Enhancing human detection of real and hyperrealistic AI-generated faces

Recent advances in generative artificial intelligence (AI) have prompted concern about the proliferation of deepfake content—synthetic images, audio, and videos—online. StyleGAN2-generated images of White human faces are especially problematic, as people often perceive them as more real than real human faces—a phenomenon known as AI hyperrealism . Across four preregistered experiments ( N = 661), we developed and tested DISCERN-AI, a novel behavioral intervention that combines feature-based instructions with inductive category learning to enhance human discrimination between real and StyleGAN2 faces. Participants consistently showed above-chance discrimination performance following the intervention, including after a 20-day delay. Overall, the results suggest that category learning regimes, such as DISCERN-AI, can be effective tools for improving human detection of synthetic facial images.