Large Language Models Show Human-Like Bias Against People With Highly Stigmatized Health Conditions

Large language models (LLMs) produced systematically different, and often more negative, judgments about fictional characters described as having stigmatized health conditions compared to healthy characters, according to a recent study. The largest gaps in these judgments appeared when characters were described as having mental health disorders or highly stigmatized physical conditions such as human immunodeficiency virus (HIV) or hepatitis B virus (HBV). These gaps emerged even though the same LLMs tended to avoid making overtly stigmatizing statements when asked directly about these conditions.

Six LLMs across 51 different scenarios were evaluated and the scores on six established stigma scales to . . .