Aligned but Biased: Comparing Human and Large Language Model Intergroup Bias

<p>Large language models (LLMs) are increasingly used to evaluate people in domains such as hiring, education, and housing. A central concern is whether these systems reproduce human intergroup bias, reduce it, or express it in new forms. Here, we compare human and LLM bias using matched stereotype, decision, and text-evaluation tasks. We fielded a nationally representative Danish survey (N = 1,000), administered parallel instruments to a five-model LLM panel, replicated published audit studies in hiring, housing, and school access, and conducted a naturalistic text-evaluation experiment using real Danish social media posts from demographically identified authors whose identities were removed before evaluation. We demonstrate that LLMs reproduce the broad human stereotype hierarchy: people with a non-Western background are consistently seen as least warm and competent. We also find that this bias is often hidden in standard audit designs. When models make decisions from explicit identity cues, such as names, they generally do not reproduce human discrimination gradients. This is often because they accept or reject nearly all cases rather than make graded individualized decisions. Bias reappears when identity is implicit. Unlabeled texts written by non-Western authors receive lower ratings on warmth, competence, and employability and more negative trait descriptions. Summarizing the text before evaluation reduces these gaps, suggesting that some bias is carried by surface features of language. These results show that aligned LLMs are not simply unbiased versions of humans: They can suppress visible discrimination while retaining stereotype structures that emerge in associative judgments and evaluations of natural language.</p>