Skip to main content
AI Watermark Removal

Technical guide

AI Text Watermark: How Statistical Watermarking Works

An AI text watermark isn't a character hidden in the output. It's a bias applied at the exact moment the model chooses its next word, invisible by design and detectable only with the matching key and enough text to score.

By ignacio-sjPublished Revised Research/proposal

Key takeaways

  • Text watermarking works by biasing token sampling at generation time, not by inserting a visible or hidden character afterward. Detecting it means scoring whether a passage's word choices match a configured pattern, not searching for a marker.
  • Google's SynthID Text is the best-documented scheme: a logits processor described in a peer-reviewed Nature paper, which Google says runs in the Gemini app (a claim this site grades Contested) and which is open-sourced in Hugging Face Transformers since v4.46.0, with a reference detector that reports watermarked, not watermarked, or uncertain.
  • It descends from a 2023 scheme (Kirchenbauer et al.) that splits each token's candidates into a favored green list and a disfavored red list. Detection on that original scheme stayed reliable after strong human paraphrasing once a detector had roughly 800 tokens, at a false-positive rate of one in 100,000, the clearest independent number in this literature. OpenAI's textGrain figures, published 2026-10-05, are the first a provider has given for a production text watermark.
  • Independent researchers have red-teamed SynthID Text specifically and found real gaps. A peer-reviewed 2025 study found it vulnerable to paraphrasing, copy-paste splicing, and back-translation; a 2026 preprint found its default scoring method grows more vulnerable, not less, as more sampling layers are added.
  • Adoption isn't universal because the tradeoffs are real. OpenAI reportedly built an internal ChatGPT text watermark around 99.9% effective, per leaked internal documents, and held it back over false positives, circumvention risk, and disproportionate impact on non-native English writers. Its first shipped text watermark, textGrain, was announced on 2026-10-05, rolling out to ChatGPT and Codex text in the EU only, with an opt-in for API customers and a detector limited to approved researchers.

Scoring a passage

What a detector actually measures

A detector holding the same secret key can recompute what each word's green list would have been, then count how often the text actually landed on green. These two illustrative passages show what a clear signal versus chance-level noise looks like; the exact words and counts are made up for illustration, not a real analyzed sample.

Likely watermarked

Theresultsclearlyshowastrongtrendacrosseverysample

8 of 10 words landed green

Well above the roughly half expected by chance alone, which is what a statistical test flags as an unlikely coincidence.

Ordinary text

Mycatknockedaglassoffthetableagaintoday

5 of 10 words landed green

Right around the chance baseline, since nothing biased these word choices toward either list.

Why length matters so much

Ten words is too short to conclude anything with real confidence; this is simplified for illustration. A 2024 follow-up to the original green-list paper found the signal stayed statistically detectable after strong human paraphrasing once a detector had roughly 800 tokens to work with, at a false-positive rate of about one in 100,000. Short passages, heavy edits, or translation all shrink how much real signal a detector has left to count.

How does an AI text watermark work?

Research/proposal

You'll see the exact moment a watermark gets applied, which explains nearly every quirk that follows.

An AI text watermark works inside the model's sampling step, biasing which word gets picked rather than adding anything to the text. A language model doesn't pick one correct next word. At every step it produces a probability distribution over plausible candidates, and sampling picks one of them.

A text watermark sits inside that sampling step. Before a token is chosen, the scheme pseudorandomly splits the vocabulary into a favored sublist and a disfavored one, seeded by a secret key combined with the preceding tokens, then nudges sampling toward the favored side. That green/red split is one design among several; others steer the random draw itself with the key.

Nothing gets added to the output. The model writes ordinary prose, just with a faint statistical lean baked into which words it happened to choose.

Keyed sampling goes back to a 2022 proposal by Scott Aaronson, which Anthropic names as the origin of the family its Claude watermark belongs to. SynthID Text uses tournament sampling, and OpenAI's textGrain couples sampling to keyed randomness through an optimal-transport problem. The green/red version traces to Kirchenbauer, Geiping, Wen, Katz, Miers, and Goldstein's 2023 paper, the scheme most later research responds to. It splits each token's candidates into a green sublist and a red sublist, softly biases sampling toward green, and detects the mark with a statistical test on green-token frequency, with no access to the original model needed at detection time.

How does Google's SynthID Text work?

Confirmed

Google published a Nature paper, open-sourced a production implementation in Hugging Face, and says it ships inside the Gemini app. Robustness is a separate question.

