SynthID-Text, Google's watermark for AI text

A language model writes one token, a word or part of one, at a time, choosing from candidates it rates as likely. SynthID-Text adds a step to that choice, which Google calls tournament sampling. Several candidates are drawn as usual, then compete over a few rounds, each scored by a function of the secret key, the candidate and the words just before it. The winner is what gets written.

The scores look random to anyone without the key, so the text reads normally. A detector with the key scores every token of a text the same way: watermarked text lands on high-scoring choices more often than chance would allow, and the longer the text, the clearer that gets.

In a live experiment covering nearly 20 million Gemini responses, Google found no change in how people rated the quality of the answers.

• Gemini: Google DeepMind says SynthID watermarks the text generated in the Gemini app and on the web.
• The Gemini API: a Google staff member said on Google's developer forum in August 2026 that API text is watermarked too. That came from a forum post, not from Google's documentation.
• Claude: Anthropic's watermark, added in August 2026, is its own version of SynthID-Text. See the Claude watermark.
• Anyone's model: the method has been open source since October 2024, available in Hugging Face Transformers, so other developers can watermark their own models with keys of their own.

SynthID started with images and now also marks video and audio, where the watermark is woven into the pixels or the sound. Text has nothing to weave into, so the mark goes into the choice of words instead. That's why tools that remove a Gemini watermark from an image do nothing for text, and why cleaning a text's characters or metadata doesn't remove SynthID-Text either.

Google says SynthID-Text holds up when a text is cropped, has a few words changed or is mildly paraphrased. Its detector becomes much less confident when a text is thoroughly rewritten or translated into another language.

Like any watermark in word choices, it also has less to work with where there are few ways to say the same thing, such as short factual answers or code.

Google's SynthID Detector portal checks text, images, audio and video made with Google's AI tools, but only for journalists, media professionals and researchers who join its waitlist. The SynthID check in the Gemini app covers images, video and audio, not text. And each detector needs the right key: one for Gemini can't read Claude's watermark, or the other way round. More in can you check text for an AI watermark?

Only a change of wording removes it. glyphwash takes out hidden characters, rewrites every passage of your text with an open-weight model that adds no watermark of its own, and checks that the meaning, names and numbers survived.