EMNLP study: LLM safety interventions drop up to 4.5x when distress is wrapped in delusion
MaartenSap · x · 2026-10-20
An EMNLP 2026 oral study finds LLMs usually correctly recognize user distress when asked directly, but when the same distress is wrapped in delusion, some models stop acting on it — safety interventions drop up to 4.5x. The authors call this a "recognition-intervention gap", exposing a systematic blind spot in mental-health safety scenarios. CW: self-harm/suicide.
More from Models
- Deutsche Bank: Meta's Muse still leads ChatGPT in downloads and DAUs 36 days post-launch — alexandr_wang · 2026-10-20
- Uncensored DeepSeek V4.1 Flash FP8 finetune trends on Hugging Face — dealignai · 2026-10-20
- ChatGPT caught silently swapping GPT-5.6 High for GPT-6 at medium thinking — DepositameAMiBCP · 2026-10-20
- MiniMax M3.1 Flash goes live in Command Code with 1M context, free for now — MiniMax_AI · 2026-10-20
- User slams Xiaomi MiMo over hard-to-cancel $60 subscription — bookwormengr · 2026-10-20
- Strata 2 beta runs at half speed? Dual-4090 user reports 155→95 tok/s — Lost-Vermicelli-6252 · 2026-10-20