A study published in Nature Medicine found that OpenAI's ChatGPT Health chatbot recommended non-emergency care for 51.6% of scenarios that actually required an ER visit. Researchers at Mount Sinai Hospital tested 60 medical scenarios with 16 demographic variations each, comparing the AI's triage recommendations against those of three physicians.

The chatbot correctly identified unmistakable emergencies like strokes with 100% accuracy, but consistently missed more nuanced crises. It under-triaged diabetic ketoacidosis and impending respiratory failure, suggesting patients see a doctor within 24–48 hours instead of going to the ER. Lead author Dr. Ashwin Ramaswamy said the bot appeared to be "waiting for the emergency to become undeniable."

The study also found ChatGPT Health over-triaged 64.8% of non-emergency cases, recommending ER visits for issues as mild as a three-day sore throat. It inconsistently referred users to the 988 suicide hotline — sometimes when unnecessary, sometimes failing to do so when needed.