Implicit Bias: Scientific Foundations

Author(s): Anthony G. Greenwald and Linda Hamilton Krieger

Source: California Law Review , Jul., 2006, Vol. 94, No. 4 (Jul., 2006), pp. 945-967
Published by: California Law Review, Inc.

Stable URL: https://www.jstor.org/stable/20439056

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

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

is collaborating with JSTOR to digitize, preserve and extend access to California Law Review

This content downloaded from

103.133.254.22 on Sun, 11 Feb 2024 10:47:39 +00:00
All use subject to https://about.jstor.org/terms
 California Law Review
VOL. 94 JULY 2006 No. 4
Copyright ( 2006 by Califomnia Law R

Implicit Bias: Scien

Anthony G. Greenwa
Linda Hamilton Krieg

The assumption that human behav

trol has taken a theoretical battering
in some ways resembles the previous
are important differences between th
mechanisms were embedded in a th
support among scientists, despite man
the middle third of the twentieth ce
gists have abandoned Freud's psychoan
processes.
Theoretical conceptions of conscious control over human behavior
were strongly re-established in the last third of the twentieth century, but
the dominance of such views has been crumbling during the past two dec
ades. Unlike the Freudian revolution, however, the new science of

Copyright C 2006 California Law Review, Inc. California Law Review, Inc. (CLR) is a
California nonprofit corporation. CLR and the authors are solely responsible for the content of their
Dublications.
t Professor of Psychology, University of Washington.
tt Professor of Law, University of California, Berkeley School of Law (Boalt Hall). Thanks to
Jennifer Eberhardt, Jerry Kang, Tom Newkirk, and Jeff Rachlinski for comments on preliminary
versions of this article, and to Ian Ayres, Gary Blasi, Jack Dovidio, John Jost, and Mahzarin Banaji for
useful discussions that preceded the writing.
1. See Matthew Hugh Erdelyi, Psychoanalysis: Freud's Cognitive Psychology (1985);
Matthew Hugh Erdelyi & Benjamin Goldberg, Let's Not Sweep Repression Under the Rug: Toward a
Cognitive Psychology of Repression, in Functional Disorders of Memory 355 (John F. Kihlstrom
& Frederick J. Evans eds., 1979); Anthony G. Greenwald, New Look 3: Unconscious Cognition
Reclaimed, 41 Am. Psychol. 766 (1992); John F. Kihlstrom, The Psychological Unconscious, in
Handbook of Personality: Theory and Research 445 (Lawrence A. Pervin ed., 1990); Howard
Shevrin & Scott Dickman, The Psychological Unconscious: A Necessary Assumption for All
Psychological Theory?, 35 Am. Psychol. 421 (1980).

945

This content downloaded from

103.133.254.22 on Sun, 11 Feb 2024 10:47:39 +00:00
All use subject to https://about.jstor.org/terms
946 CALIFORNIA LA WREVIEW [Vol. 94:945

unconscious mental process is not the product of a single brilliant th

cal mind. Rather, it is being constructed from an evolving, accum
body of reproducible research findings.2
This Article introduces implicit bias-an aspect of the new scie
unconscious mental processes that has substantial bearing on dis
tion law. Theories of implicit bias contrast with the "naive" psychol
conception of social behavior,3 which views human actors as bein
solely by their explicit beliefs and their conscious intentions to act.
is explicit if it is consciously endorsed. An intention to act is consci
the actor is aware of taking an action for a particular reason. Of
actors may dissemble and deny they are taking an action for a pa
reason, so conscious intentions based on explicit beliefs may be
verify. But a deceitful actor is nevertheless capable of asserting the
or identifying the intention that provides the basis for action, even
unwilling to do so.4 In contrast, the science of implicit cognition su
that actors do not always have conscious, intentional control over th
esses of social perception, impression formation, and judgment that
vate their actions.

2. The early stages of this modern revolution are reviewed by Greenwald, supra note
and Wilson's expos? of the inadequacies of introspective explanations of behavior was a
starting point of the modern revolution, leading to widespread understanding that the se
measures of conscious mental process that were widely used in psychological research w
suspect. See Richard. E. Nisbett & Timothy DeCamp Wilson, Telling More Than
Know: Verbal Reports on Mental Processes, 84 Psychol. Rev. 231 (1977). Wegner's an
more recent works reveal the frequency with which seemingly ordinary voluntary actions are
in ways that evade conscious scrutiny, further undermining the idea that a conscious mi
effective governor of most human behavior. See generally Daniel M. Wegner, The Ill
Conscious Will (2002); John A. Bargh et al., The Automated Will: Nonconscious Activa
Pursuit of Behavioral Goals, 81 J. Personality & Soc. Psychol. 1014 (2001).
3. "Naive psychology" refers to laypersons' intuitions about determinants and consequ
human thought and behavior, especially their own. Modern treatments were largely inspir
Heider's book, The Psychology of Interpersonal Relations, which initiated systematic invest
how laypersons' intuitions differ from scientific understanding. Fritz Heider, The Psy
Interpersonal Relations (1958).
4. Methodological investigations by social psychologists in the 1960s revealed social inf
operating within research and interview settings that would lead people to describe their expli
inaccurately in experimental studies. See Martin T. Orne, On the Social Psycholo
Psychological Experiment: With Particular Reference to Demand Characteristics a
Implications, 17 Am. Psychol. 776 (1962); Milton J. Rosenberg, The Conditions and Cons
Evaluation Apprehension, in Artifact in Behavioral Research 279 (Robert Rosenthal &
Rosnow eds., 1969); Stephen J. Weber & Thomas D. Cook, Subject Effects in Laboratory Re
Examination of Subject Roles, Demand Characteristics, and Valid Inference, 11 Psychol
(1972). Work inspired by Festinger's cognitive dissonance theory initiated modern in
understanding people's inability to identify the causes of their own thought and behavior
Festinger, A Theory of Cognitive Dissonance (1957). Nisbett and Wilson's article su
the humbling implications of the ensuing two decades of research. See Nisbett & Wilson, supra

This content downloaded from

103.133.254.22 on Sun, 11 Feb 2024 10:47:39 +00:00
All use subject to https://about.jstor.org/terms
 2006] IMPLICIT BIAS: SCIENTIFIC FOUNDATIONS 947

I
IMPLICIT COGNITION
Many mental processes function implicitly, or outside conscious at
tentional focus.5 These processes include implicit memory,6 implicit per
ception,7 implicit attitudes,8 implicit stereotypes,9 implicit self-esteem,10
and implicit self-concept.11 The meaning of implicit in these phrases is
technical, but still reasonably close to its everyday meaning. For example,
research on "implicit memory" demonstrates that even when a person can
not voluntarily ("explicitly") retrieve a memory, that person's behavior
may reveal that some previous experience has left a memory record. In
such situations, the memory is said to be expressed implicitly, and not ex
plicitly, in the behavior. For example, on the first day of one implicit
memory experiment,12 subjects were asked to pronounce each of a long list
of people's names. Some of these names were recognizably famous, while
others were not. On Day Two, these same subjects judged whether each
name on another long list was famous or not. Half of Day Two's non
famous names were repeated from Day One. The result: On Day Two,
more of the repeated non-famous names than the novel ones were judged
famous. These "false fame" judgments comprise an implicit-memory ef
fect. The names acquired some familiarity from Day One's attended-but
not-studied pronunciation even though, by Day Two, the subject often did
not consciously remember the initial exposure on Day One. This perhaps
vague feeling of familiarity for repeated names was sometimes misattrib
uted to fame, leading to greater false judgments of fame for the repeated
than the non-repeated names. Subjects presumably go through a mental

