Cross-Modal Demographic Bias in Generative AI Representations of Physicians: A Comparative Audit of Large Language Models and Text-to-Image Models.

With the rapid expansion of generative artificial intelligence (AI) in text generation and image synthesis, algorithmic bias has become an increasingly important concern in medicine, where professional authority, public trust, and representational fairness are closely intertwined. This study employed a dual-path audit framework to examine demographic bias in AI-generated physician representations across eight clinical specialties using both textual and visual analyses. The findings show that neutral prompts do not necessarily produce demographically neutral outputs. All three large language models spontaneously introduced demographic attributes, with gender and age appearing far more frequently than race, skin tone, or body type. In the visual modality, the three text-to-image models differed significantly in their distributions of perceived gender, age, and skin tone. Comparisons with benchmark data from the Association of American Medical Colleges (AAMC) further demonstrated that AI-generated outputs deviated significantly from the actual physician workforce composition at both the overall and specialty levels. These results suggest that generative AI does not simply mirror real-world physician demographics; rather, it selectively reorganizes and amplifies existing social associations, occupational stereotypes, and structural inequalities. This study provides empirical evidence of cross-modal representational bias in generative AI and underscores the need for more fine-grained fairness evaluation and responsible AI governance in medical contexts.