SynthID Text builds on that same sampling bias with a more elaborate procedure the peer-reviewed Nature paper calls tournament sampling, where candidate tokens compete across sampling rounds before one is picked.

Google reports testing it in a live experiment across the Gemini app and web experience, gathering feedback on a figure widely cited as nearly 20 million responses. It found no detectable drop in output quality.

It's also open-sourced. A production-grade implementation ships in Hugging Face Transformers from v4.46.0, with a reference Bayesian detector that returns one of three verdicts: watermarked, not watermarked, or uncertain.

What does a watermark detector need to work?

Research/proposal

Roughly 800 tokens, at one false positive in 100,000. It is the field's clearest independent number, and it may not transfer to the system you care about.

A watermark detector needs the specific scheme, its configuration, and enough text to produce a meaningful score. No detector is universal.

Short, factual answers give it the least to work with. Google's own documentation says watermark application is less effective there, precisely because there's less room to change wording without hurting accuracy.

Detection on the original scheme is a counting test. The detector counts how many of a passage's tokens land on the green list and runs a one-proportion z-test against what pure chance predicts.

That gives an interpretable p-value and needs no access to the model at detection time. A watermarked passage shows more green tokens than chance would produce, and the excess only becomes statistically clear as the passage gets longer.

The clearest number doesn't come from SynthID Text at all, since Google hasn't published an equivalent figure. It comes from a 2024 stress test of the original green list, red list scheme.

Kirchenbauer and coauthors attacked that scheme three ways:

  • Human rewriting
  • LLM-generated paraphrasing
  • Dilution inside longer mixed documents

It stayed statistically detectable after strong human paraphrasing once a detector could examine roughly 800 tokens, at a false-positive rate of one in 100,000.

Treat that as evidence about the general approach's staying power. It is not a guarantee that transfers word for word to SynthID Text or to Claude's watermark.

The first figures for a production text watermark came from OpenAI on 2026-10-05, for textGrain, at a 1% target false-positive rate. About 80% of 200-token and about 95% of 400-token English passages were detected. Replacing 10% of the words in a 400-token passage with synonyms cut detection from about 92% to 66%, and replacing 25% cut it to 17%. These are the provider's own numbers, not an independent test.

Can paraphrasing defeat SynthID Text?

Research/proposal

Outside researchers red-teamed the open-sourced scheme rather than trust the claims. Paraphrasing, copy-paste splicing, and back-translation all measurably degraded detection.

SynthID Text is robust to mild paraphrasing, but detector confidence can be greatly reduced by thorough rewriting or translation, a central weakness Google's own documentation already concedes.

Sadasivan and coauthors' widely cited 2023 paper goes further, arguing on theoretical grounds that watermarking and other AI-text detectors aren't reliable against a genuinely motivated paraphraser. Broader surveys of the field single out robustness under adversarial editing as its central unsolved challenge, not a footnote.

Because SynthID Text is open-sourced, outsiders can test it directly rather than take Google's word, and several have:

  • A peer-reviewed 2025 study (IEEE TrustCom) found paraphrasing, copy-paste splicing, and back-translation all measurably degraded detection. The same researchers' hybrid defense improved detection F1 by an average of 11.1% over the vanilla scheme.
  • A 2026 preprint went after the detection math itself, proving the default mean-score detector becomes more vulnerable as more of SynthID Text's internal sampling layers are used, while the alternative Bayesian scoring method already built into the system holds up better against that specific attack.

None of this is a working bypass anyone can point to with confidence, and none of it makes the watermark bulletproof either. Heavy rewriting and translation reliably weaken the statistical signal, sometimes a lot, but weakened and gone are different claims, and no independent study has shown a paraphrase attack that defeats SynthID Text every time.

Can someone steal a statistical watermark's secret rule?

Research/proposal

Proven undetectable without the key at COLT 2024, approximated through public API queries for under $50 the same year. Both results hold.

Christ, Gunn, and Zamir gave keyed statistical watermarking a rigorous foundation at COLT 2024. They constructed a scheme where, without the secret key, watermarked text and ordinary text are computationally indistinguishable under standard cryptographic assumptions.

That proof assumes the key never leaks. In 2024, Jovanović, Staab, and Vechev at ETH Zurich's SRI Lab tested what it is worth once a real system sits behind a public API.

Querying the API for under $50 in total cost was enough to approximate the provider's secret green-list rule. With that approximation they ran two attacks:

  • Spoofing: forging the watermark onto text the model never wrote, succeeding more than 80% of the time.
  • Scrubbing: stripping it from text the model did write, pushing success from near 0% to over 85% in settings people had assumed were safe.

