Study Finds ChatGPT Health Often Says 'Not an Emergency', Yet Patients Should Go to A&E

Health feature on OpenAI’s chatbot, ChatGPT Health, is said to frequently underestimate the severity of medical conditions in several emergency situations, according to a study published in Nature Medicine recently. In the study, researchers assessed ChatGPT Health’s ability to perform triage—the process of assessing the urgency of a patient’s condition and determining whether someone should be admitted to the emergency department (A&E) immediately or simply see a doctor within a given timeframe. The results showed that ChatGPT Health misclassified 51.6 percent of emergency cases. In other words, in more than half of the scenarios that should have prompted immediate attendance at the emergency department, the chatbot instead advised patients to delay treatment and consult a doctor within 24 to 48 hours. Some of the scenarios tested included serious conditions such as diabetic ketoacidosis (a life-threatening complication of diabetes) and respiratory failure that could be fatal if not treated promptly. “Any doctor with medical training would say such a patient should be taken to the ED immediately,” said the lead author of the study, Dr. Ashwin Ramaswamy of Mount Sinai Hospital in New York. According to Ramaswamy, in some cases the ChatGPT Health feature appeared to wait until the situation was clearly an emergency before advising hospitalisation. On the other hand, this feature, officially launched in early January 2026, also showed another error—over-triage, i.e., overestimating the severity of cases that are not emergencies. Around 64.8 percent of non-emergency cases were advised to see a doctor, whereas home treatment would have sufficed. For example, a patient with a mild sore throat for three days was still advised to consult a doctor within 24 to 48 hours, whereas self-care would have sufficed. Nevertheless, the study also found that ChatGPT Health could identify some emergencies well; for instance, clearly stroke-like symptoms were detected as emergencies. OpenAI responded to the study by saying that the research does not fully reflect how ChatGPT Health is typically used. According to the company founded by Sam Altman, the health chatbot is designed for interactive use, where users can provide additional information through follow-up questions, not just a single scenario. OpenAI also emphasised that ChatGPT Health is not intended for diagnosis or treatment, but as an additional source of health information.