Skip to main content
Category: Effects
Type: A latency difference on a sorting task. White American students responded faster when White names and pleasant words shared a key than when Black names and pleasant words shared a key. The score is that time gap, not a hiring decision, a grade, or an arrest
Origin: Anthony G. Greenwald and Mahzarin R. Banaji, “Implicit Social Cognition: Attitudes, Self-Esteem, and Stereotypes,” Psychological Review, 1995, volume 102, number 1, pages 4–27. The measure in the counted case is Anthony G. Greenwald, Debbie E. McGhee, and Jordan L. K. Schwartz, “Measuring Individual Differences in Implicit Cognition: The Implicit Association Test,” Journal of Personality and Social Psychology, 1998, volume 74, number 6, pages 1464–1480. Both PDFs were read. Devine (1989), Fazio et al. (1995), and the other papers cited as background were not opened
Also known as: Implicit attitude and implicit stereotype, the two constructs the 1995 paper defines. The implicit association test, or IAT, is the 1998 procedure, not a second title for this page. The 1995 paper uses the words “implicit bias” once, inside a paragraph on consciousness raising. It does not define that phrase. This page is not the Halo Effect, the Stereotype Effect, or In-Group Bias
Quick Answer — Implicit Bias, as this page uses the name, is a faster keypress when White names and pleasant words share one response. In Experiment 3 of the 1998 paper, 26 White American students at the University of Washington sorted names and words. The text says the advantage for White + pleasant averaged 179 ms across four contrasts. On a five-item semantic differential, the mean racial preference was 0.008. Only one of the 26 IAT scores pointed the other way. The task recorded key times and questionnaires. It did not record a decision about another person.

​
What is Implicit Bias?

Implicit Bias, on this page, is the 1998 finding that White names paired with pleasant words drew a faster key than Black names paired with pleasant words, in a sample that did not show the same gap on one explicit scale.
Implicit attitudes are introspectively unidentified (or inaccurately identified) traces of past experience that mediate favorable or unfavorable feeling, thought, or action toward social objects.
Greenwald and Banaji print that sentence as the definition of an implicit attitude. The trace is past experience. The person does not identify it, or identifies it inaccurately, by looking inward. What it mediates is a favorable or unfavorable feeling, thought, or action toward a social object. The 1998 procedure does not ask the student to report that trace. It asks for a speeded sort. When the pairing that is easier gets the same key, the sort is faster. The difference in speed is the score. A useful picture is a mailbox with two slots. One slot is labeled with a name group and with “pleasant.” The other is labeled with the other name group and with “unpleasant.” Halfway through, the labels swap. The letters are the same. Only the pairing on the slot changes. The clock runs on each letter. Confirmation Bias is a different habit, about seeking material that fits a belief already in view. This clock does not ask the student to choose an article or a memory.

​
Implicit Bias in 3 Depths

  • Beginner: White + pleasant was the faster pairing. The paper’s prose average is 179 ms. One of 26 IAT scores was positive, which in this table means a Black preference.
  • Practitioner: Read four difference scores before you quote one. The semantic differential mean is 0.008. The feeling thermometer mean is −8.73. Those two questionnaires do not agree with each other, and neither one is the latency.
  • Advanced: Figure 5 prints four callouts, +213 ms, +120 ms, +201 ms, and +198 ms. The prose says 179 ms. The male-name IAT and the female-name IAT correlate .46. The authors leave a familiarity alternative open. No later decision was scored.