5. For an overview of implicit social cognition, which encompasses the phenomena of implicit
attitudes, stereotypes, self-esteem, and self-concept, see Anthony G. Greenwald & Mahzarin R. Banaji,
Implicit Social Cognition: Attitudes, Self-Esteem, and Stereotypes, 102 Psychol. Rev. 4 (1995); see
also Anthony G. Greenwald et al., A Unified Theory of Implicit Attitudes, Stereotypes, Self-Esteem, and
Self-Concept, 109 Psychol. Rev 3 (2002).
6. See Daniel L. Schacter, Implicit Memory: History and Current Status, 13 J. Experimental
Psychol.: Learning, Memory, & Cognition 501 (1987); Larry L. Jacoby & Mark Dallas, On the
Relationship Between Autobiographical Memory and Perceptual Learning, 110 J. Experimental
Psychol: general 306 (1981).
7. See John F. Kihlstrom et al., Implicit Perception, in Perception Without
Awareness: Cognitive, Clinical, and Social Perspectives 17 (Robert F. Bornstein & Thane S.
Pittman eds., 1992).
8. See Anthony G. Greenwald et al., Measuring Individual Differences in Implicit
Cognition: The Implicit Association Test, 74 J. Personality & Soc. Psychol 1464 (1998).
9. See Laurie A. Rudman et al., Implicit Self-Concept and Evaluative Implicit Gender
Stereotypes: Self and Ingroup Share Desirable Traits, 27 Personality & Soc. Psychol. Bull. 1164
(2001).
10. See Anthony G. Greenwald & Shelly D. Farnham, Using the Implicit Association Test to
Measure Self-Esteem and Self-Concept, 79 J. Personality & Soc. Psychol 1022 (2000).
11. See Rudman et al., supra note 9.
12. See Larry L. Jacoby et al., Becoming Famous Overnight: Limits on the Ability to Avoid
Unconscious Influences of the Past, 56 J. Personality & Soc. Psychol. 326 (1989).

This content downloaded from

103.133.254.22 on Sun, 11 Feb 2024 10:47:39 +00:00
All use subject to https://about.jstor.org/terms
948 CALIFORNIA LA WREVIEW [Vol. 94:945

process resembling the following: "This name seems familiar

familiar? Perhaps it's famous." For names that subjects explic
bered seeing and pronouncing on Day One, subjects correctly
why the name seemed familiar. Therefore they did not mistaken
the familiarity to fame.

II
IMPLICIT ATTITUDES AND IMPLICIT STEREOTYPES
Implicit-memory research conducted in the 1980s led rese
develop measures of other implicit mental phenomena. Two
implicit attitudes and implicit stereotypes-are especially rele
and discrimination.

A. Implicit Attitudes
Social psychologists define an attitude as an evaluative d
that is, the tendency to like or dislike, or to act favorably or u
toward, someone or something. Explicit expressions of attitudes
quently, whenever we say we like or dislike someone or som
statement that one likes a particular presidential candidate prov
example. Attitudes can also be expressed through favorable or un
action, such as by voting for or against a particular presidential
If the voter understands that the favorable vote results from f
liefs about the candidate, the vote is an explicit attitude expressio
In other situations, a vote might function as an implicit attit
tor-that is, an action that indicates favor or disfavor toward so
but is not understood by the actor as expressing that attitude."3
ple, a voter may vote for a particular candidate even thoug
knows nothing other than the candidate's name. One of the
might influence a voter to vote for this candidate is that the
name shares one or more initial letters with the voter's nam
case, the vote can be understood, at least in part, as an implicit
of the voter's self-favorable attitude.14
As an additional, hypothetical example, consider how pe
impressions of a liked or disliked candidate's spouse, child,
Someone who knows nothing about the candidate's relative other
relative's relation to the candidate may find that they like or
relative. Not surprisingly, this attitude toward the relative is lik
the attitude toward the candidate. Evaluation of the unknown r
therefore be regarded as an implicit indicator of attitude t

13. Greenwald and Banaji define implicit attitudes as "introspectively uni

inaccurately identified) traces of past experience that mediate favorable or unfavorable f
or action toward social objects." Greenwald & Banaji, supra note 5, at 8.
14. See id. at 12.

This content downloaded from

103.133.254.22 on Sun, 11 Feb 2024 10:47:39 +00:00
All use subject to https://about.jstor.org/terms
2006] IMPLICIT BIAS: SCIENTIFIC FOUNDATIONS 949

candidate. Here, the "implicit" designation indicates that the attitude ex

pressed toward the candidate determined the attitude toward the relative,
even though the liking or disliking for the relative may be experienced as
an independent attitude.
Implicit attitudes are of greatest interest when implicit and explicit
attitudes toward the same object differ. These discrepancies are referred to
as dissociations between implicit and explicit attitudes. Dissociations are
commonly observed in attitudes toward stigmatized groups, including
groups defined by race, age, ethnicity, disability, and sexual orientation.
Researchers have used the Implicit Association Test (IAT), a procedure
described below, to reveal such attitudinal dissociations.15

B. Implicit Stereotypes
A social stereotype is a mental association between a social group or
category and a trait. The association may reflect a statistical reality, but it
need not. If the association does reflect a statistical reality, members of the
group will be more likely to display the trait than will members of other
groups. A perfect or near-perfect correlation, which might be a defining
trait-such as physical stamina for basketball players-is of little psycho
logical interest and is often not even considered part of a stereotype. It is of
greater psychological interest when the correlation between group mem
bership and trait expression is much less than perfect, but the trait never
theless distinguishes members of one group from others. Suppose,
hypothetically, that 10-15% of people over the age of seventy drive on
highways at speeds noticeably below speed limits, but that only 5% of
younger people drive this slowly. If these figures were accurate, they
would reflect a genuine association between age and driving behavior.
However, the stereotype would apply to only a small minority (10-15%) of
elderly people. Nonetheless, it may come to serve as a default assump
tion-the assumption that any elderly person is likely to drive slowly.
The first experimental demonstration of implicit stereotypes made use
of the stereotype that associates male gender with fame-deserving
achievement.16 In this experiment, which was based on Jacoby et al.'s
false-fame implicit-memory effect described in Part I,17 Banaji and
Greenwald found that the false-fame effect was substantial when the pro
nounced (i.e., attended-but-not-studied) names were male, but was

15. See Wilhelm Hofmann et al., A Meta-Analysis on the Correlation Between the Implicit
Association Test and Explicit Self-Report Measures, 31 Personality & Soc. Psychol. Bull. 1369
(2005) (discussing factors that promote dissociation of implicit from explicit attitudes); Brian A.
Nosek, Moderators of the Relationship Between Implicit and Explicit Evaluation, 134 J.
Experimental Psychol.: General 565 (2005).
16. See Mahzarin R. Banaji & Anthony G. Greenwald, Implicit Gender Stereotyping in
Judgments of Fame, 68 J. Personality & Soc. Psychol. 181 (1995).
17. See Jacoby et al., supra note 12.

This content downloaded from

103.133.254.22 on Sun, 11 Feb 2024 10:47:39 +00:00
All use subject to https://about.jstor.org/terms
 950 CALIFORNIA LA WREVIEW [Vol. 94:945

noticeably weaker when the names were female. Banaji and G

described this result as an implicit indicator of the stereotype th
ates maleness with fame-deserving achievement."8 Put more tech
implicit stereotype of this kind can be defined as "the introspect
dentified (or inaccurately identified) traces of past experience tha
attributions of qualities to members of a social category. "19
Stereotypes can involve associations of either favorable or
able traits with a group. Because the favorable-unfavorable dis
also central to the concept of attitude, it is natural to ask how st
and attitudes differ. For stereotypes, the content of the ascribed t
than its evaluative valence, is central. For attitudes, the opposite
example, in the implicit-fame experiment, it was the trait of fam
than the positivity of fame, that defined the implicit-stereotype
non.

