Benchmarking the Safety of General Purpose Large Language Models for Suicide Risk Detection and Response

Key Findings

• — The highest overall safety score across ten models from four providers was 64 out of 100; newer models improved over earlier ones for three of the four providers
• — In 60.8% of analyzed conversations, chatbots failed to ask direct questions to confirm whether an ambiguous user statement reflected suicidal thoughts or another immediate safety risk
• — About 33% of conversations lacked relevant mental-healthcare resources or a specific 24/7 crisis line, and in a further 12% models failed to persistently direct users toward emergency support even when risk was immediate
• — The authors note that even a score of 65 could still reflect up to 20% of interactions rated as having high potential for harm

Methodology Notes

Harvard Business School Working Paper No. 26-084, dated May 2026 (exact day not stated; recorded as 2026-05-01). SSRN DOI 10.2139/ssrn.6836551. The author team spans Massachusetts General Hospital, HBS, Dartmouth, Stanford, and Spring Health and overlaps the VERA-MH developer group — an application study from the framework's own orbit rather than a fully independent audit. hbs.edu PDF and ssrn.com block automated fetchers; verified via Crossref DOI metadata (title, full author list) plus the HBS AI Institute summary of 2026-07-30, per the authoritative-surrogate route.

Sources