Aversive Racism Explorer

Two equally qualified applicants, one White and one Black, are rated by a decision-maker who holds egalitarian values. When the qualifications are unambiguous, the two are recommended at the same rate; as the decision becomes ambiguous and a merit-based rationale becomes available, a gap favouring the in-group appears. Raise the ambiguity, then make the anti-discrimination norm salient and watch the gap close.

71%White45%BlackRecommendation rate
Discrimination gap: 26 points (a substantial gap). The bias is expressed only because the ambiguity supplies a non-racial reason for the choice, leaving the decision-maker’s egalitarian self-image intact.

The mapping is illustrative of the direction and rough magnitude of the aversive-racism selection studies; it is not a fitted model of any single experiment.

The Bias of Crowds

An individual’s implicit-bias score is unstable from one occasion to the next, but the average bias of the people in a place is stable and tracks that place’s structural racial disparity. Set the structural disparity of a context, then increase the number of people aggregated to watch the average settle onto the structural level while any single score stays scattered.

Individual scoresStructural levelAggregate of N
-0.50.00.51.0implicit-bias D-score
One person (score 0.76) is a noisy guide. The average of 50 people is 0.43, which has converged on the structural level 0.42. As the crowd grows, the aggregate stabilises and tracks the context’s structural disparity, which is why the bias-of-crowds account measures bias at the level of the place rather than the single mind.

Individual scores are generated by a fixed deterministic function for illustration; the point is the contrast between a noisy single estimate and a stable aggregate, not any specific data.

Implicit Association Scoring

The IAT infers an automatic association from response speed: faster responses when a racial group shares a key with “good” than with “bad” imply a positive automatic association. Edit the eight latencies in each block (milliseconds) to see how the means, the pooled standard deviation, and the D measure respond. The defaults reproduce the worked example.

Trial
Congruent (ms)
Incongruent (ms)
#1
#2
#3
#4
#5
#6
#7
#8
Congruent mean 642.5 ms; incongruent mean 720.0 ms; difference 77.5 ms. Pooled SD 126.85 ms, so D = 0.61 — a moderate automatic preference for the group paired with “good” in the congruent block.

D is relative, not absolute: it records which pairing is easier for this respondent, not a quantity of animus, and one administration is a noisy estimate.