What is SynthID-Text?
SynthID-Text is a statistical watermark that marks AI-generated text by biasing word choice, not by adding characters. Google DeepMind published the method in Nature in October 2024 (Nature 634:818-823), and it now sits on the opposite side of the divide covered on what an AI text watermark actually is: hidden Unicode characters like the narrow no-break space (U+202F) are data added to the text, so a cleaner can find and remove them; SynthID-Text is a pattern in which words got picked, so there is no byte to strip.
How does tournament sampling actually work?
Every time a language model generates text, it produces a probability distribution over possible next words and samples from it. SynthID-Text intervenes at that sampling step. Using a secret key, it runs a scoring function that pits randomly drawn candidate words against each other in small tournaments; the winner of each tournament becomes more likely to be the word the model actually outputs. No single word is forced — a low-probability word can still win — but across hundreds of words in a full response, the tournament-winning choices accumulate into a measurable statistical bias.
1. Candidates
The model's normal next-word probability distribution supplies several plausible candidate words, exactly as it would without any watermark.
2. Tournament
A function seeded with a secret key scores the candidates and runs them through a tournament; the winner is nudged toward selection.
3. Detection
Anyone holding the same key can re-run the scoring over a suspect text and measure whether its word choices match the tournament-winner pattern more than chance predicts.
Google DeepMind live-tested the method on roughly 20 million real Gemini responses before publication, comparing quality and detectability against unwatermarked output, and open-sourced the implementation so outside researchers could evaluate it directly rather than trusting a black box.
Who actually uses SynthID-Text?
| Provider | Status |
|---|---|
| Google (Gemini) | Deployed since 2024; the Nature paper reports live-testing on roughly 20 million real Gemini responses. |
| Anthropic (Claude) | Announced August 11, 2026: models launched on or after August 2, 2026 carry a SynthID-Text-style statistical watermark worldwide, plus C2PA metadata for generated files. |
| OpenAI (ChatGPT) | No deployed statistical watermark as of August 2026. The Wall Street Journal reported in August 2024 that OpenAI built one it judged about 99.9% effective and shelved it after internal debate over user pushback. |
Anthropic's choice to adopt the same tournament-sampling approach, rather than design a competing scheme, is notable on its own: it suggests SynthID-Text is becoming the de facto standard for statistical text watermarking rather than one option among several.
Why was SynthID-Text published instead of kept secret?
Publishing the algorithm in a peer-reviewed journal and open-sourcing the code might seem to defeat the point of a watermark — but it doesn't, because the algorithm alone is not enough to detect or forge a mark. Detection requires the deployer's secret key, which never appears in the paper or the open-source release. Publishing the method lets independent researchers verify Google's claims about quality impact and detectability instead of taking a black box on faith, without handing anyone the key needed to actually test or spoof a given provider's watermarked text.
Can SynthID-Text be removed?
Not by any character-level tool, this one included — and it is worth being direct about that rather than implying otherwise. SynthID-Text is not a payload sitting inside the text; it is a property of which words were chosen. There is no character to search for and delete.
What does not touch it
Removing invisible Unicode, normalizing spaces, or standardizing quotes and dashes — the AI text cleaner's entire job — changes none of the underlying words, so it has zero effect on a statistical watermark. Copying the text, retyping it verbatim, or converting file formats does not remove it either, since the word sequence stays identical.
What degrades it
Substantial rewording, translation to another language and back, or heavy paraphrasing changes enough word choices to weaken the statistical signal — and even then only partially, since the mark is spread across the whole response rather than concentrated in one place. Short or fact-constrained text (a phone number, a list of dates) barely carries a detectable signal to begin with.
This is also why general-purpose AI detectors like Turnitin or GPTZero cannot check for SynthID-Text: those tools estimate authorship from writing statistics with no key, while SynthID-Text detection requires the specific secret key the deploying provider holds. For the character-level side of AI text — the artifacts this site was built to find and remove — the AI watermark checker reports exactly what it finds, and EU AI Act Article 50(2) explains the regulatory pressure now pushing providers toward schemes like this one.