🚨🤖 How safe is AI triage at consumer scale? A structured stress test of ChatGPT Health (launched Jan 2026) evaluated triage accuracy using 60 clinician-authored cases across 21 domains (960 total responses). 📊 What emerged: Performance followed an inverted U-shape — most dangerous failures occurred at clinical extremes. ⚠️ Emergency cases: • 52% of gold-standard emergencies were under-triaged • Examples: diabetic ketoacidosis & impending respiratory failure → advised 24–48 hr follow-up instead of ED • Classical emergencies (e.g., stroke, anaphylaxis) were correctly triaged 🟡 Non-urgent cases: • 35% showed inappropriate escalation 🧠 Cognitive bias vulnerability: When family/friends minimized symptoms (anchoring bias), recommendations shifted significantly (OR 11.7) — mostly toward less urgent care. 💬 Crisis safeguards: Suicide-related prompts triggered crisis messaging inconsistently — sometimes more often when no specific method was described. 📉 Equity findings: No statistically significant differences by race or gender, though CIs could not exclude meaningful disparities. 🔎 Bottom line: Missed high-risk emergencies + inconsistent crisis activation raise safety concerns. Prospective validation is essential before large-scale consumer deployment of AI-driven triage systems. #AIinHealthcare #DigitalHealth #PatientSafety #ClinicalAI #HealthTech #MedicalEthics nature.com/articles/s41591-0…