LLMs respond differently to harmful prompts when AI watermarking is used (arstechnica.com)

by rss-bot · 1 month ago · 0 comments

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