LLMs respond differently to harmful prompts when AI watermarking is used

Source: Ars Technica • Sep 17, 2026 • arstechnica.com

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