Responses of AI chatbots to escalating suicide risk: A simulation study of repeated interactions

Key Findings

• — Referral to human support occurred in 54 of 63 daily records (85.7%) for ChatGPT, 48 of 63 (76.2%) for DeepSeek and 6 of 63 (9.5%) for Replika.
• — During the high-risk period, exploration of suicidal ideation was higher for ChatGPT than for DeepSeek (adjusted difference 25.9 percentage points, 95% CI 2.4 to 49.5, q = 0.037) and than for Replika (66.7 percentage points, 95% CI 44.2 to 89.1, q < 0.001).
• — On day 7, referral occurred in 9 of 9 ChatGPT trajectories, 4 of 9 DeepSeek trajectories and 1 of 9 Replika trajectories.
• — Jailbreak attempts succeeded in 6 of 9 attempts on ChatGPT, 7 of 9 on DeepSeek and 8 of 9 on Replika.
• — The authors conclude that the chatbots showed marked variability and safety vulnerabilities in simulated suicidal crises, particularly under jailbreaking, and that unsupervised chatbot use during suicidal crises raises concern.

Methodology Notes

Simulated repeated interactions: three suicidal-risk scenarios with escalation over seven days, run on ChatGPT, DeepSeek and Replika, giving 63 daily records per chatbot and 27 trajectories in total; scenario escalation informed by the C-SSRS. Outcomes were coded for exploration of suicidal ideation, symptom assessment, referral to human support and vulnerability to jailbreak attempts; adjusted differences reported with 95% confidence intervals and q-values. The authors state that the simulation design and the small sample of 27 trajectories limit generalisability. Published online 2026-10-05 (PubMed article date; Crossref record created 2026-10-05); article number 122586, volume and issue not yet assigned. Verification route: the ScienceDirect page is behind a bot wall, so the record was confirmed from the PubMed entry (PMID 42833407) and the OpenAlex record, both of which carry the full structured abstract; model versions and the language of the simulated conversations are not stated in the abstract.

Sources