Enhancing Human Detection of Real and AI-Generated Hyperrealistic Faces

Recent advancements in generative artificial intelligence (AI) technologies have prompted concerns about the proliferation of deepfake content—synthetic images, audio, and videos—online. StyleGan2, a powerful generative adversarial network architecture, can generate synthetic images of human faces that appear more realistic than actual human faces—a phenomenon termed AI hyperrealism. Across four preregistered experiments (total N = 661), we tested whether a novel behavioral intervention, termed DISCERN-AI, improved people’s ability to discriminate between real and synthetic StyleGan2 images. DISCERN-AI combines instructions about which visual features are diagnostic for classifying real and synthetic images with an inductive learning training protocol. In Experiment 1, participants completed a pre-test, DISCERN-AI, and a post-test. In Experiment 2, participants underwent DISCERN-AI or a control task before completing a post-test. In Experiments 3A and 3B, participants completed a shortened version of the intervention or control task, followed by an immediate post-test and a second post-test after 20 days. In all experiments, participants showed below-chance discrimination performance (AI hyperrealism) in the pre-test and control conditions. After completing DISCERN-AI, however, participants consistently showed above-chance discrimination performance, even when tested after 20 days. Together, the results provide support for an easily scalable intervention that substantially improves people’s ability to discriminate between real and synthetic StyleGan2 images.