, –
operational behaviorism, 
operationalization, in neobehaviorism, ,
, 
Oppenheim, Paul, , 
Orcutt, G. H., –
ordering (linear), and principle of transitivity,
t, –
Ostrom, Elinor, 
other minds, and empathy, 
overconfidence bias, and reliability of feeling of
understanding, , –
overdetermination, , . See also
testing-by-overdetermination
parameterization, , , –
parsers, as productive objectual models,
–
particle physics, models in, –
Pearson, Karl, , –, n,
nn–
peer review system, –, 
Peirce, Charles Sanders, , n
perception, n, , , , 
performative commitments, –
performative skills: defined, ;
distinguishing from theoretical skills, ,
–
persistence. See principle-activity pairs
Peterson, S., n
Petzoldt, Joseph, 
phenomena: applying theoretical principles to,
–; as constructed entities, –; vs.
model, indistinct boundary between, –
. See also Criterion for Understanding
Phenomena
phenomenological laws, , , 
phenomenology of understanding, , , ,
–, , 
philosophical theories of explanation, , 
philosophy of science: relevance of scientific
understanding to, ; on scientific
understanding, –, 
phronèsis, , , 
physics: explanation in, as pragmatic, ,
, –, ; necessity of abstract
mathematical models in, –; pluralism
of understanding in, ; thought

experiments in, –, ; understanding
in, –, –, . See also abstract
mathematical models; constructive theories;
principle theories
Plantinga, T., , 
Plato, 
pluralism, strategic, 
pluralism of understanding, –;
Copenhagen-Gottingen circle and, , n;
matching understanding to research type,
; in physics, ; pluralism of epistemic
interests, and theoretical diversity, –;
pluralism of theoretical approaches, in
biology, –
Polanyi, Michael, , –, 
political science, , –, n. See
also understanding in political science;
unification in international relations studies
Pooley, O., 
population biology, theoretical understanding
in, 
population genetics, abstract mathematical
models in, , –
positivism: on explanation, ; in historical
studies, . See also logical positivism
possibility, understanding of, , , , –,
, 
post-autistic economics, n
practical understanding, vs. theoretical
knowledge, n.
pragmatic approach: and epistemic value of
models, ; on explanation, –, –; on
representation, –
pragmatic nature of models: in biology, ,
–; in economics, –, ; in
engineering science, –; in science, 
pragmatic necessity, 
pragmatic understanding, ; characteristics
of, ; epistemic relevance of, –, ;
intelligibility as, –, –, ; scientific
understanding as, , , –, , –.
See also skill(s)
prediction: as feature of understanding, n;
and principle of uniform consequence,
–, t. See also expectation
Prescott, E. C., –, 
principle-activity pairs: of agency and
voluntary action, ; of causality and
intervention, ; of identity of indiscernibles
and identification, ; and intelligibility,
–, t, –; of noncontradiction and
deductive inference, ; of single value and
testing-by-overdetermination, –, t;
of subsistence and narration, t, ; of
sufficient reason and explanation, t, ; of
transitivity and linear ordering, t, –;
of uniform consequence and prediction,
–, t

principle of continuity (Leibniz), –, , 
principle of single value, , , n
principle theories: defined, ; relation to
constructive theories, –; special
theory of relativity as, –, –, 
probability, , , , , n
productive objectual models: as autonomous,
concrete objects, –, –; epistemic
value of, –, –, , n;
human role, , –, , –; and
scientific understanding, –; and skills,
, –, 
Provine, William B., 
Psillos, S., n
psychology. See behaviorism; neobehaviorism
Ptolemy, , , , 
Punnen, A. P., 
Punnett, Reginald, 
Putnam, Hilary, 
Putnam, Robert D., n
quantum theory: Einstein on, –; impact of
theory of understanding, –; and principle
of uniform consequence, 
Radder, Hans, , n, n, n, ,

