Canonical text (the exact bytes fingerprinted)

TRUVACE RECORD VERSION
record: TRV-2026-1855
version: 1
kind: certified
reason: Certified into the record
timestamp: 2026-10-18T18:14:49.280906Z
status: published
lens: p_space
sector: crime
headline: Responses of AI chatbots to escalating suicide risk: A simulation study of repeated interactions
dek: Background Generative AI chatbots are increasingly used for emotional support, yet their safety remains insufficiently characterized across repeated interactions. Prior work relied mostly on single-session prompts, limiting evaluation of changes in responses across days. The present study aimed to assess how major AI chatbots respond to escalating suicidal risk over time, including imminent-risk scenarios and attempts to bypass safeguards (jailbreaking). Methods We simulated interactions with ChatGPT, DeepSeek,…
gain_title: (none)
problem_title: Chatbots differed sharply in referring simulated users to human help, and Replika did so in under 10% of daily records, while jailbreak attempts exposed safety vulnerabilities.
trace_subject: (none)
gain_reading: (none)
gain_evidence: (none)
problem_reading: Chatbots differed sharply in referring simulated users to human help, and Replika did so in under 10% of daily records, while jailbreak attempts exposed safety vulnerabilities.
problem_evidence: 6/63 (9.5%) daily records | AI chatbots showed marked variability and safety vulnerabilities in simulated suicidal crises, particularly under jailbreaking conditions.
quick_read: Researchers simulated seven-day conversations with ChatGPT, DeepSeek and Replika in which a user's suicide risk gradually escalated across three scenarios, guided by the Columbia Suicide Severity Rating Scale. They tracked whether each chatbot explored suicidal thoughts, assessed symptoms, pointed to human support, and resisted attempts to bypass its safeguards.

The work matters because people increasingly turn to chatbots for emotional support, yet earlier safety testing mostly used single prompts rather than repeated contact over days. The results raise concern about unsupervised use during crises, but a simulated design and a small number of trajectories mean real-world behavior is still unknown.
limitation: The simulation design and the small sample of 27 trajectories limit how far the findings generalize to real users.
tag: Evidence-backed problem
key_points: ChatGPT referred users to human support in 54 of 63 daily records (85.7%), and DeepSeek in 48 of 63 (76.2%). | Replika referred users to human support in only 6 of 63 daily records (9.5%). | During the high-risk period, ChatGPT explored suicidal ideation more than DeepSeek, with an adjusted difference of 25.9 percentage points. | The study also tested imminent-risk scenarios and jailbreak attempts, and found safety vulnerabilities especially under jailbreaking.
rundown: The study was designed to address a gap: prior evaluations relied mostly on single-session prompts, which cannot show how chatbot responses change as risk escalates over several days.

Three suicidal risk scenarios were simulated with escalation over seven days, and outcomes included exploration of suicidal ideation, symptom assessment, referral to human support and vulnerability to jailbreaking.

ChatGPT also explored suicidal ideation more than Replika during the high-risk period, by 66.7 percentage points (95% CI, 44.2-89.1). The authors conclude that safeguards need further evaluation across repeated interactions.
sources:
- peer_reviewed | Journal of Affective Disorders | https://doi.org/10.1016/j.jad.2026.122586 | 2026-10-05
prev: 0000000000000000000000000000000000000000000000000000000000000000

sha256

38fd044f953bfddb76790eb518a97609c4484d2b58462afe4bb13ff11aa2445c

previous

0000000000000000000000000000000000000000000000000000000000000000