The Greenwald Implicit Association Test
Try a simple sorting task. Words appear one at a time on a screen. If the word is a flower ("rose," "tulip") or a pleasant word ("love," "peace"), press the left key. If it is an insect ("wasp," "cockroach") or an unpleasant word ("vomit," "poison"), press the right key. Most people find this easy. Now swap the rules so flowers share a key with unpleasant words and insects share a key with pleasant ones. Almost everyone slows down and makes more mistakes.
That slowdown is the basis of the Implicit Association Test (IAT), introduced in 1998 by social psychologist Anthony Greenwald of the University of Washington and his collaborators Debbie McGhee and Jordan Schwartz. The idea is that sorting is faster when the two categories sharing a key are already linked in memory. Measuring that speed difference, they argued, could reveal associations people do not report on questionnaires, or may not even know they hold.
The IAT became one of the most influential tools in social psychology and the centerpiece of a public conversation about "implicit bias." It also became one of the most argued-about. This article covers how the test works, what the original experiments found, and what more than two decades of research say about its reliability and its ability to predict behavior.
The problem it was designed to solve
By the 1990s, survey measures of overt prejudice in the United States had declined considerably, yet disparities in hiring, health care and criminal justice persisted. Social psychologists were also aware that self-reports are vulnerable to self-presentation: people may not admit to attitudes they consider unacceptable.
Researchers had already developed implicit measures based on reaction time, especially evaluative priming, in which briefly showing one word speeds or slows the evaluation of the next. In a 1995 Psychological Review article, Greenwald and Mahzarin Banaji set out a theory of "implicit social cognition," arguing that past experience can influence judgments in ways people cannot introspectively identify. The IAT was meant to measure those influences more reliably than earlier priming tasks did.
How the IAT works
The standard IAT has seven blocks of trials, a structure later refined in the scoring work of Greenwald, Brian Nosek and Banaji. A race IAT, for example, works roughly like this:
- Target practice. Sort faces (or names) as Black or White, with one category per key.
- Attribute practice. Sort words as good or bad, one category per key.
- First combined block (practice), and 4. first combined block (test). Now both tasks are interleaved. One key is used for White faces and good words, the other for Black faces and bad words.
- Reversed target practice. The face categories switch sides.
- Second combined block (practice), and 7. second combined block (test). One key is now used for Black faces and good words, the other for White faces and bad words.
The test compares speed and accuracy in the two combined arrangements. If someone is faster when White and good share a key than when Black and good share a key, the test scores that as an automatic preference for White over Black. The order of the combined blocks is counterbalanced across participants because order affects results.
Modern IATs use the D score, an algorithm introduced in 2003 that divides the latency difference by the person's overall variability and includes error penalties. A D score near zero indicates no relative preference; conventional cut-offs label scores as slight, moderate or strong.
Crucially, an IAT score is relative. It compares two pairings. It cannot say, on its own, that someone feels negatively toward a group, only that one combination was easier than the other.
The original 1998 experiments
The founding paper, "Measuring individual differences in implicit cognition: The Implicit Association Test," appeared in the Journal of Personality and Social Psychology (vol. 74, pp. 1464–1480). It reported three experiments, each designed to show a different property.
- Experiment 1 used flowers versus insects and musical instruments versus weapons, paired with pleasant and unpleasant words. These are near-universal evaluative differences, and the IAT detected them strongly: participants were much faster when flowers and instruments shared a key with pleasant words.
- Experiment 2 tested Japanese American and Korean American students using Japanese and Korean surnames. Each group tended to show a more favorable association for its own group, showing the test could detect expected individual differences.
- Experiment 3 used typical White American and Black American first names with White students who described themselves as unprejudiced. Despite their explicit reports, they on average showed faster responses when White names were paired with pleasant words. The IAT effect and explicit measures correlated only weakly.
That third result is the one that made the IAT famous: a measure seemed to detect associations that contradicted people's stated attitudes.
Project Implicit and the data flood
In 1998 Greenwald, Banaji and Nosek put demonstration IATs online. The site later became Project Implicit, a nonprofit hosted with Harvard and other universities, and millions of people have taken tests there on race, gender, age, weight, sexuality, disability and more. In a 2002 analysis of early website data, Nosek, Banaji and Greenwald reported that implicit preferences for White over Black, young over old and similar patterns were widespread, and that they were often stronger than the preferences people reported explicitly.
