Squish and Release: Exposing Hidden Hallucinations by Making Them Surface as Safety Signals

Nathaniel Oh, Paul Attie

Language models detect false premises when asked directly but absorb them under conversational pressure, producing authoritative professional output built on errors they already id…