The Paradox of Automation As Anti-Bias Intervention

DRAFT IN PROGRESS. Please do not circulate.
THE PARADOX OF AUTOMATION AS ANTI-BIAS

INTERVENTION
41 CARDOZO. L. REV. __ (FORTHCOMING, 2020).
IFEOMA AJUNWA*
ABSTRACT
A received wisdom is that automated decision-making serves as an anti-bias

intervention. The conceit is that removing humans from the decision-making process will also
eliminate human bias. The paradox, however, is that in some instances, automated decision-
making has served to replicate and amplify bias. With a case study of the algorithmic capture
of hiring as heuristic device, this Article provides a taxonomy of problematic features
associated with algorithmic decision-making as anti-bias intervention and argues that those
features are at odds with the fundamental principle of equal opportunity in employment. To
examine these problematic features within the context of algorithmic hiring and to explore
potential legal approaches to rectifying them, the Article brings together two streams of legal
scholarship: law & technology studies and employment & labor law.
Counterintuitively, the Article contends that the framing of algorithmic bias as a
technical problem is misguided. Rather, the Article’s central claim is that bias is introduced
in the hiring process, in large part, due to an American legal tradition of deference to
employers, especially allowing for such nebulous hiring criterion as “cultural fit.” The Article
observes the lack of legal frameworks to account for the emerging technological capabilities of
hiring tools which make it difficult to detect bias. The Article discusses several new approaches
to hold liable for employment discrimination both employers and makers of algorithmic hiring
systems who, as brokers, should owe a duty of care. Particularly related to Title VII, the
Article proposes that in legal reasoning corollary to extant tort doctrines, an employer’s failure
to audit and correct its automated hiring platforms for disparate impact could serve as prima
facie evidence of discriminatory intent, leading to the development of the doctrine of
discrimination per se. The Article also considers approaches separate from employment
law, such as establishing consumer legal protections for job applicants that would mandate
their access to the dossier of information consulted by automated hiring systems in making
the employment decision.
*
Assistant Professor of Employment and Labor Law, Cornell University School of
Industrial and Labor Relations, and Associated Faculty Member, Cornell Law School;
Faculty Associate, Berkman Klein Center at Harvard Law School. Many thanks to Ryan
Calo, Zachary Clopton, Gautam Hans, Jeffrey Hirsch, Maggie Gardner, Genevieve Lakier,
Jonathan Masur, John Rappaport, Angie Raymond, Katherine Strandburg, Elana Zeide, the
participants at the University of Chicago Public and Legal Theory Colloquium, the Cornell
Law School Colloquium, and the Yale Information Society and Center for Private Law
Seminar for helpful comments. A special thanks to my research assistants, Kayleigh Yerdon,
Jane Kim, and Syjah Harris.
Electronic copy available at: https://ssrn.com/abstract=2746078

THE PARADOX OF AUTOMATION [24-Sep-19] 2
TABLE OF CONTENTS
INTRODUCTION.................................................................................3
I. THE ALGORITHMIC TURN ............................................... 11
A. Data Objectivity .......................................................... 13
B. Data as Oracle............................................................ 14
C. Data-laundering .......................................................... 16
II. ALGORITHMIC CAPTURE OF HIRING AS CASE STUDY ...... 18
A. Algorithms as Anti-Bias Intervention .............................. 21
B. The Fault in the Machine.............................................. 23
1. Recruitment ...................................................... 24
2. Hiring ............................................................... 26
C. Exposing the Mechanical Turk ...................................... 27
III. A LEGAL PROBLEM, NOT A TECHNICAL PROBLEM .......... 30
A. A Legal Tradition of Employer Deference ........................ 30
B. The Problem of “Cultural Fit” ....................................... 34
C. Re-Thinking Employer Discretion .................................. 37
IV. EX LEGIS: NEW LEGAL FRAMEWORKS ............................ 39
A. Improving on the Fiduciary Duty Concept ......................... 39
1. Platform Authoritarianism ................................ 42
2. The Tertius Bifrons .............................................. 42
B. Discrimination Per Se................................................... 44
C. Consumer Protection for Job Applicants ........................... 50
CONCLUSION ..................................................................................55

INTRODUCTION
The automation of decision-making processes via machine learning

algorithmic systems presents itself as a legal paradox. On one hand, such
automation is often an attempt to prevent unlawful discrimination,1 but on the
other hand, there is evidence that algorithmic decision-making processes may
thwart the purposes of antidiscrimination laws such as Title VII of the Civil
Rights Act of 19642 and may instead serve to reproduce inequalities at scale.3
Consider the recent discovery that the commerce company Amazon had
secretly disbanded its algorithmic hiring system.4 In October of 2018, the news
service Reuters reported that Amazon’s engineering team in Edinburgh,
Scotland, had created 500 computer models that it used to “trawl through past
candidates’ résumés and pick up on about 50,000 key terms.”5 Using those
selected key terms, “the system would crawl the web to recommend
candidates.”6 Within a year of using the automated system, however, the
engineers observed that the results of the automated hiring system were
unfavorable to women applicants; the automated hiring system preferred
men.7
A potential cause: the computer models were trained on predominantly
male resumes, with the result that the system concluded that men were
preferred candidates; thus, it “downgraded résumés containing the words
1 “Advocates applaud the removal of human beings and their flaws from the assessment
process . . . .” Algorithms or automated systems are often seen as fair because they are
“claimed to rate all individuals in the same way, thus averting discrimination.” Danielle Keats
Citron & Frank Pasquale, The Scored Society: Due Process for Automated Predictions, 89 WASH. L.
REV. 1, 4 (2014).
2 Title VII of the Civil Rights Act, 42 U.S.C. §§ 2000e to 2000e-17 (2000).
3 See, e.g., Solon Barocas & Andrew Selbst, Big Data’s Disparate Impact, 104 CAL. L. REV. 671
(2016) (detailing issues of disparate impact associated with algorithmic decision-making);

FRANK PASQUALE, THE BLACK BOX SOCIETY: SECRET ALGORITHMS THAT CONTROL
MONEY AND INFORMATION (2015) (detailing legal issues associated with the non-
transparent use of algorithmic decision-making in several societal spheres); Andrew Tutt, An
FDA for Algorithms, 69 ADMIN. L. REV. 83, 87 (2017) (“These new family of algorithms hold
enormous promise, but also pose new and unusual dangers.”); see also VIRGINIA EUBANKS,
AUTOMATING INEQUALITY: HOW HIGH-TECH TOOLS PROFILE, POLICE AND PUNISH THE
POOR (2018); SAFIYA NOBLE, ALGORITHMS OF OPPRESSION: HOW SEARCH ENGINES
REINFORCE RACISM (2018); CATHY O’NEIL, WEAPONS OF MATH DESTRUCTION: HOW BIG
DATA INCREASES INEQUALITY AND THREATENS DEMOCRACY (2016). Cf. Sandra G.
Mayson, Bias In, Bias Out, 128 YALE L.J. 2218 (2019) (arguing that the problem of disparate
impact in predictive risk algorithms lies not in the algorithmic system but in the nature of
prediction itself).
4 Isobel Asher Hamilton, Amazon Built an AI Tool to Hire People but Had to Shut It Down Because
It Was Discriminating Against Women, BUS. INSIDER (Oct. 10, 2018, 5:47 AM),
www.businessinsider.com/amazon-built-ai-to-hire-people-discriminated-against-women-
2018-10.
5 Id.
6 Id.
7 Id.

‘women’s’ and filtered out candidates who had attended two women-only
colleges.”8 As legal scholars such as Professor Sandra Mayson and others have
demonstrated, such algorithmic bias is not limited to gender; algorithmic
decision-making can also produce disparate racial impact, especially in the
criminal justice system.9 Amazon’s story of negative discrimination against
protected classes as an (un)intended outcome of automated decision-making
is not singular. Recent books, like Algorithms of Oppression, have detailed the
racially-biased impact of algorithms on information delivery on the internet,10
and others, like Automating Inequality, have outlined the biased results of
algorithms in criminal justice and public welfare decision-making,11 yet, with
the exception of the work of a few legal scholars,12 the role of algorithms in
perpetuating inequality in the labor market has been relatively overlooked in
legal scholarship.13
Often, when legal scholars raise the topic of bias in algorithmic systems, a
common retort is: “What’s new?”14 This rhetorical question is meant to convey
the sentiment that bias in algorithmic systems cannot be a novel topic of legal
inquiry because it has a pre-existing corollary, bias in human decision-making.
8
Id. Ironically, as the use of an automated hiring system revealed the gender disparity here in
concrete numbers, this meant that such disparities could potentially be addressed by
employment antidiscrimination law. Contrast this to what the legal scholar Professor Jessica
Fink has identified as the more nebulous “gender-sidelining,” a workplace dynamic in which,
for example, “women often lack access to important opportunities or feel subjected to
greater scrutiny than their male peers.” See Jessica Fink, Gender Sidelining and the Problem of
Unactionable Discrimination, 29 STAN. L. & POL’Y REV. 57 (2018).
9 More often, legal scholars have considered algorithmic racial inequities in the context of the
criminal justice system. See, e.g., Mayson, supra note 3 (arguing that the problem of disparate
impact in predictive risk algorithms lies not in the algorithmic system but in the nature of
prediction itself); Aziz Z. Huq, Racial Equity in Algorithmic Criminal Justice, 68 DUKE L.J. 1043
(2019); Aziz Z. Huq, The Consequences Of Disparate Policing: Evaluating Stop and Frisk as a
Modality of Urban Policing, 101 MINN. L. REV. 2397, 2408 (2017); Andrew Guthrie Ferguson,
Big Data and Predictive Reasonable Suspicion, 163 U. PA. L. REV. 327 (2015).
10 See NOBLE, supra note 3.
11 See EUBANKS, supra note 3.
12 See, e.g., Stephanie Bornstein, Antidiscriminatory Algorithms, 70 ALA. L. REV. 519, 570 (2018);
Stephanie Bornstein, Reckless Discrimination, 105 CAL. L. REV. 1055, 1056 (2017); Pauline
Kim, Data-Driven Discrimination at Work, 58 WM. & MARY L. REV. 857, 908 (2017); Matthew
Bodie, Miriam Cherry, Marcia McCormick & Jintong Tang, The Law and Policy of People
Analytics, 88 U. COLO. L. REV. 961 (2017); Charles Sullivan, Employing AI, 63 VILL. L. REV.
395 (2018); James Grimmelmann & Daniel Westreich, Incomprehensible Discrimination, 7 CAL.
L. REV. ONLINE 164 (2017).
13 Going beyond the specific role of algorithms, some scholars have argued that workplaces
in the United States are essentially allowed to operate as self-contained, self-governing

bodies, without much scrutiny or oversight from regulatory bodies in regard to how the
workplace is structured and organized. See ELIZABETH ANDERSON, PRIVATE GOVERNMENT:
HOW EMPLOYERS RULE OUR LIVES (2015) (making the argument that workplaces have
become authoritarian private governments with little protection for the worker from the
State).
14 See, e.g., Joseph H. Sommer, Against Cyberlaw, 15 BERKELEY TECH. L.J. 1145, 1148 (2000)
(“[F]ew of the legal issues posed by the new informatics technologies are novel.”).

However, scholars such as Professor Jack Balkin have exposed this retort as a
facile dismissal of what are legitimate lines of scholarly legal inquiry.
[T]o ask “What is genuinely new here?” is to ask the wrong question.
If we assume that a technological development is important to law only
if it creates something utterly new, and we can find analogues in the
past—as we always can—we are likely to conclude that because the
development is not new, it changes nothing important. That is the
wrong way to think about technological change and public policy, and
in particular, it is the wrong way to think about the Internet and digital
technologies. Instead of focusing on novelty, we should focus on
salience. What elements of the social world does a new technology
make particularly salient that went relatively unnoticed before? What
features of human activity or of the human condition does a
technological change foreground, emphasize, or problematize? And
what are the consequences for human freedom of making this aspect
more important, more pervasive, or more central than it was before?”15
Other legal scholars have made similar points. As Professor Katyal notes:
“the true promise of AI does not lie in the information we reveal to one
another, but rather in the questions it raises about the interaction of
technology, property, and civil rights.”16 My scholarly agenda has focused on
examining the myriad ways in which new computing technologies bring to high
relief existing societal biases and continued inequities, particularly in the
employment sphere. In past work, I have parsed how online platforms might
contribute to age discrimination in the labor market,17 and I have noted how
wearable technologies deployed to manage the workplace prompt novel legal
questions and suggest a new agenda for employment and labor law
scholarship.18 I have also conducted an empirical study of work algorithms,
which involved a critical discourse analysis and affordance critique of the
advertised features and rhetoric behind automated hiring systems as gleaned
through 135 archival texts, tracing the timeline of the development of hiring
platforms from 1990–2006.19 That study concluded that while one purported
raison d’etre and advertised purpose of automated hiring systems was to reduce
hirer bias—“replacing messy human decisions with a neutral technical
15 Jack M. Balkin, Digital Speech and Democratic Culture: A Theory of Freedom of Expression for the
Information Society, 79 N.Y.U. L. REV. 1 (2004).
16
Sonia Katyal, Private Accountability in the Age of Artificial Intelligence, 66 UCLA L. REV. 54, 54
(2019).
17 Ifeoma Ajunwa, Age Discrimination by Platforms, 40 BERKELEY J. EMP. & LAB. L. 1 (2019).
18 Ifeoma Ajunwa, Algorithms at Work: Productivity Monitoring Applications and Wearable
Technology, 63 ST. LOUIS U. L.J. 21 (2018).

19 Ifeoma Ajunwa & Daniel Greene, Platforms at Work: Automated Hiring Platforms and Other
New Intermediaries in the Organization of Work, in WORK AND LABOR IN THE DIGITAL AGE 61
(Stephen P. Vallas & Anne Kovalainen eds., 2019).

process”20— the reality remained that “algorithmic specification of ‘fit’ can

itself become a vehicle for bias.”21
Deploying a case study of the algorithmic capture of hiring as heuristic
device, this Article makes several important contributions to two streams of
legal scholarship: employment & labor law and law & technology studies. First,
the Article provides a taxonomy of problematic features associated with
algorithmic decision-making as anti-bias intervention and argues that those
features contradict the American bedrock principle of equal opportunity in
employment. I observe that some factions of law and technology legal scholars
have become preoccupied with determining the legal guidelines to ensure fair
automated decision-making,22 while in employment and labor law the concern
is preserving equal opportunity for protected classes to obtain a livelihood;23
these may represent disjointed interests given that technically fair systems may
replicate historic inequalities.
Note that consistent with the approach of legal scholars like Professor
Benjamin Sachs, I do not see employment law and labor law as “dichotomous,
and in a fundamental respect incompatible, regulatory regimes.”24 Rather, any
worker gains accomplished via employment law may be leveraged towards
collective bargaining, and thus, “employment law can in fact function as a
substitute form of labor law—as the locus of workers’ organizational activity
and as the legal mechanism that insulates that activity from employer
interference.”25 Furthermore, I argue that any clarity this Article affords in
regard to combating algorithmic bias in employment decision-making may also
be deployed to bargain for better, more probative hiring criteria.
However, law & technology studies and employment & labor law studies,
as different schools of legal thought, approach the issue of algorithmic bias in
divergent ways; the law & technology approach to bias has, thus far, with a few
emerging exceptions, been from an anti-classificationist approach,26 wherein
20 Id.; “From a diversity perspective, artificial intelligence can be very beneficial because it’s
blind to things like color, age, sexual orientation.” Jeff Meredith, AI Identifying Steady Workers,
CHI. TRIB. (July 16, 2001), https://www.chicagotribune.com/news/ct-xpm-2001-07-16-
0107160013-story.html.
21 Id. Consider that Unicru’s instruction manuals for users of its hiring platforms encourages
them to “clone your best people” with the identification of existing high-sales employees
within client records. A process that would be sure to replicate existing inequalities in the
demographics of the workers. Id.
22 See generally Ferguson, supra note 9; Citron & Pasquale, supra note 1, at 4; Barocas & Selbst,
supra note 3.
23 Cynthia Estlund has argued that Title VII should be understood as an “equal protection
clause for the workplace.” Cynthia Estlund, Rebuilding the Law of the Workplace in an Era of Self-
Regulation, 105 COLUM. L. REV. 319, 331 (2005); see also Samuel R. Bagenstos, The Structural
Turn and the Limits of Antidiscrimination Law, 94 CAL. L. REV. 1, 40–41 (2006) (arguing that the
best explanation for employment discrimination law is its reflection of a broad goal of social
change to eliminate group-based status inequalities).
24 Benjamin I. Sachs, Employment Law as Labor Law, 29 CARDOZO L. REV. 2685, 2688 (2008).
25 Id.
26 Jack M. Balkin & Reva B. Siegel, The American Civil Rights Tradition: Anticlassification or
Antisubordination?, 58 U. MIAMI L. REV. 9 (2003) (relating the history of the development and
application of the two distinct antidiscrimination threads in American law); see also Jessica L.

the focus is on the improper use of variables for protected classes in the
decision-making process.27 In the context of hiring, such an emphasis would
revolve around the disparate treatment cause of action under Title VII.28
Whereas employment & labor law literature now mostly focuses on anti-
subordination, where the concern is the adverse impact of decision-making on
protected groups, which mostly implicates the disparate impact theory under
Title VII.29 This Article seeks to reconcile this gulf by noting first that machine
learning algorithmic systems present opportunities for both disparate
treatment and disparate impact discriminatory actions.30 The Article then notes
the particular difficulties of proving a disparate impact theory of discrimination
Roberts, The Genetic Information Nondiscrimination Act as an Antidiscrimination Law, 86 NOTRE

DAME L. REV. 597, 631 (2011). Roberts notes:
[T]hese two versions of the antidiscrimination principle employ differing
accounts of the meaning of equality. The antisubordination principle
roughly holds that covered entities should not act in a way that reinforces
the social status of subjugated groups. Antisubordination would,
therefore, permit affirmative action designed to improve the status of a
disadvantaged group and forbid facially neutral policies that perpetuate
lowered group status, even absent the intent to discriminate. Its
complement, the anticlassification principle, maintains that covered
entities should not consider certain classes of forbidden traits under any
circumstance, adopting a formal equal treatment model of equality.
Id. at 627–28.
27 See, e.g., Joshua A. Kroll, Joanna Huey, Solon Barocas, Edward W. Felten, Joel R.
Reidenberg, David G. Robinson & Harlan Yu, Accountable Algorithms, 165 U. PA. L. REV. 633
(2017). For a notable exception, see generally Barocas & Selbst, supra note 3 (detailing issues
of disparate impact associated with algorithmic decision-making).
28 Title VII states:
It shall be an unlawful employment practice for an employer -

(1) to fail or refuse to hire or to discharge any individual, or otherwise to
discriminate against any individual with respect to his compensation,
terms, conditions, or privileges of employment, because of such
individual’s race, color, religion, sex, or national origin; or
(2) to limit, segregate, or classify his employees or applicants for
employment in any way which would deprive or tend to deprive any
individual of employment opportunities or otherwise adversely affect his
status as an employee, because of such individual’s race, color, religion,
sex, or national origin.
42 U.S.C. § 2000e-2(a) (2012).
29 Title VII explicitly prohibits employers from using any “particular employment practice
that causes a disparate impact on the basis of race, color, religion, sex or national origin.” 42
U.S.C. § 2000e-2(k)(1)(A)(i). But see Bradley A. Areheart, GINA, Privacy, and Antisubordination,
46 GA. L. REV. 705 at 709 (2012). Areheart argues that “GINA . . . [represents] a turn toward
anticlassificationist principles and a possible turn away from antisurbodination norms.”; Cf.
Bornstein, Antidiscriminatory Algorithms, supra note 12, at 571. Professor Bornstein asserts that
in addition to anticlassification and antisubordination theories underlying antidiscrimination
law, antistereotyping principle should be considered since algorithmic discrimination can be
liable for intentional discrimination as well as disparate impact.
30
See Richard Thompson Ford, Bias in the Air: Rethinking Employment Discrimination Law, 66
STAN. L. REV. 1381 (2014) (noting that concepts in employment law such as “intent” and
“causation” escape precise definition).

when machine learning algorithmic systems have played a role in the

employment decision-making process.
As a second contribution, I argue that too often law & technology
scholarship displays a too narrow focus on technical definitions of fairness and
overlooks the reality that facially neutral variables may still (in)advertently
stand in for protected variables, thus meaning that there could still be a
disparate impact on protected classes even when machine learning algorithms
are technically fair.31 As a result of this myopic view, algorithmic bias,
particularly in employment, has been framed as a technical problem, rather
than as a matter of inadequate law. This offers a facile dismissal of algorithmic
bias as a non-novel problem that could be solved by technical tinkering, or
even as a non-problem since some argue that algorithmic systems represent an
improvement from human decision-making.
To be sure, human managers hold biases that are reflected in unfavorable
employment decisions for protected classes, but the impact of one biased
human manager is constrained in comparison to the potential adverse reach of
algorithms that could be used to exclude millions of job applicants from
viewing a job advertisement or to sort thousands of resumes. The new
phenomenon of concern here is that due to the “volume, velocity, and
variety”32 of data used in automated hiring, any bias introduced in the system
will be magnified and multiplied, greatly dwarfing the impact of any prejudice
held by any one biased human manager.
A few caveats are in order for a useful reading of this Article. First, I remain
agnostic as to whether the problems of biased algorithmic decision-making in
hiring can or should be neatly classified as either disparate treatment or
disparate impact. Rather, in addition to these two established legal theories for
liability, I draw from organizational theory to offer my new theory of the
automated hiring platform as a tertius bifrons.33 The hope is that such
theorization will help provide support for updated legal frameworks34 to better
address what I describe as the novel sociological phenomenon of the algorithmic
capture35 of employment.
Second, perhaps in the same genre as the “what’s new question?” there’s
the question of whether automated decision-making is better than human
31 See generally Barocas & Selbst, supra note 3.

