AI models reproduce racial and gender stereotypes in medicine

Research indicates that advanced artificial intelligence models continue to reproduce racial and gender stereotypes within medical contexts, potentially worsening healthcare disparities.

A study from Flinders University involving next-generation reasoning Large Language Models (LLMs)—specifically o3-mini and DeepSeek-R1—found that improved reasoning capabilities do not inherently improve representational fairness. After generating 36,000 clinical vignettes, researchers noted that these models frequently misrepresented the distribution of race and gender in medical conditions, mirroring biases previously observed in GPT-4.

In dermatology, studies have shown that AI algorithms can perform significantly worse at detecting skin cancer on darker skin tones compared to lighter skin. Additionally, researchers noted that algorithms used to allocate healthcare resources have historically relied on flawed proxies, such as healthcare spending, which can disadvantage Black patients due to systemic socioeconomic barriers.