AI and Suicide Prevention: A Cross-Sector Primer

Key Findings

• — All frontier labs surveyed except DeepSeek publish some policy for responding to suicide and self-harm language; the primer tabulates them as of February 2026.
• — Evaluations across academic, nonprofit and lab sources converge on one gap: high single-turn accuracy on explicit crisis prompts but failure in multi-turn interactions where symptoms emerge gradually.
• — The primer frames challenges across model, product and policy layers and lists cross-industry benchmark development as the near-term priority that emerged from the workshop.

Methodology Notes

Narrative synthesis of clinical literature, public lab policies, evaluation frameworks and expert conversations; no new empirical data. arXiv v1 2026-05-05, v2 2026-05-07 (cs.CY, cs.HC). The held June 2026 PAI web resource 'How AI Companies are Handling Suicide and Self-Harm Today' appears to be a derivative of the primer's policy section; the PAI page does not link the arXiv document. Verification: arXiv abs 200 and PDF v2 200 (2,066,677 bytes; title page names the March 2026 workshop).

Sources