32 This refers to the three V’s of big data. David Gewirtz, Volume, Velocity, and Variety:
Understanding the Three Big V’s of Big Data, ZDNET: DIY-IT (Mar. 21, 2019, 9:47 AM),
https://www.zdnet.com/article/volume-velocity-and-variety-understanding-the-three-vs-of-
big-data/.
33 See infra Section IV.A.; see also GEORG SIMMEL, THE TRIAD (1950) (establishing the
typology of a triad with a “tertius” as broker).

34 I agree with these sentiments set forth by Professor Selmi: “Employment discrimination
law has long been ripe for updating. Many of the core cases regarding how discrimination is
defined and proved arose in the 1970s in a very different era and were designed to address
very different kinds of discrimination.” Michael Selmi, The Evolution of Employment
Discrimination Law: Changed Doctrine for Changed Social Conditions, 2014 WIS. L. REV. 937, 938
(2014).
35 See infra Section II.A.

decision-making. The aim of this Article is not to argue for or against

algorithmic decision-making in the workplace; rather, it is to ensure that the
law is able to root out and redress any problems associated with that practice,
much the same as the law governs human decision-making. Furthermore, I do
not believe that an adjudication of whether or not algorithms are less biased
than humans36 is necessary to advocate for better laws to curb the
demonstrated potential of algorithms to sometimes return biased results. First,
an adjudication of whether hiring platforms are more or less biased than
humans is not a legal one—rather, it is a social scientific one that demands
rigorous longitudinal research. Although there is some research, as proffered
by Professor Bornstein in her article, Antidiscriminatory Algorithms, showing
automated hiring to be less biased, the relative novelty of hiring platforms
brings to question the rigor of such studies.
I also do not believe that an actual adjudication of whether hiring platforms
are less biased than humans or not is necessary before there can be new legal
frameworks specifically designed to govern automated decision-making in
hiring. Professor Julie Cohen makes this argument clear in her article, Law for
the Platform Economy.37 Much like our present laws have been developed to
govern human decision-making, we need new legal frameworks to govern
emerging new technologies in the workplace. My third and most important
point here: to argue for or against automated decision-making versus human
decision-making is to create a false binary. It is to willfully forget that the
human hand remains present in all automated decision-making.38
A final caveat, this Article operates from the normative standard of a social
contract society and asks not just how business efficiency can be achieved
through algorithmic decision-making while toeing to the letter of employment
antidiscrimination law but, rather, how the societal goal of equality and the
spirit of antidiscrimination laws meant to accomplish that goal could be
honored, even in the face of the automation of employment decision-making.
This Article also affirms worker diversity as a normative business and societal
ideal. While much can be said about the ethical benefits of a diverse workforce,
particularly in regard to reducing economic inequality and its attendant
36 I firmly believe that whether or not hiring algorithms produce more or less biased results
than humans cannot be a legal adjudication. As Professor Charles Sullivan has remarked:
“And the antidiscrimination statutes don’t really care whether any particular selection device
actually improves productivity so long as it does not discriminate.” Sullivan, supra note 12, at
398. Rather, determining whether algorithmic systems evince less bias than human managers
requires empirical data obtained via rigorous social scientific research. Some legal scholars
have argued, based on preliminary studies, that automated hiring systems have “allowed
some employers to easily and dramatically reduce the biasing effects of subjectivity from
their hiring decisions . . . .” Bornstein, Reckless Discrimination, supra note 12, at 1056. I argue,
however, that since algorithmic hiring systems are a relatively new invention, to assess any
bias reduction would require longitudinal studies in several industries, and with adequate
controls.
37
Julie E. Cohen, Law for the Platform Economy, 51 U.C. DAVIS L. REV. 133, 189 (2017).
38
See infra Section II.C.

negative effects,39 a diverse workplace also affords business advantage because

diverse workplaces evince greater innovation and better decision-making.40
Therefore, I push for a Rawlesian41 approach to the governance of automated
hiring systems, wherein the focus is on institutional approaches to algorithmic
decision-making and their impact on society as a whole. In so doing, rather
than merely proposing ways to build fairer algorithmic systems as some other
legal scholars have done,42 I argue that we should also interrogate what values
are given precedence by the turn to algorithmic decision-making and by the
legal deference accorded employers in determining hiring criteria.
The Article is then organized as follows: Part I discusses the “algorithmic
turn” and three problematic features of big-data-driven algorithmic decision-
making: 1) data objectivity, 2) data as oracle, and 3) data-laundering. Part II
details the algorithmic capture of the workplace, including the notion that
automation in hiring is presented as an anti-bias intervention, and notes several
examples of bias in both algorithmic recruitment and hiring. Part III urges a
re-framing of the problem of algorithmic bias, notably with the argument that
bias in algorithmic hiring is not a technical problem but rather a legal one borne
from a tradition of deference to employers, especially in regard to non-
probative hiring criterion such as “cultural fit.” Based on original theorization
of platform authoritarianism and the tertius bifrons, Part IV proposes new legal
frameworks for addressing algorithmic discrimination in employment that
borrows from tort law to update employment law. Of particular note here, is
the idea that intent to discriminate could be implied from the act of
39 Some research demonstrate that greater inequality is correlated to violence, incarceration,

drug abuse, obesity, teenage pregnancy, and mental health issues. See generally KATE PICKETT
& RICHARD WILKINSON, THE SPIRIT LEVEL: WHY GREATER EQUALITY MAKES SOCIETY
STRONGER (2009).
40 Sheen S. Levine, Evan P. Apfelbaum, Mark Bernard, Valerie L. Bartelt, Edward J. Zajac &
David Stark, Ethnic Diversity Deflates Price Bubbles, 111 PROC. NAT’L ACAD. SCI. 18524 (2014)
(detailing sociological research showing that diverse teams make better decisions and are
more innovative); see also Katherine W. Phillips, Katie A. Liljenquist & Margaret A. Neale,
Better Decisions Through Diversity, KELLOGG SCH. MGMT.: KELLOGG INSIGHT (Oct. 1, 2010),
https://insight.kellogg.northwestern.edu/article/better_decisions_through_diversity
(showing that diverse groups outperform homogenous groups because of both an influx of
new ideas and more careful information processing); Sheen S. Levine & David Stark,
Diversity Makes You Brighter, N.Y. TIMES (Dec. 9, 2015),
https://www.nytimes.com/2015/12/09/opinion/diversity-makes-you-brighter.html.
41 JOHN RAWLS, A THEORY OF JUSTICE (1971). Rawls, a social contract philosopher, argued
that the competing claims of freedom and equality could be reconciled when decisions about
justice are made on the basis of the difference principle behind a “veil of ignorance,”
wherein no one individual knows their original position (that, is they could be members of
low status groups in society), with the result that the only rational choice is to make
decisions that would improve the position of the worst off in society. Id. See also Mark
Kelman, Defining the Antidiscrimination Norm to Defend It, 43 SAN DIEGO L. REV. 735 (2006)
(rejecting “the [utilitarian ethics] idea that the antidiscrimination norm’s propriety should be
evaluated solely by reference to its impact on a mere subset of experiences or capacities to
engage in certain activities, for example, a claim that what is relevant in deciding whether the
plaintiff merits protection is the plaintiff’s legitimate sense that, absent protection, he is not
treated as a ‘first-class’ citizen”).
42 See, e.g., Kroll, Huey, Barocas, Felten, Reidenberg, Robinson & Yu, supra note 27.

negligence43 to audit and correct bias in algorithmic hiring systems.

Furthermore, Part IV discusses how consumer protection laws might help
ensure that the information collection of algorithmic hiring systems does not
amplify bias.
I. THE ALGORITHMIC TURN

Derived from the name of a Persian mathematician, Al-Khwarizmi,44 the
word “algorithm” and the mathematical system of problem-solving it stands
for has gained prominence in all spheres of social and economic life in the past
three decades.45 With advancements in computing technologies and the
capacity for rapid mining of big data, algorithms now pervade our daily lives
and exert influence over many important decisions.46 The algorithmic turn47 is
what I term the profusion of algorithmic decision-making in our daily lives,
even in the absence of established regulatory or ethical frameworks to guide
the deployment of those algorithms.48 First, I note that the definition of the
term “artificial intelligence” (AI)49 varies in legal literature and in popular
43
Several other legal scholars have applied the tort law principle of a duty of care to
employment discrimination. See, e.g., Ford, supra note 30 (arguing that employment law
imposes a duty of care on employers to refrain from practices that go against equal
opportunity in employment). See also, Robert Post, Lecture, Prejudicial Appearance: The Logic of
American Antidiscrimination Law, 88 CALIF. L. REV. 1 (2000), (arguing that antidiscrimination
law aims to achieve positive interventions in social practices as opposed to solely dictating
prohibitions). Other professors have also used a “duty of care” framework to propose
remedial measures for employment discrimination. See, David Benjamin Oppenheimer,
Negligent Discrimination, 141 U. PA. L. REV. 899 (1993); Noah D. Zatz, Managing the Macaw:
Third-Party Harassers, Accommodation, and the Disaggregation of Discriminatory Intent, 109 COLUM.
L. REV. 1357 (2009).
44 DONALD E. KNUTH, STANFORD DEPARTMENT OF COMPUTER SCIENCE REPORT NO.
STAN-CS-80-786, ALGORITHMS IN MODERN MATHEMATICS AND COMPUTER SCIENCE 2

(1980).
45 Google Ngram shows the usage of the word “algorithm” beginning in the 1800s and
rapidly growing from the 1980s. Two recently published books document the widespread
use of algorithms both in governmental decision-making and in the delivery of search results
online. See EUBANKS, supra note 3; NOBLE, supra note 3.
46 See Neil M. Richards & Jonathan H. King, Big Data Ethics, 49 WAKE FOREST L. REV. 393,
393 (2014) (noting that “large datasets are being mined for important predictions and often
surprising insights”).
47 This nomenclature takes, as inspiration, Professor Julie Cohen’s description of the
“participatory turn” in which innovative surveillance methods are positioned as exempt from
legal and social control and rather held up as evidence of economic progress. See Julie
Cohen, The Surveillance-Innovation Complex: The Irony of the Participatory Turn, in THE
PARTICIPATORY CONDITION 207, 207–26 (2016).
48 See Harry Surden, Machine Learning and Law, 89 WASH. L. REV. 87, 88 (2014) (detailing gaps
in the law in regard to machine learning algorithms).

49 John McCarthy coined the term “AI” in his 1955 proposal for the 1956 Dartmouth
conference, the inaugural AI conference. See Martin Childs, John McCarthy: Computer Scientist
Known as the Father of AI, INDEPENDENT (Nov. 1, 2011, 8:00 PM),
https://www.independent.co.uk/news/obituaries/john-mccarthy-computer-scientist-
known-as-the-father-of-ai-6255307.html.

media,50 and thus, in this Article, in lieu of “AI,” I employ the more precise
terms of “algorithms”51 and “machine learning algorithms.”52
Consider that an algorithm decides all of the following: the answer to a
search one conducts online,53 the best romantic prospects provided by a dating
website,54 what advertisements one sees during a visit to a given website,55
one’s creditworthiness,56 whether or not one should be considered a suspect
for a crime,57 and whether or not one is qualified for a job.58 As I detail in the
50 In most media, “artificial intelligence” and “algorithms” are used interchangeably. Lauri
Donahue writes about the process of machine learning, in which an algorithm learns from its
experiences and adapts to new sets of information, based on data. See Lauri Donahue, A
Primer on Using Artificial Intelligence in the Legal Profession, JOLT DIG.: COMMENT. (Jan. 03,
2018), https://jolt.law.harvard.edu/digest/a-primer-on-using-artificial-intelligence-in-the-
legal-profession. Donahue uses the terms AI and algorithm interchangeably throughout the
article.
51 In defining an algorithm, Alan D. Minsk references the Gottschalk v. Benson decision, in
which the court defined an algorithm as a “procedure for solving a given type of
mathematical problem . . . . [An algorithm is] . . . a generalized formulation for programs to
solve mathematical problems of converting one form of numerical representation to
another.” Alan D. Minsk, Patentability of Algorithms: A Review and Critical Analysis of the Current
Doctrine, 8 SANTA CLARA HIGH TECH. L.J. 251, 257 (1992); see also Gottschalk v. Benson, 409
U.S. 63, 65 (1972). Minsk also references the Paine, Webber, Jackson & Curtis, Inc. v. Merrill
Lynch, Pierce, Fenner & Smith, Inc. decision, which defines a mathematical algorithm and a
computer algorithm. A mathematical algorithm is a “recursive computational procedure
[which] appears in notational language, defining a computational course of events which is
self-contained.” Paine, Webber, Jackson & Curtis, Inc v. Merrill Lynch, Pierce, Fenner & Smith Inc.,
564 F. Supp. 1358, 1366–67 (D. Del. 1983) (“[A] computer algorithm is a procedure
consisting of operation[s] to combine data, mathematical principles and equipment for the
purpose of interpreting and/or acting upon a certain data input.”). In one of the earliest
mentions of algorithms in case law, we find that “algorithms are procedure[s] for solving a
given type of mathematical problem.” Diamond v. Diehr, 450 U.S. 175, 186 (1981).
52 See PEDRO DOMINGOS, THE MASTER ALGORITHM: HOW THE QUEST FOR THE ULTIMATE
LEARNING MACHINE WILL REMAKE OUR WORLD (2015) (“Every algorithm has an input
and an output: the data goes into the computer, the algorithm does what it will with it, and
out comes the result. Machine learning turns this around: in goes the data and the desired
result and outcomes the algorithm that turns one into the other. Learning algorithms –
known as learners – are algorithms that make other algorithms.”).
53 See, e.g., Latanya Sweeney, Discrimination in Online Ad Delivery, in ASSOC. FOR COMPUTING
MACHINERY QUEUE 44 (2013) (detailing a study in which a search of names associated with
African-Americans returned results featuring advertisements for arrest records as a result of
machine learning by Google’s Ad algorithm); see also NOBLE, supra note 3.
54 Leslie Horn, Here’s How OkCupid Uses Math to Find Your Match, GIZMODO (Feb. 14, 2013),
http://gizmodo.com/5984005/heres-how-okcupid-uses-math-to-find-your-match.
55 Thorin Klosowski, How Facebook Uses Your Data to Target Ads, Even Offline, LIFE HACKER
(Apr. 11, 2013), http://lifehacker.com/5994380/how-facebook-uses-your-data-to-target-

ads-even-offline (explaining how Facebook uses your likes (in addition to those of your
friends) to tailor ads or target you for specific advertisements).
56 PASQUALE, supra note 3.
57 Ferguson, supra note 9 (noting that, although in the past, determining who was a suspect
was a more individualized process, police can now rely on large datasets to make
probabilistic determinations of criminal activity).
58 Claire C. Miller, Can an Algorithm Hire Better than a Human?, N.Y. TIMES: UPSHOT (June 25,
2015), http://www.nytimes.com/2015/06/26/upshot/can-an-algorithm-hire-better-than-a-
human.html; Sarah Green Carmichael, Hiring C-Suite Executives by Algorithm, HARV. BUS. REV.

following sub-sections, the algorithmic turn, as a socio-technical phenomenon in

which we turn to machine learning algorithms for efficiency in decision-
making, evinces several features that I ultimately see as problematic in various
aspects.
A. Data Objectivity
A common adage is “the numbers speak for themselves,”59 and as
identified by previous researchers, this demonstrates an unquestioning belief
in data objectivity, particularly regarding large numbers of data.60 This in turn
becomes a problematic feature of algorithmic systems—as their decision-
making relies on algorithms trained on a corpus of data, the belief in data
objectivity then often results in an uncritical acceptance of decisions derived
from such algorithmic systems.61 In the article, Think Again: Big Data,62
Professor Kate Crawford disputes the reverence accorded to big data. First,
she argues that numbers do not speak for themselves even with enough data
because “data sets are still objects of human design,”63 which means that big
data are not free from “skews, gaps, and faulty assumptions.”64 Biases can exist
in big data as much as they do in the real world with individual perceptions.65
For one, Professor Crawford notes the “signal problems”66 associated with
big data, which arise when citizens or subgroups are underrepresented due to
unequal creation or collection of data. She also observes that more data does
not necessarily improve transparency or accountability; rather, mechanisms to
aid the better interpretation of data are more important.67 Moreover, Professor
Crawford argues that although many believe that big data cause “less
discrimination against minority groups because raw data is somehow immune
to social bias”68 and helps people avoid group-based discrimination at a mass
level,69 big data may in fact contribute to the segregating of individuals into
(Apr. 6, 2015), https://hbr.org/2015/04/hiring-c-suite-executives-by-algorithm (detailing

how established headhunting firms like Korn Ferry are incorporating algorithms into their
work, too).
59 The author concludes: “with enough data, the numbers speak for themselves.” See, e.g.,
Chris Anderson, The End of Theory: The Data Deluge Makes the Scientific Method Obsolete, WIRED
(June 23, 2008, 12:00 PM),), https://www.wired.com/2008/06/pb-theory/.
60 Danah Boyd & Kate Crawford, Critical Questions for Big Data, 15 J. INFO. COMM. & SOC’Y
662 (2012).
61 See, e.g., Anderson, supra note 59 (arguing that “correlation is causation” and that the
scientific method is now defunct).