Among the frequently reported patterns in Project Implicit data:
- On average, White participants show a pro-White implicit preference on the race IAT, while Black participants' averages are much closer to neutral.
- Many people, including women, show associations of male with career or science and female with family or liberal arts.
- Preferences for young over old and thin over fat are among the strongest averages.
Large online samples have also allowed researchers to track change. A 2019 analysis of Project Implicit data by Tessa Charlesworth and Banaji reported that implicit attitudes about sexual orientation, race and skin tone had moved toward neutrality between 2007 and 2016, with sexual orientation changing fastest. Implicit age and disability attitudes showed no meaningful change, and implicit anti-fat bias increased slightly.
What the IAT is good at
The IAT is reliably sensitive to differences between groups of people and between conditions. Averages across many participants are stable and can track cultural patterns, as the Project Implicit data show. The flower-insect effect is large and easy to replicate. Compared with earlier implicit measures such as evaluative priming, the IAT has better internal consistency, and it is harder to fake than a questionnaire, although people can learn to fake it with instruction and practice.
These strengths are about groups and averages. The problems start when IAT scores are used to say something about a single person.
Problem 1: reliability
A measure used to classify individuals needs to give similar results when the same person is tested again. For the IAT, that test-retest reliability is modest. In a 2020 Annual Review of Psychology article, Greenwald and Calvin Lai themselves cited a test-retest figure of about .50 for a single IAT, describing it as well below psychometric standards for diagnostic use with individuals. By comparison, well-established personality and ability tests often reach .80 or higher.
In practice, this means a person might be labeled "moderate preference" on one day and "slight" or "no preference" on another. Averaging several IATs taken by the same person raises reliability substantially, which is one reason researchers emphasize group-level uses.
Problem 2: predictive validity
The deeper controversy is whether IAT scores predict discriminatory behavior. Several meta-analyses have tried to answer this, with differing emphases:
- Greenwald and colleagues (2009) pooled 122 research reports and found an average correlation of about r = .27 between IAT measures and criterion measures such as behavior, judgments and physiological responses. For the specific domain of Black–White race, the IAT predicted better than explicit measures did.
- Oswald, Mitchell, Blanton, Jaccard and Tetlock (2013) focused on ethnic and racial discrimination and used different inclusion and coding choices. They found an average correlation of roughly r = .15 and concluded that the IAT was a poor predictor of discrimination and did no better than explicit measures.
- Kurdi and colleagues (2019) analyzed 217 research reports with over 36,000 participants. Implicit and explicit measures each made small, independent contributions to predicting intergroup behavior (standardized coefficients of about .14 and .11). Correlations were higher, reaching about r = .37, in studies whose IAT and behavior measures were closely matched in content and designed with that question in mind.
Greenwald, Banaji and Nosek argued in 2015 that even statistically small effects can be socially significant when they apply to many people and many decisions. Critics such as Hart Blanton and Philip Tetlock respond that small, unreliable correlations do not justify telling individuals they are biased or using the test in training as though it diagnoses individuals. Both sides agree on one point: the IAT is not a lie detector for prejudice.
Problem 3: what exactly is being measured?
An IAT effect could reflect more than personal attitudes. Researchers have proposed that it partly captures:
- Cultural knowledge. Knowing what associations are common in your culture, even if you reject them. Some researchers distinguish "personal" from "extrapersonal" associations.
- Salience and familiarity. Categories that are more familiar or salient may be easier to sort with "good," regardless of attitude.
- Task-switching ability. General cognitive speed and the ability to switch between rules influence scores, although the D algorithm reduces this.
There is also a long debate about whether people are truly unaware of their IAT-measured associations. Studies by Adam Hahn and colleagues found that people could predict their own IAT results fairly accurately for several social groups, which challenges the idea that the associations are entirely "unconscious." Many researchers now prefer the term implicit to mean "measured indirectly" rather than "unconscious."
Problem 4: can implicit bias be changed, and does it matter?
If IAT scores reflect something that causes discrimination, changing them should reduce discrimination. Two large studies tested this.
Lai and colleagues (2016) compared nine interventions in studies with 6,321 participants. All nine, including exposure to counter-stereotypical examples and practicing if-then plans, reduced implicit racial preference immediately. None was still effective after a delay of several hours to several days.
Forscher and colleagues (2019) used network meta-analysis on 492 studies with more than 87,000 participants. They found that implicit measures can be changed, but usually only weakly, and that changes in implicit measures did not lead to corresponding changes in explicit attitudes or behavior.
