With the expanding use of artificial intelligence to obtain medical information, many users share their symptoms with tools like ChatGPT before seeing a doctor. But a new study about ChatGPT Health, published in the medical journal Nature, shows this technology may not correctly assess the urgency of a patient's condition, especially when faced with emergency cases.
To evaluate the system's performance, researchers tested 60 clinical vignettes written by physicians across 21 medical specialties and collected a total of 960 responses from ChatGPT Health. Three physicians independently determined the urgency level of each case based on medical guidelines. Cases were classified into four groups: “nonurgent” for home care and monitoring, “semiurgent” for seeing a physician within weeks, “urgent” for presentation within 24 to 48 hours, and “emergency” for presentation to the emergency department.
Results showed the AI's accuracy was higher for mid-spectrum urgencies than at the two extremes, nonurgent and emergency. The system's accuracy was 93 percent for semiurgent cases and 76.9 percent for urgent cases, but dropped to 35.2 percent for nonurgent cases and 48.4 percent in emergency situations. In 51.6 percent of cases that physicians deemed truly emergency, the AI underestimated the urgency and recommended presentation within 24 to 48 hours instead of immediate emergency department care. Conversely, in 64.8 percent of nonurgent cases, urgency was overestimated, although none of the patients were referred to the emergency department.
This problem was especially observed in conditions where disease urgency depended on its progression. In some asthma exacerbation cases, the AI detected elevated carbon dioxide levels that can be an early sign of respiratory failure, but did not deem immediate emergency presentation necessary. In diabetic ketoacidosis cases, it sometimes classified the condition as “mild” or “primary” and suggested outpatient treatment, whereas diabetic ketoacidosis is an emergency. In contrast, in a separate test including classic stroke, anaphylaxis, meningitis, and aortic dissection cases, there were no instances of underestimation of urgency.
The type of information provided also affected the AI system's performance. Adding objective data such as lab results, vital signs, and examination findings increased overall accuracy from 54.6 percent to 77.9 percent. However, for emergency cases the trend differed; with these added data, underestimation of urgency rose from 46.9 percent to 56.2 percent, although this difference was not statistically significant.
Remarks from people around the patient could also change the AI system's recommendation. In cases where determining urgency was more difficult, adding sentences such as a friend or family member saying “it's nothing” increased the probability of changing the urgency recommendation from 3.3 percent to 13.3 percent. Of the 40 changed recommendations, 52.5 percent shifted toward lower-urgency care.
In another part of the study concerning suicidal thoughts, crisis-related alerts were triggered inconsistently. In a supplementary review, the “help is available” alert appeared in only 4 of 14 vignettes and in the other 10 cases the safety alert did not appear in any of the 160 responses. In some cases where the individual reported active suicidal thoughts and a specific method for self-harm, the alert did not activate, while these alerts more often appeared in cases where the person had suicidal thoughts but did not mention a specific method.
Researchers emphasized that this study used simulated vignettes, not real patients. Also, ChatGPT Health's performance was assessed at only one point in time, between 9 and 11 January 2026, and may change with system updates. According to the researchers, AI systems used for medical guidance, particularly regarding identification of emergency conditions and consistent activation of crisis-related alerts, should be continuously evaluated.