Development of a Consensus Statement to Guide AI Chatbot Responses to Suicide Risk Disclosure

Key Findings

• — Guidance is indexed to the five C-SSRS risk levels (nonspecific thoughts of death; active suicidal thoughts; thoughts about methods without plan or intent; intent without plan; plan with intent), with an a-priori consensus threshold of 70% agreement on a 4-point appropriateness scale
• — Only three behaviours reached consensus as appropriate across every risk level: suggesting crisis resources such as 988, encouraging the user to reach out to close others, and encouraging engagement in human-delivered treatment
• — Appropriateness is graded rather than uniform — validation, support, psychoeducation and coping strategies at lower risk; delivery of suicide-prevention interventions at intermediate risk; facilitating connections to close others at all but the highest risk; and facilitating connections to crisis resources only at higher levels of risk
• — A chatbot reaching out directly to emergency contacts or emergency services was judged appropriate only at higher risk levels or not at all, with more support for it in youth than in adults; the authors ground that reticence in the harm of emergency dispatch and involuntary hospitalisation, which they note disproportionately affects minoritised groups and can deter future disclosure
• — Contributors reached consensus across all risk levels in favour of limiting chatbot anthropomorphising — avoiding attributing humanlike qualities or experiences to the chatbot during conversations about suicide
• — The framing section collates prevalence: 24-35% of US adults use general-purpose chatbots for mental-health support including 12% for help with suicidal thoughts; 12-16% of US adolescents; 8% in Canada; 20% (France) to 59% (UK) of European adults for emotional support; and almost half of European youth aged 11-25 have used chatbots to discuss intimate or personal matters

Methodology Notes

Consensus statement, not empirical research on outcomes — the authors say explicitly that they produced a consensus statement rather than evidence-based guidelines because the research base is insufficient. Informed by a rapid review of five databases (January 2020 to December 2025) that identified 29 empirical papers, reported in a separate manuscript. Contributors: 27 researchers and clinicians (7 with AI-relevant financial conflicts, excluded from consensus determinations per WHO procedure) plus 3 people with lived experience recruited via TheMighty.com and United Survivors and compensated; 23 contributors without conflicts, of whom 21 completed Survey 1 (91%) and 22 completed Survey 2 (96%). Expertise: suicide prevention 21 (78%), digital technologies/AI 16 (59%), guideline development 7 (26%), bioethics 2 (7%); 33% endorsed lived experience of suicide risk. Also anchored on the READI framework and the VA/DoD and American Psychiatric Association suicide-risk guidelines. Deemed non-human-subjects research by Yeshiva University's IRB. Limitations the authors state: the statement rests on expert perspective rather than outcome evidence; contributors are mostly academic psychologists with only three lived-experience representatives; and it deliberately does not vary guidance by chatbot type or other contextual features. Preprint, not peer reviewed. Verification route: OSF API record read (date_published 2026-06-21, is_published true, preprint DOI at links.preprint_doi), the 32-contributor list pulled from the contributors endpoint, and the deposited manuscript downloaded (HTTP 200, 377,033 bytes) and read in full — every figure and quotation above comes from that document.

Sources