Tech article

LLMs respond differently to harmful prompts when AI watermarking is used

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

Ars Technica | Sep 17, 2026 | Dan Goodin

Read the original article