Generative Artificial Intelligence for Medical Image Creation in Health Professions Education: a Scoping Review.

Generative artificial intelligence (AI) can create synthetic medical images, enabling new educational applications, but current tools have not been systematically validated for educational fidelity, safety, or equity. This scoping review maps how generative AI systems are used for medical image creation in health - professions education, summarizes reported benefits and risks, and identifies evidence gaps. Searches of PubMed, Embase, Scopus, and grey literature (January 2010 to May 2025) identified 19 eligible studies. Diffusion models (e.g., Stable Diffusion, DALL - E, Midjourney) dominated recent work, with applications concentrated in dermatology, ophthalmology, and anatomy; radiology and pathology were rarely represented. Several studies reported short - term gains in learner performance or engagement, including small controlled trials, but study designs, outcome measures, and follow - up durations varied widely. Across the literature, important harms and failure modes were common: medically inaccurate or anatomically implausible outputs (including hallucinated structures), very low task accuracy in some domains (e.g., synthetic ECG generation reported 32. 7% accuracy), inconsistent evaluation metrics, and limited external validation. Multiple studies also documented demographic bias and stereotyping in generated images, indicating that uncritical use may worsen rather than solve diversity and representation problems. Overall, generative AI for image creation should be treated as an experimental adjunct requiring rigorous human - in - the - loop review, bias auditing, and alignment with educational theory. Future research should adopt standardized frameworks for synthetic - image evaluation (e.g., fidelity and cognitive load considerations), include study quality assessment, and test long - term learning transfer and clinical impact.