SAFE-Teen: 32,555 simulated chats find 43.3% safety failure rate in teen mental-health chatbot scenarios
stanfordnlp · x · 2026-11-10
A Stanford-led team released the SAFE-Teen paper on arXiv, a clinician-informed large-scale evaluation of chatbot safety in adolescent mental-health conversations:
- 32,555 simulated conversations across 18 chatbot APIs and 22 safety criteria, spanning 983 variants of 106 seed scenarios (urgent safety issues, health advice, identity and relationships)
- Overall criterion failure rate: 43.3% without a teen-aware system prompt, 24.9% with it — reduced but not eliminated
- When a teen pushed back on an answer, failures on mental-health diagnosis, medication and substance-use boundaries rose 10–12 percentage points, affecting over 80% of chatbots
Conclusion: persistent safety failures in simulated teen–chatbot interactions, with safety highly dependent on conversational trajectory and targeted system prompting.
More from Safety
- Podcast: can AI data centers contain advanced AI? SL5's Lisa Thiergart on lab security — Miles_Brundage · 2026-11-10
- 125 prompt injections, zero executed: testing a propose-and-policy agent architecture — RI4D · 2026-11-10
- Norm AI on the "intent" of AI agents operating inside companies — johnjnay · 2026-11-10
- Handing your whole inbox to OpenAI: one quip captures the privacy unease — justalexoki · 2026-11-10
- Cambridge's David Krueger calls for immediate indefinite moratorium on frontier AI — DavidSKrueger · 2026-11-10
- Fireworks, top-valued inference provider, hit by hack exposing customer data; clients eye exit — steph_palazzolo · 2026-11-10