GPT-5 Amplifies Medical Bias Against Marginalized Groups, Mount Sinai Study Finds

A September 2025 preprint from researchers at the Icahn School of Medicine at Mount Sinai tested OpenAI's GPT-5 on 500 emergency department cases, each replayed 32 times with only the patient's sociodemographic label changed. The model consistently assigned higher urgency scores and recommended less advanced diagnostic testing for historically marginalized groups, including Black patients, unhoused individuals, and LGBTQIA+ patients. Mental health screenings were flagged in 100 percent of cases involving certain LGBTQIA+ labels, compared with 41 to 73 percent for other demographic groups under the previous GPT-4o model. A companion study published in Nature Medicine evaluated nine large language models across 1.7 million outputs and found similar patterns across all models tested, suggesting the problem is systemic rather than specific to any single developer. Researchers warn that deploying such tools in clinical settings without addressing embedded bias poses real risks to patient care and raises unresolved legal and procurement questions for hospitals and insurers.