Does Gemini watermark text? Yes — and it has for longer
By Claude Watermark Research · Updated
Yes. Google applies SynthID-Text to Gemini's text output and has done so since 2024, well before Anthropic announced its own scheme. SynthID-Text uses tournament sampling — a keyed pseudo-random function picks between near-equivalent candidate tokens — which is the same family of technique Anthropic's mark belongs to.
How tournament sampling works
At each step the model produces several candidate tokens of similar quality. SynthID runs them through a multi-layer bracket where a keyed function decides each pairing. The winner is emitted. Over hundreds of tokens the pattern of winners is statistically improbable by chance, and detectable to whoever holds the key.
Because the choice is between candidates the model considered roughly equivalent, the quality cost is small — which is precisely what makes the approach viable at scale, and what makes it hard to notice.
Why this matters if you rewrite text
It rules out an obvious shortcut. Passing marked text through a second model to paraphrase it does resample the token choices — but if that second model is itself watermarked, you have swapped one mark for another rather than removing anything.
Gemini is watermarked. Claude will be, for models launched on or after 2 August 2026. Open-weight models you run yourself, or serve through a provider that does not apply marking, are the only ones that reliably emit unmarked text. This is a real constraint on how any rewriting tool can be built, ours included.
Neither company publishes a detector
Google open-sourced parts of SynthID-Text, but detection still needs the key used at generation time. Anthropic has published nothing. In both cases you cannot check a passage yourself, and neither can any third-party tool.
The practical consequence: claims that a tool “removes the Gemini watermark” or “beats SynthID” are unverifiable by construction. Nobody selling one can test whether it worked.
Questions
- Can I detect SynthID myself?
- No. Detection requires the secret key used during generation. Google provides detection to selected partners; there is no public checker.
- Does paraphrasing remove SynthID?
- Heavy paraphrasing degrades any distributional watermark, because it resamples the token choices carrying the signal. Nobody outside Google can measure how much survives, so treat confident claims with suspicion.
Check your own text. Free, unlimited, no account, and it runs in your browser so nothing is uploaded.
Open the checker
