Isabel Gauthier¹; ¹Vanderbilt University

Hyperrealism---judging AI-generated faces as real more often than genuine photographs---has been attributed to algorithmic bias that makes White AI faces statistically prototypical. The present work suggests this account is insufficient. Using poor-quality webcam photographs as real faces and ChatGPT-4o generated faces approximating their aesthetic, we studied how object recognition ability (o) predicts accuracy in AI face categorization across two studies with 68 and 101 participants. Study 1 paired faces differing in race and gender on each trial; Study 2 matched the same faces within trials on these attributes, elevating the relatively more polished appearance of AI faces as a salient cue. Hyperrealism was high in both studies and not limited to White faces. Most observers misjudged AI faces as real, doing so significantly more in Study 2. This demonstrates hyperrealism as a product of comparison context rather than fixed image properties, because the faces were the same across studies. Most observers showed high sensitivity to differences between real and AI faces (high |d'|), but their correct category assignment depended on o. O significantly predicted performance after controlling for age, gender, and general intelligence. This extends results from Chow et al. (2025), where o predicted AI face detection with StyleGAN faces. In Study 2's more challenging context, o predicted the rate of learning correct categorizations over the last 80% of trials, without receiving feedback. High-o observers appear less dependent on single salient cues and better able to integrate multiple subtle features (for instance, looking for different kinds of imperfections to identify real faces). A high o was not necessary to discriminate AI faces from real ones, but it predicted correct inferences about category membership. These results suggest that o could help identify who would benefit most from training to detect AI-generated content.

Acknowledgements: This work was supported by the David K. Wilson Chair Research Fund at Vanderbilt University and NSF award 2316474 to Isabel Gauthier.