Download as pdf or txt
Download as pdf or txt
You are on page 1of 4

Taking Aim at Biased Algorithms

Author(s): Rachel Levy


Source: Math Horizons, Vol. 25, No. 1 (September 2017), pp. 5-7
Published by: Mathematical Association of America
Stable URL: http://www.jstor.org/stable/10.4169/mathhorizons.25.1.5
Accessed: 09-11-2017 20:07 UTC

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
http://about.jstor.org/terms

Mathematical Association of America is collaborating with JSTOR to digitize, preserve and


extend access to Math Horizons

This content downloaded from 128.122.230.132 on Thu, 09 Nov 2017 20:07:38 UTC
All use subject to http://about.jstor.org/terms
Taking Aim
at Biased Algorithms
C
Rachel Levy
athy O’Neil is a mathematician, data
scientist, author, activist, and blogger (at
mathbabe.org). She has worked in higher
education, on Wall Street, and for various
start-up companies.
Her 2016 book Weapons of Math Destruction aims to
help the public better understand the impact that algo-
rithms have on everyone’s life. Many students trained in
the mathematical sciences take jobs directly addressed
in the book, such as in data science and finance. O’Neil
raises new issues about the impact of mathematicians’
work on society. As topics in the news heighten public
awareness and concern about the power and role of
Adam Morganstern
algorithms, mathematicians have an opportunity to Cathy O’Neil.
provide new tools to foster transparency, equity, and
benevolence. are somehow dealing with objective truth, but every
I chatted with Cathy O’Neil in January. This inter- step along the way requires human intervention. This
view has been edited for length and clarity. misconception that algorithms are somehow revealing
objective truth is the most important clarification that I
Rachel Levy: What are the main take-home messages of
want to make.
Weapons of Math Destruction?
In fact, it is the opposite. These algorithms are often
Cathy O’Neil: My book describes the way mathemat-
propagating historical biases and past mistakes. It is a
ics and the trust of mathematics is used against the
particular shame when the brand of mathematics is be-
public. The public trusts and fears mathematics. So
ing deployed to protect something that is fundamentally
when marketers, salespeople, or even scientists represent
immoral and corrupt.
algorithms as mathematical and when they represent
machine learning as sophisticated and opaque, people do RL: What are some examples of weapons of math de-
not question them. Therefore, those automated systems struction (WMDs)?
remain almost entirely unaccountable, and bad things CO: Weapons of math destruction are algorithms that
can happen. Algorithms are not mathematics. They are important, opaque, and destructive. There are ex-
have mathematical attributes, but they are ultimately amples all across normal everyday life for Americans.
working with human-curated data on a human-curated Examples include assessments for teachers that
agenda. are happening in many large cities. They are called
We—the developers—embed our biases in the algo- “value-added models” and have been addressed by the
rithms. Not to mention we have chosen what data to use American Statistical Society. Judges use predictive po-
to train our algorithms, and the data we have chosen is licing and recidivism-risk algorithms for parole, bail, and
a social construction. There’s a belief that algorithms sentencing. Political microtargeting algorithms inform

www.maa.org/mathhorizons : : Math Horizons : : September 2017 5


This content downloaded from 128.122.230.132 on Thu, 09 Nov 2017 20:07:38 UTC
All use subject to http://about.jstor.org/terms

mh-05-07-levy-final.indd 1 9/29/17 1:00 PM


the kind of information we get from candidates. And
some kinds of tailored advertising, which is an area I
“ Algorithms are not
mathematics. They have
mathematical attributes, but
worked in until I realized that I wanted no part of it.
The same technology that shows me luxury yarn and they are ultimately working
knows that I am very vulnerable to such advertisements
with human-curated data on a


is used by predatory for-profit colleges and payday lend-
ers to target single black mothers who want a better life
for their children. human-curated agenda.
These algorithms don’t affect everyone equally. The
working class and working poor are affected more often factors such as tuition cost. The model has spun off an
by these algorithms than highly educated and well-off industry of side effects, including growing administra-
people. Well-off people can get through their lives fairly tions and unnecessary expenditures, not to mention
unscathed by the kinds of corporate and government rising tuitions. That’s just one example of many that
surveillance that I worry about in the book. I urge illustrates how important it is to have a good definition
people to look at algorithms and weapons of math de- of success.
struction through the lens of class and race. The algorithms that are both important and nonde-
The people building and deploying the algorithms structive are the ones that help the people who are the
are often well intentioned but naive. They are technolo- least lucky or are suffering the most. There are colleges
gists—sometimes mathematicians, computer scientists, using algorithms to find struggling students, especially
or statisticians—and they have an arm’s-length perspec- struggling freshmen, and connecting them with advising
tive on the targets of the algorithms that they build. support.
They do not acknowledge or understand the kind of It is important that the people who are struggling
effects they are creating. the most and raising the largest number of red flags are
RL: Can you name some algorithms that are designed and not being punished. There is an example of this that I
blogged about—Mount St. Mary’s College. They seemed
used in ways that are relatively fair and just? What quali-
to be doing this identification and expelling students
ties make them so?
before the U.S. News survey was due.
CO: One of my favorite examples is sports. The amount
You asked for a good algorithm, and I came up with
of sports data is blossoming. The data comes from
one that is being used to do good but could easily be
the games, and the games are on public view, such as
used to do harm. Therein lies the conundrum. It really
national television. The top radio shows serve the pur-
depends on the usage. A good rule of thumb is to ask:
pose of cleaning the data—they’ll talk endlessly about
“Is it helping or punishing the people who are the worst
whether a play should have been an error or a base hit.
off ?”
We have transparency.
Algorithms make moral decisions all the time, and
One of the most difficult aspects of building algorithms
the people programming the algorithms should not be
fairly is to have a well-defined and agreed-upon defini-
the ones making those moral decisions. Right now, we
tion of success. In sports, success is clearly defined: A
are conflating the job of building the algorithm with
sports team wants to win games. You could argue they
answering the moral questions.
want to make as much profit as possible, but there is at
least a correlation between the two. If you look under RL: What important things have happened since the
the covers of many of the algorithms that are destruc- hardcover version of your book was published that we can
tive or ineffective, you’ll see that their definition of suc- read about in the upcoming paperback version?
cess is ambiguous, incorrect, or so hard to measure that CO: The run-up to the election largely happened after
unintended consequences abound. I finalized my book. So the biggest gaping hole is the
In the book I talk at length about the U.S. News & way propaganda, fake news, hoaxes, and the Facebook
World Report college-ranking model. For the last 30 algorithm have further destroyed our concept of truth
years, colleges have been trying to get ranked higher, and our efforts at democracy. I talk about the Facebook
and the definition of success in this model is an arbi- algorithm and political microtargeting in the hardcover
trary set of attributes that do not include important version, but I don’t come out and say that the Facebook

