Perceiving Emotion in AI-generated and Real Human Faces: Comparative Judgments of Valence, Intensity, and Authenticity

Abstract

The study focuses on influences of perceptual cues (facial expression and image origin) and contextual cues (source attribution) on emotion perception of AI-generated and real human facial expressions. Using a 2 (image origin: AI vs. real) × 2 (emotion: angry vs. smiling) × 3 (instruction: AI vs. real vs. none) mixed factorial design, valence, intensity and authenticity ratings were collected. Results indicated that perceptual cues highly influenced the three ratings. Emotion expression strongly influences authenticity. Smiling expressions were rated as more authentic than angry expressions. Image origin significantly influenced valence and intensity. Real faces were more consistently categorized as negative for angry expressions and more emotionally intense than AI-generated faces. No significant main effects of instruction-based contextual cues were found except for a limited interaction for valence, with group differences for angry expressions. Emotion authenticity ratings did not differ between AI-generated and real faces. Overall, the findings support that bottom-up perceptual cues may provide the primary affective signals to top-down contextual cues which plays a modulatory role under specific conditions. These findings have implications for understanding social perception in human–AI interaction and the increasing perceptual plausibility of AI-generated facial expressions.