ChatGPT Health Misses Critical Triage Recommendations

Researchers at Mount Sinai report in Nature Medicine on 23 February 2026 a stress test of OpenAI's ChatGPT Health using 60 clinician-authored vignettes across 21 domains and 16 conditions (960 responses). The system under-triaged 52% of gold-standard emergency cases and showed high failure rates at non-urgent (35%) and emergency (48%) extremes, sometimes directing diabetic ketoacidosis to 24–48-hour evaluation. Crisis-intervention messaging fired inconsistently for suicidal ideation, and authors call for prospective validation before consumer-scale deployment.