The Faces of Generative AI: Predictors of FACES

Facial Appearance as Core Expression Scales (FACES) was designed to assess maxillofacial surgery patients' perceptions of their faces. We used FACES to study ratings of non-patient participant's own faces, and faces produced by the generative AI application DALL·E.

DALL·E was used to generate 16 photo-realistic faces of males and females aged 20, 30, 40, and 50 of differing apparent ethnicities. Four stem descriptions were used for four sets of four images, for example "The face of a 40-year-old male (or female) of mixed South American and African ancestry wearing something dark in the style of a professional photo portrait." This was followed by descriptions taken from all seven FACES items, such as "the face is (not) like I want others to see me." Image generation instructions differed only in "not" being used for half of the images.

Participants (n = 333) rated images using FACES to test the hypothesis that FACES can distinguish between faces generated by positively and negatively worded instructions. That hypothesis was confirmed for all 8 image pairs.

We also found that Rosenberg Self-Esteem and State Self Esteem Scale scores predicted FACES ratings of participants' own faces. However, DALL·E was unable to depict realistic maxillofacial anomalies.