​
Origin

Implicit Bias, as a defined entry, is not the title of either paper. The 1995 title is implicit social cognition. The 1998 title is the implicit association test. The 1995 review starts from implicit memory. The signature, in their words, is that traces of past experience affect performance even though the earlier experience is unavailable to self-report or introspection. Their template says an implicit C is the introspectively unidentified, or inaccurately identified, trace of past experience that mediates R. For an attitude, R is favorable or unfavorable feeling, thought, or action toward social objects. For a stereotype, the same template is aimed at attributions of qualities to members of a social category. A footnote says the implicit-explicit pair overlaps labels such as unaware-aware and unconscious-conscious, and that those dichotomies are not the same kind of claim. Dissociation, a clash between an indirect score and a direct score, is interesting. They say it is not required before an attitude can be called implicit. They also counted how attitudes were measured. In a 1989 census of the European Journal of Social Psychology, the Journal of Personality and Social Psychology, the Journal of Experimental Social Psychology, and the Personality and Social Psychology Bulletin, 47 studies included an attitude measure. All 47 used at least one direct measure. Six, 13 percent, used some indirect measure. The review’s point is that a direct question was the default. An indirect measure, in their definition, is one where the person is not told that the inferred object is what is being assessed. They write that indirect measures are theoretically required for implicit cognition, not only practically useful for avoiding a polite answer. The same review says judgment latencies vary a lot within one person, so a usable score needs many similar trials, and that latency measures had mostly served group comparisons rather than individual differences. The 1998 paper is the attempt to make an individual-difference score. An IAT, in the abstract, measures the differential association of two target concepts with an attribute. Flowers versus insects is the easy demonstration. Experiment 3 is the race demonstration. In-Group Bias is a different page. The 1998 discussion calls the Experiment 3 pattern an implicit in-group preference. It does not hand you that other page’s measure.

​
Key Points

Four splits keep a key time, a questionnaire, a sign, and an alternative account from becoming one verdict.
1

The score is a pairing, and the prose average is 179 ms

The student classifies names as Black or White and words as pleasant or unpleasant. White + pleasant means those two categories share a key. Black + pleasant means the other pairing shares a key. The 1998 text says White students were faster on White + pleasant, and that this IAT effect averaged 179 ms over the four contrasts in Figure 5. Separate tests give F(1, 21) = 41.94 for male names and F(1, 21) = 28.83 for female names, with p = 10^-6 and p = 10^-5. The figure itself labels the four contrasts +213 ms, +120 ms, +201 ms, and +198 ms. This page does not replace 179 with an average of those labels. The error bars are standard deviations of the 13 people in each order, not confidence intervals.
2

Positive means a Black preference, and the two questionnaires split

Table 5 scores every index so that a positive number is a preference for Black relative to White. The feeling thermometer mean is −8.73, SD 15.06, d = −0.58, t(25) = −2.96, p = .01, on a range from −99 to 99. The semantic differential mean is 0.008, SD 0.559, d = 0.01, t(25) = 0.07, p = .95, on a range from −6 to 6. The male-name IAT mean is −.181, SD .139, d = −1.30, t(25) = −6.61, p = 10^-7. The female-name IAT mean is −.145, SD .141, d = −1.03, t(25) = −5.26, p = 10^-5. Those IAT means are natural logarithms of latency, not milliseconds. The authors say d = M / SD, and they treat .2, .5, and .8 as small, medium, and large. They also say the pro-White effect size on the IAT was about twice the thermometer effect size. The semantic differential, in their words, showed a virtual absence of racial preference.
3

Nineteen of 26 did not report a White preference on one scale

Figure 6 plots the semantic differential against the average of the two IAT measures. The text says 19 of 26 White subjects explicitly endorsed either Black-White indifference, a zero on that scale, or a Black preference, a positive score. All but one of those 19 had a negative IAT score. Only one of all 26 had a positive IAT score. The paper calls that pattern discouraging evidence of unconscious forms of prejudice, and also says the IAT may reveal an explicitly disavowed preference. The five explicit measures still correlate with the two implicit measures at an average r = .14. Attitude questionnaires with the IAT average r = .17. The Modern Racism Scale and the Diversity and Discrimination scales average r = .12 with the IAT. The male-name IAT and the female-name IAT correlate .46.
4

Order did not carry the result, and familiarity is still on the page

Half the students did White + pleasant first (n = 13) and half did Black + pleasant first (n = 13). Order effects were F(1, 21) = 0.03 and 2.01, p = .86 and .17, for male names and female names. The weak direction is a slightly larger effect when the authors’ “compatible” pairing comes first. They say that label assumes White + pleasant is the compatible pairing for these students. A different alternative does not depend on order. The discussion says these students were much less familiar with the African American names than with the White names, and that greater liking for familiar items could explain the IAT. They say the same alternative cannot cover the flower and insect test, where the negative category had the more frequent words. They point to preliminary work, not opened here, and they do not claim Experiment 3 has closed the question.

