Even advanced reasoning models such as OpenAI’s o3-mini and DeepSeek-R1 generate clinical vignettes with substantial demographic skews, under- or over-representing racial and gender groups relative to real-world U.S. epidemiological baselines in the majority of tested conditions. 

Aggregate performance metrics (accuracy, AUC, sensitivity, etc.) can mask clinically important failures concentrated in specific patient subgroups, a pattern previously observed in population-health algorithms and medical-imaging systems. 

Meaningful safety requires pre-specified subgroup validation alongside aggregate results, continuous post-deployment monitoring for drift, treating insufficient subgroup data as uncertainty rather than assuming fairness, and applying the same lifecycle scrutiny to both regulated devices and general-purpose models used in clinical workflows.

New reasoning models still reproduce medical stereotypes. Safety evaluation should expose subgroup failures before aggregate performance hides them.

Last week, Eurasia Review highlighted a study with an unsettling result for anyone assuming that better artificial intelligence reasoning will automatically produce fairer medical output. Researchers generated 36,000 clinical vignettes with OpenAI’s o3-mini and DeepSeek-R1 across 18 medical conditions. In 14 of 18 conditions for o3-mini and 16 of 18 for DeepSeek-R1, at least one racial group was represented more than 20 percentage points away from published U.S. epidemiological baselines. Gender representation also diverged substantially across many conditions.

The study deserves one important boundary. It did not test diagnosis on real patients, and neither model was being evaluated as a cleared medical device. Generated vignettes are not clinical outcomes. What the study does establish is narrower and still consequential: stronger reasoning performance did not erase demographic defaults. A model can become more capable on one axis while retaining a weakness on another.

That........

© Eurasia Review