Skip to content
Figure 1Strength of watermark evidence by text length, unedited output (worked example)
Strength of watermark evidence by text length, unedited output (worked example)5 z10 z400 words9.8 z200 words6.9 z100 words4.9 z50 words3.5 z25 words2.4 z
Illustrative z-scores for the same watermark strength. Evidence grows with length; a z of about 4 is a common bar for a confident call.

Text watermarking for language models: what it can show

Detectors guess from the outside. A watermark is an attempt to work from the inside: the system that generates the text deliberately leaves a statistical trace that a matching test can find later. It is the most principled approach to identifying machine text, and it has hard limits that are worth understanding before anyone relies on it.

Quick answer

A text watermark biases the generator's word choices towards a secret subset of words. A detector holding the key counts that subset's share and runs a significance test. On long, unedited text from the watermarking system this gives strong evidence. It says nothing about other systems' text, weakens with editing or paraphrase, and is weak on short texts.

Key figures

Mechanism
biased word choice from a keyed list
Detection
a significance test, reported as a z-score
Needs
the generator's cooperation and key
Strong on
long, unedited output
Weak on
short, paraphrased or mixed text
Proves
a statistical pattern, not a person's intent

How a statistical watermark works

In the scheme Kirchenbauer and colleagues described in 2023, each time the model picks the next word, the preceding word is used with a secret key to split the vocabulary into a green list and a red list. The generator adds a small bonus to green words, so its output uses them more often than chance. A person reading the text notices nothing. A detector with the key recomputes the lists for every word and counts green words. In unwatermarked text about half are green; in watermarked text many more are, and a standard significance test turns the count into a z-score.

The appeal for testers is that the detection step is a well-defined statistical test with a false positive rate that can be set in advance, unlike the learned scores discussed in how AI text detectors work. With a high enough z threshold, the chance that ordinary human writing trips it is tiny.

What a watermark test can and cannot show

A watermark answers one narrow question well
QuestionCan the watermark answer it?Why
Did this long passage come from the watermarking system, unedited?Often, with strong evidenceMany words carry the signal
Did a person use any machine at all?NoOther generators leave no watermark
Is this short reply machine-written?RarelyToo few words for a confident test
Was this text paraphrased from watermarked output?SometimesParaphrasing removes much of the signal
Did the writer intend to deceive?NoA statistical pattern says nothing about intent

Removal and robustness

Because the signal lives in word choice, changing the words erodes it. Kirchenbauer and colleagues' follow-up study on reliability found that watermarks remained detectable after human rewriting and machine paraphrasing if enough text was observed, while noting that paraphrasing weakens the signal and more text is needed to recover confidence. Work on adversarial testing of detectors shows the same tension for every detection approach: the more the text is reworked, the less any fingerprint survives.

Rule of thumb

Treat a strong watermark result on a long, unedited passage as meaningful evidence about where that passage came from. Treat the absence of a watermark as meaning almost nothing.

Testing a watermark scheme

  • Measure the false positive rate on a large, varied set of human writing at the chosen z threshold, including formulaic and non-native text.
  • Measure detection by length band, because the evidence depends on how many words are available.
  • Measure detection after light editing, paraphrasing and translation, holding the threshold fixed, and report the drop.
  • Check text quality with and without the watermark, since a strong bias towards green words can make writing worse.

A worked run, with invented but realistic counts: 2,000 human-written documents of 300 to 900 words scored at a z bar of 4 produced 1 flag, a false positive rate of 0.05 percent with an upper bound near 0.3 percent. At a bar of 3 the same set produced 6 flags, 0.3 percent, and the one document flagged at both bars was a legal template full of repeated fixed phrases. Detection on 2,000 unedited watermarked documents was 1,960 at z of 4 and fell to 1,140 after a paraphrase pass, a drop of 41 points. Both halves of that record belong in the report, along with the version of the paraphraser.

