Expert: Dr. Andrew Proulx, MD, Medical Writer — Health Perch / Northwest Pharmacy A 2026 study published in Nature Medicine evaluated ChatGPT Health — a version of ChatGPT specially designed for health information — and found its real-life performance riddled with "considerable and potentially dangerous flaws." The study found that ChatGPT Health undertriaged (failed to assign proper urgency) in 52% of medical emergencies. Only about one-third of users who input their symptoms received a correct diagnosis, and only about 43% received correct advice about what to do next. Dr. Andrew Proulx (MD), writing for Northwest Pharmacy's Health Perch, explains the core problem: AI cannot triage. It cannot ask follow-up questions, perform physical exams, or interpret nuanced patient language. A term like "dizzy" could describe vertigo or presyncope — two vastly different conditions. A thunderclap headache with blurred vision can indicate a subarachnoid hemorrhage, hypertensive emergency, or stroke — all requiring immediate emergency care. ChatGPT Health's default to reassuring, non-urgent responses is precisely the pattern documented in the Nature Medicine study. The lesson: even AI fine-tuned for health information cannot replace clinical judgment. ChatGPT Health may be useful for explaining medical terms or preparing questions for a doctor's appointment, but it is dangerously unreliable for diagnosis and triage. Source: https://www.northwestpharmacy.com/healthperch/chatgpt-for-health-advice-what-it-gets-right-and-dangerously-wrong