Comparing ChatGPT with human judgements of social traits from face photographs

Facial first impressions of social traits play an influential role in our everyday lives. With the advent of artificial intelligence techniques, researchers have begun to employ such tools in the prediction of human impressions formed from the face alone. ChatGPT's latest version features the ability to interpret images as input, and so begs the question: does the chatbot's judgements of social traits from face images align with human judgements? To this end, I carried out a series of studies utilising a pre-existing face image set and its accompanying norming data. In Study 1a, with a focus on three core trait dimensions (attractiveness, dominance, and trustworthiness), I presented ChatGPT with pairs of faces which had been rated as high versus low on a given trait. For the majority of pairs, the chatbot's responses aligned with human judgements. In Study 1b, I found that ChatGPT's ratings of attractiveness showed medium to large associations with those provided by human observers. Finally, I investigated the possibility of biases in the chatbot's perceptions. While Study 2 found no support for an extreme form of race bias in judgements of social traits, the results of Study 3 providing evidence of an attractiveness halo effect – more attractive faces were also judged to be more confident, intelligent, and sociable. Taken together, these results suggest that ChatGPT's responses align with human judgements of social traits, including the presence of a halo effect. As such, I discuss some of the implications for ChatGPT's use across several domains.