LLMs respond differently to harmful prompts when AI watermarking is used

内容摘要

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