Generative AI text-to-image systems are increasingly used in medical education due to their speed and apparent visual realism, yet they introduce distinct safety and equity risks. Although empirical evaluations are accumulating, evidence on representational bias and clinical fidelity remains fragmented. We conducted a PRISMA-guided systematic review to synthesize findings from 36 empirical studies evaluating AI-generated images in medical teaching, assessment, and patient education contexts. Most studies (80.6%) evaluated DALL·E-based tools. Representational bias was pervasive: 75% of studies reported significant demographic skew, particularly in studies that examined race (66.7%) and gender (58.3%), often within the same study; generated clinicians were often depicted as predominantly white and male. Clinical fidelity limitations were reported in 17 studies (47.2%), ranging from anatomical hallucinations to plausible but incorrect depictions of medical equipment. Importantly, bias and fidelity were often coupled; high visual plausibility can mask clinically or socially consequential distortions, potentially anchoring incorrect mental models. We conclude that AI-generated images should not be treated as neutral educational resources. Safe integration requires a shift from passive adoption to active governance, including expert curation and the incorporation of visual AI literacy into medical curricula.
Alon et al. (2026) studied this question.