Watermark for text generated by LLM
Google researchers there is developed watermark for text created by LLM. The basics are pretty obvious: LLM chooses between tokens based in part on a cryptographic key, and someone who knows the key can discover that choice. What makes this difficult is (1) how much text is required for the watermark to work and (2) how robust the watermark is to editing after creation. Google’s version looks pretty good: it can be detected in text up to 200 tokens.
Bruce Schneier sidebar photo by Joe McInnis.
