Heaven Dixon¹, Sofia DiCostanzo¹, Keith A. Schneider¹, Adrian W. Gilmore¹; ¹Department of Psychological & Brain Sciences, University of Delaware

The increasing realism of AI-generated faces presents a challenge for the human visual system, raising questions about our ability to distinguish artificial from real facial stimuli. Research examining individuals’ ability to differentiate real faces from hyper-realistic AI-generated faces suggests that behavioral performance is often at or near chance levels. Despite this behavioral limitation, EEG evidence indicates that real and artificial faces can be decoded at a neural level, with significant classification emerging as early as ~150 ms following stimulus onset (Moshel et al. 2022). However, the identity of brain regions that allow for accurate classification is not presently known. To address this gap, we used functional magnetic resonance imaging (fMRI) to examine neural responses to face stimuli in healthy adults aged 18–35. Participants viewed images of real and AI-generated faces and made a binary judgment about whether each face was real or artificially generated. AI-generated faces were produced using StyleGAN3, and the real face stimuli were drawn from the real-world face images used to train the model. Consistent with recent behavioral findings, participants performed at chance when distinguishing between real and AI-generated faces. Critically, early ventral visual regions exhibited stronger responses to real faces compared to AI-generated faces (q < .05). These findings suggest that early visual processing stages retain sensitivity to facial authenticity, even when subsequent behavioral responses do not achieve above-chance accuracy.

Acknowledgements: NSF Grant No. 2225805