III
RESPONSE BIAS AND IMPLICIT BIAS
The term "bias," sometimes referred to as "response bias," denotes a
displacement of people's responses along a continuum of possible judg
ments. Response bias need not indicate something unwise, inappropriate,
or even inaccurate. For example, instructors may vary in their response
bias in grading, such that some assign a relatively high grade to average
student performance while others assign a lower grade to the same per
formance. Instructors who differ in response bias on the grading dimension
may nevertheless be equally sensitive to differences among students. Con
sider an instructor who is biased to grade leniently and assigns grades ex
clusively between A (highest) and C (lowest). This instructor's grades may
be perfectly correlated with those of a severe-grading instructor who limits
grades to the B-to-D range. If these two instructors graded the same work,
each of the lenient instructor's grades would be exactly one letter grade
above those of the more severe instructor. Unless there are established
standards that associate specific performances with specific grades, one
could not accuse either instructor of being less "accurate" than the other.
A more widely recognized form of bias does affect response accuracy
and bears a pejorative connotation. Imagine a particular instructor who dif
ferentially assigns grades to two identically performing students when one
student is male and the other is female, or when one is White and the other
is Black. In this case, the fairness and accuracy of judgments are both
compromised. Attitudes and stereotypes are plausible causes of such dis
criminatory biases. If, among equally qualified job applicants, one favors
members of one race over those of another, this plausibly reflects an

18. Banaji & Greenwald, supra note 16, at 186-87.

19. See Greenwald & Banaji, supra note 5, at 15.

This content downloaded from

103.133.254.22 on Sun, 11 Feb 2024 10:47:39 +00:00
All use subject to https://about.jstor.org/terms
 2006] IMPLICIT BIAS: SCIENTIFIC FOUNDATIONS 951

attitudinal bias: one may have a more favorable attitude toward one race
group than toward the other. If, among equally qualified renters, one as
sumes that members of one race will be more conscientious in paying rent
than those of another, this may be a bias rooted in stereotype. If, among
equally qualified candidates for a management position, men are consid
ered preferable to women, it could be due to operation of a stereotype that
treats leadership as a trait more frequently found among men than
women.20
Implicit biases are discriminatory biases based on implicit attitudes or
implicit stereotypes. Implicit biases are especially intriguing, and also es
pecially problematic, because they can produce behavior that diverges from
a person's avowed or endorsed beliefs or principles. The very existence of
implicit bias poses a challenge to legal theory and practice, because dis
crimination doctrine is premised on the assumption that, barring insanity or
mental incompetence, human actors are guided by their avowed (explicit)
beliefs, attitudes, and intentions.21
Biases can be either favorable or unfavorable. Ingroup bias designates
favoritism toward groups to which one belongs. There is a widespread in
tuition that it is often acceptable to be biased in favor of at least some of
the groups to which one belongs. In this view, bias is a problem only when
it is directed against some group. Thus it may be considered acceptable to
be biased in favor of one's siblings, children, schoolmates, and friends.
Interestingly, the intuition that biases in favor of one's smaller in
groups (such as family and friends) are acceptable typically does not ex
tend to believing that biases favoring one's larger ingroups (one's race,
sex, ethnicity, religion, or age group) are appropriate. Is there a boundary
encompassing ingroups toward which favorable biases can be considered
acceptable? The illegality of some kinds of biased behavior toward certain
groups (regardless of one's membership)-such as those defined by race,
sex, ethnicity, religion, and age-provides a non-psychological boundary.
Psychologically, the small size of some ingroups is no doubt significant
because many people feel more obliged to help others when they are one of
only a few people who can possibly be helpful,22 as may often be the case
for family members.

20. Discriminatory biases are plausibly stereotype-based when they oppose the bias that might be
expected as an attitude effect. For example, gender biases that discriminate against women are
plausibly stereotype-based, given that research has found that attitudes toward women are often more
favorable than attitudes toward men. See Alice H. Eagly & Antonio Mladinic, Gender Stereotypes and
Attitudes Toward Women and Men, 15 Personality & Soc Psychol. Bull. 543, 551-55 (1989).
21. See generally Linda Hamilton Krieger & Susan T. Fiske, Behavioral Realism in Employment
Discrimination Law: Implicit Bias and Disparate Treatment, 94 Calif. L. Rev. 997 (2006).
22. This psychological truth was demonstrated very clearly by Darley and Latan?, who found that
a solitary witness to a simulated epileptic seizure was considerably more likely to intervene than was
one of a group of such witnesses. See John M. Darley & Bibb Latan?, Bystander Intervention in
Emergencies: Diffusion of Responsibility, 8 J. Personality & Soc. Psychol. 377 (1968). The effect

This content downloaded from

103.133.254.22 on Sun, 11 Feb 2024 10:47:39 +00:00
All use subject to https://about.jstor.org/terms
952 CALIFORNIA LA WREVIEW [Vol. 94:945

Perhaps fortunately, the situations in which people wish

in favor of their smaller, important ingroups-such as in pro
their own children-are often those for which no question
crimination against others arises. Nevertheless, a positive
any ingroup necessarily implies a relative negativity toward
tary outgroup. In some circumstances, this relative favoring
puts members of other groups at a discriminatory disadv
one allows favoritism toward a family member or friend to
ing, job assignment, rental, or admissions decision.

IV
THE IMPLICIT ASSOCIATION TEST
The recent development of the Implicit Association Test (IAT) has
accelerated research on implicit bias. The IAT's general method can be
adapted to measure a wide variety of the group-valence and group-trait as
sociations that underlie attitudes and stereotypes. The IAT is an implicit
measure because it infers group-valence and group-trait associations from
performances that are influenced by those associations in a manner that is
not discerned by respondents.23
The most widely used IAT measure assesses implicit attitudes toward
African Americans (AA) relative to European Americans (EA).24 In this
"Race IAT," respondents first practice distinguishing AA from EA faces
by responding to faces from one of these two categories with the press of a
computer key on the left side of the keyboard and to those of the other
category with a key on the right side of the keyboard. Respondents next
practice distinguishing pleasant-meaning from unpleasant-meaning words
in a similar manner. The next two tasks, given in a randomly determined
order, use all four categories (AA faces, EA faces, pleasant-meaning
words, and unpleasant-meaning words). In one of these two tasks, the IAT
calls for one response (say, pressing a left-side key) when the respondent
sees AA faces or pleasant words, whereas EA faces and unpleasant words
call for the other response (right-side key). In the remaining task, EA faces

of being in a unique position to help is so strong that the presence of multiple bystanders can result in
less likelihood of any help being given than when only a single bystander is present. Id.
23. The I AT was first reported by Greenwald, McGhee, and Schwartz in 1998. See Greenwald et
al., supra note 8. Although other implicit measures have been developed and have been used
extensively in research, see Russell H. Fazio & Michael A. Olson, Implicit Measures in Social
Cognition Research: Their Meaning and Use, 54 Ann. Rev. Psychol. 297 (2003), the IAT that has
been used most widely, and this Article focuses on it. The statement that respondents do not discern the
influence of associations on their IAT performance is properly limited to respondents who have not
become aware of the way in which the procedure assesses association strengths.
24. The Race IAT uses these formal race category labels, instead of ?Black and White,?
because the color-name labels carry associative connotations of good and bad that are unrelated to race,
and these connotations might interfere with the measurement of race-valence associations.

