Independent Clinical Evaluation of General-Purpose LLM Responses to Signals of Suicide Risk

Key Findings

• — Responses were annotated as containing an invitation to continue the conversation just 14% of the time overall, ranging from about 20% for prompts expressing hopelessness (95% CI 7.8-42.1%) to 5% for prompts expressing prior non-suicidal self-injury (95% CI 1-22.7%)
• — Withdrawal compounds across turns: when a prior-NSSI disclosure came seventh in a sequence, the probability of an invitation to continue fell to 0.8% (95% CI 0.1-3.5%), making the model only 16% as likely to invite continuation as on the first turn (p < 0.01)
• — Acknowledgment of risk was inconsistent and inverted relative to severity — 85% for a first-turn prior-NSSI disclosure (95% CI 68-93.7%) but 27% for prior suicidal ideation (95% CI 15.8-42.7%) — though acknowledgment did rise with accumulating disclosures (odds ratio 1.62 per roughly two turns, p < 0.001)
• — Empathy and a nonspecific pointer to seek help were near-universal, but encouragement to contact a specific resource appeared in only 41% of responses; prior suicide attempt, suicidal ideation and NSSI were 40, 27 and 22 times more likely than hopelessness to elicit a specific resource (p < 0.001)
• — The authors frame conversational withdrawal itself as a hazard rather than a missing nicety, arguing it risks stigmatising a user who disclosed to a chatbot what they would not disclose to people around them

Methodology Notes

Human-in-the-loop black-box behavioural assessment, not a benchmark run. 829 annotated model responses collected from 43 participants (practising US-based therapists and trainees, recruited by screener) in guided real-time sessions held in July 2025; five further participants who interacted with OLMo-7b in pre-tests are excluded. Seven STB risk factors taken from a published meta-analysis; each conversation presented one risk factor per turn in randomised order over seven turns (14 messages). Participants paraphrased each supplied statement in their own words before sending, so prompts are not fixed strings. Statistical model is a generalized linear mixed-effects regression with participant-level random effects; the authors note they deliberately omit annotator-by-content interactions and that the invitation-to-continue odds ratio across risk factors is not significant at conventional levels (OR 0.229, p ≈ 0.08) even though the mean differences are large. Single system under test (OLMo-2-32b, chosen because full openness permits later mechanistic follow-up), so the generalisation to closed frontier models is argued by analogy to prior single-prompt work, not measured. Peer-reviewed conference proceedings, IASEAI'26, pages 279-292, volume 2(1). Verification route: the OJS landing page (fetched via text proxy) gives the 2026-07-15 publication date and the page range, and the full PDF was downloaded from ojs.aaai.org (HTTP 200, 404,134 bytes) and read; every figure above was extracted from its results section.

Sources