A new Biomedicines study finds that practicing gastroenterologists perform only slightly better than chance when distinguishing real from AI-generated colonoscopy images. In tests with 32 physicians reviewing real, augmented, and fully synthetic polyp images, overall accuracy was just 61.2%, with no significant difference between residents and senior doctors.
While all CycleGAN-generated images were flagged as synthetic, diffusion-generated ones fooled experts nearly 80% of the time—raising concerns about the need for automated authenticity safeguards.