Unmasking bias in artificial intelligence: sex and racial representation in critical care medicine through text-to-image generators

Background The development of artificial intelligence (AI) has revolutionised medicine, impacting areas such as radiology, patient education, and clinical training. However, concerns about AI perpetuating societal biases are increasing. This study examines whether AI-generated images of critical care medicine physicians align with real-world demographic data, focusing on female and non-White representation. Methods We conducted a cross-sectional study comparing 300 AI-generated physician images of adult and paediatric critical care medicine attendings and trainees from DALL-E 3, Midjourney 6.0, and Stable Diffusion 3.0 with demographic data from the Association of American Medical Colleges (AAMC) and Graduate Medical Education. Images were assessed by two reviewers with assistance from the Chicago Face Database. Discrepancies were resolved by consensus. AI and real-world data were compared using χ 2 tests. Results For adult critical care medicine, the proportion of female physicians was 27.3% in real-world data, while AI platforms showed 78.3% (P<0.001) on DALL-E 3, 9.3% ( P <0.001) on Midjourney 6.0, and 19.7% ( P =0.003) on Stable Diffusion 3.0. For paediatric critical care medicine, where the AAMC reported a female representation of 49.6%, the AI-generated figures were: DALL-E 3 at 77.3% ( P <0.001), Midjourney 6.0 at 20.7% ( P <0.001), and Stable Diffusion 3.0 at 60.3% ( P <0.001). Additionally, AI overrepresented non-White physicians, with DALL-E 3 showing a striking 90.0% ( P <0.001) compared to the AAMC's 45.6%. Conclusions Significant discrepancies exist between AI-generated critical care medicine physician images and real-world demographics. Midjourney 6.0 underrepresents, while DALL-E 3 overrepresents female physicians. DALL-E 3 overrepresents non-White physicians, while other platforms more accurately depict demographics in paediatric critical care medicine. These findings highlight potential biases that have important implications for education, clinical decision-making, and public perception.