In media coverage, individuals’ membership in minority groups, such as their country of origin, is often explicitly mentioned. In contrast, majority membership is usually not highlighted. Does this reflect prejudice against minorities? A study with over 900 participants suggests that it does not. Instead, a fundamental cognitive principle of differentiation drives this pattern: People tend to notice and communicate rare or striking characteristics more readily than common and redundant ones. Artificial intelligence shows the same tendency and even amplifies it. This is reported by a team from the Social Cognition Lab at Ruhr University Bochum, Germany, led by Dr. Anna Schulte in the journal Social Psychological and Personality Science on December 1, 2025.

The research team conducted five studies involving more than 900 participants, as well as an analysis of six large language models such as ChatGPT. In one of the studies, U.S. American participants received a fictitious FBI press release about a criminal incident, including a suspect description (age, sex, weight, height, country of origin, clothing, other characteristics). The suspect’s membership in either a social majority or minority was systematically varied through the country of origin: The suspect either came from the United States (majority) or from one of the countries with the largest immigrant groups in the U.S. (Mexico, India, China, Philippines, El Salvador, Vietnam, Cuba, Dominican Republic, Guatemala, Korea; random assignment).

Participants were asked to summarize the information for a news article. The researchers analyzed whether the suspect’s origin was mentioned or not. “The results showed that the country of origin was mentioned more than three times as often when the suspect belonged to a minority, regardless of whether the participants themselves belonged to a minority,” reports Anna Schulte.

Because these findings could also be explained by motivational factors such as prejudice, the study was repeated with positive events. Instead of criminal incidents, participants read about lottery winnings or scientific breakthroughs. Here, the effect was even more pronounced: The subject’s origin was mentioned almost four times as often when they belonged to a minority. “This study is particularly central because it shows that the primary driver is the communicative focus on distinct characteristics, not a specifically negative portrayal of minorities,” says Anna Schulte.

The research team also had six different AI language models perform the same task. They created 1,000 negative and 1,000 positive scenarios and presented them to the models as prompts, asking them to summarize the information for a news article. The result: AI models mentioned minority group membership even more frequently than humans did, both in negative and positive contexts. Why the models exhibit this tendency more strongly is not yet fully understood. “The findings suggest that AI models adopt statistical patterns from their (human-generated) training data and overgeneralize existing communication tendencies. Further research is needed here,” says Anna Schulte.

“The results indicate that no deliberate disparagement of minority groups is taking place,” concludes Schulte. Instead, the researchers see a fundamental cognitive principle behind the excessive emphasis on minority characteristics. Nonetheless, this tendency results in minority groups being disproportionately represented in media contexts – which are usually negative. “We refer to this phenomenon as the ‘minority dilemma’,” the researcher explains.

Media professionals who are aware of the effect could try to mitigate it by either always or never mentioning a subject’s origin. Both approaches, however, have drawbacks: Since people are generally interested in distinct information, omitting such details may undermine trust in a news source. Conversely, consistently reporting all characteristics of all individuals may create the impression that the news source provides irrelevant and redundant information, which can also undermine trust. “One possible intervention would be to report other distinct characteristics instead, such as a person’s birthplace. This could provide sufficient distinctiveness and informativeness even for minority group members. We plan to systematically investigate such measures in a follow-up project,” says Schulte. Another important implication concerns the use of AI in news production: Since AI models not only reproduce existing biases in training data but even amplify them, media professionals should be aware of these risks when relying on AI for text generation.