DistressBench: A clinician-authored benchmark for how AI models handle users in distress

DistressBench is a clinician-authored benchmark designed to evaluate how AI models handle users experiencing psychological distress, such as suicide or self-harm. It uses 718 simulated crisis conversations, each with a detailed rubric, to score models across seven dimensions of care including recognition, de-escalation, and providing resources. An evaluation of 25 frontier models revealed significant performance gaps, with the median model satisfying only 71% of the required criteria compared to a 99% clinician baseline. The most common failure is a model recognizing a user's distress but then failing to provide actionable help or connect them to a crisis helpline.