Deepfake smiles matter less—the psychological and neural impact of presumed AI-generated faces

Abstract

High-quality AI-generated portraits ("deepfakes") are becoming increasingly prevalent. Understanding the responses they evoke in perceivers is crucial in assessing their societal implications. Here we investigate the impact of the belief that depicted persons are real or deepfakes on psychological and neural measures of human face perception. Using EEG, we tracked participants' (N = 30) brain responses to real faces showing positive, neutral, and negative expressions, after being informed that they are either real or fake. Smiling faces marked as fake appeared less positive, as reflected in expression ratings, and induced slower evaluations. Whereas presumed real smiles elicited canonical emotion effects with differences relative to neutral faces in the P1 and N170 components (markers of early visual perception) and in the EPN component (indicative of reflexive emotional processing), presumed deepfake smiles showed none of these effects. Additionally, only smiles presumed as fake showed enhanced LPP activity compared to neutral faces, suggesting more effortful evaluation. Negative expressions induced typical emotion effects, whether considered real or fake. Our findings demonstrate a dampening effect on perceptual, emotional, and evaluative processing of presumed deepfake smiles, but not angry expressions, adding new specificity to the debate on the societal impact of AI-generated content.