As AI tools increasingly enter mental health care, a new preprint examines whether they reflect or distort societal stigma. The study, posted on arXiv, uses GPT-based large language models to estimate mental health stigma and compares those estimates with human perceptions, focusing specifically on differences across gender groups.

The results point to a double problem: AI estimates often diverge from what humans actually report, and the size of that gap differs by gender. This suggests that LLMs are not simply neutral mirrors of public opinion—they carry their own systematic skews that could reinforce or amplify existing disparities in how stigma is perceived and experienced.

The authors argue this misalignment matters for real-world applications. If AI systems are used to assess or respond to mental health stigma, they may act on distorted assumptions rather than ground truth, potentially leading to biased recommendations or interventions. The study underscores the need for careful evaluation of AI models before they are deployed in sensitive mental health contexts.