This content downloaded from

103.133.254.22 on Sun, 11 Feb 2024 10:47:39 +00:00
All use subject to https://about.jstor.org/terms
2006] IMPLICIT BIAS: SCIENTIFIC FOUNDATIONS 953

share a response with pleasant words and AA faces with unpleasan

words.25
The implicit-attitude measure produced by this IAT is based on rela
tive speeds of responding in the two four-category tasks. This measure al
lows an inference about attitudes (category-valence associations) because it
is easier to give the same response to items from two categories when those
two categories are cognitively associated with each other. For American
respondents taking the Race IAT, response speeds are often faster when
EA, rather than AA, is paired with pleasant words.26 This frequently ob
served pattern supports the interpretation that EA-pleasant is a stronger
association than AA-pleasant. Researchers have described this result as
showing implicit attitudinal preference for EA relative to AA.27
Research comparing IAT (implicit) measures with parallel survey
type self-report (explicit) measures has found systematic variations in the
agreement between these two types of measures. There is substantially
greater agreement between the two types of measures when implicit and
explicit attitudes have been shaped by the same experiences, which is
likely to be the case for attitudes toward consumer brands, sports teams,
and political candidates.28 When implicit and explicit measures of attitudes
or stereotypes disagree-for example, when a Race IAT shows preference
for EA and a self-report measure shows impartiality-there is said to exist
a dissociation between the two.

V
PREDICTIVE VALIDITY OF THE IAT
Researchers have extended the IAT into increasingly diverse d
applying its general method to a wide variety of groups

25. Various nonessential aspects of the IAT procedure, such as the hand assigned to
category and order of performing the two four-category tasks, are randomized or cou
avoid their systematically influencing findings.
26. Brian A. Nosek et al., Harvesting Implicit Group Attitudes and Beliefs from a D
Web Site, 6 Group Dynamics: Theory, Research, and Practice 101,105 (2002) (repor
from a dataset with N = 192,364).
27. Because each task involves two associations, the complete description of this in
association strengths is that the combined strength of the EA-pleasant and AA-unpleas
is stronger than the combined strength the of AA-pleasant and EA-unpleasant ass
association-strength interpretation of the IAT has been widely, although not universally,
recent discussion of alternative interpretations, see Brian A. Nosek, Anthony G
Mahzarin R. Banaji, The Implicit Association Test at Age 7: A Methodological and Conc
in Automatic Processes in Social Thinking and Behavior (John A. Bargh ed.,
2006). We consider one of these alternative interpretations?that the IAT measures cult
infra Part V.
28. See Hofmann et al., supra note 15; Nosek, supra note 15; Anthony G. Green
Understanding and Using the Implicit Association Test: I An Improved Scoring Alg
Personality & Soc. Psychol. 197 (2003).

This content downloaded from

103.133.254.22 on Sun, 11 Feb 2024 10:47:39 +00:00
All use subject to https://about.jstor.org/terms
 954 CALIFORNIA LA WREVIEW [Vol. 94:945

categories.29 Perhaps because of the importance of attitude as a

construct in psychology, more attention, thus far, has been give
gating implicit attitudes than to investigating implicit stereoty
ognition of the importance of understanding relations b
measures and behavior, many studies that have used an IAT at
ure have also included a measure of one or more social behav
theoretically expected to be related to attitude or stereotype m
examined behaviors have ranged widely, including (as just a
ples) voting for political candidates, measures of warmth and di
interracial interactions, and measures of brain activity when
ages of members of a racial group. Analyses of the data the
whether individual differences in implicit attitudes or stereotyp
by the IAT correlate with (i.e., are predictive of) individual diff
behavior.
A recent meta-analytic review by Poehlman, Uhlmann, Greenwald,
and Banaji identified and summarized sixty-one such studies.30 This re
view's method appraised the value of IAT measures by assessing the rele
vant body of research in the aggregate, rather than as isolated research
findings. To do this, the researchers averaged the available correlational
measures of relations between IAT measures and behaviors of interest
within groups of studies that tested related hypotheses, as well as over the
entire group of eighty-six independent findings from the sixty-one studies.
For comparison, parallel analyses examined the aggregated correlations of
the same behavioral measures with self-report (explicit) measures, which
most of the studies had also obtained.
Both the implicit (IAT) and the parallel explicit measures displayed
predictive validity, meaning that both types of measures, on average, were
significantly correlated with measures of behavior, as expected. To be
clear, this does not mean that statistically significant correlations were
found in all studies, but that averages of the correlational results of the col
lected similar tests clearly showed the expected relationships. Predictive
validity was greater (meaning that the average correlation was larger) for
self-report (explicit) measures than for IAT measures. However, within the
critical group of studies that focused on prejudicial attitudes and stereo
types-in other words, within the studies of implicit bias-predictive va
lidity was significantly greater for the IA T measures.
Importantly, implicit measures of bias have relatively greater predic
tive validity than explicit measures in situations that are socially sensitive,
like racial interactions, where impression-management processes might

29. See Nosek, supra note 15.

30. See T. Andrew Poehlman et al., Understanding and Using the Implicit Association
Test: III. Meta-Analysis of Predictive Validity (2005) (unpublished manuscript, on file with
authors).

This content downloaded from

103.133.254.22 on Sun, 11 Feb 2024 10:47:39 +00:00
All use subject to https://about.jstor.org/terms
2006] IMPLICIT BIAS: SCIENTIFIC FOUNDATIONS 955

inhibit people from expressing negative attitudes or unattractive stereo

types. Additionally, implicit measures have relatively greater validity in
predicting spontaneous behaviors such as eye contact, seating distance, and
other such actions that communicate social warmth or discomfort.

VI
How PERVASIVE IS IMPLICIT BIAS?
Since 1998, IAT measures of implicit attitudes have been available on
the Internet for self-administered demonstration use.3" These web
accessible demonstrations, which allow users to interactively experience
the IAT, have accumulated sufficient data to allow researchers to draw
conclusions about the pervasiveness of implicit and explicit biases.32
Table 1 displays results for a dozen data sets, comparing the degree of
favoritism toward advantaged versus disadvantaged groups revealed by
implicit versus explicit measures. Two differences between the implicit and
explicit measures are readily apparent in these data. First, the explicit
measures generally show much greater evidence for attitudinal impartiality
or neutrality. Averaged across the dozen topics, 42% of respondents ex
pressed exact or near-exact neutrality on explicit measures. On the IAT
measures, however, only 18% of respondents demonstrated sufficiently
small implicit bias to be judged implicitly neutral. Second, the IAT meas
ures consistently revealed greater bias in favor of the relatively advantaged
group (averaging almost three-quarters of respondents across all the topics)
than did the explicit measures (for which an average of slightly over one
third of respondents showed bias favoring advantaged groups).
Table 1 also shows a bias index, computed as the percentage of re
spondents showing favorability to the advantaged group minus the percent
age showing favorability to the disadvantaged group. Whereas this index
averaged only 20% for explicit measures, it averaged the very large value
of 64% for IAT measures. The broad generalization justified by the data in
Table 1 is that implicit attitude measures reveal far more bias favoring ad
vantaged groups than do explicit measures.
It is important to note that these data came from voluntary visitors to
the IAT website-a self-selected sample, which is different from a repre
sentative sample that can be obtained by selecting and recruiting respon
dents randomly from a defined population. As a result, the data in Table 1
cannot be interpreted as representing the attitude distribution of some spe
cific population of interest, such as adult residents of the United States.

