Lillie C. del Real¹, Jessica N. Goetz¹, Mark B. Neider¹; ¹University of Central Florida

Humans typically classify deepfake face images at near chance levels (50%). However, they also appear to experience synthetic and real faces differently. For example, humans attribute higher ratings of certain perceptual judgments (e.g., attractiveness) to synthetic faces compared to real faces. Could these percepts serve as guided heuristics to aid in distinguishing between real and synthetic faces? Across two experiments, our goal was to identify percepts that were indicative of face type and prompt participants to use those percepts to inform classification of real and synthetic faces. In Experiment 1, 124 participants were randomly assigned to one of four conditions in which they rated real and synthetic StyleGAN2 faces. There were three heuristic conditions where participants rated each face on one facial percept (attractiveness, symmetry, and unusualness) and an unguided condition where participants rated only on the face ground truth (real/synthetic). Participants in the heuristic conditions rated synthetic faces more attractive, symmetrical, and usual than real faces (all ps < .001) without knowing the face ground truth, suggesting these percepts could serve as indicators for deepfake classification. In Experiment 2, 96 naïve participants were assigned to one of the four conditions where they conducted the same rating task as Experiment 1. Participants then used their rating of the face to inform their real/synthetic classification. Results suggested that larger rating differences between real and synthetic faces, specifically in the symmetry, unusualness, and unguided conditions, were associated with higher classification accuracy (all ps < .001). Thus, when making perceptual judgments of faces, individuals who utilized a wider range of the ratings scale tended toward higher deepfake detection accuracy. While guided heuristics emphasizing face attributes may serve as useful cues for classification, the effectiveness of those heuristics may depend on one’s ability to differentiate between synthetic and real faces in perceptual judgment space.