Quantifying Race and Gender Bias in Otolaryngology Clinical Vignettes Generated by Large Language Models

Introduction: Large language models (LLMs) are increasingly used to generate clinical vignettes for medical education, and prior studies have demonstrated demographic bias in AI-generated content, though the magnitude of such bias and its consistency across models remain incompletely characterized. This study quantifies race and gender bias in LLM-generated otolaryngology clinical vignettes.

Methods: Three LLMs (GPT-4o, Claude Sonnet 4.5, and Gemini 2.5) were prompted in independent sessions to generate clinical vignettes for 15 otolaryngology diagnoses between November 2025 and January 2026. Each diagnosis was generated 220 times per model (9,900 total vignettes). Generated patient race and sex were compared with demographic distributions from national epidemiological databases. Bias was assessed using chi-square goodness-of-fit tests to compare LLM-generated and reference distributions. The magnitude of deviation was quantified using Cramer's V. Analyses were conducted separately for API-generated and browser-generated vignettes. Between-model differences in bias magnitude were evaluated using non-parametric tests with adjustment for multiple comparisons.

Results: Across diagnoses, all LLMs demonstrated statistically significant deviations from epidemiologic race and sex distributions (p < 0.05). Cramer's V values indicated small-to-moderate bias across models, with higher effect sizes observed in API outputs compared with browser outputs. Although Cramer's V values varied, no statistically significant differences in bias magnitude were observed between LLMs after adjustment for multiple testing.

Conclusions: LLMs generating otolaryngology clinical vignettes exhibit measurable race and sex bias relative to real-world epidemiology. Bias magnitude is similar across models, suggesting that demographic bias reflects a property of generative modeling rather than model-specific behavior. Cramer's V provides a framework for evaluating AI-generated medical content and guiding future bias-mitigation strategies.