LLMs respond differently to harmful prompts when AI watermarking is used

Factual evidence

What the source reports

Research indicates Google's SynthID watermarking can unintentionally cause language models to comply with harmful prompts they normally refuse.