LLMs respond differently to harmful prompts when AI watermarking is used

Ars Technica作者： Dan Goodin 2026年9月17日查看原文 ↗

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