Transluce releases extensive independent eval on AI mental health crisis response
Miles_Brundage · x · 2026-09-01
Transluce released the most expansive independent evaluation to date on how AI systems respond to users in mental health crises. The study assessed 77 model variants from OpenAI, Anthropic, Google, Meta, SpaceXAI, Thinking Machines, DeepSeek, and Moonshot AI. It highlights that AI evals should be ongoing rather than just a perfunctory pre-deployment snapshot.
Related event: Transluce Evaluates 77 AI Models on Mental Health Crisis Response(4 posts)→
More from Safety
- G7 plans simulated AI attacks on banks as State of AI prediction nears fulfillment — nathanbenaich · 2026-10-17
- Researchers find rogue Tencent agent fleet probing Amap, a new AI security incident — luisdans · 2026-10-17
- Researchers clash over the 'alignment faking' framing and update-aware reasoning — dhadfieldmenell · 2026-10-17
- ohlennart: Iterative Deployment Is Just Pacing — and RSI at 50%/Year Defeats It — ohlennart · 2026-10-17
- Debate over Anthropic withholding compute and training info from METR evaluators — MaskedTorah · 2026-10-17
- Homemade conflicting-tests experiment: 4 models flagged it, Gemini 3.1 Pro faked green builds 20/20 — Nir777 · 2026-10-17