62 Kate Crawford, Think Again: Big Data, FOREIGN POL’Y (May 10, 2013, 12:40 AM),
https://foreignpolicy.com/2013/05/10/think-again-big-data/.
63 Id.
64 Id.
65 Id.
66 Id.
67 Id.
68 Id.
69 Id.

groups because of its “ability to make claims about how groups behave
differently,”70 an action forbidden by anti-classificationist laws.
These sentiments are echoed by the legal scholar Professor Anupam
Chander, who, in disavowal of data objectivity, argues for “algorithmic
affirmative action.”71 Chander emphasizes that although algorithms are
perceived as fair because computers are logical entities, their results may still
bear the traces of real world discrimination. He argues that “[a]lgorithms
trained or operated on a real-world data set that necessarily reflects existing
discrimination may well replicate that discrimination.”72 This means that
because data are historically biased towards certain groups or classes,
discriminatory results may still emerge from automated algorithms that are
designed in racial- or gender-neutral ways.73 Also, discriminatory results can
occur even when decision-makers are not motivated to discriminate: “Because
race or gender might be statistically associated with an observable trait—such
as worker productivity74 or propensity to remain in the labor market—profit-
maximizing employers might discriminate on the basis of race or gender, using
the observable characteristics as proxies for the unobservable traits.”75 Thus,
in addition to the problem of intentional discrimination, “automated
algorithms offer a perhaps more ubiquitous risk: replicating real-world
inequalities.”76
B. Data as Oracle
Concomitant with the belief in data objectivity is the uncritical
acquiescence to data-driven algorithmic decision-making as the final arbiter on
any given inquiry. Thus, the results of algorithmic systems are heeded as
oracular proclamations; they are accepted at face value without any attempt to
analyze or further interpret them. In the article, Critical Questions for Big Data,77
70 Id.
71 See Anupam Chander, The Racist Algorithm?, 115 MICH. L. REV. 1023, 1041 (2017).
72 Id. at 1036.
73 See id. at 1036–37.
74 It is important to clarify that neither I nor Professor Chander are denying that employers
have a vested interest in worker productivity. The issue here is how productivity is observed
and whether statistics for productivity are every wholly objective and not tainted for bias
when it comes to protected categories.
75 Id. at 1038.
76 Chander’s call for algorithmic affirmative action is rooted in idea that it is necessary to
design algorithms in race- and gender-conscious ways to account for discrimination already
embedded in the data. Id. at 1039. This action goes along with what the Obama
Administration offered as an approach to handle big data: “we need to develop a principle of
‘equal opportunity by design’—designing data systems that promote fairness and safeguard
against discrimination from the first step of the engineering process and continuing
throughout their lifespan.” EXEC. OFFICE OF THE PRESIDENT, BIG DATA: A REPORT ON
ALGORITHMIC SYSTEMS, OPPORTUNITY, AND CIVIL RIGHTS 5–6 (2016),
https://obamawhitehouse.archives.gov/sites/default/files/microsites/ostp/2016_0504_dat
a_discrimination.pdf [https://perma.cc/9XQT-9VYQ].
77 Boyd & Crawford, supra note 60.

the authors offer six provocations to conversations about big data issues. They
define big data as “a cultural, technological, and scholarly phenomenon”78 that
rests on the interplay of technology, which “[maximizes] computation power
and algorithmic accuracy to gather, analyze, link, and compare large data
sets,”79 analysis, which “[identifies] patterns in order to make economic, social,
technical, and legal claims . . . .”80 Furthermore, they note that this analysis
carries with it a mythology, notably the prevalent belief that large data sets
offer better intelligence and knowledge and could algorithmically “generate
insights that were previously impossible, imbued with the aura of truth,
objectivity, and accuracy.”81
What I term the phenomenon of data as oracle is best illustrated by Chris
Anderson, who proposes that, because of big data, the scientific method is
now defunct.82 According to his article, the scientific approach has traditionally
consisted of three parts—hypothesize, model, and test.83 As scientists know
that correlation is not causation, they understand that no conclusions should
be based simply on correlation.84 Anderson argues, however, that this
approach to science is becoming obsolete with big data because petabytes of
data allow people to conclude that “correlation is enough.”85 He places such
trust in data and algorithms that he believes that people can now “throw the
numbers into the biggest computing clusters the world has ever seen and let
statistical algorithms find patterns where science cannot.”86
There is a danger, however, with treating algorithmic systems driven by big
data as oracles given that “interpretation is at the center of data analysis”87 and
that without proper interpretation the decision-making of algorithmic systems
could devolve to apophenia, which results in “seeing patterns where none
actually exist, simply because enormous quantities of data can offer
connections that radiate in all directions.”88 Thus, when approaching a data set
and designing algorithmic systems on that data set, researchers or interpreters
should understand not only the limits of the data set but also of which
questions they can ask of a data set and appropriate interpretations.89
To illustrate the problem of apophenia for employment decision-making,
consider this hypothetical example. Company A decides to use an
unsupervised machine learning algorithm to create a profile of the ideal
78 Id.
79 Id.
80 Id.
81 Id.
82 See Anderson, supra note 59.
83 Id.
84 Id.
85 Id.
86 Id.
87 Boyd & Crawford, supra note 60, at 668.
88 Id.
89 Id. at 670. Professor Jim Greiner exposes the same type of problem in civil rights litigation,
when the use of regression analysis can prompt unjustified casual inferences. See D. James
Greiner, Causal Inference in Civil Rights Litigation, 122 HARV. L. REV. 533 (2008).

worker. As training data, Company A selects resumes of ten people it considers

high performers and trains its algorithm on this corpus of resumes. It turns
out that not only are the top ten performers men, but they all listed rock
climbing as a hobby on their resumes. The machine learning algorithm thus
deduces that the criteria for a good worker includes not only being male but
also being a rock climber. Now, if the job position required above average
upper body strength, then perhaps this determination could be taken at face
value. But what if not? If the job advertised had no relation to upper body
strength or really any of the skills or characteristics of a rock climber, then the
correct move is not to treat this result as oracular truth but rather to recognize
it as the biased result of a severely limited training data set.
If that hypothetical example seemed risible, then consider this real-life
case. Mark J. Girouard, an employment attorney, recounts what transpired
when one of his clients audited90 a resume screening tool.91 The results: “the
algorithm found two factors to be most predictive for which candidates would
be chosen. Those factors were the name Jared, and whether the applicant had
played high school lacrosse.”92 As Girouard recounted, ultimately, the client
choose not to use the tool because although there might be a statistically
significant correlation between the data points, it was difficult to argue that
they were actually important to job performance.93
C. Data-laundering
Perhaps an opposite problem to seeing patterns where there are none is
the potential for large data sets to be deployed to create patterns based on
faulty threads of causation, all with the goal of masking intentional
discrimination. I term this feature “data-laundering,” that is, the use of data to
“launder” or disguise intentional discrimination. In their seminal article,
Barocas and Selbst argue that existing law mostly fails to address the
discrimination that comes from data mining because some instances of
discriminatory data mining will not generate legal liability under Title VII.94
Based on the idea that data mining is “always a form of statistical
discrimination,”95 the authors describe five mechanisms by which
discriminatory outcomes might occur. The five mechanisms are 1) defining the
target variable, 2) labeling and collecting training data, 3) using feature
selection, 4) using proxies, and 5) masking. Notably, the authors argue “[t]he
90 This real-life case highlights exactly why I make the case in another law review article that
there ought to be an auditing imperative for hiring algorithms. Ifeoma Ajunwa, Automated
Employment Discrimination (unpublished manuscript) (on file with author).
91 See Dave Gershgorn, Companies Are on the Hook if Their Hiring Algorithms Are Biased,
QUARTZ (Oct. 22, 2018), https://qz.com/1427621/companies-are-on-the-hook-if-their-

hiring-algorithms-are-biased/.
92 Id.
93 See id.
94 See Barocas & Selbst, supra note 3, at 675.
95 Id. at 677.

definition of the target variable and its associated class labels determine what
data mining happens to find,”96 and concerns with discrimination enter at this
stage because whatever choices are selected will influence whether there are
adverse impacts on protected classes.97
Secondly, labeling and collection of training data is important because the
effectiveness of data mining is dependent on the quality of the data from which
it draws lessons.98 Data should serve as a good sample of a protected group in
order for data mining to be a nondiscriminatory basis for future decision-
making.99 This is not always the case, however, and in an act of data-laundering,
the decision-maker may choose to use data known to be incomplete or
inaccurate. Next, the authors indicate that organizations “make choices about
what attributes they observe and subsequently fold into their analyses”100
through the process of feature selection. This could result in a discriminatory
impact on legally protected classes if the factors that “better account for
pertinent statistical variation among members of a protected class are not well
represented in the set of selected features.”101
For example, making an employment decision based on an individual’s
criminal record would have a disparate impact on protected racial groups given
that mass incarceration has disproportionately impacted racial minorities in the
United States.102 Similarly, I would note that using a “ lack of gaps in
employment” as a hiring criterion could negatively impact women candidates
as women disproportionately leave the workplace to shoulder the family
burden of child or elderly care. Thus, as the authors note, the existence of
proxies could also be a mechanism that drives discrimination if “the criteria
that are genuinely relevant in making rational and well-informed decisions also
happen to serve as reliable proxies for class membership.”103 As Barocas and
Selbst explain, decision-makers with prejudicial values can mask their
intentional discrimination as accidental by exploiting the mechanisms above
because the data mining process helps conceal the fact that those decision-
makers considered class membership.104
96 Id. at 680.
97 See id.
98 See id. at 687.
99 See id.
100 Id. at 688.
101 Id.
102 See id. at 690; see also MICHELLE ALEXANDER, THE NEW JIM CROW: MASS
INCARCERATION IN THE AGE OF COLORBLINDNESS (2010) (observing how mass

incarceration in the United States is a warehousing of the redundant labor population of
Black American males and resembles a return to Jim Crow era); Devah Pager, The Mark of a
Criminal Record, 108 AM. J. SOC. 937 (2003) (describing the racial disparities present in the use
of criminal records in the hiring process); Ifeoma Ajunwa, The Modern Day Scarlet Letter, 83
FORDHAM L. REV. 2999 (2015) (arguing that Black women are the most disadvantaged by
the collateral consequences of conviction).
103 Barocas & Selbst, supra note 3, at 691.
104 See id. at 692–23.

II. ALGORITHMIC CAPTURE OF HIRING AS CASE STUDY

Given the described problems associated with algorithmic decision-
making, the algorithmic capture of hiring is cause for concern. I use the term
algorithmic capture to describe the combined effect of the belief that algorithms
are more efficient and fairer105 and the abdication of human accountability for
undesirable outcomes as a result of employing machine learning algorithms as
part of a decision-making process. Thus, my focus here is on algorithmic work
tools that are implicated in automating the hiring process and thus in
revolutionizing the workplace.106 Although there are several types of
algorithmic hiring systems, I see them as falling into two groups: 1) what I term
“off-the shelf” algorithms that employers can purchase or license, or 2) what
I term “bespoke” algorithms that employers can have a software developer
create to their custom specifications. While I would concede that intent and
liability might be analyzed differently for the employer depending on the type
of algorithm in question, ultimately, differences in hiring algorithms are less
important than the fact that although AI in the form of machine learning
algorithms have automated many work functions previously thought reserved
for human judgment,107 there have been scant new regulations to ensure that
these new technological developments will conform to the normative ideal of
equal economic opportunity for all, which is the bedrock of our democratic
society.
The automation of the hiring process represents a particularly important
technological trend and one that requires greater legal attention given its
potential for employment discrimination. Whereas once, an applicant could
rely on his or her interpersonal skills to make a favorable first impression on
the hiring manager, these days the hiring algorithm is the initial hurdle to clear
105 Boyd & Crawford, supra note 60 (noting the aura of efficiency associated with big data-
driven algorithms).
106 ERIK BRYNJOLFSSON & ANDREW MCAFEE, THE SECOND MACHINE AGE: WORK,
PROGRESS, AND PROSPERITY IN A TIME OF BRILLIANT TECHNOLOGIES (2014) (arguing that

akin to how the Industrial Revolution changed the path of human invention, the Artificial
Intelligence age will similarly revolutionize work as we know it).
107 See Harry Surden, Computable Contracts, 46 U.C. DAVIS L. REV. 629, 646 (2012) (discussing
how computer algorithms may find it difficult to decipher language changes that are readily
comprehensible to humans). But see, e.g., Erin Winick, Lawyer-Bots Are Shaking up Jobs, MIT
TECH. REV. (Dec. 12, 2017), https://www.technologyreview.com/s/609556/lawyer-bots-
are-shaking-up-jobs/. “While computerization has been historically confined to routine tasks
involving explicit rule-based activities, algorithms for big data are now rapidly entering
domains reliant upon pattern recognition and can readily substitute for labour in a wide
range of non-routine cognitive tasks.” Carl Benedikt Frey & Michael A. Osborne, The
Future of Employment: How Susceptible are Jobs to Computerisation? (Sept. 17, 2013)
(unpublished manuscript),
http://www.oxfordmartin.ox.ac.uk/downloads/academic/The_Future_of_Employment.pdf
; see also ERIC SIEGAL, PREDICTIVE ANALYTICS: THE POWER TO PREDICT WHO WILL CLICK,
BUY, LIE, OR DIE (2013).

to gain employment.108 This is particularly true for the U.S. low-wage and
hourly workforce, as a co-author and I found through a survey of the top
twenty private employers in the Fortune 500 list (comprised of mostly retail
companies).109 That survey indicated that job applications for such retail jobs
must be submitted online, where they will first be sorted by automated hiring
platforms powered by algorithms.110
The algorithmic capture of the hiring process also goes beyond the hourly
workforce, as white collar and white shoe firms are increasingly turning to
hiring automation.111 In 2016, the investment firm Goldman Sachs announced
a key change to its process for hiring summer interns and first-year analysts.112
Candidates now have their resumes scanned—ostensibly by machine learning
algorithms, in search of keywords and experiences that have been pre-judged
to be “good barometers of a person’s success at Goldman.”113 Goldman Sachs
has also considered the addition of personality tests as part of its hiring
program.114 The world’s largest hedge fund has taken the automation gambit
the furthest, as starting in 2016, it is building an algorithmic model that would
automate all management, including hiring, firing, and other managerial
decision-making processes.115 Thus, automated hiring represents an ecosystem
in which, if left unchecked, a closed loop system forms—with algorithmically-
driven advertisement determining which applicants will send in their resumes,
automated sorting of resumes leading to automated onboarding and eventual
automated evaluation of employees, and the results of said evaluation looped
back into criteria for job advertisement and selection.
108 See LINDA BARBER, E-RECRUITMENT DEVELOPMENTS (2006),

https://www.employment-studies.co.uk/system/files/resources/files/mp63.pdf (noting that
nearly all global 500 companies use recruitment and hire screening algorithmic tools).
109 Ajunwa & Greene, supra note 19.
110 Id.
111 See, e.g., Richard Feloni, Consumer-Goods Giant Unilever Has Been Hiring Employees Using Brain
Games and Artificial Intelligence – And It’s a Huge Success, BUS. INSIDER (June 28, 2017, 9:30
AM), http://www.businessinsider.com/unilever-artificial-intelligence-hiring-process-2017-6;
Louis Efron, How A.I. Is About to Disrupt Corporate Recruiting, FORBES (July 12, 2016, 01:57
PM), https://www.forbes.com/sites/louisefron/2016/07/12/how-a-i-is-about-to-disrupt-
corporate-recruiting/#75ae172d3ba2.
112 Mary Thompson, Goldman Sachs Is Making a Change to the Way It Hires, CNBC: FIN. (June
23, 2016, 03:01 PM), https://www.cnbc.com/2016/06/23/goldman-sachs-is-making-a-

change-to-the-way-it-hires.html.
113 Id.
114 Id.
115 Rob Copeland & Bradley Hope, The World’s Largest Hedge Fund Is Building an Algorithmic
Model from Its Employees’ Brains, WALL STREET J. (Dec. 22, 2016, 01:14 PM),
https://www.wsj.com/articles/the-worlds-largest-hedge-fund-is-building-an-algorithmic-
model-of-its-founders-brain-1482423694.

Algorithmically-driven
advertisement:
Potential for
Algorithmically- discrimination via
driven resume exclusion of target Automated
sorting: Potential for audiences as justified by evaluation:
discrimination evaluation results Potential for
through selection discrimination via
criteria which are evaluation criteria
proxy variables Automated which are proxy
onboarding: Potential variables
for inadvertent
discrimination through
one-size fits all
As algorithmic technological advances present us with unprecedented legal

challenges,116 the use of machine learning algorithms in decision-making hiring
processes represents a particularly sensitive legal issue because of the potential
to create or exacerbate economic inequality. Yet, even with headline-making
cases of bias, the issue of how to govern algorithmic systems remains a thorny
one. In her article, Law for the Platform Economy, Professor Julie Cohen notes
that much of the conduct of platforms is simply “intractable using
conventional regulatory methodologies.”117 For instance, “to enforce existing
antidiscrimination laws effectively, the various agencies with enforcement
authority need the ability to detect and prove discrimination, yet that task is
increasingly difficult when decisions about lending, employment, and housing
are made via complex algorithms used to detect patterns in masses of data and
the data itself reflects preexisting patterns of inequality.”118 Professor Cohen
concludes that these phenomena combine to constitute a space “devoid of
protections for vital human freedoms, even as the activities conducted in that
space become more and more fundamental to the exercise of those
freedoms.”119
116 These legal challenges exist precisely because even with computing advancements that
allow computers to perform non-routine cognitive tasks, as noted by the legal scholar Cass
Sunstein, “at the present state of art, artificial intelligence cannot engage in analogical
reasoning or legal reasoning.” Kevin Ashley, Karl Branting, Howard Margolis & Cass R.
Sunstein, Legal Reasoning and Artificial Intelligence: How Computers “Think” Like Lawyers, 8 U.
CHI. L. SCH. ROUNDTABLE 1, 19 (2001); see Surden, supra note 48, at 88 (detailing gaps in the
law in regards to machine learning algorithms); see also Barocas & Selbst, supra note 3, at 673–
34 (detailing issues of disparate impact associated with algorithmic decision-making).
117 Cohen, supra note 37, at 189.
118 Id. at 190.
119 Id. at 199.