31. Interactive demonstrations of more than a dozen versions of the IAT are available at
https://implicit.harvard.edu.
32. These demonstration tests were not set up to conduct research but were nevertheless obliged
to record data to enable computations of results that were reported to web visitors. The accumulated
data provided by the site's many visitors has proved to be a remarkably rich archive.

This content downloaded from

103.133.254.22 on Sun, 11 Feb 2024 10:47:39 +00:00
All use subject to https://about.jstor.org/terms
956 CALIFORNIA LA WREVIEW [Vol. 94:945

Even so, the greater favoritism to advantaged groups found in IAT

ures than in explicit measures would almost certainly be found with r
sentative samples. Strong evidence for this assertion comes
examination of the Race IAT data for the wide-ranging demograph
groups shown in Table 2.
Table 2 shows that, with one notable exception, the percentage
spondents who display implicit race bias varies relatively little a
groups categorized by varied age, sex, and educational attainment. Afr
Americans constitute the only subgroup of respondents who do no
substantial implicit pro-EA race bias on the Race IAT. Approxim
equal percentages of African Americans displayed implicit bias in the
AA and pro-EA directions. Significantly, among African American
results showed considerably greater favoritism to the dominant Euro
American group than did the results from self-report measures,
showed very strong favoritism toward African Americans. The r
shown in Table 2 strongly suggest that any non-African America
group of the United States population will reveal high proportions of
sons showing statistically noticeable implicit race bias in favor o
relative to AA.

This content downloaded from

103.133.254.22 on Sun, 11 Feb 2024 10:47:39 +00:00
All use subject to https://about.jstor.org/terms
2006] IMPLICIT BIAS: SCIENTIFIC FOUNDATIONS 957
Table 1. Distributions of Responding on Self-report (Explicit) and IAT (Implicit) Measures

% biased toward disadvantaged (dis) and advantaged (adv)

Disadvantaged Advantaged N groups, and % neutral (neu)
group group Self-report (Explicit) IAT (Implicit)

dis |Neu adv index dis neu adv index

IAT Demonstration Web Site Tests

Afr. American Eur. Ameri- 22074 11.3%

_________
154.6% 134.0%1
can4 44
23%
J
10.6% 24.1% 165.3% 5

Old Young 11528 16.7% 36.8% 46.6% 30% 4.7% 14.3% 81.1% 76%
TAT Research Web Site Tests

Afr. American can

Eur.2Ameri- 211 11.8% 56.4% 31.9% 20% 12.3% 18.5% 69.2% 57%

Asians Whites 211 16.4% 56.9% 26.7% 10% 11.3% 25.9% 62.8% 51%

Canadian American 218 24.1% 39.5% 36.4% 12% 13.3% 21.7% 65.0% 52%

Foreign places Amercan 178 20.9% 36.6% 42.4% 22% 9.6% 14.0% 76.4% 67%
places

Gay people Straight

people
217 14.3% 45.7% 40.0% 26% 8.3% 22.9% 68.8% 60%

Muslims Jews 144 10.4% 49.3% 40.3% 30% 11.1% 20.7% 68.2% 57%

Old people Young 236 27.4% 39.2% 33.5% 6% 5.5% 15.6% 78.9% 73%
___________ people

Poor Rich 211 36.7% 37.6% 25.7% -11% 1.4% 4.3% 94.3% 93%

Fat people Thin people 239 13.4% 42.4% 44.2% 31% 13.1% 20.8% 66.1% 53%

Japan USA 263 19.9% 19.9% 60.2% 40% 6.2% 15.2% 78.7% 73%

AVERAGES (12 data sets, 19.5% 42.4% 38.1% 20% 9.2% 18.0% 72.8% 64%
unweighted)I I I I II I
The implicit data reported in this table were obtained from IAT measure
ant and unpleasant words were classified together with the items represent
table. The explicit data listed in this table were obtained from self-report
tudes. The "index" column reports a bias index for each topic, computed as
advantaged group minus the percentage favoring the disadvantaged group.
index, the more pervasive is the bias. The bias index's values for IAT me
higher values than for the self-report measures, indicating that implicit bi
explicit bias. The race and age data from the IAT demonstration website ar
The data from the IAT research website were reported by Nosek, supra no

This content downloaded from

103.133.254.22 on Sun, 11 Feb 2024 10:47:39 +00:00
All use subject to https://about.jstor.org/terms
958 CALIFORNIA LA WREVIEW [Vol. 94:945

Table 2. Percentages Favoring European American (EA) Relative to African American

report (Explicit) and IAT (Implicit) Measures

Self-report (Explicit) IAT (Implicit)

Subcategories N Percent favoring Percent favoring
AA |neither l EA lIndex |AA |neither |EA |Index

Education Level

thru high school 3869 9.9% 57.9% 32.2% 22% 9.8% 26.2% 64.0% 54%
grad__ _ _ __ _ _ _ _ _ _ _ _

at least some college 13028 11.3% 54.1% 34.

at least some grad 3829 12.5% 53.5% 34.

school

Race and Ethnicity

Black (inc. multira- 2277 58.9% 36.2% 4.8% -54% 34.1% 33.6% 32.4% -2%
cial)__ _ _ _ _ _ _ _ _ _ _

Hispanic (not Black) 1204 15.0% 59.7% 25.3% 10% 10.2% 29.2% 60.5% 50%

Asian & Pacific 1080 9.6% 57.5% 32.9% 23% 7.7% 24.8% 67.5% 60%
Islander

White 14805 3.4% 56.0% 40.7% 37% 6.8% 21.7% 71.5% 65%

Age
under 25 13823 9.7% 55.7% 34.5% 25% 9.4% 23.7% 66.9% 58%

25-44 5403 14.9% 53.9% 31.2% 16% 12.8% 24.4% 62.8% 50%

45 and older 1743 12.3% 47.1% 40.6% 28% 12.6% 25.6% 61.8% 49%

Sex

Female 13060 12.3% 57.8% 29.8% 17% 11.4% 25.2% 63.4% 52%

Male 7971 9.6% 49.4% 41.0% 31% 9.2% 22.2% 68.6% 59%

Political Ideology = = = =
Conservative 3053 4.8% 44.0% 151.2%I 46% 6.5% 19.9%J 73.6%J 67%

Middle 10612 11.0% 54.0% 35.0% 24% 10.3% 23.8% 65.9% 56%
Liberal 6427 14.8% 59.9% 25.3% 11% 12.9% 26.0% 61.1% 48%

The finding of high levels of the bias index for all demographic subgroups other than Black (i.e., Afri
can American) indicates the pervasiveness of pro-EA bias. Even though the bias index was lower in
groups of Hispanics and political liberals than in other groups, it was still quite high among those
groups.

