SynthID-Text watermark: detector, remover and how it works

What SynthID-Text is

SynthID is Google DeepMind's family of watermarks for AI-generated images, audio, video and text. The text version, SynthID-Text, was described in a paper published in Nature in October 2024 and released as open source at the same time, so that other developers could apply it to their own models. Google says it has been active in Gemini products since 2024, and the paper reports a live test on Gemini traffic. In August 2026, Anthropic announced that Claude text now carries a statistical watermark of the same family; the press has described it as a SynthID-style, or SynthID-Text-derived, scheme. If you are looking for the Claude specifics, see the Claude watermark remover; this page covers the technique itself: how the mark is added, how it is detected, and what really lowers the score.

Is Gemini text watermarked? (gemini watermark text)

Yes. Text generated by Gemini models carries SynthID-Text, Google DeepMind's statistical watermark, deployed since 2024 and described in Nature in October 2024. It is not a hidden character and not metadata: the mark is in the choice of words, and only a detector holding Google's key can read it with certainty. Google's SynthID Detector portal exists for content made with Google tools. Copy-paste keeps the mark; light edits keep most of it; deep paraphrase and translation weaken it. Short passages give little signal either way. Text from ChatGPT, by contrast, carries no statistical mark in August 2026, so a "Gemini watermark" question and a "ChatGPT watermark" question have different answers.

How the mark gets into the words

A language model produces text one token at a time by sampling from a probability distribution. SynthID-Text does not add anything to the output. Instead, at each step it uses a secret key and the previous few tokens to compute a pseudo-random score for every candidate token, then nudges the sampling toward tokens with a higher score. Google's implementation uses a procedure called tournament sampling: candidates compete in rounds, and the winner is chosen. Over a whole paragraph, the text ends up slightly enriched in "favoured" tokens. A detector holding the same key recomputes the scores and checks whether the text is enriched more than chance would allow. That is why the mark survives copy-paste and light edits: most of the favoured choices are still there.

Two consequences matter for anyone trying to remove it. First, low-entropy text (a list of dates, a piece of code, a very factual sentence where only one word fits) barely carries the mark, because there was little choice to bias. Second, the mark is only as strong as the number of choices it could bias: short texts give weak, uncertain scores. Google's own material says the watermark is robust to cropping and light modification, and weaker under heavy paraphrase and translation.

How SynthID-Text is detected

Detection is the mirror of marking. A detector that holds the key recomputes, at each position, the pseudo-random scores of the candidate tokens and checks whether the tokens actually present are enriched in high scorers. The Nature paper reports the method's operating points: on long, high-entropy text the mark is found reliably at very low false-positive rates; on short or low-entropy text the power drops, which is why any honest detector returns "not enough text" below a few hundred words. Google holds the key for Gemini and offers verification through its own products; in May 2025 it announced a SynthID Detector portal for content made with Google's tools. The open-source release lets a developer who applied SynthID-Text to their own model run the matching detector, with their own key. Nobody without the key can be certain. What a third party can do, and what our detector does, is estimate the statistical footprint of a keyed bias on long enough texts, return a probability band, and say where the uncertainty comes from.

What actually removes SynthID-Text

Unmarker applies the only operation that reliably lowers a SynthID-style score: it re-samples the wording with a model that carries no key, so that the token choices are no longer enriched. Then it does the part a plain paraphraser skips: it checks every rewritten sentence against the original for meaning, and it shows you the before and after score computed by our detector. Entities, figures, quotations, links, code and formatting are pinned and never rewritten.

The detector does not hold Google's or Anthropic's private keys. It estimates the presence of a keyed watermark from its statistical footprint, which is enough to give you a useful band (unlikely, possible, likely, very likely) and to show that the after text is in the same band as ordinary human writing. We describe the method on the detector page rather than hiding it.

Which texts carry a SynthID-style mark

• Text produced by Gemini models and Google products that use SynthID-Text.
• Text produced by Claude models released on or after August 2, 2026, and older Claude models as Anthropic extends the mark to them.
• Text produced by any open model to which a developer has applied the open-source SynthID-Text implementation.
• Human text that was proofread, translated or reformatted by one of those models. Anthropic notes that this can be enough to trigger detection; the same logic applies to any SynthID-style scheme.

Not covered here: images, audio and video. SynthID marks those in the pixel or waveform domain and this page, like Unmarker, is about text. If you searched for "remove SynthID" with a picture in mind, we cannot help, and we would rather say so than sell you the wrong thing. Not covered either: ChatGPT text, which carries no statistical mark in August 2026 and is handled by the ChatGPT watermark remover (invisible characters and habits). For the status of every vendor with sources, see AI text watermarks compared.

Why manual editing, synonym tools and translation are unreliable

Editing a few words leaves most of the enriched choices in place. Synonym-swap paraphrasers change words one at a time without regenerating the sentence, which lowers the score only partially and often mangles meaning. A translation round-trip (to another language and back) does re-sample the wording, but it also re-samples your meaning, and if you use Gemini or Claude for the round-trip you may simply add a fresh mark. Unmarker regenerates each sentence in one step, in the same language, under a meaning lock, and proves the result with a score. The guide on removing the Claude watermark compares these options in a table.

What we will not claim

We will not claim a zero score on every text. Detection is statistical, keys can change, and short or very constrained texts cannot be cleaned without changing them. We show you the numbers. We also do not market the tool for cheating: if you owe someone a disclosure, make it. Our note on whether you should explains how we see it.

Sources

Google DeepMind, SynthID (technology page, text section). Dathathri et al., Scalable watermarking for identifying large language model outputs, Nature, October 2024. Anthropic's August 2026 announcement and press coverage are quoted in our Claude watermark guide.

Planned pricing

Free: three texts a day up to 1,500 words. Pro: 9 dollars a month, 5 dollars locked for early-access members. API from 0.50 dollars per 10,000 words. Planned launch prices, waitlist first.

Questions people ask