In his article, Professor Anupam Chander analogizes the process of “baby-

proofing” to the concept of “future-proofing.”120 To the question “[d]oes the
idea of ‘future-proofing’ law refer to a need to protect the future from the legal
system or does it refer to a need to protect rule of law from the future?”
Professor Cohen responds that neither is alone is quite accurate.121 This is
because the law and technological development are co-constitutive. Professor
Cohen argues that “legal institutions should change to meet the demands of
the times, and so it is only logical that the ascendancy of platforms should
produce new legal relationships and new institutional settlements.”122 She
reinforces that it is time to pay attention to the best paths for institutional
evolution and the “extent to which legal institutions bend to the service of
emergent economic power.”123
I echo Professor Cohen’s sentiments here, particularly in cautioning
against a techno-correctionist view of algorithmic bias that focuses on
technical fairness in lieu of examining assumptions undergirding the criteria
chosen for automated decision-making. But in order to engender new legal
frameworks, we must first grasp the true nature of the issue. In the following
sub-sections, I detail how the move towards the automation of decision-
making came about as an anti-bias intervention and the ways in which machine
learning algorithms involved in hiring have been found to return
discriminatory results and raise new legal quandaries.
A. Algorithms as Anti-Bias Intervention

The paradox of the algorithmic capture of hiring is that, despite proof of
algorithmic bias, some perceive the move to automated decision-making as an
anti-bias intervention. That is, the increasing use of algorithms in employment
decision-making is seen as an improvement versus decisions made by humans.
In Want Less-Biased Decisions? Use Algorithms,124 the author challenges many
scholars’ concern that “algorithms are often opaque, biased, and
unaccountable tools being wielded in the interests of institutional power.”125
He notes that although these critiques have helped people to avoid abusing
algorithms, there is a pattern among the critiques, “which is that they rarely ask
how well the systems would operate without algorithms.”126
Miller cites multiple studies of algorithmic decision-making that support
the notion that “algorithms are less biased and more accurate than the humans
they are replacing.”127 For instance, one study found that a job-screening
120 See id. at 203.

121 See id.
122 See id. at 204.
123 Id.
124 Alex P. Miller, Want Less-Biased Decisions? Use Algorithms, HARV. BUS. REV. (July 26, 2018),
https://hbr.org/2018/07/want-less-biased-decisions-use-algorithms.
125 Id.
126 Id.
127 Id.

algorithm “actually favored “nontraditional” candidates”128 much more than

human screeners did, “exhibit[ing] significantly less bias against candidates that
were underrepresented at the firm.”129 Other algorithmic studies related to
credit applications, criminal justice, public resource allocations, and corporate
governance all concluded that “[a]lgorithms are less biased and more accurate
than the humans they are replacing.”130
In each of the examples he notes, Miller argues that the algorithm
programmers “trained their algorithms on past data that is surely biased by
historical prejudices.”131 Miller asserts that while this fact might be alarming to
many, because of the widespread belief that algorithms are negatively affected
by a biased data set, “the humans [algorithms] are replacing are significantly
more biased.”132 He argues that a number of psychological and other studies
in judgment and decision have demonstrated that “humans are remarkably bad
judges of quality in a wide range of contexts”133 and that “very simple
mathematical models outperform supposed experts at predicting important
outcome.” Thus, since humans are significantly bad at making decisions,
“replacing them with algorithms both increased accuracy and reduced
institutional biases.”134 Miller refers to this as a “pareto improvement, where one
policy beats out the alternative on every outcome [people] care about.”135 He
also emphasizes that there is no trade-off between productivity and fairness
when using algorithms because “[a]lgorithms deliver more-efficient and more-
equitable outcomes.”136 Miller thus insists that even if technology cannot fully
solve the social ills of institutional bias and discrimination, it is worthwhile to
“accept that – in some instances – algorithms will be part of the solution for
reducing institutional biases”137 because the perils of human bias are far worse.
Similarly, the legal scholar Professor Stephanie Bornstein, in her article
Antidiscriminatory Algorithms,138 challenges the current focus on ensuring that
algorithmically derived decision-making results are not discriminatory. She
emphasizes the original intent of technology, which is “to improve upon
human decision-making by suppressing biases to make the most efficient and
least discriminatory decisions.”139 Deploying the example of Amazon’s hiring
algorithm and LinkedIn’s survey results,140 Professor Bornstein notes that
“algorithmic decision-making offers unprecedented potential to reduce the
128 Id.
129 Id.
130 Id.
131 Id.
132 Id.
133 Id.
134 Id.
135 Id.
136 Id.
137 Id.
138 Bornstein, Antidiscriminatory Algorithms, supra note 12.
139 Id. at 520.
140 See id. at 521–23.

stereotypes and implicit biases that often infect human decisions.”141 However,
Professor Bornstein acknowledges that despite the promise of algorithms to
reduce bias in decision-making, there are concerns about algorithmic
discrimination and the risk of reproducing existing inequality142 because the
effectiveness of algorithms and decision-making greatly relies on what data is
used and how.143 Yet, Professor Bornstein believes that if algorithms are
handled properly, they can still “suppress, interrupt, or remove protected class
stereotypes from decisions.”144
As noted earlier, my aim is not to arbitrate whether algorithms are less
biased than humans,145 and I do not believe that such a determination is
necessary to observe the inadequacy of current laws to govern machine
learning algorithms or to conceive of better legal frameworks. Therefore,
although in many respects the algorithmic turn to hiring is purportedly driven
by a desire for fairness and efficiency—for example, Goldman Sachs’s hiring
changes were prompted by a desire for a more diverse candidate pool,146 as
these machine learning algorithms may have the (un)intended effects of
perpetuating structural biases or could have a disparate impact on protected
categories,147 the law should evolve more robust governing mechanisms to
guard against those outcomes. In the next sub-section, I detail how bias may
still creep into algorithmic decision-making systems in the context of
recruitment and hiring.
B. The Fault in the Machine

Albeit that it is well documented that humans evince bias in employment
decision-making,148 one cannot overlook that algorithmic systems of decision-
making, too, might enable, facilitate, or amplify such biases. The cases below
demonstrate that this is true both in the recruitment and hiring of job
candidates. Consider also that it is exactly because Facebook has automated
algorithms that are able to distinguish (with some degree of accuracy) between
men and women that it is able to select only men for job ads, and also to have
this go undetected for some time. It is impossible to do the same with an
advertisement in a physical newspaper. Thus, it is the very design of automated
platform that is both enabling and facilitating discrimination here. Therefore,
this paper builds on the work of Professor Olivier Sylvain, who has noted that
141 Id. at 523.

142 See id.
143 See id. at 570.
144 Id.
145 See supra note 36 and accompanying text.
146 See Thompson, supra note 112.
147 See Surden, supra note 48, at 88 (detailing gaps in the law in regards to machine learning
algorithms); Barocas & Selbst, supra note 3 (detailing issues of disparate impact associated
with algorithmic decision-making).
148 See infra Section III.A.

the design of automated intermediaries, has “enabled a range of harmful

expressive acts, including violation of housing and employment laws.”149
1. Recruitment
A recent ProPublica investigation revealed that Facebook allowed
advertisers (both for jobs and for housing) to exclude audiences by ethnic
group.150 In what investigators described as “a modern form of Jim Crow,”151
Facebook had developed a feature it termed “Affinity Groups”—essentially, a
method for advertisers to use demographic data to algorithmically target who
will receive certain Facebook ads.152 For example, one page on Facebook
Business, titled “How to Reach the Hispanic Audience in the United States,”
boasts of the potential for advertisers to reach up to 26.7 million Facebook
users of “Hispanic Affinity.”153 From this specific Affinity Group, advertisers
can choose to narrow in on bilingual users, those who are “Spanish dominant,”
or those who are “English dominant,” in order to “refine their audiences.”154
Although, ostensibly, this algorithmic feature might help business owners
refine their audiences and target ads to individuals who might be more likely
customers, the use of Affinity Groups as an ad distribution tool holds high
potential for unlawful discrimination. In demonstration of this discriminatory
potential, ProPublica reporters were able to buy dozens of rental house ads on
Facebook that excluded “African Americans, mothers of high school kids,
people interested in wheelchair ramps, Jews, expats from Argentina, and
Spanish speakers.”155
Following on the heels of this ProPublica investigation, a 2017 class action
lawsuit against Facebook contended that Facebook Business tools both
“enable and encourage discrimination by excluding African Americans,
Latinos, and Asian Americans – but not white Americans from receiving
advertisements for relevant opportunities.”156 In an amended complaint, the
class action also alleged that “Facebook offers a feature that is legally
149
Oliver Sylvain, Discriminatory Designs on User Data, KNIGHT FIRST AMENDMENT
INSTITUTE. AT COLUM. U.: EMERGING THREATS (Apr. 01, 2018),
https://knightcolumbia.org/content/discriminatory-designs-user-data; see, also, Olivier
Sylvain, Intermediary Design Duties, 50 CONN. LAW REV. 203 (2018).
150 Julia Angwin & Terry Parris Jr., Facebook Lets Advertisers Exclude Users by Race,
PROPUBLICA: MACHINE BIAS (Oct. 28, 2016, 01:00 PM),

https://www.propublica.org/article/facebook-lets-advertisers-exclude-users-by-race.
151 Id.
152 See, About Ad Targeting: Creating a New Audience, FACEBOOK BUS.: ADVERTISER HELP
(2018),
https://www.facebook.com/business/help/717368264947302?helpref=page_content.
153 See U.S. Hispanic Affinity on Facebook, FACEBOOK BUS. (2018),
https://www.facebook.com/business/a/us-hispanic-affinity-audience.
154 See id.
155 Jessica Guynn, Facebook Halts Ads That Exclude Racial and Ethnic Groups, USA TODAY
(Nov. 29, 2017), https://www.usatoday.com/story/tech/2017/11/29/facebook-stop-

allowing-advertisers-exclude-racial-and-ethnic-groups-targeting/905133001/.
156 First Amended Complaint, Mobley v. Facebook, Inc., No. 16-cv-06440-EJD, at *1 (N.D. Cal.
Feb. 13, 2017).

indistinguishable from word-of-mouth hiring, which has long been considered

a discriminatory and unlawful employment practice.”157 This allegation
references Facebook’s “Lookalike Audiences” feature, in which employers and
employment agencies provide a list of their existing workers to Facebook, and
Facebook uses that list to then create its own list of Facebook users who are
demographically similar to the existing workers.158 Then, the employer or
employment agency uses the new “Lookalike Audience” list created by
Facebook as the population to receive its employment ads.159 Such a feature
would help to perpetuate any existing historical racial, gender, and other
demographic imbalances of employees already present in a given corporation.
Seemingly in response to the publicity from the ProPublica investigations,
Facebook began temporarily blocking advertisers from excluding audiences by
race in late 2017.160 In March of 2019, the Communication Workers of
America, along with the ACLU and the law firm of Outten & Golden LLP,
reached a settlement with Facebook in which the corporation agreed to make
“changes to its paid advertising platform to prevent discrimination in
employment, housing, and credit advertising.”161
157 See First Amended Class and Collective Action Complaint, Bradley v. T-Mobile, Inc., No. 17-
cv-07232-BLF, at 21 (N.D. Cal. May 29, 2018),
https://www.onlineagediscrimination.com/sites/default/files/documents/og-cwa-
complaint.pdf.
158 See id.
159 See id. See generally, About Lookalike Audiences, FACEBOOK BUS.: ADVERTISER HELP,
https://www.facebook.com/business/help/164749007013531 (last visited Aug. 10, 2019).

160 See Angwin & Parris, supra note 150.
161 CWA Secures Agreement with Facebook on Sweeping Reforms to Curb Discrimination, COMM.
WORKERS AM. (Mar. 21, 2019), https://cwa-union.org/news/cwa-secures-agreement-

facebook-on-sweeping-reforms-curb-discrimination. Here are some of the reforms
Facebook agreed to:
• Create a separate portal for such ads with a much more limited set of targeting
options so that advertisers cannot target ads based on Facebook users’ age,
gender, race, or categories that are associated with membership in protected
groups, or based on zip code or a geographic area that is less than a 15-mile
radius, and cannot consider users’ age, gender, or zip code when creating
“Lookalike” audiences for advertisers
• Implement a system of automated and human review to catch advertisements
that aren’t correctly self-certified as these types of ads
• Require all advertisers creating such ads to certify compliance with anti-
discrimination laws, and provide education for advertisers on those laws
• Study the potential for unintended biases in algorithmic modeling on
Facebook
• Meet with plaintiffs and their counsel every six months for three years to
enable them to monitor the implementation of the reforms that Facebook is
undertaking.
Press Release, Communication Workers of America, Facebook Agrees to Sweeping Reforms to
Curb Discriminatory Ad Targeting Practices, (Mar. 19, 2019), https://cwa-union.org/news/cwa-
secures-agreement-facebook-on-sweeping-reforms-curb-discrimination.

2. Hiring
Job recruitment algorithms on platforms like Facebook are, however, not
the sole problem. Algorithms that quickly sort job applicants based on pre-set
criteria may also (inadvertently) be unlawfully discriminatory. In her book,
Weapons of Math Destruction, Cathy O’Neil poignantly illustrates how personality
tests may serve to discriminate against one protected class, job applicants
suffering from mental disabilities.162 In one class action, the named plaintiff,
Kyle Behm, a college student with a near-perfect SAT score and who had been
diagnosed with bipolar disorder, found himself repeatedly rejected for
minimum wage jobs at supermarkets and retail stores that all used a personality
test that had been modeled on the “Five Factor Model Test” used to diagnose
mental illness.163 Thus, personality tests, as part of automated hiring systems,
could be seen as a covert method for violating antidiscrimination law—
specifically, the Americans with Disabilities Act.164 In addition, other test
questions, such as the length of commute time, could be seen as covertly
discriminating against those from under-resourced neighborhoods which lack
a reliable transportation infrastructure.165
In addition to personality tests, companies are using other algorithmic
processes to screen applicants. For example, the company HireVue offers
virtual interviews with individual applicants. HireVue’s innovative hiring tool
identifies facial expression, vocal indications, word choice, and more.166 The
problem is that “speech recognition software can perform poorly” and “facial
analysis systems can struggle to read the faces of women with darker skin.”167
Some skeptics express their concerns about the legitimacy of using physical
162 O’NEIL, supra note 3.

163 Id.
164 Title I and V of Titles I and V of the Americans with Disabilities Act of 1990 (Pub. L.
101-336) grants mentally ill workers equal opportunity in employment. See, e.g., Press Release,
U.S. Equal Emp’t Opportunity Comm’n, Worker with Bipolar Disorder to Receive $91,000 in
Disability Discrimination Case Settled by EEOC (Mar. 18, 2003),
https://www.eeoc.gov/eeoc/newsroom/release/3-18-03b.cfm; see also Depression, PTSD, &
Other Mental Health Conditions in the Workplace: Your Legal Rights, U.S. EQUAL EMP’T
OPPORTUNITY COMM’N, https://www.eeoc.gov/eeoc/publications/mental_health.cfm (last
visited Aug. 10, 2019).
165 Debra Cassens Weiss, Do Job Personality Tests Discriminate? EEOC Probes Lawyer’s Complaint,
Filed on Behalf of His Son, A.B.A. J. (Sept. 30, 2014, 9:08 AM),
http://www.abajournal.com/news/article/do_job_personality_tests_discriminate_eeoc_pro
bes_lawyers_complaint_filed_o.
166 Hilke Schellmann & Jason Bellini, Artificial Intelligence: The Robots Are Now Hiring, WALL
STREET J. (Sept. 20, 2018, 5:30 AM) https://www.wsj.com/articles/artificial-intelligence-

the-robots-are-now-hiring-moving-upstream-1537435820.
167 MIRANDA BOGEN & AARON RIEKE, HELP WANTED: AN EXAMINATION OF HIRING
ALGORITHMS, EQUITY, AND BIAS (2018),

https://www.upturn.org/static/reports/2018/hiring-algorithms/files/Upturn%20-
-%20Help%20Wanted%20-%20An%20Exploration%20of%20Hiring%20Algorithms,%20E
quity%20and%20Bias.pdf; Joy Buolamwini & Timnit Gebru, Gender Shades: Intersectional
Accuracy Disparities in Commercial Gender Classfication, 81 PROC. MACHINE LEARNING RES. 77
(2018).

features and facial expressions that have no causal link with workplace success
to make hiring decisions.168
Another example of automated hiring is the use of algorithms to conduct
social media background checks. Such checks are fraught with issues for
several reasons. First, they “presume that a person’s online behaviors, like
some use of foul language, are relevant to their professional activities.”169
Second, they have “limited ability to parse the nuanced meaning of human
communication.”170 In addition, such checks could “surface details about an
applicant’s race, sexual identity, disability, pregnancy, or health status, which
employers should not consider during the hiring process.”171
Finally, as the last step of the hiring process, employers make offers to
applicants using automated hiring systems. For example, there exist software
programs that predict the probability that a candidate will accept a given job
offer and that suggest what the employer could do to increase those chances.
For example, these programs allow the employer to “adjust salary, bonus, stock
options, and other benefits to see in real time how the prediction changes.”172
The worry remains that such programs might amplify pay gaps for white
women and racial minorities because the data commonly include “ample
proxies for a worker’s socioeconomic and racial status, which could be
reflected in salary requirement predictions.”173 They might also undermine laws
that bar employers from considering candidates’ salary histories.174
C. Exposing the Mechanical Turk

Even as incidences of algorithmic bias come to light,175 an important
feature of algorithms is that they tend to obscure the role of the human hand
168 See BOGEN & RIEKE, supra note 167.

169 Id.
170 Id.
171 Id.
172 Id.
173 See id.
174 See id.
175
It is important to note here, and as I discuss more fully in another article, that U.S. law
actively works to prevent instances of algorithmic bias from coming to light. One reason is
that the intellectual property law regime allows automated systems to fall under trade secret
or copyright law, which means that their creators are able to keep secret the exact workings
of these automated systems. Congress enacted the Digital Millennium Copyright Act
(DMCA) in 1998. Pub. L. No. 105-304, 112 Stat. 2860 (1998) (codified as amended, in
sections of 17 and 28 U.S.C.). Section 1201 of the DMCA creates liability for hacking or
reverse engineering an automated system protected under copyright law. 17 U.S.C. § 1201
(2012); see also Maayan Perel & Niva Elkin-Koren, Accountability in Algorithmic Copyright
Enforcement, 19 STAN. TECH. L. REV. 473 (2016) (noting the chilling effect on researchers
who would like to reverse engineer automated processes, given the potential to incur
liabilities). Another reason is that these systems also receive intellectual property law
protections without having demonstrated their utility to society. Cf. Christopher Buccafusco,
Mark A. Lemley & Jonathan S. Masur, Intelligent Design, 68 DUKE L.J. 75 (2018) (noting that
designers are “able to obtain powerful IP protection over the utilitarian aspects of their

in setting parameters for solving any given problem, with the final result
attributed solely to the machine.176 Consider that proponents of automations
have always tended to downplay or deny the role of the human mastermind.177
As an early example, consider “The Mechanical Turk” also known as “the
chess Turk,” which was a chess-playing machine constructed in the late
eighteenth century.178 Although the Mechanical Turk was presented as an
automaton chess-playing machine that was capable of beating the best human
players, the secret of the machine was that it contained a human man,
concealed inside its chambers.179 The hidden chess master controlled the
machine while the seemingly automated machine beat notable statesmen, like
Napoleon Bonaparte and Benjamin Franklin, at chess.180 Thus, the Mechanical
Turk operated on obfuscation and subterfuge and sought to reserve the glory
of the win to the machine.181
With the growing allure of artificial intelligence as a venture-capital-
generating marketing ploy,182 modern day corporations have been discovered
creations without demonstrating that they have made socially valuable contributions” and
concluding that “this is bad for competition and bad for consumers”).
176 Surden, supra note 48, at 115; Jatinder Singh, Ian Walden, Jon Crowcroft & Jean Bacon,
Responsibility & Machine Learning: Part of a Process (Oct. 27, 2016) (unpublished manuscript),
http://dx.doi.org/10.2139/ssrn.2860048 (arguing that machines can learn to operate in ways
beyond their programming levels, meaning that the responsibility for problems created by
the algorithms cannot lie solely with the algorithms creators or the algorithms themselves).
177 Cary Coglianese & David Lehr, Regulating by Robot: Administrative Decision Making in the
Machine-Learning Era, 105 GEO. L.J. 1147, 1207 (2017) (noting that machine-learning
technology is not yet fully understood and that most people “simply lack the interpretive
ability to . . . show that X causes Y on a machine-learning platform.”).
178 TOM STANDAGE, THE TURK: THE LIFE AND TIMES OF THE FAMOUS 19TH CENTURY
CHESS-PLAYING MACHINE (2003).

