AI-generated face images of emotional expressions

Abstract Recently developed artificial intelligence (AI) tools are now capable of generating realistic face images that are indistinguishable from photographs. However, little is known regarding how these synthetic images portray emotional expressions. Here, with only simple prompting, we generated a set of images depicting seven facial expressions using ChatGPT. These images were intermixed with posed facial expression photographs drawn from a standardised database and we collected participants’ judgements regarding which emotions were portrayed, along with their rated intensities. Our results showed that synthetic images were more accurately classified as their intended emotions, as well as more intense displays, in comparison with posed photographs. In addition, AI-generated images depicted action units associated with particular emotions to a similar extent to posed photos, although this pattern varied across emotions. Finally, both types of images were correctly classified by ChatGPT with few or no errors. Taken together, these findings provide insight into how AI tools represent, and subsequently depict, human emotions. Further, this approach may facilitate the creation of novel, individually-tailored image sets for use in future research by incorporating more complex and specific prompting.