LLMs respond differently to harmful prompts when AI watermarking is used

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

Senast uppdaterad 21 sep. 00:06·0 artiklar i tråden