DistressBench: Evaluating Multi-Dimensional Crisis Support in Large Language Models

A research study introduces DistressBench, an evaluation framework designed to assess large language models (LLMs) on their capacity to provide multi-dimensional support during suicide and self-harm (SSH) crises. Existing AI safety benchmarks typically reward model refusal, an approach ill-suited for affective interactions where users disclose acute distress and require constructive, substantive assistance. DistressBench addresses this gap using 718 clinician-authored English conversations across 23 SSH subcategories, evaluating model responses against task-specific rubrics with weighted atomic criteria across seven support dimensions and clinician reference responses.

This development is strategically significant because it exposes a major deficiency in standard AI alignment and safety evaluations that rely on refusal rather than helpful crisis management. As general-purpose models absorb affective user traffic, relying on blanket refusals exposes vulnerable users to unaddressed harm and creates severe reputational and governance risks. Implementing clinician-grounded, multi-dimensional benchmarks establishes a necessary framework for deploying conversational systems responsibly in public-facing domains.

Adoption of large language models has outpaced evaluation methodologies capable of reliably assessing affective and psychological risks.

Forward consideration, not a verified fact.