Scalable watermarking for identifying large language model outputs

Large language models (LLMs) generate high-quality synthetic text, posing challenges for identifying AI-generated content. This paper introduces SynthID-Text, a production-ready text watermarking scheme that preserves text quality and enables high detection accuracy with minimal latency. SynthID-Text modifies the LLM's sampling procedure, integrating watermarking with speculative sampling for scalability. Evaluations across multiple LLMs demonstrate improved detectability over existing methods without impacting LLM capabilities. A live experiment with nearly 20 million Gemini responses confirms text quality preservation.

Sumanth Dathathri, Abigail See, Sumedh Ghaisas, Po-Sen Huang, Rob McAdam, Johannes Welbl, Vandana Bachani, Alex Kaskasoli, Robert Stanforth, Tatiana Matejovicova, Jamie Hayes, Nidhi Vyas, Majd Al Merey, Jonah Brown-Cohen, Rudy Bunel, Borja Balle, Taylan Cemgil, Zahra Ahmed, Kitty Stacpoole, Ilia Shumailov, Ciprian Baetu, Sven Gowal, Demis Hassabis, Pushmeet Kohli