Worked example: green-list counts behind the z-scores

The z-scores in the figure come from a simple calculation, shown here with invented but realistic counts. Suppose the key splits the vocabulary in half, so an unwatermarked text lands on the green list about 50 percent of the time, and the watermarking generator lands on it about 74 percent of the time. For a passage of 400 scored tokens the expected green count by chance is 200 with a standard deviation of 10, so 298 green tokens gives z of 9.8. The same passage paraphrased by a second tool keeps 232 green tokens, z of 3.2, below the bar of 4. An honest human passage of the same length lands on 206, z of 0.6.

Invented but realistic counts. The last two rows give the same answer for different reasons, which is the limit of the method: absence of the watermark says nothing about how the text was produced.
TextScored tokensGreen tokens (expected by chance)z-scoreCall at z of 4
Watermarked, unedited400298 (200)9.8Watermark found
Watermarked, 50-token excerpt5037 (25)3.5Inconclusive
Watermarked, then paraphrased400232 (200)3.2Inconclusive
Human-written400206 (200)0.6No watermark
Machine text from another system400197 (200)-0.3No watermark

Common mistakes with watermark tests

  • Reading no watermark as human-written. Text from any system without this key, and paraphrased text from this one, both return no watermark.
  • Lowering the z bar to catch more edited text. Every point off the bar raises the false positive rate on honest writing by a calculable amount; measure it before moving it.
  • Scoring excerpts. The evidence grows with length, and a 50-token excerpt of a strongly watermarked text is inconclusive by design; score the whole text or report the token count with the result.
  • Assuming the key is stable. Providers rotate keys and schemes; a test that worked last year may not read this year's output, so record which key version each run used.
  • Skipping the human false positive measurement because the theory gives a rate. Real writing has repeated phrases, lists and quoted text that the independence assumption behind the z-score does not model.

Watermarks and content credentials

Watermarks for text are related to, but different from, provenance records for images and media such as the C2PA content credentials specification, which attach signed metadata rather than hiding a pattern in the content. Both answer where something came from; neither answers who is responsible for how it was used, which is why the guide to proving you wrote it focuses on process evidence. The two also fail differently. A credential is either valid or broken, and any edit outside a signing tool breaks it, so testing one is a matter of checking signatures and manifests, as the guide to verifying content credentials describes. A watermark degrades gradually with editing, so testing one is statistical. A pipeline that carries both, for example a generator that watermarks its text and a publishing tool that signs the file, needs both kinds of test, and the guide to testing content credentials pipelines covers the signing half.

Three ways to ask where a text came from; only the credential is verified rather than estimated
ApproachWhere the evidence livesSurvives editing?False positive controlWhat absence means
Statistical watermarkIn the word choices of the textPartly; degrades with editsSet by the z bar, then measuredNothing
Content credentialIn a signed manifest attached to the fileNo, unless re-signed by a toolCryptographic; a forged one failsNothing, or stripped on upload
Learned detector scoreInferred from the finished textWeaklyMeasured on a benchmark, often highWeak evidence either way

Common questions

What is a text watermark for AI?

A hidden statistical pattern that a text generator deliberately embeds in its word choices, which a detector holding the key can test for later.

Can watermarks prove a text was written by AI?

They can give strong evidence that a long, unedited passage came from the watermarking system. They cannot show whether other generators were used or what a writer intended.

Can a watermark be removed?

Editing and paraphrasing weaken it because the signal is in word choice. Research found it can survive moderate rewriting if enough text remains, but heavy rewriting removes most of it.

Do watermarks cause false positives?

Rarely at a high threshold, because the test's false positive rate can be set in advance. It must still be measured on real human writing.

Why can't short texts be checked reliably?

The test counts words that match the secret list; with only a few dozen words, chance variation is too large for a confident result.

Do text watermarks survive translation?

Mostly not. The signal lives in the specific words the generator chose, and translation replaces all of them. Published work found some signal survives light rewriting; a full translation is far beyond that.

Sources

  1. Kirchenbauer et al., A Watermark for Large Language Models
  2. Kirchenbauer et al., On the Reliability of Watermarks for Large Language Models
  3. C2PA, Content Credentials technical specification 2.1

This guide is part of the testing AI systems hub. It is best read alongside how ai text detectors work and adversarial testing of ai text detectors, which cover the neighbouring questions.