Study finds next-generation AI medical models still perpetuate racial and gender stereotypes

A Flinders University study published in August 2026 reveals that next-generation reasoning AI models, o3-mini and DeepSeek-R1, still perpetuate racial and gender stereotypes in medicine. Analyzing 36,000 clinical vignettes, researchers found that these newer models show no improvement over GPT-4, frequently overrepresenting Black patients in stereotyped conditions and exaggerating majority genders. Experts warn this risks reinforcing biased assumptions in clinical workflows.

Medical AI bias evaluation

• ▪Flinders University researchers evaluated the next-generation reasoning large language models o3-mini and DeepSeek-R1 by generating 36,000 unique clinical vignettes describing fictional patients.
• ▪A study published in the Journal of Medical Internet Research in August 2026 found that advancements in artificial intelligence reasoning do not inherently improve representational fairness in medical contexts.

Racial stereotype persistence

• ▪The o3-mini and DeepSeek-R1 models overrepresented Black populations in clinical vignettes for stereotypically associated conditions like sarcoidosis, systemic lupus erythematosus, preeclampsia, and essential hypertension.
• ▪Qualitative analysis of DeepSeek-R1's reasoning traces showed the model explicitly invoked disease-demographic associations when selecting patient demographics without referencing quantitative epidemiological data.

Gender stereotype persistence

• ▪The Flinders University study found significant gender misrepresentation in 56% of evaluated medical conditions for the o3-mini model and 67% of conditions for the DeepSeek-R1 model.
• ▪The o3-mini and DeepSeek-R1 models frequently misrepresented gender distributions in medical conditions, with outputs consistently exaggerating the majority gender and skewing toward healthcare stereotypes.

GPT-4 comparison results

• ▪The o3-mini and DeepSeek-R1 models showed no improvement in racial representation compared to GPT-4, with significant racial misrepresentation occurring in 78% of conditions for o3-mini and 89% for DeepSeek-R1, compared to 67% for GPT-4.
• ▪The o3-mini and DeepSeek-R1 models exhibited higher median misrepresentation of Black populations in stereotyped conditions, at 44% and 31% respectively, compared to 15% in GPT-4.

Clinical workflow monitoring needs

• ▪Flinders University researchers warned that consistently overrepresenting demographic groups in clinical AI outputs risks reinforcing narrowed demographic assumptions in diagnostic reasoning.
• ▪The Flinders University study authors recommended continuous monitoring of potential biases and awareness of demographic defaults before integrating large language models into clinical workflows.

1 source