1 of 1
Story summary
- ChatGPT Health, launched January 2026 by OpenAI, was evaluated in a structured stress test that revealed safety concerns and called for validation before consumer deployment.
- The stress test used 60 clinician-authored vignettes across 21 clinical domains to assess triage recommendations.
- The system under-triaged 52% of gold-standard emergencies, including diabetic ketoacidosis.
- Triage recommendations shifted toward less urgent care when family or friends downplayed symptoms.
- Crisis intervention messages were inconsistently activated.
