How AI betrays itself through unnatural eye contact in group photos

Generated images now hide pixel-level artifacts so well that traditional detection fails, especially in photos with multiple people. This work exploits a different weakness: AI struggles to make eye gaze, head position, and pupil alignment coherent between interacting individuals. Using a custom dataset with controlled eye perturbations and a novel caption-training method, the approach improves detection accuracy from 67.8% to 71.5% on multi-person images and transfers across different AI generators. Code will be released.