Together these studies suggest that brief interventions do not produce lasting changes in IAT scores, and that changing IAT scores is not, by itself, a demonstrated path to changing behavior.
Variants and rival implicit measures
The IAT's popularity inspired many related tasks, partly to fix its known weaknesses.
- Single-Category IAT. The standard IAT always compares two groups, so it cannot say whether someone has a positive or negative association with one group on its own. The single-category version pairs only one target category with good and bad.
- Brief IAT. A shorter version designed for situations where testing time is limited.
- Go/No-go Association Task. Introduced by Nosek and Banaji in 2001, it asks people to respond only to items from a target category and an attribute, and to withhold responses to everything else, measuring associations through accuracy under time pressure.
- Affect Misattribution Procedure (AMP). Developed by Keith Payne and colleagues in 2005, it briefly flashes a prime, such as a face, followed by an ambiguous Chinese character that people unfamiliar with Chinese rate as pleasant or unpleasant. Feelings evoked by the prime tend to spill over onto the rating.
These measures are not interchangeable. Correlations between different implicit measures of the same attitude are often modest, which is one reason researchers increasingly treat "implicit bias" as a family of related phenomena rather than a single hidden attitude.
Faking and context effects
Because the IAT relies on reaction times, early researchers hoped it would be immune to faking. Later studies showed this is only partly true. When participants were told how the test works and instructed to fake, many could shift their scores, for example by deliberately slowing down in the block they wanted to look worse on. Researchers have developed methods to detect such slowing, but naive participants who simply try to "be unbiased" usually cannot change their scores much.
IAT scores are also sensitive to context. Seeing admired members of a group before the test, or changing the specific images and words used, can shift average results. To supporters, this shows that associations are flexible and context-dependent. To critics, it suggests that a single score is a snapshot of a person in a moment, not a stable trait.
The theory behind the test
Two broad ideas shape how researchers interpret IAT results. The original view, associated with Greenwald and Banaji's implicit social cognition framework, is that the test measures associations in memory built up through experience. A newer family of dual-process models distinguishes fast, associative processes from slower, deliberative ones and treats IAT scores as a partial window onto the former.
Keith Payne, Heidi Vuletich and Kristjen Lundberg proposed a "bias of crowds" model in a 2017 Psychological Inquiry paper, in which IAT scores mostly reflect the associations prevalent in a person's environment at the time of testing. On this account, IAT scores are weak predictors of individual behavior but useful indicators of cultural or regional patterns, and aggregated scores are more meaningful than individual ones. This view fits both the low test-retest reliability and the stable group averages.
Implications for implicit bias training
Many companies, police departments and universities adopted implicit bias training in the 2010s, sometimes including taking an IAT. Research on such programs has generally found that short, one-time sessions have little lasting effect on behavior. That does not mean bias does not exist or that disparities are not real; field experiments such as résumé audit studies continue to document discrimination. It means that the IAT's role in explaining that discrimination, and in fixing it, is less certain than early public discussion suggested. Researchers have increasingly emphasized structural changes, such as standardized interviews, clear evaluation criteria and accountability, which do not depend on changing what is in an individual's head.
Common misconceptions
"Banaji and Nosek co-authored the original IAT paper." They became Greenwald's key collaborators and co-founded Project Implicit, but the 1998 paper's authors were Greenwald, McGhee and Schwartz.
"An IAT score shows whether you are racist." The test's own developers caution against using it to diagnose individuals. A single score has modest reliability, is relative rather than absolute, and predicts individual behavior weakly.
"The IAT measures unconscious attitudes people cannot know about." Evidence that people can often predict their own results complicates that claim. "Implicit" refers mainly to the indirect method of measurement.
"The IAT has been debunked." Also an overstatement. Group-level IAT effects are robust and replicate well. The debates concern individual-level reliability, predictive validity and interpretation.
Why the IAT still matters
The IAT changed social psychology. It provided a quick, reliable way to measure associations in large samples, created one of the largest datasets in the field and pushed researchers to think carefully about the gap between what people say and what they do. It also became an object lesson in how a research tool can be stretched beyond what its evidence supports once it enters public debate.
The honest summary is this: the IAT detects real, culturally widespread associations at the group level, but it is a noisy instrument for individuals and only a modest predictor of behavior. For related classic research on group dynamics and prejudice, see the Robbers Cave experiment, the Clark doll experiment and the blue eyes, brown eyes exercise. For the broader psychology of biased thinking, see cognitive biases explained and how priming changes what you notice.
