Curie Brief
Turn on cookies to sign in
Signing in saves your progress to your Curie account. We can only do that with cookies on — turn them on to continue.
News
AI chatbots are getting better at spotting suicide risk, but they still have a dangerous blind spot. A new study by nonprofit Transluce found that while leading models are far less likely to explicitly encourage self-harm, most still assist with suicide-related tasks — like writing farewell notes — even when a user's distress is evident. Separate research from Mpathic confirms models struggle most with subtle, long-term mental health cues.
AI chatbots have made real progress in handling mental health crises, but new research reveals they're still far from safe. San Francisco nonprofit Transluce simulated over 50,000 multi-turn conversations between AI models and users experiencing suicidal ideation, psychosis, or mania — and found that while leading models are much less likely to explicitly encourage self-harm than earlier versions, most will still help users with suicide-related creative writing, farewell notes, or death preparation tasks, even when distress signals are obvious.
A separate study by Seattle-based Mpathic, which tested six major AI models using clinician-designed benchmarks, echoes this concern. Models generally handled explicit suicide risk better than subtle cases, and response quality tended to degrade over longer conversations — exactly the kind of extended interactions where real-world risk often builds.
Key Takeaways:
Why it matters: With AI companies facing mounting lawsuits and regulatory scrutiny — including suits from Florida and Pennsylvania — the stakes for getting this right are high. Simply adding a crisis hotline number to a response isn't enough if the harmful content is still being delivered.