6 September 2017 : : Math Horizons : : www.maa.org/mathhorizons


This content downloaded from 128.122.230.132 on Thu, 09 Nov 2017 20:07:38 UTC
All use subject to http://about.jstor.org/terms

mh-05-07-levy-final.indd 2 9/29/17 1:00 PM


algorithm is a WMD because I didn’t have enough pressure on companies or or-
evidence for the destruction part. I think we do now, so ganizations to deal with these
I will update that. issues soon.
I discuss political polling and predictive polling,
RL: My students are calling for
which are total wastes of time and destructive to the
more focus on the impact of our
kinds of information we should be interested in learn-
work on society. Are there ways
ing. They also change people’s voting patterns in bad
that your book could be used in
ways. For example, if you think that your candidate is
college courses?
a shoo-in or has no chance, you might not vote. So it
CO: Hey, that’s great. It could be used for discussion—
can suppress votes on election day. In addition, be-
hopefully also with case studies so there could be some
cause early polling of unknown candidates is so dis-
technical assignments and projects. A large obstacle is
couraging and considered evidence that they will not
that the WMD algorithms I discuss are proprietary and
make it, we continue to have a very tightly controlled
not discoverable through the Freedom of Information
two-party system.
Act. So you would have to study some kind of replica of
RL: What are some of the main concerns that people the algorithms and see how the attributes come through
have raised about your book? in those replicas. You could assign different groups the
CO: I haven’t gotten much negative feedback, but I same problem and see how different their answers are.
have a thick skin about trollish comments on Amazon. Modeling is a craft, not a science. It is almost never
The biggest complaint is that the book is political. I’m proof-worthy.
not a journalist. I’m more like an activist, so I don’t It is challenging to get data sets that are anything like
take that badly. I don’t want it to be so vitriolic that the ones I talk about in my books, such as the data on
people who differ politically from me can’t read it. So people who have applied for a job or committed a crime,
far, the consensus has been that it is readable, even to because with human data there are privacy concerns.
people who disagree with me.
I have gotten overwhelmingly positive feedback, espe- RL: What are you hoping to accomplish in the next five
cially from young technical people considering jobs in years?
data science. They are emailing and offering to work for CO: I want there to be an academic field that deals with
my company because they want to do good with their algorithm accountability. It is interdisciplinary and in-
data science skills. It is very pleasant to see that there is cludes the mathematical sciences, computer science, soci-
a burgeoning society of good nerds. There is not much ology, information science, and law. Existing data science
I’d rather do than hang out with people like that. institutes are not yet addressing basic questions of policy,
ethics, discrimination, fairness, meaning, interpretability,
RL: What is the goal of your new company? transparency, and accountability in algorithms. I want
CO: My company is called O’Neil Risk Consulting and the field up-front and center. It could be a department
Algorithmic Auditing (ORCAA). An organization or a or at least tenure lines, journals, and conferences—a real
company like a police department that is currently using community of people tackling the question from different
an algorithm might worry that it is unfair. They may angles.
not intend it to, but their algorithm might be racist, Algorithms are not going away. Any part of our
sexist, or have disparate impact. We could audit that culture that makes decisions in structured ways will be
algorithm for fairness, for legality, or even for meaning. replaced by algorithms, so this is extremely urgent. It
Our goal is to come up with a standard for evidence is the difference between being able to push back about
in the field of algorithms. We were “told trust me, it is unfair decisions and living in an Orwellian machine
a machine learning algorithm and it works.” That era where decisions are made for us. n
is over. We need to put the science into data science
and have a standard for what it means for an algo- Rachel Levy is a professor of mathematics and associate
rithm to work well and how to measure bias. dean for faculty development at Harvey Mudd College.
Is a résumé-filtering algorithm filtering out people She serves as vice president for education for SIAM and
from some groups unfairly? Regulators may not be on received the MAA’s Alder Award for teaching.
top of this question, but private lawsuits might put http://dx.doi.org/10.4169/mathhorizons.25.1.5

www.maa.org/mathhorizons : : Math Horizons : : September 2017 7


This content downloaded from 128.122.230.132 on Thu, 09 Nov 2017 20:07:38 UTC
All use subject to http://about.jstor.org/terms

mh-05-07-levy-final.indd 3 9/29/17 1:00 PM

You might also like