Entering the era of photorealistic AI-generated medical images

Abstract Artificial intelligence (AI)-generated medical images are reaching photorealistic quality, which could help medical education within visually driven specialties such as ophthalmology and dermatology. Such technology may make customizable teaching materials more accessible. However, as real and generated images become harder to distinguish, several concerns emerge. AI hallucinations may pollute real medical image databases, evidence-based medicine could be harmed, risks of fraud increase, and learners’ non-technical clinical skills and visual diagnostic ability may decline. To address these challenges, we propose a three-stage human-machine collaborative framework. Stage 1 uses AI-generated disease images for static feature recognition, Stage 2 simulates disease trajectories for dynamic clinical reasoning, and Stage 3 integrates supervised human-machine collaborative diagnosis for real-world decision-making. Our previous work demonstrated that training with AI-generated images significantly improved medical students’ diagnostic accuracy for infectious keratitis from 42.68% to 71.27% (n = 37, P < .001) and diagnostic confidence from 1.6 to 3.1 on a 5-point scale (n = 13, P < .001). This framework may help students develop essential skills for AI-integrated healthcare, including identifying AI hallucinations and maintaining diagnostic autonomy. Medical educators must shift from passively implementing AI tools to actively governing their use. This requires redefining educators as both technical instructors and ethical supervisors, building better detection tools for AI-generated materials, and adding AI education within core curricula. We aim to prepare young doctors for a future AI-integrated healthcare environment.