This content downloaded from

103.133.254.22 on Sun, 11 Feb 2024 10:47:39 +00:00
All use subject to https://about.jstor.org/terms
2006] IMPLICIT BIAS: SCIENTIFIC FOUNDATIONS 959

VII
WHY IS IMPLICIT BIAS SO PERVASIVE?
This question can be divided into three parts: First, how are implicit
attitudes and stereotypes acquired? Second, what mental representations
underlie implicit attitudes and stereotypes? Third, do the representations
underlying implicit attitudes and stereotypes differ from those underlying
explicit attitudes and stereotypes? Answers to these questions could ex
plain both the weak relations observed between IAT and explicit measures
and the substantially greater bias apparent in implicit attitudes than in ex
plicit ones. It may be several years before thorough research-based answers
to these questions are available. These answers will require, in part, re
search that examines the formation of implicit attitudes and stereotypes in
young children. To be used with preschool children, the IAT needs modifi
cations, the most substantial of which is to replace printed-word stimulus
items either with pictures or with spoken words.33
In a recent review article, Rudman wrote, "The hypothesized causal
influences on attitudes include early (even preverbal) experiences, affective
experiences, cultural biases, and cognitive consistency principles. Each
may influence implicit attitudes more than explicit attitudes, underscoring
their conceptual distinction."34 Rudman's proposal that early experiences
and affective experiences may be reflected more in implicit attitudes than
in explicit attitudes may explain why implicit attitudes generally reveal
more bias, as Tables 1 and 2 show. As Rudman also noted, influences of
cultural factors on the IAT can also explain why people often display im
plicit attitudes that appear more concordant with their general cultural mi
lieu than with the experiences of their individual upbringing.35 As an
example, African Americans' implicit racial attitudes, rather than showing
strong ingroup favoritism, are (on average) remarkably close to indicating
racial neutrality.36 This can be seen in Table 2, which also shows that this
pattern for African Americans' implicit attitudes contrasts sharply with
their explicit racial attitudes, which are strongly polarized in the ingroup
favorable (pro-AA) direction. This could indicate that African Americans'
implicit racial attitudes show an influence of the United States's pro
European-American culture. There is no evidence of this influence on Af
rican Americans' explicit attitude responses. The observation of approxi
mate racial neutrality of African Americans' implicit attitudes is especially

33. Research with IAT procedures that have been adapted for use with preschool children is
being actively pursued in the laboratories of Mahzarin R. Banaji and Andrew L. Meltzoff.
34. Laurie A. Rudman, Sources of Implicit Attitudes, 13 Current Directions in Psychol. Sci.
79, 79 (2004).
35. Id. at 80.
36. See Gary Blasi & John T. Jost, System Justification Theory and Research: Implications for
Law, Legal Advocacy, and Social Justice, 94 Calif. Law Rev. 1119, 1136 (2006) (discussing this and
related observations).

This content downloaded from

103.133.254.22 on Sun, 11 Feb 2024 10:47:39 +00:00
All use subject to https://about.jstor.org/terms
960 CALIFORNIA LA WREVIEW [Vol. 94:945

impressive because it is a blatant exception to the general pa

plicit attitudes that reveal more bias than explicit attitudes.
If implicit attitude and stereotype measures are indicator
cial-cognitive content of one's broad cultural environment, t
data indicate that (for as-yet-unclear reasons) explicit measures
and stereotypes do not reflect the social-cognitive content of th
those who provide the measures. If true, this conclusion wou
provide a discouraging assessment of the value of explicit measu
provides a perspective on one of the most reasonable and p
tiques that has been offered of the IAT. The essence of this crit
IAT measures should be interpreted as indicating modal beliefs o
that respondents understand to be generally endorsed by oth
cultural beliefs).37
The view that the IAT provides a measure of one's unders
cultural beliefs external to oneself implies that individual di
IAT measures are indictors merely of differences in the clarity
those external, cultural beliefs are perceived. If that were th
IAT measures should have no more relation to interesting form
behavior than would differences among people in the clarity of
perceptions, such as their perceptions of symbols on an eye char
dicting this expectation, however, meta-analytic evidence for th
validity of IAT measures indicates that IAT measures successf
a variety of types of social behavior.38 Failure to explain thi
validity of the IAT constitutes a notable weakness of the cul
interpretation of IAT measures.39

37. Olson and Fazio describe cultural beliefs as "extrapersonal associations." M

& Russell H. Fazio, Reducing the Influence of Extrapersonal Associations on the Impl
Test: Personalizing the IAT, 86 J. Personality & Soc. Psychol. 653, 653 (2004).
Hilton call such beliefs "environmental association^]." Andrew Karpinski & James L. H
and the Implicit Association Test, 81 J. Personality & Soc. Psychol. 774, 775 (200
Tetlock refer to them as "[sjhared cultural knowledge." Hal R. Arkes & Philip E. Tetlo
of Implicit Prejudice, or "Would Jesse Jackson Fail' the Implicit Association Test
Inquiry 257, 275 (2004).
38. See Poehlman et al., supra note 30; see also supra Part V.
39. Those who regard the IAT as reflecting cultural beliefs rather than implicit a
additional challenge. Their views include these two propositions: (a) the IAT reflects
and (b) the IAT assesses something different from what explicit measures assess. It
from these two propositions that (c) explicit measures do not measure cultural beliefs
endorsed by many, including those who advocate the cultural-beliefs critique
that (d) explicit measures assess views that respondents avow or endorse. Juxtapo
one arrives at the seemingly paradoxical conclusion that people's endorsed beliefs do n
cultural beliefs. It is genuinely puzzling to arrive at this conclusion. What might av
explicit measures assess other than the average levels of beliefs that are dominant i
Proponents of the cultural-belief interpretation of the IAT have not yet addressed th
implications of their interpretation.

This content downloaded from

103.133.254.22 on Sun, 11 Feb 2024 10:47:39 +00:00
All use subject to https://about.jstor.org/terms
2006] IMPLICIT BIAS: SCIENTIFIC FOUNDATIONS 961

VIII
Do IMPLICIT BIASES PRODUCE DISCRIMINATORY BEHAVIOR?
As Parts V and VII described, evidence that implicit attitudes produce
discriminatory behavior is already substantial40 and will continue to accu
mulate. The dominant interpretation of this evidence is that implicit attitu
dinal biases are especially important in influencing nondeliberate or
spontaneous discriminatory behaviors.
A study by McConnell and Leibold,41 which was one of the first ex
perimental investigations to relate an IAT race attitude measure to dis
criminatory behavior, provides a good illustration. In this study, the
behavior of White undergraduate students was videotaped while they were
being interviewed separately by White and Black experimenters.42 These
subjects also completed a race attitude IAT measure. Subjects whose Race
IAT scores indicated strong implicit preference for White relative to Black
hesitated less and made fewer speech errors when speaking to the White
experimenter than to the Black experimenter. They also spoke more to and
smiled more at the White experimenter than the Black experimenter. These
subtle and spontaneous behaviors suggested higher levels of comfort inter
acting with the White experimenters.43
This result of the McConnell and Leibold experiment is especially
important in light of findings that were reported by Word, Zanna, and Coo
per44 well before the IAT was developed. In the first of their two studies,
Word et al. showed that when interviewing both Black and White job ap
plicants, White students showed greater indications of nonverbal discom
fort and spent less time speaking with the Black applicants-two indicators
that McConnell and Leibold had found to be predicted by the Race IAT. In

40. See Poehlman et al., supra note 30.

41. See Allen R. McConnell & Jill M. Leibold, Relations among the Implicit Association Test,
Discriminatory Behavior, and Explicit Measures of Racial Attitudes, 37 J. Experimental Soc.
Psychol. 435 (2001).
42. Subjects did not know in advance about the videotaping, but they received a subsequent
description and explanation, at which time their signed consent to use the videotape as a source of
research data was requested.
43. Other published studies have likewise found correlations of IAT-measured racial associations
with indicators of subtle or spontaneous discriminatory actions. See Leslie Ashburn-Nardo et al., Black
Americans' Implicit Racial Associations and Their Implications for Intergroup Judgment, 21 Soc.
Cognition 61 (2003); Kurt Hugenberg & Galen V. Bodenhausen, Ambiguity in Social
Categorization: The Role of Prejudice and Facial Affect in Race Categorization, 15 Psychol. Sci. 342
(2004); Kurt Hugenberg & Galen V. Bodenhausen, Facing Prejudice: Implicit Prejudice and the
Perception of Facial Threat, 14 Psychol. Sci. 640 (2003); Jennifer A. Richeson et al., An fMRI
Investigation of the Impact of Interracial Contact on Executive Function, 6 Nature Neuroscience
1323 (2003). Several similar unpublished results involving Race IAT measures were included in the
Poehlman et al. meta-analytic review, along with similar results from studies of implicit biases toward
other stigmatized groups, such as Germans' implicit attitudes toward Turks. See Poehlman et al., supra
note 30.
44. See Carl O. Word et al., The Nonverbal Mediation of Self-Fulfilling Prophecies in Interracial
Interaction, 10 J. Experimental Soc. Psychol. 109 (1974).

