Human beings have a tendency to infer personality or character traits from other people’s facial features, and these biases—ungrounded in any actual relationship between faces and behavior—lead to unfair outcomes.
Steven Lehr and colleagues explored whether AI models, which are trained primarily on text but have the ability to “see” images, share these biases. The authors asked GPT-4o to make over 4,500 forced-choice judgments between computer-generated faces, and presented thousands of additional forced choices to GPT-5, Gemini 3 Flash Preview, and Claude Sonnet 4.5. In some of the experiments, models were asked to choose which of two faces was more competent or more trustworthy. Other trials used related traits, including asking which face was more confident, smart, hardworking, lazy, inept, careless, warm, helpful, sincere, selfish, hypocritical, or aggressive. Some experiments asked LLMs to judge which computer-generated human face would be more likely to be a serial killer, to be arrested for human trafficking, or to defraud the public using a Ponzi scheme. Finally, LLMs were asked to choose between faces in the contexts of hiring a university president, investing in a tech startup, or selecting a financial manager. In all these cases, the models were willing to weigh in and in the majority of cases chose the same face that a human would typically see as more trustworthy or confident. Across studies, GPT-4o selected the face that would be expected based on human ratings 74.88% of the time. GPT-5 showed notably more bias than its predecessor. When GPT-5 advised on consequential decisions, the model recommended the more competent-looking individual fully 97.04% of the time, as compared to GPT-4o’s 75.19%. Models from other companies produced similar results.
If LLMs were free of human face-to-character bias, they could be used as a tool to help eliminate this form of bias in contexts such as job candidate selection or parole decisions. According to the authors, AI in its present form is instead likely to worsen unfairness if used in such contexts.
PNAS Nexus
Like humans, language models demonstrate face-to-character biases
18-Aug-2026
S.L. is affiliated with Cangrade, Inc., a company that works on de-biasing machine learning models. However, Cangrade, Inc. does not currently create generative AI models, did not fund this research, and is not expected to profit in any way from the results or their publication, other than by association. The other authors do not declare any competing interests.