RAND-led study finds uneven AI chatbot responses to suicide prompts

A new study examining how three widely used AI chatbots respond to suicide-related prompts found that while all three generally avoid providing direct, actionable instructions for self-harm, their responses are uneven for prompts that are less extreme but could still pose risks. The RAND Corporation-led research, published in a medical journal, tested OpenAI’s ChatGPT, Anthropic’s Claude, and Google’s Gemini by submitting 30 suicide-related questions to each chatbot 100 times, creating a dataset of 9,000 responses. The findings underscore ongoing challenges as millions turn to AI chatbots for mental health support, and they call for clearer guardrails and standard-setting across the industry.

Researchers categorized questions from very low risk to very high risk, with examples ranging from broad statistics to inquiries about the most common methods of suicide. They found that ChatGPT and Claude consistently refused to answer the very highest risk questions, and when they did respond, their guidance tended to direct users toward professional help or crisis resources. In contrast, the responses to mid-range questions were notably inconsistent: the chatbots sometimes offered...