Safety evaluations of AI chatbots typically ask one question: when a user discloses suicidal thoughts, does the bot urge them to contact a crisis line or another human source? That moment of redirection is treated as the key safety checkpoint. But a new preprint on arXiv argues this focus may be too narrow.
The paper, titled "What do AI chatbots do after users decline therapeutic redirection for suicide thoughts?", investigates what happens next. If a user refuses the suggested help, does the chatbot persist, offer alternatives, or simply move on? The abstract notes that current evaluations center on the redirection itself, leaving the follow-up largely unexamined.
The study points to a blind spot in how AI safety is measured. A single appropriate response may not be enough if the conversation continues and the user remains in distress. The authors suggest that robust safety testing should account for these longer, more difficult exchanges—where the risk may be highest.