​
Applications

These four moves follow splits the two papers print. They are not a claim that an office, a class, a household, or another year copies 179 ms or 1 of 26.

Put both questionnaires next to the clock

Table 5 is four rows. The semantic differential is flat. The thermometer is not. A sentence that says “they reported no preference” has kept one row and dropped the other. Ask which scale, and ask whether the number is the mean or a count of people at zero.

Read the sign before you name the direction

In this table a positive score is a Black preference. A negative IAT is the White-preference direction. The millisecond callouts on the figure are unsigned gaps between pairings. They are not the signed log score. Quote the table note with the number.

Name the response that was timed

The recorded acts are a keypress and a questionnaire placed in an unmarked envelope. The consent form said the task might reveal an objectionable attitude and that the student could withdraw. Nothing in Experiment 3 is a callback, a sentence, or a grade. If the decision you care about was not timed, the 179 ms does not stand in for it.

Keep the familiarity line in the note

The names were chosen because other introductory students judged them more likely to belong to one group. Examples include Brandon, Ian, and Jed, beside Darnell, Lamar, and Malik, and Betsy, Katie, and Nancy beside Ebony, Latisha, and Tawanda. The authors treat uneven familiarity as a live alternative for this experiment. Do not file that alternative as settled by a paper this page did not open.

​
Case Study

The counted case is Experiment 3. The subjects were 14 female and 12 male White American students from introductory psychology at the University of Washington. They received optional course credit. Before the task, the consent procedure said the experiment could reveal attitudes they might find objectionable, and that they could withdraw. After the computer tasks they answered five questionnaires in the booth and put them in an unmarked envelope. The sort used 25 pleasant words and 25 unpleasant words, with two unpleasant words changed from the earlier experiments. It also used two sets of 50 first names, male names in one subexperiment and female names in the second. Other students had judged the names as more likely to belong to White Americans or to Black Americans. Words were lowercase. Names were uppercase. The procedure otherwise matched the previous experiment in the same paper: one subexperiment, then the other. The paper does not say that a person met, hired, taught, or punished anyone. The text then gives the latency result. Faster responses on White + pleasant. A prose average of 179 ms across the four contrasts drawn in Figure 5. Large F tests for the male-name and female-name versions. No significant order effect. Table 5 puts the log-latency IAT at d = −1.30 and d = −1.03, the thermometer at d = −0.58, and the semantic differential at d = 0.01. The authors say the IAT showed a stronger relative White preference than either questionnaire, and that the semantic differential showed almost none.
Indeed, only one of the 26 White subjects had a positive IAT score.
That sentence follows the count of 19 of 26 at zero or on the Black-preference side of the semantic differential. The authors call the pattern evidence that an IAT can show a preference the explicit scale disavows, and they call it discouraging because it suggests unconscious forms of prejudice. They also say the small correlations do not, by themselves, damage construct validity, because they did not expect implicit and explicit scores to move together. In the same stretch they refuse a stronger claim: the result cannot be assumed just because someone suspects that White Americans in general have an automatic negative association to African American names. Familiarity with the names is the alternative they print. The lesson inside this file is narrow. On this task, a near-zero semantic differential and a negative IAT sat together for most of these 26 students, and the clock’s prose average was 179 ms. The paper does not say why the names were unevenly familiar, and it does not follow any student out of the booth. Self-Fulfilling Prophecy is a belief that changes a later outcome. This file has no later outcome.

