ChatGPT Health Safety Study Finds Inappropriate Guidance in Over Half of Emergency Cases

People use OpenAI’s ChatGPT Health to assess symptoms and seek advice on whether to obtain medical care, but generative AI is not a clinical diagnostic tool. Underestimating a medical emergency or mental health crisis could delay treatment. About 40 million people worldwide use AI healthcare tools each day, making triage accuracy, prompt quality and safeguards a public health concern.

A new Nature Medicine study found that ChatGPT Health misdiagnosed or failed to provide appropriate emergency-care guidance in 52% of high-certainty emergency cases. The strength of its suicide-crisis alerts was even inversely correlated with clinical risk. The study also found that the quality of users’ questions affected results, while misinformation has been ranked as the top health technology risk for 2026.