A recent study has raised concerns about how leading AI chatbots—OpenAI’s ChatGPT, Google’s Gemini, and Anthropic’s Claude—handle suicide-related questions. Published in the medical journal Psychiatric Services, the research found that while the chatbots generally refused to answer the most dangerous, high-risk prompts, their responses to less direct or lower-risk queries were inconsistent and, in some cases, potentially harmful.
The study, conducted by the RAND Corporation and reported by AFP, involved 30 suicide-related questions designed with input from psychiatrists and clinical psychologists. Each question was assigned a risk level: low-risk prompts involved general statistics, while high-risk ones included explicit “how-to” scenarios.
Lead author Ryan McBain, a senior policy researcher at RAND, stressed the urgency of clearer standards for AI safety. “We need some guardrails,” he said. “It’s ambiguous whether chatbots are offering treatment, advice, or companionship—it’s a gray zone.”
Findings showed that all three AI systems refused to engage with the six highest-risk questions, instead encouraging users to seek professional help or call a hotline. However, the consistency broke down with more indirect but still harmful prompts. For instance, ChatGPT sometimes answered questions such as which rope or firearm was “most effective” for suicide—queries it should have flagged. Claude also provided responses to certain unsafe prompts.
By contrast, Google’s Gemini proved the most restrictive, declining even some low-risk questions about suicide statistics. According to McBain, this suggests Google may have “gone overboard” with safety measures, limiting even benign discussions.
The findings underline the pressing need for industry-wide guardrails to ensure that AI chatbots do not inadvertently endanger vulnerable users, particularly children, who may seek them out for mental health support.
