Gemini watermark

What the evidence says

Google DeepMind announced text watermarking for the Gemini app and web experience in May 2024, and the Nature paper published that October describes it as already running in production — by the authors’ account, the first deployment of a generative text watermark at scale. Note the scope: Google documents the app and web experience. It does not document watermarking for text returned by the Gemini API, AI Studio or Vertex AI, so this page does not claim it.

The method is called tournament sampling. Candidate tokens drawn from the model’s own distribution compete in successive rounds, scored by pseudorandom functions derived from a key and the preceding context. SynthID-Text does not affect training and modifies only the sampling procedure, and detection runs without the underlying model.

That last point settles the question this database exists to answer. The Nature paper separates generative watermarking, which shapes the text as it is written, from edit-based watermarking, which alters text after the fact. SynthID-Text is the former and never edits a finished string, so no character is inserted. Google publishes no statement about Unicode specifically; the conclusion follows from the published mechanism rather than from a denial.

Quality was measured rather than asserted. Across nearly twenty million Gemini responses, the thumbs-up rate differed by 0.01 percent and the thumbs-down rate by 0.02 percent between watermarked and unwatermarked models, both statistically insignificant. The reported latency cost was 0.57 percent per token.

The limits are specific, and finer than the usual summary. Cropping, changing a few words and mild paraphrasing are survivable; a thorough rewrite or a translation into another language greatly reduces detector confidence. Watermarking is also less effective on factual answers, where there is less room to vary wording without losing accuracy. There is still no public way to check a passage: the SynthID Detector portal accepts images, video and audio, and text detection, announced as imminent in May 2025, is absent from the product page as of this check.

Sources

1. Scalable watermarking for identifying large language model outputsNature 634(8035), 818–823 · 2024-10-23 · Peer-reviewed
2. SynthID: identifying AI-generated textGoogle DeepMind · 2024-05-14 · Provider documentation
3. SynthID: Tools for watermarking and detecting LLM-generated textGoogle AI for Developers · 2025-04-09 · Provider documentation
4. SynthID Detector: a new portal to help identify AI-generated contentGoogle · 2025-05-20 · Provider documentation