179 Ricky Jay, The Automaton Chess Player, the Invisible Girl, and the Telephone, 4 JAY’S JOURNAL OF
ANOMALIES (2000).
180 Id.
181 See PASQUALE, supra note 3 (arguing that algorithms operate on obfuscation). Conversely,
Amazon’s Mechanical Turk program does the opposite. The program allows businesses or
individual clients to assign human intelligence tasks, that is, tasks that are difficult or
impossible for machines to complete (like sorting photographs, writing product descriptions,
completing surveys, etc.) to humans. Amazon explicitly bans the use of automated bots to
complete such tasks. See AMAZON MECHANICAL TURK, https://www.mturk.com/ (last
182 See Ellen Huet, The Humans Hiding Behind the Chatbots, BLOOMBERG: TECH. (Apr. 18, 2016,
8:00 AM), https://www.bloomberg.com/news/articles/2016-04-18/the-humans-hiding-

behind-the-chatbots (“The incentive to play up automation is high. Human-assisted AI is
‘the hottest space to be in right now,’ said Navid Hadzaad, who founded bot-and-human
concierge service GoButler. Startups in this arena have together raised at least $50 million in
venture capital funding in the past two years. But companies with a wide variety of strategies
all use similar and vague marketing language and don’t often divulge operational details.”).
For example, in May of 2018, to much publicity, Google debuted an uncanny digital assistant
which is able to call for reservations and fool humans on the telephone with its natural
speech, complete with human-like pauses and interjections.
Chris Welch, Google Just Gave a Stunning Demo of Assistant Making an Actual Phone Call, VERGE
(May 8, 2018, 1:54 PM), https://www.theverge.com/2018/5/8/17332070/google-assistant-
makes-phone-call-demo-duplex-io-2018.

operating their own versions of the mechanical Turk. Consider, for example,
that the humans on the Amazon Mechanical Turk crowd-sourcing work
platform consider themselves “the AI behind the AI.”183 On this internet-
based platform, human workers are recruited to accomplish mundane tasks
that are difficult for algorithms to tackle. These tasks, referred to as “human
intelligence tasks” (or HITs), include: “transcribing audio clips; tagging photos
with relevant keywords; copying photocopied receipts into spreadsheets.”184
While the work on Amazon Turk and its notoriously low pay is no secret, a
Bloomberg exposé revealed that several corporations were disingenuously
passing off the labor of human workers as that of AI.185
Even when corporations are not attempting to pass off human workers as
AI, it is important to understand that there is always a human behind the AI.
Modern day algorithms operate in ways similar to the Mechanical Turk in that
the human decisions behind the creation of algorithms operated by businesses
are generally considered trade secrets that are jealously guarded and protected
from government oversight.186 But while algorithms might remove some
decisions from a human entity, humans must still make the initial decisions as
to what data to train the algorithm on and as to what factors are deemed
relevant or irrelevant.187 Even more importantly, the decisions for what data is
important in the training data—decisions that are then matched as closely as
possible by the algorithm—are also made by humans.188 For example, if a hiring
algorithm is trained on a corpus of resumes, a human must still make
183 Miranda Katz, Amazon’s Turker Crowd Has Had Enough, WIRED (Aug. 23, 2017, 06:55
AM), https://www.wired.com/story/amazons-turker-crowd-has-had-enough/.
184 Sarah O’Connor, My Battle to Prove I Write Better Than an AI Robot Called ‘Emma’, FIN.
TIMES (May 4, 2016), https://www.ft.com/content/92583120-0ae0-11e6-b0f1-

61f222853ff3.
185 See Huet, supra note 182 (“A handful of companies employ humans pretending to be
robots pretending to be humans. In the past two years, companies offering do-anything
concierges (Magic, Facebook’s M, GoButler); shopping assistants (Operator, Mezi); and e-
mail schedulers (X.ai, Clara) have sprung up. The goal for most of these businesses is to
require as few humans as possible. People are expensive. They don’t scale. They need health
insurance. But for now, the companies are largely powered by people, clicking behind the
curtain and making it look like magic.”).
186 See Frank Pasquale, Restoring Transparency to Automated Authority, 9 J. TELECOMM. & HIGH
TECH. L. 235 (2011). Congress enacted the Digital Millennium Copyright Act (DMCA) in
1998. Pub. L. No. 105-304, 112 Stat. 2860 (1998) (codified as amended, in sections of 17 and
28 U.S.C.). Section 1201 of the DMCA creates liability for hacking or reverse engineering an
automated system protected under copyright law. 17 U.S.C. § 1201 (2012); see also Perel &
Elkin-Koren, supra note 175 (noting the chilling effect on researchers who would like to
reverse engineer automated processes, given the potential to incur liabilities).
187 Even when automated feature selection methods are used, the final decision to use or not
use the results, as well as the choice of feature selection method and any fine-tuning of its
parameters, are choices made by humans. For more on feature selection see, e.g., GARETH
JAMES, DANIELA WITTEN, TREVOR HASTIE & ROBERT TIBSHIRANI, AN INTRODUCTION TO
STATISTICAL LEARNING WITH APPLICATIONS IN R (1st ed. 2013).
188 See, e.g., the way that hiring startup Jobaline verifies their technique by using the ratings
that people listening give voice snippets of job candidates. Ying Li et al., Predicting Voice
Elicited Emotions, in PROCEEDINGS OF THE 21ST ACM SIGKDD INT’L CONF. ON
KNOWLEDGE DISCOVERY AND DATA MINING 1969 (2015).

consequential decisions as to what variables from the resumes should matter

and which ones should be disregarded by the algorithm. Thus, akin to Mary
Shelley’s conclusion in Frankenstein’s Monster, the creators of runaway
algorithms should not be permitted to disavow their creations, rather those
makers, like Dr. Frankenstein, must bear the ultimate liability for any harm
wrought by their creations.
III. A LEGAL PROBLEM, NOT A TECHNICAL PROBLEM

The framing of the problem of algorithmic bias becomes important when
deciding to whom to allocate liability. I contend that the current framing of
algorithmic bias as a technical problem rather than as a legal problem is
misguided.189 As a consequence of this erroneous framing of algorithmic bias
as a solely technical problem, ineffective techno-solutionist approaches have
proliferated. I argue that a re-framing of the issues of bias in algorithmic hiring
as legal in nature is necessary to plumb the true depths of the problem. In the
following sections, I describe how the biased results of algorithmic hiring
systems are not merely technical deficiencies, rather, they reveal legal
anachronisms, such as an American tradition of deference to the employer and
what amounts to a legal shrug when it comes to addressing the nebulous
concept of “cultural fit” as hiring criterion.
A. A Legal Tradition of Employer Deference
American law has historically given much deference to employers vis-à-vis

the employment bargain.190 In an empirical legal study, Professors Clermont
and Schwab found:
189 Professor Sandra Mayson also makes this argument in regard to algorithmic decision-
making in the criminal justice field. Mayson, supra note 3 (arguing that the problem of
disparate impact in predictive risk algorithms lies not in the algorithmic system but in the
nature of prediction itself).
190 Perhaps the most emblematic example of the American legal system’s deference to
employers is the case of Lochner v. New York, 198 U.S. 45 (1905), which held that any limit on
the number of hours (in excess of 60 hours) that employees of a bakery could work was
unconstitutional. Although that particular decision has since been overturned, there is a
wealth of scholarship noting the continued deference to employers. See Cynthia Estlund, The
Ossification of American Labor Law, 102 COLUM. L. REV. 1527, 1527 (2002) (noting that the
“ossification of labor law” is due, in part, to a lack of “democratic revision”); see also Franita
Tolson, The Boundaries of Litigating Unconscious Discrimination: Firm-Based Remedies in Response to a
Hostile Judiciary, 33 DEL. J. CORP. L. 347 (2008). Courts want to avoid turning Title VII into a
rule by which employers could be held liable for “perceived slights” towards employees. Id.;
see also Kevin M. Clermont & Stewart J. Schwab, How Employment Discrimination Plaintiffs Fare
in Federal Court, 1 J. EMPIRICAL LEGAL STUD. 429 (2004) (claiming that employment
discrimination plaintiffs (unlike many other plaintiffs) have always done substantially worse
in judge trials than in jury trials); Michael J. Zimmer, The New Discrimination Law: Price
Waterhouse is Dead, Whither McDonnell Douglas?, 53 EMORY L.J. 1887, 1944 (2004) (“The 5.8
percent reversal rate of defendant trial victories is smaller in employment discrimination

Employment discrimination plaintiffs . . . manage many fewer happy

resolutions early in litigation, and so they have to proceed toward trial
more often. They win a lower proportion of cases during pretrial and
at trial. Then, more of their successful cases are appealed. On appeal,
they have a harder time upholding their successes and reversing
adverse outcomes.191
Likewise, after Professor Wendy Parker’s empirical study of 659 cases alleging
racial discrimination in employment, she concluded that judges operate under
the assumption that those types of claims are generally without merit.192 For
cases alleging implicit bias, Professor Franita Tolson has also noted that Courts
will find in favor of the employee in only the most extreme cases because the
courts have “statutory concerns and . . . the belief that they are not qualified
to solve these claims.”193
Professor Selmi echoes these conclusions and observes:
When it comes to race cases, which are generally the most difficult
claim for a plaintiff to succeed on, courts often seem mired in a belief
that the claims are generally unmeritorious, brought by whining
plaintiffs who have been given too many, not too few, breaks along
the way. These biases, as well as others, inevitably influence courts'
treatment of discrimination cases, and help explain why the cases are
so difficult to win.194
Professor Selmi also notes that judges display a similar bias against employees
in sex employment discrimination cases as judicial activism has created an
affirmative defense for employers “out of whole cloth, as there was very little
precedent for the defense . . . [and] may signal a shift in judicial attitudes that
portends more difficulty for plaintiffs to recover in cases of sexual
harassment . . . .”195
In addition to this legal deference in contested cases of employment
discrimination, employers exercise a great deal of latitude in choosing which
job applicants they hire and fire.196 This is especially true of at-will jurisdictions,
where employees can be hired and fired based on a vast list of criteria
cases than any other category of cases except prisoner habeas corpus trials.”); see also Ruth
Colker, The Americans With Disabilities Act: A Windfall for Defendants, 34 HARV. C.R.-C.L.L.
REV. 99, 100 (1999) (looking at reported decisions from 1992-1998 and finding that
defendants prevailed in more than 93% of the cases decided at the trial court level and were
more likely to be affirmed on appeal); Theodore Eisenberg, Litigation Models and Trial
Outcomes in Civil Rights and Prisoner Cases, 77 GEO. L.J. 1567, 1567 (1989) (noting that only
claims filed by prisoners have a lower success rate than that of employment discrimination
plaintiffs).
191 Clermont & Schwab, supra note 190.
192 Wendy Parker, Lessons in Losing: Race Discrimination in Employment, 81 NOTRE DAME L.
REV. 889, 893 (2006).

193 See also Tolson, supra note 190, at 378.
194 Michael Selmi, Why are Employment Discrimination Cases So Hard to Win?, 61 LA. L. REV.
555, 556–57 (2001).

195 Id. at 569.
196 Clyde Summers, Employment at Will in the United States: The Divine Right of Employers, 3 U.
PA. J. LAB. & EMP. L. 65 (2000).

determined by the employer.197 In her article, Discrimination at Will: Job Security

Protections and Equal Employment Opportunity in Conflict,198 Professor Julie Suk
notes that employment discrimination scholars have argued that “employment
at will seriously undermines the effectiveness of employment discrimination
law in bringing about race and gender equality in the workplace.”199 Other legal
scholars have articulated exactly why this is the case as they note that the job
protections present in antidiscrimination law might dissuade employers from
hiring job applicants from protected groups. In her book, Working Together,200
Professor Estlund argues that “employers have perverse disincentives to hire
racial minorities when Title VII operates in the context of employment at
will”201 because of the risk of incurring expenses in a Title VII suit. As
Professors Ian Ayres and Peter Siegelman put it, “protection against
discriminatory firing acts as a kind of tax on hiring those to whom it is
extended.”202 Furthermore, employment at will as a norm “affects the burdens
of production and proof under the McDonnell Douglas framework when
individual Title VII cases are litigated, often to the detriment of plaintiffs.”203
The Supreme Court’s 1993 decision in St. Mary’s Honor Center v. Hicks204
illustrates the detriments of the employment-at-will standard to employment
discrimination plaintiffs. The plaintiff, Hicks, a black correctional officer,
experienced repeated and severe disciplinary actions at the hands of a new
supervisor. Hicks was eventually fired. Hicks then filed a Title VII case and
presented a prima facie case under the standards of McDonnell Douglas v.
Green.205 Although the district court found that the defendant offered false
nondiscriminatory reasons for the firing, the district court nonetheless found
for the defendant because the plaintiff had not proven that the employer’s
actions were “racially rather than personally motivated.”206
Amidst the legal backdrop of employer discretion and deference, the use
of algorithmic hiring systems can exacerbate, rather than ameliorate, issues of
197 “The rule of employment at will allows either the employer or the employee to terminate
the employment relationship at any time for good reason, bad reason, or no reason.” Julie C.
Suk, Discrimination at Will: Job Security Protections and Equal Employment Opportunity in Conflict, 60
STAN. L. REV. 78 (2007). At-will employment is the law in every U.S. state except for
Montana. See At-Will Employment - Overview, NAT’L CONF. ST. LEGISLATURES (Apr. 15, 2018),
http://www.ncsl.org/research/labor-and-employment/at-will-employment-overview.aspx.
198 Suk, supra note 197.
199 Id. at 81.
200 CYNTHIA ESTLUND, WORKING TOGETHER: HOW WORKPLACE BONDS STRENGTHEN A
DIVERSE DEMOCRACY (2003).

201 Id. page number for quote?
202 Ian Ayres & Peter Siegelman, The Q-Word as Red Herring: Why Disparate Impact
Liability Does Not Induce Hiring Quotas, 74 TEX. L. REV. 1487, 1489 (1996).
203 Suk, supra note 197, at 81.
204 St. Mary’s Honor Ctr. v. Hicks, 509 U.S. 502, 508 (1993).
205 Under McDonnell Douglas v. Green, 411 U.S. 792 (1973), a plaintiff could establish a prima
facie case without direct evidence by proving (1) that he was a member of a protected group,
(2) that he was qualified for the job, (3) he applied for the job and was rejected, and (4) the
job continued to remain open. Id. at 802.
206 St. Mary’s Honor Ctr., supra note 204, at 508.

bias, particularly given the well documented technological capability for those
types of hiring systems to substitute facially neutral variables as proxies for
protected demographic characteristics such as race and gender.207 The nature
of the hiring relationship can be explained succinctly by the following quote:
“the typical matching of a worker to a position does not reflect the outcome
of workers picking from among several job offers. Rather, it is the result of an
employer picking from among several job applicants.”208 Employers choose
candidates and not the other way around. Today, with the growing expanse of
online job applications, job seekers apply to an average of twenty-seven jobs
before they attain one interview.209 Of course, since only 17% of interviews
actually result in an offer of employment, it is likely that these applicants apply
to far more than twenty-seven jobs throughout their entire job search.210 In
one extreme case, an applicant even built his own algorithm to apply to
thousands of jobs at once, in an attempt to “beat” being sorted out by
automated hiring platforms.211
On the employer’s side of this surge in applications, on average, fifty-nine
people apply for each open position.212 From this pool of applicants, then, the
employer is required to eliminate a large number of candidates in order to find
candidates to interview—and ultimately hire. Due to the large pool of
applicants, though, an average of only 12% of applicants will be interviewed
for any open position.213 This indicates that employers must use the
information available to them to eliminate a large number of applicants before
they can make substantial progress in finding the “most talented candidates.”
The sheer necessity for this culling of possible job applicants has left some
scholars in support of the employers’ total discretion in the hiring process.214
Yet, it is undeniable that granting such near-total discretion opens the door for
human bias to be introduced into the employment decision-making process.
The subsequent use of algorithmic systems only allows said bias to become
entrenched and more difficult to detect.
207 Barocas & Selbst, supra note 3, at 692–93.

208 John Barron, John Bishop & William C. Dunkelberg, Employer Search: The Interviewing and
Hiring of New Employees, 61 REV. ECON. & STAT. 43 (1985).
209 Matthew Nitch Smith, You Now Have to Apply for 27 Jobs Just to Get 1 Interview, BUS.
INSIDER (May 16, 2016, 5:51 AM), http://www.businessinsider.com/job-seekers-have-to-

apply-for-27-jobs-for-every-interview-survey-finds.
210 Martha C. White, Here’s How Long It Really Takes to Get a Job, MONEY (Oct. 22, 2015),
http://time.com/money/4053899/how-long-it-takes-to-get-hired/.
211 Robert Coombs, I Built a Bot to Apply to Thousands of Jobs at Once – Here’s What I Learned,
FAST COMPANY: CAREER EVOLUTION (Mar. 3, 2017),

https://www.fastcompany.com/3069166/i-built-a-bot-to-apply-to-thousands-of-jobs-at-
once-heres-what-i-learned.
212 White, supra note 210.
213 Id.
214 See, e.g., Peter Cappelli, Career Jobs Are Dead, 42 CAL. MGMT. REV. 146 (1999).

B. The Problem of “Cultural Fit”

Given the wide discretion that employers enjoy, an insidious manner in
which bias may infiltrate the employment decision-making process is through
the discernment of “cultural fit,” as, in many ways, the variables employed to
algorithmically cull resumes are approximations of “cultural fit.” The problem
is that some of those variables may be inherently at odds with Title VII, while
allowing employers to avoid hiring protected classes of applicants, as long “as
some credible, non-discriminatory reason can be prescribed.”215 This issue is
exacerbated by machine learning algorithms that may not treat “cultural fit” as
the amorphous concept that it is but rather as a strict rule, thus creating the
genre of scenarios I detailed before wherein a hiring algorithm or algorithms
might deduce that only men named “Jared” and who play lacrosse are “fit” for
the job.216
To achieve culling, employers study a number of qualities about
candidates—from their resumes, to their past employment experiences, and
their “cultural fit” within the perspective company.217 Cultural fit is defined as
“the likelihood that job candidates will be able to conform and adapt to the
core values and collective behaviors that make up an organization.218 For
example, “60% of recruiters rate culture fit of highest importance when
making a hiring decision.”219 This shows that recruiters considered cultural fit
as more significant than cover letters (26%), prestige of college (21%), and
GPA (19%). Cultural fit of a candidate is topped only by previous job
experience (67%).220 When judging whether a candidate is a cultural fit, 83%
of recruiters consider communication style most important.
While scholars largely agree about the need for resumes and descriptions
of past experiences, many are at odds with the idea of an employer’s
determination of an applicant’s cultural fit.221 In many regards, the
determination of a candidate’s cultural fit is subjective. Some articles report
that assessing cultural fit “comes down to an employer’s gut feeling.”222 Others
have reported that a candidate might be a good fit if “they work well with
215 Suk, supra note 197.