This content downloaded from

103.133.254.22 on Sun, 11 Feb 2024 10:47:39 +00:00
All use subject to https://about.jstor.org/terms
962 CALIFORNIA LA W REVIEWr [Vol. 94:945

Word et al.'s second study, White interviewers were carefully tr

control these same subtle aspects of their behavior in their interactio
White job applicants who were unaware of the interviewer's tr
White interviewees who encountered these trained interviewers perf
worse in the interview and were more uncomfortable and distant in
interaction style. Such interviewees also judged their interviewer to
friendly. The combination of the McConnell and Word findings
that implicit bias may affect interviews in ways that can disad
Black job applicants.
Another noteworthy result is the finding that the Race IAT, wh
ministered to White American subjects, predicts activation of th
dala-a presumed indicator of fear or other negative emotional a
response to photographic images of unfamiliar African American fa
related finding was the report by Richeson et al. that IAT measures
lated with evidence of self-regulatory or executive control activity
posure to African American faces.46

Lx
WHAT CAN BE DONE TO ATTENUATE THE INFLUENCE
OF IMPLICIT BIASES ON BEHAVIOR?
In their 1995 review of then-available evidence, Greenwald and
Banaji suggested that attentional focus could attenuate automatic influ
ences on social judgment, if those automatic influences were relatively
weak.47 Applying this principle, and assuming that implicit biases consti
tute "weak automatic influences," one might expect that getting people to
think more about, or to attend more closely to, their objectives in an inter
racial interaction might eliminate the effects of implicit bias. However,
Poehlman et al.'s review of the relevant predictive validity evidence for
IAT measures suggests a limitation of this conclusion.48 Although this re
view found that the predictive validity of explicit measures was indeed
greater for more deliberative behavior, it also found that prediction of be
havior by IAT measures was not reduced when the examined behavior was
more deliberative.
Consider the application of these findings to a hypothetical situation
in which racially different applicants are being evaluated for jobs, educa
tional program admissions, loans, or medical treatments. If an interviewer
in these situations devotes more deliberate effort to evaluating the candi
dates on explicit performance criteria, the interviewer may make better

45. See Elizabeth A. Phelps et al., Performance on Indirect Measures of Race Evaluation Predicts
Amygdala Activation, 12 J. Cognitive Neuroscience 729 (2000).
46. See Richeson et al., supra note 43.
47. See Greenwald & Banaji, supra note 5, at 17.
48. See Poehlman et al., supra note 30.

This content downloaded from

103.133.254.22 on Sun, 11 Feb 2024 10:47:39 +00:00
All use subject to https://about.jstor.org/terms
 2006] IMPLICIT BIAS.' SCIENTIFIC FOUNDATIONS 963

decisions overall but may still demonstrate the effects of implicit bias.
Thus, Poehlman et al.'s conclusions suggest caution in assuming that im
plicit bias can be reduced merely by increased deliberative effort on a deci
sion. Because no studies have yet directly tested this hypothesis, the
question of how to attenuate the impact of implicit biases on subtle but im
portant aspects of interpersonal interaction still awaits an answer.

x
How CAN IMPLICIT BIASES BE ALTERED?
In the first few years after the development of the IAT, m
searchers working with the test were impressed that, when they re
administered the same IAT to themselves, their measures of implic
remained remarkably similar over time. This was in part a welcome
vation, because it indicated that IAT measures might be identically
istered on multiple occasions to the same person without losin
validity as research measures (in contrast with, for instance, int
tests). The consistency of IAT measures over time also suggested
bility of implicit attitudes and stereotypes measured by the IAT.
Subsequent research, however, has shown that conclusion to
mature, as one of the first experiments that sought to influence
formance illustrates. Starting with the assumption that media e
may influence the race-valence associations measured by the IA
gupta and Greenwald asked White and Asian-American undergradua
dents to complete a preliminary task in which they identified a
photographs of well-known and admired African Americans (sc
artists, political leaders), mixed with photographs of somewhat
known but thoroughly disreputable European Americans (terro
serial murderers).49 A subsequent Race IAT measure revealed that t
tograph-identification task reduced the level of automatic preferen
European American (relative to African American). This reductio
plicit bias persisted over a twenty-four hour delay.50
Blair summarized a number of similar studies and concluded
plicit biases are malleable.5" For example, implicit gender stereo
feminine weakness were reduced by imagining examp

49. See Nilanjana Dasgupta & Anthony G. Greenwald, On the Malleability of A

Attitudes: Combating Automatic Prejudice with Images of Admired and Disliked Individu
Personality & Soc. Psychol. 800 (2001).
50. The opposite type of preliminary exposure, consisting of photographs of admirabl
Americans and disreputable African Americans, had no noticeable impact on Race IAT sco
at 803-04. This observation suggested that the ordinary media environment encountered by t
undergraduate research subjects might have been functioning as the equivalent of these
Black) exposures. See id. at 805; cf. Jerry Kang, Trojan Horses of Race, 118 Harv. L.
(2005).
51. See Irene V. Blair, The Malleability of Automatic Stereotypes and Prejudice, 6 Personality
& Soc. Psychol. Rev. 242 (2002).

This content downloaded from

103.133.254.22 on Sun, 11 Feb 2024 10:47:39 +00:00
All use subject to https://about.jstor.org/terms
 964 CALIFORNIA LA WREVIEW [Vol. 94:945

counter-stereotypic (i.e., strong) women,52 and implicit anti-Black ra

tudes were reduced by having African American experimenters admi
the research procedure.53
In studies using the Race IAT, these effects are typically modest,
ing the form of reduction, but not elimination, of implicit biases. Alt
the necessary research has not yet been done, caution is warran
speculating that repeated interventions of the types demonstrated to
fective in these experiments will have enduring effects on levels of im
bias. Some suggest that interventions like those used in the malle
studies temporarily activate a subtype of a larger category, such that
subtype temporarily replaces the larger category. For example,
Dasgupta and Greenwald experiment,54 the preliminary exposure t
rable Blacks may have activated the relatively attractive subtype of A
American celebrities. Once activated, this subcategory would temp
function as a mental replacement for the larger (and presumably
negatively valenced) African American category. If this interpreta
correct, it seems unlikely that even repeated interventions will pr
cumulative effects in a larger societal environment that reinforces pr
ing racial attitudes and stereotypes.
This skeptical appraisal does not imply that long-term changes in
plicit biases are impossible. For example, research has shown that
person forms a new personal connection with a member of a prev
devalued outgroup, implicit attitudes toward that group may chan
matically and rapidly.55 For example, when a son or daughter ma
member of a racial or ethnic minority or when a close friend is paral
an accident and begins using a wheelchair, a favorable implicit att
may rapidly replace a pre-existing negative implicit bias.56

