Same Symptoms, Different Care

Short summary

Peer-reviewed research from Mount Sinai shows GPT-5 and other LLMs amplify sociodemographic biases in clinical triage, assigning different care recommendations based solely on patient race, income, or LGBTQIA+ status despite identical clinical presentations. A companion Nature Medicine study across nine models and 1.7M outputs confirmed bias is a category-wide property, not a single-vendor issue. The findings raise urgent procurement, design, and legal questions for hospitals adopting AI diagnostic tools.

• •GPT-5 showed no bias improvement over GPT-4o; LGBTQIA+ labels flagged for mental health eval 100% of the time vs 41-73% for other groups
• •Nature Medicine study across 9 LLMs and 1.7M outputs confirmed bias is category-wide, not vendor-specific
• •Implications span clinical design, institutional procurement, and legal liability for healthcare AI adoption

Generated with AI, which can make mistakes.