rational choice theory (RCT): and adequacy
of understanding, ; feminism views
on, –; unification of international
relations studies under, –
rationalism, , , 
rationality, –, , , 
RCT. See rational choice theory
realism, , , , , , , n, n,
, , n, , 
Reber, A. S., 
reductio ad absurdum, and understanding,
–
reduction, , , n
reductionism, n, 
reenactment and understanding, in historical
science, , , , 
Reichenbach, H., 
relativism, –, 
relativistic effects, physical reality of, –
relativity, theory of: as principle theory,
–. See also general theory of relativity;
special theory of relativity
relativity principle, in theory of relativity, ,
, 
reliability assessment, environmental factors
in, –
reliability of feeling of understanding, –,
, ; anecdotal evidence, –; clinical
research on, –, –; conditional
nature of, , –, , n; contextual
factors, –, ; manipulation and,
; methods of improving, ; and
inde x
overconfidence bias, –; sources of
unreliability, –. See also illusion of
depth of understanding
representation(s): and explanation, , , ,
, –, ; in historical studies, –
. ; internal, –; and interpretation,
; mathematical, , , n, n,
, ; models and, , n, –, ,
n, –, , , –, n,
n, –, , , , , –,
–, ; modular, ; perception and,
; representational vs. productive account
of models, –; simulations and, ;
and understanding, , , , n, ,
, , , . See also model(s)
representativeness of models: vs. accuracy,
–; in biology, factors affecting,
; and epistemic value, , –;
and generation of new knowledge, ;
pragmatic view of, –; vs. productive
account, –, n
research commitments, in biology, –
revolution(s), scientific, , , –, 
Rheinberger, Hans-Jörg, , , n
Ricoeur, P., , , , 
Riggs, Wayne, –
Rips, L. J., 
Romine, G., 
Rosenberg, A., n
Rouse, J., 
Rozenblit, L., –
rule(s): epistemic activities and, n; vs.
exemplars, ; internalization of, ; models
and, , , , , , –, –;
within scientific communities, –, ;
skills and, –, , , , –, ,
, , , , –
Ryle, Gilbert, –, 
Salmon, Wesley C., , , , n, n,
–, –, , n
Schaffer, S., n
Schilpp, P. A., 
Schlick, Moritz, 
Schram, S., n
Schrödinger, Erwin, , , –, n, 
Schurz, G., 
science(s): affective and emotional component,
; aims of, , , , , , , , , ,
, , , , , ; applied vs.
pure, , –, , ; behavioral, 
(see also behaviorism; neobehaviorism);
cognitive, ; computer, , ,
; division of cognitive labor in, ;
empirical, –; engineering. See
engineering sciences; explanation as goal
of, –; historical. See historical science;
as intersubjective enterprise, –;
laboratory, , –, –, n;
ind e x
life, , ; natural, , , , , , ,
, , , , –; normal, –;
physical, , ; political, , –,
n; revolutionary, , , –, ;
role of scientific understanding in, ;
social. See social sciences; theoretical, ;
unity of, ; values in, , , 
scientific communities: and epistemic
skills in biology, ; refinement of
individual judgment by, –, ; scientific
understanding as social achievement, ;
skills as asset of, , ; and understanding
in biology, –
scientific method: as hardwired human
function, ; internalized, as intellectual
virtue, n
scientific practice, , n, , , , ,
, ; accounts of, , ; complexity
and, ; and explanation, ; and
intelligibility, , , , , ; models
and, , –, , , ; role of
skills in, , ; in social sciences, ;
tacit knowledge and, – (see also tacit
knowledge); theory and, –, , ;
and understanding, , , , , , ,
, , . See also skill(s)
scientific theories: development of, skills and,
–; evaluation of, ; vs. model, n;
virtues of, , n
scientific understanding: establishing limits of,
; as explanatory understanding, , ;
interpretive structures and, , –;
judgment and, , –; as knowledge
about how to perform an epistemic activity,
, –, ; objectivism on, –, ,
–; vs. other forms of understanding,
; as pragmatic, , , –, , –;
productive objectual models and, –;
role in science, ; simulation-based
understanding as, –; as skill, –,
; as social achievement, ; unification
and, , , –; variety of avenues to,
. See also understanding
scientific world view, comprehensibility as
basic assumption of, 
Scriven, Michael, , , n
semantic conception of models, –, 
Shanks, D. R., 
Shapin, S., n
Shapiro, Ian, , , 
Shomar, T., 
Simon, H. A., –, n
simplicity: of models, –; of theories, ,
, , , , , 
simulacrum account of explanation, –
simulation-based understanding: as ability,
–; black-boxing of theory in, –,
–; Great Deluge Algorithm, –,

