ChatGPT Health misses many medical emergencies, study finds

A Nature Medicine safety evaluation found ChatGPT Health under‑triaged more than half of cases needing immediate care and sometimes failed to show crisis help for suicidal intent, prompting expert concern.

An independent safety study published in Nature Medicine tested ChatGPT Health with realistic clinical scenarios. Researchers compared the platform's recommendations to assessments from three independent doctors using clinical guidelines. The study found the system under‑triaged a substantial share of cases that clinicians judged required immediate hospital care and also showed inconsistent crisis banner behaviour for suicidal statements. OpenAI has said it welcomes independent research and continues to update the model. Key findings: - In scenarios judged by doctors to require immediate hospital care, the platform recommended staying home or booking routine appointments in a majority of cases (reported as 51.6%). - In tests involving suicidal ideation, a crisis intervention banner appeared when the patient statement stood alone but vanished after researchers added routine lab results (zero out of 16 attempts in that setup). - The system's recommendations changed in some scenarios when details such as comments from others or test results were added, and researchers reported instances where severe breathing difficulty was directed to a future appointment rather than urgent care. Summary: The study indicates gaps in how ChatGPT Health recognises certain urgent medical situations and in how safety guardrails respond to contextual details. OpenAI says the model is continuously refined and that independent evaluations are welcome, while experts call for clearer safety standards and independent auditing. Undetermined at this time.