OpenAI's new chatbot, ChatGPT Health, underestimated the severity of medical emergencies, according to a study published in Nature Medicine. The study tested ChatGpt Health's ability to triage, or assess, medical cases based on real-life scenarios. The researchers found that the chatbot under-triaged 51.6% of emergency cases, instead of recommending the patient to go to the emergency room within 24 to 48 hours. The bot also over-traged 64.8% of nonurgent cases. The chatbot is not intended for diagnosis or treatment, but is designed for follow-up questions rather than giving one response per case. The company is currently working to improve the safety and reliability of the model before it is made more widely available.