A new study published in the Journal of Medical Internet Research has raised important concerns about demographic bias in advanced clinical artificial intelligence.
Researchers evaluated two newer “reasoning” large language models—OpenAI’s o3-mini and DeepSeek-R1—to determine whether their enhanced reasoning capabilities reduced racial and gender stereotyping in generated medical cases. The models produced 36,000 clinical vignettes covering 18 common medical conditions, with patient demographics compared against published United States epidemiological data.
Despite their improved reasoning performance, both models frequently generated demographic distributions that differed substantially from real-world disease patterns. At least one racial group was misrepresented by more than 20% in 78% of conditions generated by o3-mini and 89% of conditions generated by DeepSeek-R1. Gender misrepresentation above this threshold occurred in 56% and 67% of conditions, respectively.
The models particularly overrepresented Black patients in conditions commonly associated with Black populations, including sarcoidosis, systemic lupus erythematosus, pre-eclampsia and hypertension. Analysis of DeepSeek-R1’s reasoning suggested that it explicitly relied on broad disease–demographic associations when selecting patient characteristics, rather than applying quantitative epidemiological data.
These findings demonstrate that improved reasoning capability does not automatically produce more representative or equitable healthcare AI. In fact, reasoning processes may amplify stereotypical associations when models generate “typical” patients rather than patients reflecting true population diversity.
For clinicians and health organisations, the message is clear: clinical AI must be evaluated not only for accuracy, but also for fairness, demographic representation and potential downstream effects on clinical reasoning. Ongoing bias auditing, transparent validation and human oversight will be essential as increasingly capable AI systems enter healthcare education and clinical workflows.
Read more: https://www.jmir.org/2026/1/e82256