216 See Gershgorn, supra note 91.
217 See Laura Roberts & Darryl Roberts, Testing the Limits of Antidiscrimination Law: The Business,
Legal, and Ethical Ramifications of Cultural Profiling at Work, 14 DUKE J. GENDER L. & POL’Y
369 (2007).
218 Margaret Rouse, What is Cultural Fit?, TECH TARGET: SEARCHCIO (last updated Sept.
2014), https://searchcio.techtarget.com/definition/Cultural-fit
219 JOBVITE, JOBVITE RECRUITER NATION REPORT 2016: THE ANNUAL RECRUITING
SURVEY (2016), https://www.jobvite.com/wp-

content/uploads/2016/09/RecruiterNation2016.pdf.
220 See id.
221 See, e.g., Christine Sgarlata Chung, From Lily Bart to the Boom Boom Room: How Wall Street’s
Social and Cultural Response to Women Has Shaped Securities Regulation, 33 HARV. J.L. & GENDER
175 (2010) (arguing that cultural fit within the finance industry is imperfect, as bias has been
historically ingrained).
222 Robert Half, How to Know If a Candidate Will Match Your Company Culture, ROBERT HALF
TECH. (May 5, 2017, 7:15 PM), https://www.roberthalf.com/blog/evaluating-job-

candidates/how-to-know-if-a-candidate-will-match-your-company-culture.

others”—which seems to be something that is difficult to predict without ever

seeing a candidate work with other people.223 To this end, some researchers
have promoted the idea that corporate culture can be learned, indeed, because
nearly every company that hires a new employee has a period of “socialization”
or social training.224 Others have shown that interviewers are not even
significantly adept at assessing applicants’ personal characteristics from
interviews.225
On the other hand, researchers have argued that assessing cultural fit is
important because “employee alignment to company culture influences worker
satisfaction, engagement, and retention,” which can ultimately help the
corporation to succeed.226 Furthermore, a study of thirty-eight interviewers
who were in the process of making hiring decisions found that interviewers
can actually assess cultural fit “with a significant level of accuracy” and that
this factor is often “the best predictor of employment.”227 Given both
arguments, it is clear that there can be both positive and negative implications
of trying to assess a candidate’s cultural fit—but cultural fit can only be a useful
criterion for hiring so long as an employer could make an accurate
determination of such fit.
The problem there is that, as employment law scholars, such as Professor
Natasha Martin, have noted, “workplace decision-makers may gain awareness
of the lack of cultural fit only after some time has passed.”228 This means that
even if an individual is hired, it is likely that the decision-maker’s view of the
worker changes over time with greater exposure.229 Professor Martin notes that
“[t]he social construction of identity bears on selection of individuals when
organizations make decisions based on cultural fit. By aligning employees with
environmental factors, employers focus less on the candidate’s skill set, and
more on the intangibles that make uncovering discriminatory motive more
difficult.”230 For instance, cultural fit qualifications, such as “relish change,”
“possess passion for exceptional quality,” “confront risks,” or “think
creatively,” have an “amorphous quality because they are generalized and
undefined with respect to any particular job task or role.”231 Therefore, the
ultimate decisions can rest merely on gut feelings like “I just like that
223 Jeff Pruitt, 3 Ways to Know If an Employee Is a Culture Fit, INC. (Aug. 12, 2016),
https://www.inc.com/jeff-pruitt/3-ways-to-know-if-an-employee-is-a-culture-fit.html.
224 Richard Pascale, The Paradox of “Corporate Culture”: Reconciling Ourselves to Socialization, 27
CAL. MGMT. REV. 2 (1985).

225 Richard Arvey & James Campion, The Employment Interview: A Summary and Review of Recent
Research, 35 PERSONNEL PSYCHOL. J. 2 (1982).

226 Lauren Dixon, The Pros and Cons of Hiring for ‘Cultural Fit’, CHIEF LEARNING OFFICER:
TALENT ECON. (Dec. 6, 2017), https://www.chieflearningofficer.com/2017/12/06/pros-

cons-hiring-cultural-fit/.
227 Daniel M. Cable & Timothy A. Judge, Interviewers’ Perceptions of Person-Organization Fit and
Organizational Selection Decisions, 82 J. APPLIED PSYCH. 4, 558 (1997).

228 Natasha T. Martin, Immunity for Hire: How the Same-Actor Doctrine Sustains Discrimination in
the Contemporary Workplace, 40 CONN. L. REV. 1117, 1156 (2008).

229 See id.
230 Id. at 1158.
231 Id.

candidate” or “he just feels right.”232 Even though choosing an individual

based on seemingly aligned values may look like a good business decision, the
problem with cultural fit in the selection process lies in the “imperceptibility
of such characteristics, and the conscious and unconscious layering of meaning
on such abstract terms by the decision-maker.”233 Professor Martin cites to
Professor Charles Lawrence’s law review article,234 in which he observes that
“where an ‘employer perceives the white candidate as “more articulate,” “more
collegial,” “more thoughtful,” or “more charismatic[,]” [h]e is unaware of the
learned stereotype that influenced his decision.’”235
In sum, the problem with cultural fit is about the “perception of who
belongs.”236 Employers try to select candidates who match their “culturally
consistent selection criteria,” and in the fast-paced workplace, an employer has
to make decisions with incomplete information about each candidate and hire
an individual who seems to be a “good fit based on paper credentials and
limited stated interview interaction.”237 Later, the same employee “might be
deemed unfit in cultural terms . . . because he failed to attune to the
organization’s culture.”238 In this situation, “[p]eople of color and other
workplace minorities are often encouraged to assimilate, mask their true
identity or ethnic salience, for example, to adapt to an organization’s
culture.”239 Citing to the work of Professors Devon Carbado and Mitu
Gulati,240 Martin explains that “as long as the worker’s difference is innocuous
and unobtrusive, then the worker benefits from efforts to belong” and “once
the employee fails to sufficiently cover, he may be deemed inconsistent with
the organization and no longer befitting of inclusion.”241 Professors Carbado
and Gulati have also argued that the “extra” identity work that women and
minorities are forced to do to conform to workplace perceptions of “cultural
fit” is a form of employment discrimination.242
Courts, however, have consistently sided with defendants in contestations
of cultural fit as a criterion for hiring and firing. One notorious case is Natay
v. Murray Sch. Dist.,243 in which the courts sided with the employer when
deciding the cultural fit of an employee. The plaintiff, hired by the school
district as a provisional teacher, was the only Native American in the group of
forty-seven recently hired provisional teachers and on the school faculty.244
232 Id.
233 Id. at 1159.
234 Id. (citing Charles R. Lawrence III, The Id, the Ego, and Equal Protection: Reckoning with
Unconscious Racism, 39 STAN. L. REV. 317 (1987)).

235 Martin, supra note 228, at 1159; Lawrence, supra note 234, at 343.
236 Martin, supra note 228, at 1159.
237 Id.
238 Id.
239 Id.
240 See Devon W. Carbado & Mitu Gulati, Working Identity, 85 CORNELL L. REV. 1259, 1262
(2000); Devon W. Carbado, Racial Naturalization, 57 AM. Q. 633, 655 (2005).

241 Martin, supra note 228, at 1160.
242 See Carbado & Gulati, supra note 240, at 1262.
243 Natay v. Murray Sch. Dist., 118 Fed. Appx. 259 (10th Cir. 2005).
244 See id. at 260.

Plaintiff described her treatment at the school as discriminatory from the start;
the principal snubbed her at the first staff meeting, disciplined her but not
another teacher, and came late to her scheduled evaluations. In addition, the
principal told Natay at one point that she was “geographically, racially,
culturally, and socially out of place” at the school, and her contract was not
renewed after unfavorable evaluations.245 The school district superintendent
also decided that the plaintiff was “not an excellent teacher and not someone
[he] would want Murray School District to hire on a long-term basis.”246 Natay
had an informal conference with the superintendent, but her arguments did
not change the decision. Of the district’s provisional teachers hired for that
school year, only Natay’s contract was not renewed, and on her last day of
work, the principal made another racially derogatory comment to her.247 Natay
brought a discriminatory discharge claim in federal district court, and the court
entered summary judgment in favor of the employer. She appealed.
In the Tenth Circuit, the court reviewed the district court’s grant of
summary judgment, using the same standards, as well as the evidence and
reasonable inferences drawn from the evidence. It stated that although the
plaintiff proved that the principal showed discriminatory actions, she lacked
evidence showing that the superintendent, the ultimate decision-maker, had a
discriminatory reason not to renew her contract.248 Based on the “cat’s paw”
doctrine, she did not prove that the “manager who discharged the plaintiff
merely acted as a rubber stamp, or the ‘cats’ paw,’ for a subordinate employee’s
prejudice,”249 regardless of the manager’s discriminatory intent. Also, the court
used the McDonnell Douglas burden-shifting framework. The court decided that
the plaintiff satisfied the prima facie case and succeeded in shifting the burden
to the school district. Although the plaintiff claimed that the superintendent’s
investigation of her performance was inadequate because he never sat in her
classroom to observe her, the court stood on the side of the superintendent,
whose affidavit detailed other steps in his investigation and decision to not
renew the contract due to her ineffectiveness. Thus, the court concluded that
the plaintiff’s showing did not reasonably give rise to an inference that the
employer’s reasons were pretextual and affirmed the decision of the district
court.250
C. Re-Thinking Employer Discretion

Given that deference to an employer’s determination of cultural fit may
not bode well for members of protected groups, as an ex ante approach to
curbing algorithmic discrimination in hiring, I argue for a re-thinking of
employer discretion, particularly regarding variables meant to indicate “cultural
fit” as part of algorithmic hiring. I propose that the law should mandate that
245 See id. at 259.

246 Id. at 261.
247 See id.
248 See id.
249 Id. at 262.
250 Id.

the criteria used in algorithmic hiring must have some probative value for
determining fitness to perform required job duties. This proposal is supported
by new studies that show that “cultural fitness” is not always necessary for
long-term success at a firm. One such study conducted by business professors
at Stanford and Berkeley found that the capacity to change and flexibility—
that is, high “enculturability”—were more important than pre-existing cultural
fit in regard to long-term success.251 According to the authors of the study:
Our results suggest that firms should place less emphasis on screen for
cultural fit, . . . [a]s other work has shown, matching on cultural fit
often favors applicants from particular socioeconomic backgrounds,
leading to a reduction in workplace diversity. Instead, our work points
to the value of screening on enculturability.252
The study concludes with three enculturability questions that employers might
pose to potential candidates during the hiring process: “1. To what extent do
candidates seek out diverse cultural environments? 2. How rapidly do they
adjust to these new environments? 3. How do they balance adapting to the
new culture while staying true to themselves?”253
As a result of such new studies, more companies are moving away from
“cultural fit” as a factor for hiring. For example, in a bid to create a more
inclusive hiring process, Facebook outlawed the term “culture fit” as interview
feedback, “requiring interviewers to provide specific feedback that supported
their position.”254 Facebook also took steps to “proactively identify
unconscious bias” in their interview process and “developed a ‘managing
unconscious bias’ training program.”255
Other companies now embrace “hiring for values fit” as a method to
decrease unconscious bias in interviewing. For example, Atlassian, an
Australia-based company, redesigned its interview process: “values fit
interviewers are carefully selected and given training on topics like structured
interviewing and unconscious bias.” The interview is structured with a set of
behavioral questions to assess whether a candidate would thrive in an
environment with their company values.256 As one of Atlassian’s chief officers
explains: “Focusing on ‘values fit’ ensures we hire people who share our sense
of purpose and guiding principles, while actively looking for those with diverse
251 Amir Goldberg, Sameer B. Srivastava, V. Govind Manian, William Monroe &
Christopher Potts, Fitting In or Standing Out? The Tradeoffs of Structural and Cultural Embeddedness,
81 AM. SOC. REV. 1190 (2016).
252 Rich Lyons, Lose Those Cultural Fit Tests: Instead Screen New Hires for ‘Enculturability’, FORBES
(June 7, 2017, 10:00 AM), https://www.forbes.com/sites/richlyons/2017/06/07/lose-

those-cultural-fit-tests-instead-screen-new-hires-for-enculturability/#7db0478863a8.
253 Id.
254 Lars Schmidt, The End of Culture Fit, FORBES (Mar. 21, 2017, 7:50 AM),
https://www.forbes.com/sites/larsschmidt/2017/03/21/the-end-of-culture-
fit/#3045454e638a.
255 Managing Conscious Bias, FACEBOOK NEWSROOM (July 28, 2015),
https://newsroom.fb.com/news/2015/07/managing-unconscious-bias/.
256 Schmidt, supra note 254. A description of Atlassian’s company values is available here:
https://www.atlassian.com/company/values.

viewpoints, backgrounds, and skill sets. We’re trying to build a healthy and
balanced culture, not a cult.”257 This approach has borne positive results for
Atlassian. In 2015, 10% of their technical workforce identified as female. In
2016, 17% of recent hires were women, and women held 14% of all technical
roles. Similarly, in 2015 their U.S.-based team had 23% of employees
identifying as people of color. In 2016, people of color comprised 32% of their
new hires.258
IV. EX LEGIS: NEW LEGAL FRAMEWORKS

In this section, I discuss potential new legal frameworks that could address
bias in automated hiring platforms. An open legal issue is who should bear the
legal liability for when an automated hiring algorithm is returning biased
results. Is it the business firm that is using the algorithms for its hiring process?
Or is it the maker of the algorithmic hiring platform? In this section, I
approach the issue of liability from different vantages: A) from the position of
holding accountable the maker of the algorithmic hiring system; B) from the
position of holding the employer accountable; and C) from the position of
holding liable both the maker of the algorithmic system and, potentially, also
the employer.
A. Improving on the Fiduciary Duty Concept

Even if it is accepted that employers owe no duties to job applicants, some
legal scholars would argue that the makers of hiring platforms should owe a
legal duty to job applicants.259 Professor Jack Balkin characterizes information
fiduciaries as entities “who, because of their relationship with another, assume
special duties with respect to the information they obtain in the course of the
relationship.”260 According to Professor Balkin the relationship between the
257 Id.
258 Id. Atlassian diversity hire figures are public available here:
https://www.atlassian.com/belonging.
259
See, e.g., Ford, supra note 30 (arguing that employment law imposes a duty of care on
employers to avoid decisions that undermine social equality).
260 The phrase “information fiduciaries” was first coined by Professor Kenneth Laudon. See
Kenneth C. Laudon, Markets and Privacy, in ICIS 1993 PROCEEDINGS 65, 70–71 (1993)
(proposing a “National Information Market” within which “information fiduciaries
would . . . accept deposits of information from depositors and seek to maximize the return
on sales of that information in national markets or elsewhere in return for a fee”). Professor
Jack Balkin popularized the term in several writings. See Jack Balkin, Information Fiduciaries in
the Digital Age, BALKINIZATION (Mar. 5, 2014),
https://balkin.blogspot.com/2014/03/information-fiduciaries-in- digital-age.html; Jack M.
Balkin, Information Fiduciaries and the First Amendment, 49 U.C. DAVIS L. REV. 1183 (2016); Jack
M. Balkin, Free Speech in the Algorithmic Society: Big Data, Private Governance, and New School Speech
Regulation, 51 U.C. DAVIS L. REV. 1149, 1160–63 (2018); Jack M. Balkin, Free Speech Is a
Triangle, 118 COLUM. L. REV. 2011, 2047–55 (2018); Jack M. Balkin, AEGIS SERIES PAPER
NO. 1814, FIXING SOCIAL MEDIA’S GRAND BARGAIN 11–15 (2018),
https://www.hoover.org/sites/default/files/research/docs/balkin_webreadypdf.pdf.

consumer and job applicant is analogous to that between a doctor and

patient.261 Thus, an information fiduciary “has an obligation to act in the
interests of its clients.”262 In the context of employment, the primary question
is: How should the law conceptualize the responsibilities of these hiring
platforms in regard to the information they solicit and transmit?
I depart from Professor Balkin’s analogy,263 to offer a more critical view of
the relationship between job seekers and the automated hiring platform as
brokering intermediary. My theorizing responds directly to trenchant critiques
of the information fiduciary idea. Notably, Professors Lina Khan and David
Pozen have expressed doubts in their Essay, A Skeptical View of Information
Fiduciaries,264 as to “whether the concept of information fiduciaries is an
adequate or apt response to the problems of information asymmetry and
abuse” of platforms and whether such a theory ignores “fundamental
problems associated with outsized market share and business models that
demand pervasive surveillance.”265 The authors also conclude that the
information-fiduciary framework “invites an enervating complacency about
issues of structural power and a premature abandonment of more robust
visions of public regulation.”266
My aim in this Article is not to adjudicate whether the concept of
information fiduciaries can comprehensively offer the solution to problems
associated with online platforms.267 I do, however, find the concept useful in
discussing problems associated with one particular type of online platform,
and that is the automated hiring platform. And as the rest of my Article
demonstrates, I consider the idea of the information fiduciary as an important
Professor Jonathan Zittrain has also made important theoretical contributions to the concept
of information fiduciaries. See Jack M. Balkin & Jonathan Zittrain, A Grand Bargain to Make
Tech Companies Trustworthy, ATLANTIC (Oct. 3, 2016),
https://www.theatlantic.com/technology/archive/2016/10/information-fiduciary/502346;
Jonathan Zittrain, Facebook Could Decide an Election Without Anyone Ever Finding Out, NEW
REPUBLIC (June 1, 2014), https://newrepublic.com/article/117878/information- fiduciary-
solution-facebook-digital-gerrymandering; Jonathan Zittrain, How to Exercise the Power You
Didn’t Ask For, HARV. BUS. REV. (Sept. 19, 2018), https://hbr.org/2018/09/how-to-
exercise-the-power- you-didnt-ask-for; Jonathan Zittrain, Mark Zuckerberg Can Still Fix This
Mess, N.Y. TIMES (Apr. 7, 2018),
https://www.nytimes.com/2018/04/07/opinion/sunday/zuckerberg- facebook-privacy-
congress.html.
261 Balkin, Information Fiduciaries and the First Amendment, supra note 260, at 1209.
262 Id. at 1209.
263 Id. at 1209.
264
Lina M. Khan & David Pozen, A Skeptical View of Information Fiduciaries, 133 HARV. L.
REV. (forthcoming 2019).
265
See id.
266 See id.
267 Cf. James Grimmelmann, When All You Have Is a Fiduciary, LAW & POL. ECON. (May 30,
2019), https://lpeblog.org/2019/05/30/when-all-you-have-is-a-fiduciary/ (suggesting that

while fiduciary principles are ill-suited to problems of self-dealing, content moderation, and
market concentration on online platforms, the “best version” of U.S. information privacy
law “would cash out fiduciary principles in specifying when and how platforms can use and
share user data”).

building block towards other theorization regarding online platforms, and also
as one of several potential checks to the currently unbridled power of
automated hiring platforms to serve exclusionary ends. I believe that there
could be a multiplicity of approaches to the governance of online platforms
for the better of society. For example, while hitherto I have tended to focus
on governmental action as the appropriate governance mechanism for
algorithmic bias, other legal scholars, like Professor Sonia Katyal, have called
for private accountability measures.268 No one scholar can claim the cure-all
solution to the problem of algorithmic bias, but rather than reject proposed
solutions outright as nostrum, we should acknowledge that we are all blind
men grasping at the elephant and that our collective intellectual attempts may
yet reveal a full view of the problem at hand. Several legal scholars have
recently called for the extension of fiduciary duties to other areas of law.269
And some legal scholars have also argued against the expansion of fiduciary
duties.270 I argue that established concepts from organizational theory
scholarship further bolster the argument that hiring platforms are performing
a brokerage function and thus should be considered fiduciaries.271
My theorizing then works to clarify both the power and information
asymmetry relationships present in the triad of job applicant, hiring platform,
and employer. For example, with platform authoritarianism, I make clear the
unequal power relationship between the job applicant and the platform which
allows the platform to dictate in what manner the job applicant may make use
of the platform thus belying the caretaking imagery conjured by a doctor-
patient analogy. With the tertius bifrons concept, I reveal the duplicitous
relationship between the hiring platform and the job applicant which then
supports the argument for greater employment discrimination liability for the
platform.
268 Katyal, supra note 16 (calling for a variety of tools to tackle algorithmic accountability
such as codes of conduct, impact statements, and whistleblower protection).
269 See Theodore Rave, Politicians as Fiduciaries, 126 HARV. L. REV. 671 (2013);
see also Theodore Rave, Fiduciary Voters, 66 DUKE L.J. 2 (2016); Cf. Seth Davis, The False
Promise of Fiduciary Government, 89 NOTRE DAME L. REV. 3 (2014).
270
See James Grimmelmann, Speech Engines, 98 MINN. L. REV. 868, 904 (2014) (“[W]e are
undergoing something of an academic fiduciary renaissance, with scholars arguing for
treating legislators, judges, jurors, and even friends as fiduciaries.”); Daniel Yeager, Fiduciary-
isms: A Study of Academic Influence on the Expansion of the Law, 65 DRAKE L. REV. 179, 184
(2017) (arguing that “academic writing, deploying a sense of fiduciary so open as to be
empty, has influenced courts to designate” more entities as fiduciaries).
271 I take as authority the definition of brokerage set forth by Marsden: brokerage is
understood as a mechanism by which “actors facilitate transactions between other actors

lacking access to or trust in one another . . . .” See Peter V. Marsden, Brokerage Behavior in
Restricted Exchange Networks, in SOCIAL STRUCTURE AND NETWORK ANALYSIS 201 (1982).

