When people try to infer how well someone will do on a task, they often rely on group characteristics. For example, people who graduated from university may be perceived as more conscientious, leading employers to prefer those with degrees even for jobs that do not require specific knowledge related to the degree. In many cases, these inferences are likely to be accurate. However, we examine how such beliefs can emerge when they are false. Across four experiments with more than 9,000 participants, we show that people look at the demographics of "top performers" on a task and use that information to learn what characteristics are associated with success. However, when doing so, they fail to adjust for the population from which the top performers emerge.

When participants see that four out of five top performers are men, they are more likely to select male candidates for a task. This is true even when the top performers emerge from a group that has four times as many men as women, reflecting the gender imbalance in some industries, and hence the information would suggest that there are no group differences. We also document this phenomenon when top performers emerge from a population that is representative of the racial makeup of the United States. While, in our data, non-White performers complete more tasks than White performers, they are sparsely represented among the top performers because they make up only approximately 20% of the US population. As a result, participants are less likely to hire (and hence discriminate against) non-White participants, falsely believing that they perform worse.

While much prior research has focused on biases and stereotypes as a source of discrimination, we show that errors in statistical reasoning can also lead to discrimination. Unlike stereotypes, this kind of error suggests that any numerically smaller group is disadvantaged, because they are less visible among top performers. Which demographic or other social group is a numeric minority may then differ depending on the particular context, suggesting a need for more targeted interventions.