52. See Irene V. Blair et al., Imagining Stereotypes Away: The Moderation of Implicit Stere
Through Mental Imagery, 81 J. Personality & Soc. Psychol. 828 (2001).
53. See Brian S. Lowery et al., Social Influence Effects on Automatic Racial Prejudice
Personality & Soc. Psychol. 842 (2001).
54. See Dasgupta & Greenwald, supra note 50.
55. Olsson, Ebert, Banaji, and Phelps recently reported that an implicit indicator of expect
outgroup racial bias was absent for college student subjects who had interracial dating experi
Andreas Olsson et al., The Role of Social Groups in the Persistence of Learned Fear, 309
(2005).
George Orwell gave a remarkable, albeit fictional, model for this type of influence in a scene from
Nineteen Eighty-Four. After 20 minutes of haranguing a crowd of Oceanians with vilification of the
Eurasian enemy, the orator receives a piece of paper and "without pausing in his speech" continues his
tirade against the (new) enemy, Eastasia:
Without words said, a wave of understanding rippled through the crowd. Oceania was at war
with Eastasia! ... The banners and posters with which the square was decorated were all
wrong! ... There was a riotous interlude while posters were ripped from the walls, banners
torn to shreds and trampled underfoot.... But within two or three minutes it was all
over.... The Hate continued exactly as before, except that the target had been changed.
George Orwell, Nineteen Eighty-Four 180-82 (1949).
56. See Greenwald et al., supra note 5.

This content downloaded from

103.133.254.22 on Sun, 11 Feb 2024 10:47:39 +00:00
All use subject to https://about.jstor.org/terms
2006] IMPLICIT BIAS: SCIENTIFIC FOUNDATIONS 965

Another way in which malleability of implicit bias has been tested

with IAT measures has been to ask research subjects to try to respond to
the IAT so as to produce a specific result-for example, asking subjects
who ordinarily show implicit preference for European American to attempt
to produce a Race IAT result showing preference for African American.
Relatively few subjects succeed at this faking assignment, partly because
few can spontaneously come up with a faking strategy.57 Although some
success in faking has been produced by instructing subjects to deliberately
respond slowly on the IAT component task for which they can ordinarily
respond rapidly, it remains apparent that the IAT is far more resistant to
faking than are explicit (self-report) measures that are designed to assess
attitudes or stereotypes.58

XI
IS IMPLICIT BIAS A PROBABLE CAUSE OF DISPARATE OUTCOMES?

[Wihen you have eliminated all which is impossible, then what

ever remains, however improbable, must be the truth.
Sherlock Holmes59

The argument that implicit bias is a probable cause of race discrimina

tion sometimes requires inference by a process of elimination. This is a
reasoning device endorsed not only by Sherlock Holmes, but also by the
Supreme Court. Specifically, in Furnco Construction Co. v. Waters,60 a
1978 employment discrimination case, the Court wrote:
[W]e know from our experience that more often than not people do
not act in a totally arbitrary manner, without any underlying
reasons.... Thus, when all legitimate reasons for [a negative
outcome] have been eliminated as possible reasons for the
employer's actions, it is more likely than not the employer, who we

57. See Rainer Banse et al., Implicit Attitudes Towards Homosexuality: Reliability, Validity, and
Controllability of the IAT, in 48 Zeitschrift f?r Experimentelle Psychologie 145 (2001); Boris
Egloff & Stefan C. Schmukle, Predictive Validity of an Implicit Association Test for Assessing Anxiety,
83 J. Personality & Soc. Psychol. 1441 (2002); Melanie C. Steffens, Is the Implicit Association Test
Immune to Faking?, 51 Experimental Psychol. 165 (2004).
58. Cfi Do-Yeong Kim, Voluntary Controllability of the Implicit Association Test (IAT), 66 Soc.
Psychol. Q. 83 (2003). Researchers may be able to detect such faking by noting when a subject is
responding unusually slowly in a task. See id. at 93. By comparison, it is harder for researchers to
detect faking on self-report measures; faking attitudes and beliefs on self-report measures typically
requires no more than modifying the position on which a pencil mark is placed in responding to a
survey questionnaire.
59. Sir Arthur Conan Doyle, The Sign of Four (1890), reprinted in Sherlock
Holmes: The Complete Novels and Stories 87,111 (1930) (emphasis removed).
60. 438 U.S. 567(1978).

This content downloaded from

103.133.254.22 on Sun, 11 Feb 2024 10:47:39 +00:00
All use subject to https://about.jstor.org/terms
 966 CALIFORNIA LA WREVIEW [Vol. 94:945

generally assume acts only with some reason, based his decision
an impermissible consideration such as race.61
Whether adjudicating an individual allegation of discrimination or
ing to understand broad patterns of disadvantage in society, if on
evidence of disparate impact-for example, in the form of syst
disadvantageous outcomes to African Americans in health care, ed
employment, housing, or criminal justice-one may begin to id
eliminate possible causes. Conceivable explanations that cannot
nated remain worth considering.
For sake of argument, let us assume that in attempting to und
whether implicit race bias has played a role in probation recomme
in a particular criminal court system, a researcher has elimina
ceivable non-race-related ("racially neutral") explanations on th
sound research evidence. Let us also assume that none of the relev
sionmakers has reported consciously holding negative racial at
stereotypes. Finally, let us assume that no test of implicit bia
administered to these decision makers. With this set of assumptio
reasonable to infer that the observed racial disparity is being
least in part, by implicit bias? Not only is it reasonable, it sh
garded as highly probable. This conclusion is justified by three
tions.
The first consideration is the observed pervasiveness of implicit bias,
as was clearly demonstrated by the data summarized in Tables 1 and 2. The
second consideration comes from the available evidence that (1) implicit
biases are predictive of discriminatory behavior and (2) implicit-bias
measures do a significantly better job than explicit-bias measures in pre
dicting behavioral indicators of discrimination.62 The third consideration is
provided by findings that implicit bias plays a causal role in discrimination.
The most important piece of this evidence at present is the finding that sub
tle discriminatory behaviors, of the types known to be predicted by IAT
measures of implicit race bias, play a significant role in determining the
outcomes of job interviews.63 The absence of another type of evidence also
supports this causal interpretation. Specifically, if-in the absence of both
racially neutral causes and explicit bias-racially disparate impact could be
shown to occur when implicit bias is shown to be absent, this would pro
vide evidence against a causal role of implicit bias in disparate impact. No
such evidence now exists.
In summary, a substantial and actively accumulating body of research
evidence establishes that implicit race bias is pervasive and is associated
with discrimination against African Americans. Consequently, when

61. Id. at 511.

62. See supra Parts IV and IX.
63. See McConnell & Leibold, supra note 41 ; Word, supra note 44, at 111-12.

This content downloaded from

103.133.254.22 on Sun, 11 Feb 2024 10:47:39 +00:00
All use subject to https://about.jstor.org/terms
2006] IMPLICIT BIAS: SCIENTIFIC FOUNDATIONS 967

racially neutral causes and explicit bias can be rejected as causal explana
tions for racially disparate outcomes, implicit race bias must be regarded a
a probable, even if not definitively established, cause. More direct confir
mations of the causal role of implicit bias may emerge in the next fe
years, as researchers increasingly include measures of implicit bias in their
studies of relevant domains in which racially disparate impact is a known
phenomenon.

This content downloaded from

103.133.254.22 on Sun, 11 Feb 2024 10:47:39 +00:00
All use subject to https://about.jstor.org/terms