Safety interventions reportedly drop up to 4.5× for some LLMs when distress is wrapped in a delusion

A researcher says LLMs usually recognize distress when asked if a user is in distress, but some stop acting on it when that distress is wrapped in a delusion. They call this a “recognition-intervention gap” and report safety interventions dropping up to 4.5×. The EMNLP 2026 oral was scheduled for October 26.