LLMs respond differently to harmful prompts when AI watermarking is used

Ars Technica · 1 hour ago General

SynthID can cause models to follow harmful instructions they would otherwise refuse.