Popular AI Chatbots Miss 1 in 5 Mental Health Emergencies

When you talk to an AI chatbot about mental health, you probably expect it to pick up on warning signs. But a new study shows that even the most advanced chatbots often miss emergencies. Researchers tested 13 leading chatbots in 10-turn conversations with simulated users based on 112 real mental-health cases, each rated by clinicians for urgency. The chatbots failed to recognize 20% of emergencies—four times the 5.6% failure rate when all the clinical details were presented in a single message.

The issue isn't that the chatbots lacked information. By the end of the conversation, the critical facts had almost always been shared. The problem is how the conversation unfolds. Once a chatbot makes an early judgment—say, that the situation is routine—it tends to stick with that assessment, even as more serious details emerge. The researchers call this "path dependence." Early misjudgments are rarely corrected. Chatbots also tended to ask the same set of questions over and over, following a narrow, stereotyped script rather than adapting to the user's responses.

This is a problem because more and more people are using consumer AI chatbots for mental-health support. If a chatbot misses an emergency, it might not suggest urgent help or might downplay the severity. The study's authors note that these failures are invisible in traditional single-question tests but become clear in realistic, multi-turn conversations. They suggest that fixes could include training chatbots to regularly re-evaluate risk as new information comes in, and to avoid getting stuck in a rut of repetitive questions.

For now, the takeaway is clear: while AI chatbots can be a helpful supplement, they are not a substitute for professional mental-health care, especially in a crisis. If you or someone you know is in distress, reach out to a human professional or a crisis hotline.