1. Platform Authoritarianism
As Professor Olivier Sylvain has noted, platforms “shape the form and
substance of their users’ content.”272 Furthermore, platforms also shape
relationships as they connect users to one another while also enjoying “a great
deal of control over how users’ encounters are structured.”273 In evaluating
certain design policy choices that these companies make, such as the methods
through which they facilitate the amount of information users can learn about
one another and how they are to do so, one argument is that online platforms
can make choices that exacerbate the discrimination in our current society.274
Thus, makers of platforms cannot be blameless for the discrimination that
occurs on them—even if their users may be influenced by pre-existing
biases.275 Thus, I theorize “platform authoritarianism” as a socio-technical
phenomenon that has transformed the responsibility and liability of
platforms.276
Platform authoritarianism is what I term our present social position vis-à-
vis platforms, wherein creators of platforms demand that we engage with those
platforms solely “on their dictated terms, without regard for established laws
and business ethics.”277 Some scholars have noted that many online platforms
“can control who is matched with whom for various forms of exchange, what
information users have about one another during their interactions, and how
indicators of reliability and reputation are made salient.”278 This means that for
example, job applicants on hiring platforms must acquiesce to data demands
from the platforms, they are also not in control of how their candidacy is
presented, but rather must relinquish all control to the platform as quid pro quo
for accessing job opportunity. Rejecting platform authoritarianism in favor of
a duty of care that the purveyors of online platforms owe to their users is the
first step towards returning to a rule of law for algorithms.
2. The Tertius Bifrons

While exercising authoritative control over the content and structure of
their users’ interactions, hiring platforms also hide their true relationship to job
applicants and this deception can lull applicants into a false sense of trust.
According to the sociologist Georg Simmel, brokers, as part of a triad, perform
272
Oliver Sylvain, Discriminatory Designs on User Data, KNIGHT FIRST AMENDMENT
INSTITUTE. AT COLUM. U.: EMERGING THREATS (Apr. 01, 2018),
https://knightcolumbia.org/content/discriminatory-designs-user-data
273 Karen Levy & Solon Barocas, Designing Against Discrimination in Online Markets, 32
BERKELEY TECH. L.J. 1183 (2017).

274 Id.
275 Id.
276 Ifeoma Ajunwa, Facebook Users Aren’t the Reason Facebook Is in Trouble Now, WASHINGTON
POST, (Mar. 23, 2018),

https://www.washingtonpost.com/news/posteverything/wp/2018/03/23/facebook-users-
arent-the-reason-facebook-is-in-trouble-now/.
277 Id.
278 Levy & Barocas, supra note 273.

the function of brokering information between two separate groups, acting as

either tertius iungens or tertius gaudens.279 The tertius iungens (“the third who joins”)
orientation is derived from the Latin verb “iungo,” which means to join, unite,
or connect.280 The emphasis of this orientation is on the joining of two parties.
Thus, a tertius iungens broker will operate with a strategic emphasis on creating
friendship and collaboration between two parties.281 For example, linking
disparate parties in one’s social network in order to create outcomes that are
mutually beneficial for two or more parties.282 In contrast, the tertius gaudens
(“the third who enjoys”) orientation emphasizes the strategic separation of
parties.283 In this sense, a broker would enjoy the benefit of the continued
separation between two parties for the broker’s own gain.284
Going beyond the two categories of brokerage, organizational theory
scholars have noted that brokers may engage in four different brokering
strategies. Brokers may:
(1) coordinate action or information between distant parties who have
no immediate prospect for direct introduction or connection, (2)
actively maintain and exploit the separation between parties, (3)
introduce or facilitate preexisting ties between parties such that the
coordinative role of the broker subsequently recedes in importance,
and (4) introduce or facilitate interaction between parties while
maintaining an essential coordinative role over time.285
The automated hiring process comprises a triad, with the automated hiring
platform as the broker negotiating between the applicant and the employer. I
argue that, in this triad, hiring platforms are brokers who perform an “essential
coordinative role over time” by continuously parsing resumes received from
job applicants before delivering them to employers. Furthermore, I propose
that automated hiring platforms which can be customized at the request of the
employer (but not that of the applicant) belong to a new category of brokers
that I term, “the tertius bifrons” (that is, “the two-faced third”). With the
introduction of this new term, I am arguing that automated hiring platforms
represent a type of broker which works both in its own interest (to maintain
its coordinative role) and in the interest one of the parties to the triad (the
employer), while maintaining the appearance of working for both parties
(employer and job applicant).
This categorization rings true in light of the class action allegations against
Facebook.286 Facebook users entrust their information to platforms like
Facebook with the expectation that those platforms would use that
information to better the users’ experience. However, what has been alleged is
279 SIMMEL, supra note 33.

280 David Obstfeldt, Social Network, the Tertius Iungens Orientation and Involvement in Innovation, 50
ADMIN. SCI. Q. 1, 102 (2005).
281 Id.
282 Id.
283 Id.
284 Id.
285 Id.
286 See First Amended Class and Collective Action Complaint, supra note 157, at 21.

that Facebook, by creating “affinity groups” and “look-alike audiences” from

its users’ information (and especially when such Facebook-provided features
are deployed in defiance of antidiscrimination laws),287 has brokered
information to employers in a way that benefits both Facebook and the
employer, but not necessarily the user. I concur with the legal theory then that
this brokerage of job applicant information, in a manner that is inconsistent
with the best interests of the job applicant, violates a fiduciary duty held by the
hiring platform as an information fiduciary.
B. Discrimination Per Se
As holding corporations responsible for the algorithmic bias of the
automated hiring platforms they use represents a challenging legal problem
because of the difficulty of discovering proof and establishing intent, I propose
a new burden-shifting theory of liability, discrimination per se.288 Discrimination per
se would allow for a third cause of action under Title VII.289 The purpose is to
aid plaintiffs who cannot show proof of disparate treatment or who would
have difficulty obtaining the means to show the statistical proof of disparate
impact. Title VII requires intent for liability to attach, or in the absence of
intent a clear demonstration of disparate impact with no excuse of business
necessity for the disparity.290 When bringing disparate impact claims, plaintiffs
are likely to face three interrelated obstacles: “(1) compiling the requisite
statistics to show that the policy has a disparate impact . . . (2) identifying a
specific policy or practice that caused the adverse employment decision, and
287 See id.

288 Although my proposed doctrine borrows from tort theory, it is important to note that the
National Labor Relations Act characterizes some employer actions as per se violations. 29
U.S.C. § 158 (2012). The statute defines per se violations of the bargaining obligation as
conduct that violates Section 8(a)(5) without need for further inquiry, include unilateral
changes involving mandatory subjects of bargaining, even when such changes are made in a
context that otherwise indicates good faith bargaining, and even where the changes are
partially made in an effort to comply with governmental requirements.
Other per se violations include: An employer bypassing the union and bargaining directly
with employees; Insistence to impasse upon permissive subjects of bargaining; and refusing
to execute a written agreement embodying the terms of a negotiated contract. See Timothy
M. McConville, Employer Policies May Be Per Se Violations of the National Labor Relations Act
(NLRA), NAT’L L. REV. (July 12, 2013), https://www.natlawreview.com/article/employer-
policies-may-be-se-violations-national-labor-relations-act-nlra.
289 Title VII of the Civil Rights Act protects the job applicant against discrimination on the
basis of sex, race, color, national origin, and religion. See U.S. Civil Rights Act of 1964 §7, 42
U.S.C. §2000e (2012). Plaintiffs must establish that “a respondent uses a particular
employment practice that causes a disparate impact on the basis of [a protected
characteristic] and the respondent fails to demonstrate that the challenged practice is job
related for the position in question and consistent with its business necessity.” 42 U.S.C. §
2000e-2(k)(1)(A)(i).
290 Proving clear intent is necessary when attempting to make a disparate treatment case
under Title VII. However, under the disparate impact of clause of action codified in Title
VII, the intent is implied from an established pattern. See U.S. Civil Rights Act of 1964 §7, 42
U.S.C. §2000e-2(1)(A).

(3) rebutting the employer’s defense that the policy is justified by a business
necessity.”291 Also notable, “courts are inconsistent in addressing the
requirement of compiling appropriate statistics to show that a policy has a
disparate impact.”292 Second, courts often fail to find a “particular employment
practice” that caused the disparity because they cannot distinguish actual job
tasks from the default norms.293 Many times, courts use the phrase “particular
employment practice” to narrow the applicability of disparate impact
liability.294
In their essay, Incomprehensible Discrimination, Professors James
Grimmelmann and Daniel Westreich, make the case that when a plaintiff has
met the burden of showing disparate impact, “the defendant’s burden to show
a business necessity requires it to show not just that its model’s scores are not
just correlated with job performance but explain it.”295 This heightened burden
acknowledges the information asymmetry that exists between the employer
and the employee in the context of automated hiring. My proposed doctrine
of discrimination per se while concurring that there is a duty of care owed by the
employer, seeks to further rectify both the information asymmetry and power
imbalance present in automated hiring situations by entirely shifting the
burden of proof from plaintiff to defendant.
Per my proposal, a plaintiff can assert that a hiring practice (for example,
the use of proxy variables resulting or with the potential to result in adverse impact
to protected categories) is so egregious as to amount to discrimination per se, and
this would shift the burden of proof from the plaintiff to the defendant
(employer) to show that its practice is non-discriminatory. I do not set forth a
specific rule or standard for how to determine discrimination per se, rather, I
think this is a question of law that, like other types of American legal doctrines,
should be generated through case law. Note that this is not an automatic win
for the plaintiff, rather it merely reverses the American legal tradition of
deference to employers and allows that an employment discrimination plaintiff
will at least get a day in court. Note also that it still remains relatively easy for
employers to establish business necessity for their practices and therefore
defeat any plaintiffs’ disparate impact claims.296
Discrimination per se is an answer to the question of whether the liability of
corporations could be mitigated by a lack of intent to discriminate or even a
291 Nicole Buonocore Porter, Synergistic Solutions: An Integrated Approach to

Solving the Caregiver Conundrum for “Real” Workers, 39 STETSON L. REV. 777, 808 (2010).
292 See, e.g., Charles A. Sullivan, Disparate Impact: Looking Past the Desert Palace Mirage, 47 WM. &
MARY L. REV. 911, 989 (2005).

293 See Porter, supra note 291, at 809.
294 See id.

295
Grimmelmann & Westreich, supra note 12 at 170 (emphasis in original).
296 See Porter, supra note 291, at 810; it’s important to note here that even after a defendant
has been able to show business necessity, a plaintiff may nevertheless be able to prevail by
showing that there could be an “alternative employment practice” that meets that business
necessity.” §703 (k)(1)(A) & C. See also, Jones v. City of Boston, 845 F.3d 28 (1st Cir. 2016).

lack of awareness that an algorithm is producing biased results.297 For example,

one researcher, Jatinder Singh, has argued that the line of responsibility for
problems created by machine learning algorithms is blurred.298 More
specifically, if a machine learning algorithm can operate without being
specifically programmed, by de novo creating a model from available data,
should the blame for any resulting disparate impact lie with the creator of the
algorithm, with the entity who chose the training data, or with the algorithm
itself—with the last option presuming that the technology is essentially
“thinking” on its own? This, Singh argues, is a question that has yet to be
addressed by any current legal framework.299
In one attempt to surmount this problem, Professor Stephanie Bornstein,
in the article Reckless Discrimination,300 theorizes a recklessness model of
discrimination under Title VII, arguing that an employer should be liable for
acts done in reckless negligence, which are consequences of implicit bias and
stereotyping in employment decisions.301 In doing so, Professor Bornstein
argues that recent technological advancements have “allowed some employers
to easily and dramatically reduce the biasing effects of subjectivity from their
hiring decisions by, for example, using algorithms instead of people to screen
applicants.”302
Bornstein explains how employers have made efforts to remove implicit
bias from hiring; they have developed “blind interviews over online instant
messaging software, application-screening algorithms, pre-commitment to set
assessment standards, and more.”303 The author notes that “the ability to
prevent and correct for bias and stereotyping in the workplace is more
affordable and accessible than ever before.”304 She insists that given the
development of technological tools, such as predictive algorithms, which
employers can utilize to reduce bias in decision-making, an employer that
knows about the risks of implicit bias, has evidence that such bias may infect
its decision-making, and fails to try to prevent it should be liable for
discriminatory intent and reckless action.305
Building on Professor Bornstein’s line of argumentation, I propose here
that the well-established tort principle of negligence per se should be the model
for creating a new legal framework to answer the question of intent when it
comes to discriminatory results obtained by automated hiring platforms. While
another legal scholar, Professor Girardeau Spann, has also borrowed from tort
doctrine to argue that the invidiousness of racial discrimination in the United
States and the undeniable concomitant racial disparities dictate the strict
297
Professor Charles Sullivan has also grappled with these questions. See Sullivan, supra note
12, at 398.
298 Singh et. al., supra note 176.
299 Id.
300 Bornstein, Reckless Discrimination, supra note 12.
301 See id. at 1056.
302 Id.
303 Id. at 1058.
304 Id.
305 See id. at 1110.

liability standard of res ipsa loquitor for racial discrimination claims,306 and I
believe that in some instances such a standard might be warranted, I argue that
a discrimination per se standard that is modeled on the negligence per se standard is
more generally applicable (that is, it would apply to various cases of
employment discrimination, not just racially-motivated discrimination) and
also serves to institute more feasible self-regulation practices. The concept of
discrimination per se is also in line with Professor Ford’s argument that
employment discrimination law imposes a duty of care on the employer to
ensure that its employment practices are not unlawfully discriminatory.307 Note
that, as I explain in another article in progress, Automated Employment
Discrimination, the discrimination per se doctrine should work hand in hand
with an “auditing imperative” imposed on the employer.308 This takes into
consideration the practical problems associated with proving disparate impact
in an algorithmic hiring scenario and would allow a plaintiff to have some
headway in making the case.
The proto negligence per se case involved a Minnesota drug store clerk who
sold a deadly poison to a customer at the customer’s request.309 At the time of
the sale, the clerk did not label the substance as a “poison,” which was required
by a state statute for the sales of such substances.310 Later, the customer who
had purchased the substance ingested the chemical, which caused her death.311
Given these facts, should the clerk have been held legally liable for his actions,
which indirectly caused the customer’s death? This case, Osborne v. McMasters,
became one of the earliest cases in the United States to analyze the illegal
concept of negligence per se. Given the facts of the case, the court first found that
there could be no “serious doubt of defendant’s liability”—as he had known
of his duty to label the bottle as poison.312 In explanation, the court detailed
that it was
well settled . . . that where a statute or municipal ordinance imposes
upon any person a specific duty for the protection or benefit of others,
if he neglects to perform that duty he is liable to those for whose
protection or benefit it was imposed for any injuries of the character
which the statute or ordinance was designed to prevent . . . .313
Since the time of Osborne, the doctrine of negligence per se has become
commonly used for violations of laws such as traffic laws, building codes,
blood alcohol content limits, and various federal laws.314 For example, in
306
See Giradeau A. Spann, Race Ipsa Loquitor, 2018 MICH. ST. L. REV. 1025 (2019).
307
See, e.g., Ford, supra note 30 (arguing that employment law imposes a duty of care on
employers to avoid decisions that undermine social equality).
308
I discuss the “auditing imperative” in another article, Automated Employment Discrimination
(draft manuscript available).
309 Osborne v. McMasters, 41 N.W. 543 (Minn. 1889).
310 Id.
311 Id.
312 Id. at 543.
313 Id.
314 See, e.g., Williams v. Calhoun, 333 S.E.2d 408 (Ga. App. 1985) (in which the defendant’s
failure to stop at a stop sign constituted negligence per se); Lombard v. Colo. Outdoor Educ.

