Implicit Stereotypes: The Pairings a Culture Teaches by Repetition

An implicit stereotype is a pairing that has been practiced. If a culture puts two things side by side often enough, in headlines, in job titles, in who appears in which photograph, the link between them becomes fluent, and fluent links are used by default when there is no time or no reason to think. That is why this page is filed under what memory keeps: a stereotype is a generalization stored in place of the specifics it was built from, and it is retrieved whether or not the person retrieving it agrees with it.

Anthony Greenwald and Mahzarin Banaji set out the term in 1995 for traces of past experience that a person cannot identify by introspection but that shape the qualities attributed to members of a social category. The key word is unidentified. People can usually report what the stereotypes in their culture are; what they cannot report is whether one has just affected a judgment they made. Patricia Devine had shown in 1989 that knowledge of the cultural content is widely shared, including by people who score low on measures of prejudice and reject the content when asked.

What follows from a score on an indirect measure is the contested part, and the page would mislead by skipping it. Three meta-analyses set the bounds. The 2009 one by Greenwald, Poehlman, Uhlmann and Banaji reported an average correlation of .274 with behavioral criteria. The 2013 one by Oswald, Mitchell, Blanton, Jaccard and Tetlock reported about .14 across the racial and ethnic domains, .15 for race and .12 for ethnicity. The 2019 one by Forscher, Lai and Axt reported that the procedures which shift these scores leave behavior where it was. One short session of a sorting task is a noisy number, and it cannot support a statement about an individual's conduct or character.

Two things hold at once. Averaged across many people, the pairings show up reliably and track what the surrounding culture repeats. For one person on one day, the measure is too noisy to describe them. Both are reasons to work on the decisions rather than on the self-assessment.

Greenwald and Banaji's "Implicit social cognition: Attitudes, self-esteem, and stereotypes" appeared in Psychological Review in 1995. It was a review rather than an experiment, and its contribution was a definition and a research program: attitudes, self-esteem and stereotypes could all be measured indirectly, through their effect on a task, rather than by asking. The definition of an implicit stereotype that most later papers quote comes from that article.

The method arrived three years later, when Greenwald, McGhee and Schwartz published the Implicit Association Test in 1998 and established the pattern it relies on, that a practiced pairing is sorted faster than a crossed one. Their experiments and the figures from them are set out on the page for implicit association. What this page takes from that work is the move it made possible: stereotype content could be measured through what a person does on a task instead of through what they are willing to say about a group.

Brian Nosek, Banaji and Greenwald then took the task to the public. Their 2002 paper in Group Dynamics reported data from the demonstration website they had opened, where visitors could complete tasks on group attitudes and stereotypes, including a task sorting men's and women's names against words about work and about family. Across 38,797 interpretable sessions the gender-career task showed a gap of 95 milliseconds in favor of the culturally repeated pairing, an effect size of 0.72. The demo above uses adapted examples from the same four categories, men's and women's names against words about paid work and about home and family, rather than the exact stimuli the paper's appendix lists. The website is a self-selected sample rather than a survey of any population, which the authors noted, and it is a demonstration rather than an assessment.

The scrutiny since has been directed at what a score means for a person, not at whether the sorting difference exists. The 2009 and 2013 meta-analyses disagree about how strongly scores relate to behavior, and both estimates are small. Schimmack pressed the question further in 2021, arguing that the test's construct validity has never been established. Indirect measures are useful for studying averages and cultural patterns and poor at describing individuals.

A practiced pairing operates where a judgment is quick, unstructured and unexplained. The rows below are illustrations, not findings from the studies cited on this page.

A pairing laid down by repetition is not undone by noticing it. That is what the 2019 meta-analysis showed: the procedures that moved scores on indirect measures left conduct where it was. The work that pays goes into what gets repeated and into how decisions are made.

• Structure the decision. Written criteria, fixed interview questions, scores recorded before discussion: each one narrows the gap a practiced pairing could operate in.
• Mask what can be masked. Blind review of work samples removes the cue that starts the association, while leaving other cues in place, so it narrows the problem rather than ending it.
• Ask for the basis of each judgment. A rating with a reason attached can be checked; a general impression cannot.
• Count outcomes rather than attitudes. Who was shortlisted, promoted, funded or stopped is measurable, and it needs no theory about anyone's associations.
• Watch what gets paired in your own output. Which photograph illustrates which story, which name goes in which example: repetition is where these pairings come from.
• Read any score you are given as a fact about the test rather than about yourself. One sitting is noisy, and the published links to behavior are weak at the level of one person.

Four terms in this area are used interchangeably in everyday speech and mean different things in the research.

• It is not implicit association, which names the kind of link and the way it is measured. Implicit stereotypes are the content: the particular group and trait pairings a person has absorbed.
• It is not stereotyping, which is the judgment a person makes when they apply a group-level belief to an individual. That judgment can be made deliberately and reported afterwards; the pairings on this page operate without being noticed.
• It is not prejudice, which is an attitude toward a group with feeling attached, usually hostility. Speed of pairing is not warmth or hostility, and the two kinds of measure agree only weakly.
• It is not a diagnosis or a test result about a person. Neither the demo above nor a reaction-time task from one session can support a claim about how anyone behaves.