; as model type, ; as scientific
understanding, –; validity vs.
representational accuracy in, –
simulations, –, –, , ; black-
box schemes in, –; representational
accuracy of, ; undermining of link
between theory and understanding in,
–, –; uses of, 
skill(s): as community asset, , ;
epistemic relevance of, –, –; in
evaluation of scientific theories, ; in
historical studies, ; model use as, ;
role in scientific practice, ; scientific
understanding as, –, ; simulation-
based understanding and, ; social skills
as epistemic skill, –; and subjectivity,
; and surplus meaning in theoretical
terms, –, ; as tacit knowledge,
, ; vs. theoretical understanding, ;
understanding as, , –, , , ,
, –, –, ; understanding
without explanation through, , , ,
. See also epistemic skills; judgment;
pragmatic understanding; tacit knowledge
skill(s), development of: methods of, ;
through use and development of models,
, –, 
Smith, L. D., , 
Smith, S., , –
Sobel, D., n
social commitments, 
social sciences, ; epistemic interests and,
, –; vs. historical studies, ;
policy research and, nn–; positivism
and, ; scientific practice in, ;
understanding in, , –; usefulness
of, n; vs.natural sciences, . See also
political science; unification in international
relations studies; unification in social
sciences
social skills, –
sono-luminescence, –
Sosa, Ernest, n
space-time, as context for intelligibility, 
special theory of relativity: constructive
theories of, , –, –, –;
debate over proper method of explanation,
; Minkowski’s explication of, –;
physical reality of relativistic effects, –;
as principle theory, –, –, ;
relation between constructive and principle
theories of, –
specific heat ratio, –
Spence, Kenneth W., 
Stachel, J., , 
Stanford, Kyle, n
Star, S. L., n
stellar models, 

Steyvers, Mark, –, –
Stokey, N . L., 
Strachan, H., , 
strategic pluralism, 
structuralist view of models, –
structure-oriented behavior tests, , t
Stueber, K. R., –, –
Suárez, M., –, , n, n, 
subject, necessity of, in understanding, , 
subjective certainty, and truth, –. See also
feeling of understanding
subjectivity: in historical science, –,
, , –; skills and, ; in
understanding, –, 
subsumption model, 
subsumption under laws, , 
superconduction, model of, as abstract,
–
surplus meaning, in theoretical terms, –
surprise, feeling of, compared to feeling of
understanding, –
symmetry: breakdown of, in superconductors,
; natural laws and, , , , n;
representation and, 
synoptic judgment, –, 
synthetic a priori, as ontological principle, 
tacit (embodied) knowledge: in biology
research, –, –, , –;
from causal information, –; defined,
; distinguishing theoretical knowledge
from, ; from exemplars, –, ; and
judgments about explanatory quality, ;
from manipulation/causal information, ,
; skills as, ; understanding as, . See
also skill(s)
tacit (embodied) understanding, –
Taylor, Georgette, n
technology, –. See also language
technology
Teller, P., n
testing-by-overdetermination, and principle of
single value, –, t
theoretical commitments, –
theoretical knowledge: vs. practical
understanding, n; role of, , –,
–, , –; and tacit knowledge,
–, –, , –
theoretical model, as model type, 
theoretical principles: applying to particular
phenomena, –; in construction of
models, 
theoretical skill, defined, 
theoretical understanding: defined, ;
demonstration of, ; understanding
through models and, –
theory: model as interpretation of, –,
; necessity of interpretive structures
for, –, ; plurality of, and
inde x
understanding, –; as representation of
scientists’ understanding, ; as ultimate
aim of science, –; and understanding,
–. See also constructive theories;
principle theories; scientific theories
theory, understanding of, vs. understanding of
phenomena, 
theory construction, understanding and,
–
theory vs. model, in economic understanding,