​
Boundaries and Failure Modes

Implicit Bias, as these PDFs let us count it, is a set of key times and questionnaires from 26 White American students in one psychology pool. Two uses overshoot the papers. One exports 179 ms, d = −1.30, or 1 of 26 onto an office, a class, a household, or another year. The other treats the authors’ phrase “unconscious forms of prejudice” as if the thermometer, the familiarity paragraph, and the figure’s four labels had not been printed. The 1995 paper is a review, not a second sample of these students. It says consciousness raising may have some value in attenuating implicit bias when the source is not properly identified, because attention can weaken a weak cue. The same section says the authors make no attempt to review the research on blinding, consciousness raising, or affirmative action. Experiment 3 gave no training, hid no name, and applied no compensatory rule. A workshop that quotes this sentence as its own result has left the page. Blinding, in that 1995 discussion, means the decision maker never receives the cue. The authors call it foolproof in principle and often impossible in practice, because a stigmatized mark tends to travel with other marks that a screen does not remove. Their illustration is a report by Allmendinger and Hackman (1993) that orchestras had candidates play behind a partition. That paper was not opened. No audition count from it is printed here. The authors’ own caution stays: a screen removes the cue in front of you, and it does not erase an advantage built by earlier unequal experience. The 1998 general discussion says the three experiments are encouraging and do not settle the method. The authors write that self-presentation in this booth may be private self-presentation, a person managing a view of the self, rather than impression management aimed at the experimenter. They do not report a test-retest correlation over days. The .46 between male-name and female-name IATs is two scores from one session. Naive Realism is the sense that one’s own view is simply the way things are. Nothing in these tables measures that sense. A last limit is the neighbor. The Halo Effect appears in the 1995 review as one place where an attitude toward a person spreads onto an attribute. The Stereotype Effect is a different page about a belief applied to a member of a category. Experiment 3 timed a pairing of names with pleasant and unpleasant words. It did not score a trait list, a height estimate, or a sentence. Later meta-analyses of whether an IAT predicts behavior, and of whether a procedure changes the score, were not opened. No coefficient from them is printed here.

​
Common Misconceptions

Three readings turn a log score, a flat scale, or a 1995 verb into a different study.
The semantic differential mean is 0.008, d = 0.01, p = .95. The feeling thermometer mean is −8.73, d = −0.58, p = .01. Both are explicit, and they do not match. Nineteen of 26 were at zero or on the Black-preference side of the semantic differential. That count is not a claim about the thermometer. The IAT rows are a third and a fourth score. “No preference” fits one row.
Table 5 is a difference of log latencies and a difference of ratings. A negative sign means the White-preference direction in this scoring. The 179 ms is the prose average of four pairing contrasts. The students were not asked to hire, grade, or punish. The consent warning said an attitude might be revealed. It did not say a discriminatory act had been observed.
The consciousness-raising sentence says the strategy may have some value when the source of implicit bias is not properly identified. The authors say they are not reviewing the research on that strategy. Experiment 3 administered no workshop. Order of pairings was not a significant effect. Familiarity of the names is left open. A training claim needs a study that trained someone and then counted a later act. These two papers are not that study.
These six pages sit beside Implicit Bias. Each one asks a different question.

Stereotype Effect

A belief about a category is applied to a member. The 1995 paper defines an implicit stereotype with the same template. Experiment 3 timed names and pleasant words, not a trait list.

Halo Effect

An attitude toward a person spreads onto another attribute. The 1995 review treats some halo findings as implicit attitude effects. This sample did not rate a second trait.

In-Group Bias

A preference for one’s own group. The 1998 discussion uses that idea for the White preference on the IAT. The score here is still a latency in one pool of 26.

Confirmation Bias

A search that favors what one already thinks. The IAT does not ask the student to choose which material to read. It asks for a speeded sort.

Naive Realism

One’s own view feels like a plain report of the facts. The 1998 authors talk about private self-presentation. They do not score that feeling.

Self-Fulfilling Prophecy

A belief changes a later outcome. Experiment 3 ends at the envelope. No later outcome is in the file.

​
One-Line Takeaway

Before you treat one IAT number as a person’s verdict, write the sign, put the semantic differential next to the thermometer, and say that Experiment 3 timed a key. The prose average was 179 ms, and one of 26 IAT scores was positive.