ChatGPT's social trait judgments align with human impressions (www.psypost.org)

A new study in Computers in Human Behavior: Artificial Humans finds that ChatGPT’s judgments of facial social traits—attractiveness, dominance and trustworthiness—largely align with average human impressions. Using neutral-expression photos from the Chicago Face Database, researcher Robin S.S. Kramer presented 360 paired comparisons and found ChatGPT agreed with human ratings over 85% of the time (especially for attractiveness). When asked to rate individual faces, ChatGPT’s scores correlated moderately with mean human ratings (~0.52) and showed reasonable internal consistency (test-retest ~0.64 versus human average ~0.74). The model also reproduced a strong “attractiveness halo”: in 92.5% of trials where one face was judged more attractive, ChatGPT also favored that face for traits like intelligence or sociability. Technically, these results are consistent with ChatGPT having learned image–text associations from large training sets rather than an explicit visual model—its image inputs are converted into text-like representations and scored via learned associations. The paper notes limited evidence of overt racial preference in extreme cross-group pairings (ChatGPT chose the higher-rated face in 58/60 such tests) but warns this cannot rule out subtler biases. Important caveats include the controlled, neutral-photo stimulus set, ChatGPT’s non-deterministic outputs, and the difficulty of benchmarking AI–human agreement when humans themselves disagree. Implications include the potential for LLMs to mirror and amplify human social stereotypes, influence generated face content, and affect downstream decisions (hiring, sentencing) — all motivating more nuanced bias auditing and limits on image-sharing in research.