Frequently Asked Questions
Q: Who created the Implicit Association Test? A: Anthony Greenwald, with Debbie McGhee and Jordan Schwartz, introduced it in a 1998 paper in the Journal of Personality and Social Psychology. Mahzarin Banaji and Brian Nosek became central collaborators and helped launch Project Implicit.
Q: How does the IAT work? A: You sort words or images into categories using two keys. In one arrangement, for example, White faces share a key with good words; in another, Black faces do. The test compares how fast and accurately you respond in the two arrangements, and the difference is scored as a relative association.
Q: If I score high on the IAT for racial bias, does that mean I am racist? A: No. The IAT measures relative associations, a single score has a test-retest reliability of only about .50, and individual scores predict behavior weakly. Its developers say it should not be used to diagnose individuals.
Q: How reliable is the IAT? A: Its internal consistency is good, but its test-retest reliability for a single administration is about .50, below the standard usually expected for individual assessment. Averaging several tests improves reliability.
Q: Does the IAT predict discrimination? A: Weakly. Meta-analyses report average correlations with behavior of roughly .15 to .27 depending on methods, with stronger links when the IAT and behavior measures are closely matched. Researchers disagree about whether such small effects matter socially.
Q: Can implicit bias be reduced? A: Brief interventions can shift IAT scores immediately, but a 2016 study of nine interventions found none lasted beyond a few hours to days. A 2019 meta-analysis found that changes in implicit measures did not translate into changes in behavior.
Q: Can the IAT be faked? A: It is harder to fake than a questionnaire, but studies show that people who understand the test and are instructed how to fake it, for example by deliberately slowing down in one block, can alter their scores.
Sources
- Greenwald AG, McGhee DE, Schwartz JLK (1998). Measuring individual differences in implicit cognition: The Implicit Association Test. JPSP 74:1464–1480 (PubMed)
- Greenwald AG, Lai CK (2020). Implicit social cognition. Annual Review of Psychology 71
- Greenwald AG et al. (2009). Understanding and using the IAT: III. Meta-analysis of predictive validity. JPSP 97:17–41
- Oswald FL et al. (2013). Predicting ethnic and racial discrimination: A meta-analysis of IAT criterion studies. JPSP 105:171–192 (PubMed)
- Kurdi B et al. (2019). Relationship between the Implicit Association Test and intergroup behavior: A meta-analysis. American Psychologist 74:569–586
- Forscher PS et al. (2019). A meta-analysis of procedures to change implicit measures. JPSP 117:522–559 (PMC)
- Lai CK et al. (2016). Reducing implicit racial preferences: II. Intervention effectiveness across time. J Exp Psychol: General 145:1001–1016
- Greenwald AG, Banaji MR, Nosek BA (2015). Statistically small effects of the IAT can have societally large effects. JPSP 108:553–561
- Association for Psychological Science: The IAT — How and when it works
- Payne BK, Vuletich HA, Lundberg KB (2017). The bias of crowds: How implicit bias bridges personal and systemic prejudice. Psychological Inquiry 28
- Fiedler K, Bluemke M (2005). Faking the IAT: Aided and unaided response control on the Implicit Association Tests. Basic and Applied Social Psychology 27
- Hahn A et al. (2014). Awareness of implicit attitudes. J Exp Psychol: General (PMC)
- Charlesworth TES, Banaji MR (2019). Patterns of implicit and explicit attitudes: I. Long-term change and stability from 2007 to 2016. Psychological Science 30
- Project Implicit
Read next:
Follow new articles
Add SciHub101 to your feed reader to see new science and technology stories.
About the Author
SciHub101 is an independent science and technology publication founded and edited by Buğra Sözeri, a software engineer and app developer based in Türkiye. AI tools assist with drafting and research. Source documentation in the existing archive is uneven, and articles are being reviewed and corrected. The Editorial Team byline identifies editorial responsibility; it does not certify specialist or independent human review. Where we compare products, anything built by our own team is labelled as such — see the editorial guidelines and About page for how corrections and disclosures work.
Related Articles
Complete Science Experiments Guide 2026 — Every Landmark Experiment by Field
Every landmark science experiment in one hub: Pasteur's swan-neck flask, Galvani's frog legs, Eratosthenes, Michelson–Morley, Hafele–Keating and more than 140 others, grouped by field and era.