Familiarity With Famous Faces and the Detection of AI-Generated Images

Abstract

Artificial intelligence tools can generate face images that are indistinguishable from real photographs. This finding extends to familiar faces, where prior familiarity provides only limited improvements in detecting synthetic images. Here, we investigated whether familiarity benefits detection across multiple images of individual famous faces. In Experiment 1, performance on an in-person image-sorting task (real and ChatGPT-generated images of Keira Knightley) showed little evidence of an association with prior familiarity. In Experiment 2, accuracy in an online task (real and Grok-generated images of Taylor Swift) also showed no association with familiarity. Across both experiments, image-based analyses indicated substantial overlap between real and synthetic images in representational space, with preliminary evidence that distance from the identity centroid may predict perceived realism. Taken together, these findings suggest that greater familiarity with a famous face provides minimal benefit when detecting synthetic images of that individual, highlighting the need for alternative approaches to improving detection.