The cryptography was not wrong. The attack broke the assumption the proof rests on, not the proof itself.

A 2025 follow-up from the same lab (Gloaguen and colleagues, ICLR 2025) runs statistical tests that detect the presence and identify the family of the three most popular watermarking schemes from black-box queries alone, with no logits access.

Why don't all AI providers watermark their text?

Reported

Why isn't a technique this well documented running in the products you use? OpenAI's answer reportedly involved 99.9% effectiveness and 30% of its users.

Watermarking a model's text output requires the provider's cooperation. There's no way to add it from outside once the text exists.

OpenAI reportedly built an internal ChatGPT text-watermarking system rated around 99.9% effective, according to leaked internal documents reported by the Wall Street Journal, years before Anthropic's Claude rollout. Per that reporting, it held the system back over three concerns:

  • Circumvention risk
  • False positives
  • Disproportionate impact on non-native English writers

Internal survey data reportedly showed roughly 30% of users would use ChatGPT less if it shipped.

That held until 2026-10-05, when OpenAI announced textGrain, a different text watermark. It is shipping narrowly: ChatGPT and Codex text in the EU, rolling out over the coming weeks, an opt-in for API customers that stays off by default, and a detector limited to approved researchers. OpenAI says it is not making text watermarking a global default at launch. The AI watermark status tracker follows each provider's position.

There's also a structural limit no policy decision fixes. Once open-weight models ship, whoever runs the model controls decoding, so the original provider can't apply sampling-time watermarking to them at all.

SynthID Text is open-sourced partly for that reason: anyone operating a deployment can switch it on. Whether they do is a property of that deployment, not of the underlying model.

FAQ

Do zero-width spaces or other invisible Unicode characters prove AI text watermarking?

No. Invisible Unicode characters are usually formatting artifacts or copy-paste residue, and even where they're inserted deliberately, that's a different, older steganography technique, not the mechanism behind SynthID Text or any other confirmed provider watermark. Real statistical text watermarks shape which words get chosen during generation, leaving no discrete character to find.

Can open-source or self-hosted models use text watermarking?

Yes. SynthID Text is open-sourced and ships as a production-grade implementation in Hugging Face Transformers, v4.46.0 and later, including a reference Bayesian detector. Whether a specific deployment actually applies it depends entirely on whoever operates that system.

How much watermarked text does a detector need to be confident?

The best documented figure comes from a 2024 stress test of the original green list, red list scheme, not SynthID Text specifically: it stayed statistically detectable after strong human paraphrasing once a detector had roughly 800 tokens to examine, at a false-positive rate of one in 100,000. Shorter passages and heavier rewriting both make detection harder.

Has anyone verified SynthID Text's robustness independently of Google?

Yes, more than once. A peer-reviewed 2025 study found it vulnerable to paraphrasing, copy-paste splicing, and back-translation, and proposed a hybrid defense improving detection F1 by an average of 11.1%. A 2026 preprint separately proved its default mean-score detector grows more vulnerable as more of its internal sampling layers are used, while the alternative Bayesian scoring already built into the system held up better.

If watermarking works, why hasn't every major AI provider shipped it?

Because the tradeoffs are real, not theoretical. OpenAI reportedly built an internal ChatGPT text watermark rated around 99.9% effective and held it back over false positives, circumvention risk, and disproportionate impact on non-native English writers, with internal survey data reportedly showing about 30% of users would use the product less if it shipped. On 2026-10-05 OpenAI announced a different system, textGrain, but only for EU ChatGPT and Codex text, rolling out over the coming weeks, and opt-in API use. Anthropic and Google made different calls, but the underlying tension hasn't gone away.

Does a statistical text watermark change how the text reads?

Not in a way a casual reader would notice. It only shifts which words get chosen among options the model already considered reasonable, so no visible character or label gets added. The difference shows up only in the statistics of token choices when a detector with the right key checks for it.

Can someone actually steal a provider's secret watermarking key?

Not the literal key, but researchers got close enough for it to matter. ETH Zurich's SRI Lab showed that querying a provider's API for under $50 was enough to approximate the secret green-list rule, forge the watermark onto text the model never wrote, and strip real watermarks with better than 85% success. That is a research demonstration against tested academic schemes, not a confirmed attack on any specific shipped product.

Next steps

Sources and citation status