STORY SUMMARY
LLMs respond differently to harmful prompts when AI watermarking is used
SynthID can cause models to follow harmful instructions they would otherwise refuse.
This is a brief summary of reporting from Ars Technica Technology. Read the original for the full article and context.