LLMs respond differently to harmful prompts when AI watermarking is used

Summary excerpt

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