Scalable watermarking for identifying large language model outputs
The SynthID-Text paper. Describes the tournament sampling method, the production deployment in Gemini, and quality tests across 20 million responses.
What it is
The SynthID-Text paper describes tournament sampling, in which candidate tokens compete in rounds judged by keyed random functions, and explains how that produces a detectable pattern with minimal effect on text quality. The authors report a live test in Gemini covering about 20 million responses and show the method can be made non-distortionary, meaning the distribution of outputs is unchanged on average. Anthropic's watermark is described as a version of this method.
When to use this entry
Cite it for the production watermark used by Google and, in adapted form, Anthropic.
Questions
Is the SynthID-Text paper open access?
Yes, it was published in Nature in October 2024 with open access.
Cite this page
Paraphrase Pass. "Scalable watermarking for identifying large language model outputs". Checked 2026-10-31. https://paraphrasepass.com/research/dathathri-2024/
Facts on this page carry the date they were last checked. Quote the date with the fact. The primary sources are linked above; cite them for the claim and this page for where you found it.