theory of relativity: as principle theory,
–. See also general theory of relativity;
special theory of relativity
theory selection: criteria for, ; intuition
and, ; judgment and, ; practical
factors affecting, ; psychological factors
affecting, n
thermodynamic phase transition, model of, as
abstract, –
Thissen, W. A. H., 
thought experiments: Galileo’s, –, ;
in historical science, –; in physics,
–, 
Tierney, M. J., n
Tolman, Edward C., , –, , ,
–
Tolman, Richard, 
top-down theories. See principle theories
Torretti, Roberto, n, n, n
traveling salesman problem, –
Trout, J. D., , nn–, , ; on feeling of
understanding, –, –, –, n,
n, n, –, , ; objectivism
of, ; on understanding, , ; on
understanding vs. feeling of understanding,
n
truth: as byproduct of understanding, ; and
coherence, n; empirical, ; feeling of
understanding and, , , ; historical
science and, ; vs. intelligibility, n,
, , , , ; knowledge and, ;
necessary, , , n; representation and,
–, n, ; science and, , ,
, , ; theory of, , , –, n;
understanding and, , n
Tuchman, Barbara, , 
Turing, Alan, , 
Turing test, for economic models, –,
–
two-rockets problem (Bell), , 
understanding: adequacy of, –; as
cognitive achievement, , , , –;
in complex systems, ; confusing
with feeling of understanding, ; and
construction of theories, –; degrees
of, ; epistemic relevance of, –;
inherent subjectivity of, ; integrated, ;
ind e x
intelligibility and, , , ; vs. knowledge,
; as knowledge about coherence, ;
from mathematical explanations, –;
multiplicity of forms, ; prediction as feature
of, n; psychological vs. cognitive,
–; regulation of desire for, ; as spark
of insight, ; subjective dimensions of,
–, ; types of, –; understanding
vs. explanation, –, , ; uses of term,
–; vagueness of concept, ; variety
of avenues to, , ; without feeling
of understanding, , n.. See also
scientific understanding
understanding, definition of: as ability to
coordinate theoretical and tacit knowledge,
–; difficulty of formulating, , ;
as grasping of relations between facts, ;
historical variation in, ; impossibility
of, ; as skill, , –, , –,
–, ; as skill to infer and intervene,
, , 
understanding, demonstration of:
through anticipation, –; through
communication, ; through successful
interaction, 
understanding, simulation-based. See
simulation-based understanding
understanding in biology: as ability to
coordinate tacit and theoretical knowledge,
–, , –; scientific communities
required for, –; types of, –
understanding in economics: explanation
and, –; Lucas’s program of general-
equilibrium and, –; models and,
–, –
understanding in engineering sciences:
criteria for, –; defined, ; goals
of, –; interpretive structures and,
–; vs. intuitive skills, ; as pragmatic
(functional), –, , n
understanding in historical science: empathy
and, –, –; explanation as
intrinsic to, –, ; intelligibility
and, , ; judgment and, –;
reenactment and, , , , ;
subjective dimensions of, –; through
colligatory concepts, , –; through
intentionial explanation, –
understanding through models: with abstract
mathematical models, , , , –;
in biological science, –, –;
of complex systems, ; with economic
models, –, –; with gray-box
models, , –; by model building,
, –, , –, ; and
theoretical understanding, –
understanding of phenomenon: confusing
understanding of theory with, –;
as pragmatic, ; scientific skill and,

–; understanding of model as, –;
understanding of theory as prerequisite for,
, , , –
understanding in physics, –, –, 
understanding in political science: epistemic
interests and adequacy in, –; pluralistic
view of, –; and rational choice theory,
–, –; synthesis and, –;
unification and, 
understanding in psychology: behaviorist
methodology, –, –; models and,
–
understanding in social sciences, , –
understanding of theory: confusing
understanding of phenomenon with,
–; as prerequisite for understanding
phenomenon, , , , –; simulation
modeling and, –
understanding without explanation, ; of
causation, –; of necessity, –; of
possibility, –; of unification, –
unexpected events, –
unification: in biological sciences, –;
and evaluation of theory, ; explanation
and, , ; models and, ; and scientific
understanding, , , , , , , –,
, n, , , , –; skills
and, ; and Supreme Court, ;
understanding of, without explanation,
–
unification in international relations studies:
advantages of, ; critique of, ; under
rational choice theory, –; and
scientific understanding, –; under
theoretical synthesis, –
unification in social sciences, –;
critique of, –, , –; goals of,
–
Universe, origin of, unintelligibility of, 
unobservable: entities, n, ; world,
, n
validity of models in economics: assessment of,
–, –, –; defined, 
Van Bouwel, Jeroen, , –, , n
van Daalen, C. E., 
van Fraassen, Bas C., –, , , , , ,
, n, n, n
Varičak, Vladimir, 
Verbraeck, A., 
visualizability, , , , , , , ,
–, 
visualization: explanation and, , ,
, , –, –, , –;
models and, , ; ontological principles
underlying, 
voluntary action, and principle of agency, 
von Ignatowski, W., 
von Neumann, J., 
 inde x

von Ranke, Leopold, – why-regress, 

von Wright, G. H.,  Wilhelmson, R., 
Voutilainen, A., ,  Winman, A., 
Winsberg, E., , , 
Waddington, C. H.,  Winther, R. G., n
Walsh, W. H., –, , ,  Wittgenstein, Ludwig, , n, 
Walt, Stephen, – Woodward, James, , , , , , n,
Waskan, J. A.,  –, 
Watson, H. W., n Woody, A., 
Watson, John B., ,  Woolgar, S., n
weather forecasting, ; accuracy vs. worldlines (Minkowski), 
representativeness in,  world-points (Minkowski), 
Weber, E., –, n, –,  Wright, Sewall, n
Weber, Max, 
Weingast, Barry,  Ylikoski, Petri, –, n, , , , ,
Whewell, William, ,  , , n, , , n
White, K. Preston, –
white-box models: assessment of, –; Zabrucky, K. M., , 
complexity of, as limitation, , –; Zagzebski, L., n
development of alternatives to, – Zhang, J., 
why-questions, , , , , , ; model Zuriff, G. E., , 
type appropriate to, t; type of answer to,
, t