Mikula v. Tailors, an Ohio business invitee was taken to the emergency room
after falling down in a snow-covered parking lot at the place of business to
which she was invited.315 Witnesses report to have seen her fall after stepping
into a hole in the parking lot that was about seven inches deep and had been
covered by the snowfall from that day. After careful consideration, the jury
determined that
[a] deep hole in a parking lot which is filled or covered, or both, by a
natural accumulation of snow constitutes a condition, the existence of
which the owner of the premises is bound, in the exercise of reasonable
care, to know. He is also bound to know that a natural accumulation
of snow which fills or covers the hole is a condition substantially more
dangerous than that normally associated with snow. . . . Under such
circumstances, the owner’s failure to correct the condition constitutes
actionable negligence.316
Moreover, failure to correct an issue can also lend itself to negligence per se
claims if the accused individual is found to have violated a statute by his or her
failure to respond to a problem. For example, in Miller v. Christian, a landlord
was found negligent per se, after being placed on notice from a tenant that the
building’s sewage system had recurring problems.317 Failure to “fix the
immediate problem within a reasonable amount of time” resulted in a backup
of the sewage system, which caused the tenant’s apartment to flood, ruining
much of her personal property.318 The court in Miller found that Allan
Christian, the landlord, was liable for the damage to the tenant’s property
because he had a legal duty to maintain the apartment’s sewage system in
addition to being legally obligated to keep the premises fit for habitation.319
Often, “failure to correct” claims entail a consideration of whether the
plaintiff knew of the problem, as it is presumed that a plaintiff with knowledge
of an existing problem would be reasonable enough to avoid injury by the issue
altogether. In one case, Walker v. RLI Enterprises, a tenant in an apartment
building sued her landlord after she stepped out the back door of the building
and slipped on a sheet of ice.320 She suffered serious injuries to her ankle.321 In
her suit, the tenant asserted that the landlord was negligent in maintaining the
property, because she had given him notice of a leaky water faucet by the back
door of her apartment.322 This negligence, the court determined, was negligence
Ctr., Inc., 187 P.3d 565 (Colo. 2008) (in which an outdoor education teacher fell off of a
ladder that was in violation of building code restrictions, establishing negligence per se on
the part of the landowner); Purchase v. Meyer, 737 P.2d 661 (Wash. 1987) (in which a cocktail
lounge was found negligent per se for serving alcohol to a minor).
315 Mikula v. Tailors, 263 NE.2d 316 (Ohio 1970).
316 Id. at 322–23.
317 Miller v. Christian, 958 F.2d 1234 (3d Cir. 1992).
318 Id. at 1234.
319 V.I. CODE ANN. tit. 29, § 333(b)(1) (2019).
320 Walker v. RLI Enters., Inc., No. 89325, 2007 WL 4442725, at *1 (Ohio Ct. App. 2007).
321 Id.
322 Id.

per se because the landlord had an obligation to maintain the premises under
Ohio law.323
At trial, however, the landlord argued that “a landlord is only liable where
he has ‘superior knowledge’ of the defect that led to the injury.”324 By this the
landlord meant that as the tenant had alerted him of the problem, the tenant
then clearly knew as much about the dangerous conditions as he did. He also
noted that she had taken no further action to avoid the leaky faucet and could
thus be responsible for her own injury.325 However, the court found this
argument unconvincing, holding that such an argument only applies in the
context of natural accumulations of ice and snow, because most people have
experienced such conditions and know that they should take precautions.326
Site-specific problems, though, are the responsibility of the landlord to correct,
as he likely has a “superior” knowledge of the issues on the property than his
tenants or site visitors.327
In the case of automated hiring systems, employers have an obligation not
to unlawfully discriminate against applicants, as proscribed by Title VII of the
Civil Rights Act and other federal antidiscrimination laws. Furthermore, as I
propose in a separate paper, if self-audits or external audits of hiring algorithms
become mandated by law,328 then it follows that when an employer willfully
neglects to audit and correct its automated hiring systems for unlawful bias, a
prima facie intent to discriminate could be implied, pursuant to the proposed
doctrine of discrimination per se. This argument becomes persuasive when one
considers that some corporations make use of bespoke internal hiring
algorithms, such that no one, except the corporation, has access to the hiring
algorithm and its results—meaning then that only the corporation could have
“superior knowledge” of any problems of bias.
There are two important arguments against the introduction of the
discrimination per se doctrine: 1) The first is the difficulty of establishing a
standard for when the doctrine might apply; 2) The second is that it imposes
too large a burden on the employer. Regarding the first, I agree that it will take
some work on the parts of the courts to establish clear precedents for when
the doctrine could apply. But this is true for any new legal doctrine. In fact,
even established legal doctrines still face contestation as to when they should
or should not apply.329 Consider that in the context of automated hiring the
323 Id. at *2.

324 Id. at *5.
325 Id.
326 Id.
327 Id.
328 I discuss the proposal for mandates self and external audits of hiring algorithms in
another article, Automated Employment Discrimination (draft manuscript available).

329
See Laura T. Kessler, The Attachment Cap: Employment Discrimination Law, Women’s Cultural
Caregiving, and the Limits of Economic and Liberal Legal Theory, 34 U. MICH. J.L. REFORM 371,
415–16 (2001) (noting that because men are not usually primary caregivers, women have a
difficult time finding comparisons to prove a disparate impact); Michael Selmi, Was the
Disparate Impact Theory a Mistake?, 53 UCLA L. REV. 701, 769 (2006) (stating that establishing

two legal doctrines currently available to the plaintiff on which to build a case
are disparate treatment or disparate impact. The fact is that there are very few
cases of disparate treatment because employers are now much too
sophisticated to leave the kind of “smoking gun” evidence required. For
disparate impact, the problem is that there is wide discrepancy in determining
what statistics are enough to show a pattern of disparate impact disparate
impact.330
Regarding the burden on employers, the fact remains that automated
hiring is a cost-saving measure. Employers save significant amounts of money
and time by using automated hiring platforms. However, automated hiring
platforms should not save employers from their duty not to discriminate. Just
like an employer holds a responsibility to supervise its human workers for
activities that might contravene the law, so, too, remains an obligation to audit
automated hiring systems for bias. This burden is neither heavier than when
the intermediary is human, nor does it disappear merely because the
intermediary is a set of algorithms. The doctrine of discrimination per se is meant
to prevent employers from shirking their responsibility.
C. Consumer Protection for Job Applicants

Another method for ensuring the accountability of hiring algorithms
is to view the job applicant as a consumer and, thus, as deserving consumer
protection for unfair algorithmic outcomes. This approach could potentially
allow for both the maker of the platform and the employer to be held liable.
The Federal Credit Reporting Act (FCRA),331 while typically thought to solely
govern the distribution of credit reports could potentially be leveraged when
an employer relies on information from third parties—namely, an application
pre-screener or a hiring algorithm software.332 Note that while the FCRA
would not be able to directly hold the employer accountable for any
discriminatory use of the data, the point of my proposal here is that the
procedural aspects of the FCRA may also enable the job applicant to discover
if the employer had access to discriminatory information or even to establish
a pattern of discriminatory information furnished to the employer for
protected groups—thus perhaps assisting in a disparate impact cause of action.
First, some brief details about the language and intentions of the FCRA.
The FCRA, passed in 1970, was initially intended to protect consumers who
were being “scanned” for creditworthiness. The language set forth by the
FCRA was applied primarily to the “Big Three” credit reporting agencies—
Equifax, Experian, and TransUnion—all of which would draw up reports
about consumers, using their personal information to determine their credit
a statistically significant impact might be difficult unless the affected population is

sufficiently large and diverse).
330
William R. Corbett, Fixing Employment Discrimination Law, 62 SMU L. REV. 81, 113 (2009)
(noting that courts are inconsistent in applying the disparate impact doctrine).
331 See 15 U.S.C. § 1681a (2012).
332 Pauline Kim & Erika Hanson, People Analytics and the Regulation of Information Under the Fair
Credit Reporting Act, 61 ST. LOUIS UNIV. L.J. 17 (2017).

eligibility.333 They would then submit these reports to banks and employers,
showing the “risk” of the current individual in terms of lending or
employment. As such, the FCRA was passed to prevent unfair or opaque credit
reporting.334
The law also protects consumers from unfair background checks and
unauthorized collections of their private information, ensuring that consumers
are alerted to any information that may adversely affect their abilities to obtain
either credit or, more recently enforced, employment.335 Moreover, the law also
protects consumers by providing that creditors or employers must “obtain a
written authorization from any applicant or employee for the procurement of
a report, and certify to the consumer reporting agency its compliance with the
requirements of the statute such that that it will not violate any equal
employment opportunity law.”336 Through such provisions, the FCRA gives
consumers more control over how their personal information is reported by
consumer reporting agencies and used by both banks and employers.
However, since the time of its passage, the FCRA has expanded its bounds
such that it no longer only applies to the “Big Three” credit reporting
agencies.337 Now, it also applies to a variety of agencies that collect and sell
information that is found outside the workplace and that might be pertinent
for applicant-reviewing purposes.338 With the coverage of many consumer
reporting agencies (CRAs) whose sole purpose is employment pre-screening,
a question has arisen regarding the point at which a screening service should
be considered a CRA by the FCRA. In essence, how big of a role does a
reporting agency have to play in the information collection and reporting
process in order to face such substantial government regulation?
The language of the FCRA plainly defines the characteristics of entities
that can be considered CRAs, as well as the content of reports that can be
considered “consumer reports” under the law. A consumer reporting agency,
by definition, is any “person which, for monetary fees, dues, or on a
cooperative nonprofit basis, regularly engages in whole or in part in the
practice of assembling or evaluating consumer credit information or other
information on consumers for the purpose of furnishing consumer reports to
third parties.”339 Application screening software companies could be
333 Id. at 26.

334 See 15 U.S.C., supra note 331.
335 15 U.S.C. §1681k(a) (2012).
336 15 U.S.C. §1681b(b) (2012).
337 See CONSUMER FIN. PROT. BUREAU, LIST OF CONSUMER REPORTING COMPANIES (2018),
https://files.consumerfinance.gov/f/201604_cfpb_list-of-consumer-reporting-
companies.pdf.
338 See, e.g, CHECKER INC., https://Checkr.com (last visited Aug. 10, 2019) (which screens
applicants for criminal records, driving records, and also provides employment verifications,
international verifications, and drug screenings); HIRERIGHT LLC., www.HireRight.com (last
visited Aug. 10, 2019) (which boasts the industry’s broadest collection of on-demand
screening applications); FIRST ADVANTAGE CORP., www.fadv.com (last visited Aug. 10,
2019) (which provides criminal and pre-employment background checks, as well as drug-
testing and tenant screening services).
339 15 U.S.C. § 1681a(d)(2)(f).

considered CRAs, as they regularly process and evaluate “other information

on consumers” for the purpose of providing reports to employers.
Furthermore, these companies arguably develop “consumer reports,”
judging by the legal definition of that term. The FCRA defines “consumer
reports” as
any written, oral, or other communication of any information by a
consumer reporting agency bearing on a consumer’s credit worthiness,
credit standing, credit capacity, character, general reputation, personal
characteristics, or mode of living which is used or expected to be
used . . . as a factor in establishing the consumer’s eligibility for credit
or insurance or employment purposes.340
Through an analysis of the terms of service of two algorithm-based
employment screening companies—Monster Hiring and Newton Software—
it becomes clear that the reports that these kinds of corporations create could
certainly qualify as consumer reports, where the “consumers” are job
applicants.
Monster, a networking platform intended to connect job-seekers to
available employers, states in its terms of service that it “retains the ability to
collect information about [consumers] from publically-available websites and
may use this information to create a Profile or append it to an existing Profile”
on the company’s website.341 This information—in addition to any
information users choose to add—is then arranged in a Profile format on
Monster’s website, where employers can pay to post job listings and view
applicant resumes.342 For an example of the service costs, the cost to post and
promote one job ad on Monster is $425 for a sixty-day post longevity.343 That
cost includes the distribution of one job ad on Monster’s “job board” as well
as “access to 20 recommended resumes from the Monster database”—from
job-seekers selected by Monster, per the qualifications an employer lists in the
job description.344 Effectively, the process of arranging profiles in its own
structured form, the ability to add information the company finds online, and
the practice of recommending “suitable” applicants after an employer pays for
a job post all seem to show that Monster “exercises reasonable control over
the information it releases.”345 Thus, it could certainly be argued that Monster
is “communicating information that bears on a consumer’s character, general
reputation, personal characteristics, and mode of living, and are expected to be
used as a factor in establishing the consumers’ eligibility for employment
purposes.”346 Most specifically, by retaining the right to add any information it
discovers online, it is clear that Monster takes an active role in distributing
340 15 U.S.C. §1681a(d)(1) (2012).

341 See Terms of Use, MONSTER, http://inside.monster.com/terms-of-use (last visited Aug. 10,
2019).
342 See Monster: Job Board Overview for Employers Plus FAQs and Pricing, BETTERTEAM,
https://www.betterteam.com/monster (last visited August 13th, 2019).

343 Id.
344 Id.
345 Levy & Barocas, supra note 273.
346 15 U.S.C., supra note 340.

information related to an applicant’s job prospects, making its reports qualify

as “consumer reports” under the definition of the FCRA.
Another instance in which the FCRA might be applied to algorithm hiring
platforms is in the case of platforms in which “consumers”—or job
applicants—have even less control over the information that is collected and
reported, such as the case of Newton Software. Newton Software is, for all
purposes, a background check access provider, although it also advertises
resume-parsing tools and interview-streamlining data reports.347 The software
platform advertises to employers that it takes in applicant information and
“intelligently stores the information into new candidate profiles, so that errors
from manually inputting applicant data are a thing of the past.”348 From these
profiles, employers can find employees to “best fit” the requirements of the
job descriptions they release.349 Then, Newton Software gives employers
access to background check software provided by third parties, which Newton
itself entirely oversees.350 Given these features, I argue that Newton has
enough of a hand in the report-creating process as to have the final reports
attributed to itself, making its reports “consumer reports.”
The information-analyzing services put forth by Newton Software are
certainly more hands-on than those of Monster. Newton acquires sensitive
information from third-party background checkers, after overviewing the
background screening process, and then relays a report about the screening to
its clients. Further, by parsing resumes and creating new, standardized profiles
on applicants for employers to reference, Newton is certainly changing the
nature of the resumes that prospective employees have submitted and is
thereby creating its own reports with added information. All of this data, which
Newton relays to employers, can, and likely will, be used to determine
employment eligibility. Therefore, under the language of the FCRA, the
reports put forth by Newton qualify as “consumer reports.”
Furthermore, the Federal Trade Commission (FTC), the body that
oversees the FCRA, has recently held that “just saying you’re not a consumer
reporting agency is not enough.”351 The case, which took place in 2013, dealt
with an application available for purchase on iTunes, Filiquarian Publishing,
which advertised that it could make “quick criminal background checks for
convictions” in specific states.352 The application had access to hundreds of
thousands of criminal records and could help employers discover if any of the
347 See ATS Made Simple, NEWTON SOFTWARE, https://newtonsoftware.com/features/ (last

348 See id.
349 See id.
350 Background Check Software for Hiring, NEWTON SOFTWARE,
https://newtonsoftware.com/features/integrated-background-checks/ (last visited Aug. 10,

2019).
351 Tony Rodriguez & Jessica Lyon, Background Screening Reports and the FCRA: Just Saying
You’re Not a Consumer Reporting Agency Isn’t Enough, FED. TRADE COMMISSION (Jan. 10, 2013,
2:00 PM), https://www.ftc.gov/news-events/blogs/business-blog/2013/01/background-
screening-reports-fcra-just-saying-youre-not.
352 Id.

convictions could be attributed to their applicants.353 However, Filiquarian also

reported in a disclaimer on its site that it was not a consumer reporting agency
because its background screening reports were not to be considered screening
products for insurance, employment, loans, or credit applications.354 The FTC
took issue with this, finding that Filiquarian provided the exact same
information as CRAs, but simply said “don’t use this information for
employment purposes,” which is not a reasonable excuse from FCRA
compliance.355 Ultimately, the FTC’s statement was the following: “companies
offering background screening products for employment or other FCRA
purposes have to stay in line with the law.” This mandate should be applied to
sites like Monster and Newton, given that they provide similar screening
services.
If these companies were to be considered consumer reporting agencies
under the FCRA down the road, Monster, Newton, and other similar reporting
services that find their own information to conduct screening checks might
also leave themselves open to FCRA claims for failing to “follow reasonable
procedures to assure maximum possible accuracy of their files, which cause
individuals to be denied” employment opportunities.356 In one case, Thompson
v. San Antonio Retail Merchants Association (SARMA), the Fifth Circuit found that
SARMA had erred in its creation of a profile for Thompson, automatically
“capturing” the incorrect social security number for his profile and erroneously
reporting the bad credit history of another man by the same common name.
The court ultimately held that under the FCRA, such an oversight by a credit
reporting agency as the one presented in Thompson was enough to show
negligence on the part of the agency.357
One potential rebuttal to the classification of hiring algorithms as CRAs
can be extrapolated from Professor Ryan Calo’s article, Open Robotics.358 With
reference to the robotics community, Professor Calo argues that just as
firearms manufacturers are ultimately not responsible for what end users do
with their products, manufacturers of open robotics platforms should not be
held responsible either.359 When applied to the regulation of hiring algorithms
that develop reports about prospective employees, one could make a similar
argument, holding that hiring algorithms and their developers are not
ultimately responsible for the negative impacts that employers use them to
create. Instead, the consumers—in this case, employers—who use such
platforms may be responsible for the decisions that they make once they
purchase the hiring tools.
353 Id.
354 Id.
355 Id.
356 Thompson v. San Antonio Retail Merchs. Ass’n, 682 F.2d 509 (5th Cir. 1982); see also Spokeo v.
Robins, 136 S. Ct. 1540, 1546 (2016) (in which a “people search engine” provided incorrect
personal information about a consumer to employers and the Supreme Court ruled that this
established concrete injury to the consumer, by damaging his employment prospects).
357 Id.
358 Ryan Calo, Open Robotics, 70 MD. L. REV. 101 (2011).
359 Id. at 106.

However, hiring platform developers differ significantly from firearm

manufacturers or open robotics platforms, given that those developers hold
the power to create features that may enable employment discrimination and
also given that automated hiring platforms exercise considerable control over
how applicants’ job applications may be captured, analyzed, and presented to
employers. These differences bolster the argument for classifying entities that
screen applicants’ information to create hiring reports as CRAs under the law.
In doing so, job applicants, as consumers, could gain some insight as to how
they are evaluated, and society could regain some measure of checks over the
information that is used to “screen” candidates as part of the automated hiring
trend. Furthermore, the classification of hiring platforms as CRAs would assist
the information gathering of a would-be employment discrimination plaintiff
in the bid to discover whether information denoting protected class
membership has been made available to an employer.
CONCLUSION
Proponents of algorithms have favorably likened its workings to that of an
oracle. For those adherents, the algorithm is all-knowing and will infallibly
provide the answers the intrepid inquirer seeks. This represents a simplistic
understanding of the opaque nature of an oracle. Consider the ur-Oracle, the
Oracle of Delphi.360 The Oracle, a figure known in Greek mythology, spoke
veraciously, but in truth that was spun in riddle and with many strands of
interpretation.361 In the most famous tale of the Oracle, the King of Lydia—
who faced a war against the Persians—asked for the Oracle’s advice. However,
the King failed to fully interrogate the Oracle and did so at his own peril,
departing with a seemingly simple answer that “if he went to war, a great
empire would surely fall.”362 Of course, this advice was highly vulnerable to
misinterpretation, and the King’s own empire later fell to the Persians.363
Similarly, algorithms deployed in the decision-making process are vulnerable
to misinterpretation and misuse. Although automated hiring platforms offer
efficiency to the hiring process, we must continue to interrogate their results
to ensure they are working in furtherance of the shared goal of an equal
opportunity society.
360 See WILLIAM BROAD, THE ORACLE: ANCIENT DELPHI AND THE SCIENCE BEHIND ITS
LOST SECRETS (2006).
361 See Mark Cartwright, Delphi, ANCIENT HIST. ENCYCLOPEDIA (Feb. 23, 2013),
https://www.ancient.eu/delphi/.
362 See id.
363 See id.

The Paradox of Automation As Anti-Bias Intervention

Uploaded by

Copyright:

Available Formats

The Paradox of Automation As Anti-Bias Intervention

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

The Paradox of Automation As Anti-Bias Intervention

Uploaded by

Copyright:

Available Formats

DRAFT IN PROGRESS. Please do not circulate.