Implicit Association: What the IAT Measures and What It Does Not

Most people arrive here looking for the implicit bias test, so start with the mechanism. If two concepts are strongly linked in memory, sorting them with the same hand is easy; if they are weakly linked or opposed, the sorting takes longer and produces more errors. Put flowers and pleasant words on one key and the task runs smoothly; put insects there instead and it stalls. The Implicit Association Test turns that difference in milliseconds into a score.

The appeal of the method is that it does not ask people what they think. Attitudes measured by questionnaire are shaped by what the respondent is willing to say and can observe about themselves, while a sorting speed is harder to manage deliberately. Greenwald, McGhee and Schwartz showed in 1998 that the two kinds of measure can come apart: in their third experiment the sorting task showed a clear in-group preference while the questionnaire ratings showed a much weaker one, and one of them showed almost none.

What a score from such a task supports has been argued for twenty-five years, and the page would mislead if it skipped that. Greenwald, Poehlman, Uhlmann and Banaji's 2009 meta-analysis put the average correlation between IAT scores and behavioral measures at .274. Oswald, Mitchell, Blanton, Jaccard and Tetlock's 2013 meta-analysis of criterion studies in the racial and ethnic domain put it at .148, a weak predictor at the individual level. Schimmack argued in 2021 that the construct validity of the test has not been established, and Forscher, Lai and Axt's 2019 meta-analysis found that procedures shifting scores on implicit measures do not reliably change behavior.

Two things follow, and they do not cancel each other. The sorting-speed difference is a robust finding that tracks exposure to cultural pairings. An individual's score from one sitting is a noisy number that cannot support a statement about that person's character or conduct, which is why the organization running the public version calls it educational rather than diagnostic. This page does not administer a test about race, gender or any other social group; the demo above uses the flower, insect, instrument and weapon categories of the original experiment.

Anthony Greenwald, Debbie McGhee and Jordan Schwartz published "Measuring individual differences in implicit cognition: The Implicit Association Test" in the Journal of Personality and Social Psychology in 1998. Three experiments built up the case for the method in order of how sensitive the subject matter was.

Experiment 1 used categories almost everyone evaluates the same way. Thirty-two students at the University of Washington sorted 150 words, 25 in each of six categories: flower names, insect names, musical instruments, weapons, pleasant-meaning words and unpleasant-meaning words. Each measure ran through five steps: sort the target categories, sort pleasant from unpleasant, do both at once in one pairing, swap the target keys, then do both in the other pairing. Responses were faster when flowers shared a key with pleasant words than when insects did, and faster for instruments than for weapons. Results for the two category pairs were indistinguishable, so the authors collapsed them: pooled across both, the gap averaged 129 milliseconds among students who met the crossed pairing first and 223 among those who met the matching pairing first, which they reported as an order effect. Error rates averaged just under 5 percent.

Experiment 2 tested whether the method could pick up a difference between groups of people rather than between categories. Seventeen Korean American and 15 Japanese American students sorted Korean and Japanese surnames against the same pleasant and unpleasant words. Each group sorted its own ethnicity faster when paired with pleasant words, and the effect tracked how immersed the student was in their ethnic culture.

Experiment 3 took the sensitive case. Twenty-six White American students sorted first names judged more likely to belong to Black or to White Americans, again against pleasant and unpleasant words, then completed five questionnaires about racial attitudes in private. The sorting task showed an in-group preference, with IAT effect sizes averaging 1.13, against 0.30 for the two attitude-rating measures: a feeling thermometer, which registered a much weaker preference in the same direction, and a semantic differential, which registered almost none. That dissociation is the finding the test became famous for, and the one that has drawn the most scrutiny since, because what a difference in sorting speed means about a person is what the later meta-analyses tried to settle.

Implicit associations are formed by exposure and show up in speed and ease rather than in opinions. The rows below are illustrations, not findings from the studies cited on this page.

The evidence points away from self-insight as the lever. Forscher, Lai and Axt found that procedures which change scores on implicit measures do not reliably change behavior, so the useful moves are about decisions and processes rather than about adjusting an inner state.

• Change the decision, not the mind. Masked review, structured interviews with fixed questions, written criteria scored before discussion: these constrain where an association could enter.
• Record the basis of each decision. An association shows up most easily in judgments with no stated reason, so asking for the reason narrows the gap it could operate in.
• Treat your own score as information about the method. One sitting of a reaction-time task is too noisy for a personal conclusion, and the meta-analytic correlations with behavior are weak.
• Be skeptical of training that promises to remove it. Short interventions move test scores without moving conduct, which is a reason to spend the budget on process instead.
• Look at outcomes rather than attitudes. Counts of who was shortlisted, promoted or stopped are measurable, and they do not require any theory about what anyone associates with what.
• Keep the exposure in view. Associations come from what is repeatedly paired in the surrounding culture, so what a workplace or a publication puts side by side is itself a lever.

The word implicit does a lot of work in this area, and it is worth separating the measure from the content it is used to study.

• It is not stereotyping, which is the use of a group-level belief to fill in what an individual is like. An implicit association is a link between concepts; stereotyping is a judgment made with one.
• It is not implicit stereotypes as such. That term names the content, the particular pairings a person has absorbed about a group, while implicit association names the kind of link and the way it is measured.
• It is not prejudice, which is an attitude of hostility or dislike toward a group. A fast pairing is not a feeling, and the correlation between the two kinds of measure is modest at best.
• It is not a test result about a person. An IAT score is a difference in average response time from one short session; the public version is offered as education, and the research literature does not support using it to assess, select or diagnose anyone.