Anthropic explains how Claude’s text watermark works and what it can detect
Image generated by the Inferama Editorial Team
01

What Anthropic has announced

Anthropic describes a technique that embeds an imperceptible statistical signal in text generated by compatible Claude models. The company says the method is intended to help estimate whether Claude contributed to a passage, as part of its commitments on transparency for AI-generated content.

Anthropic’s page, published on November 13, 2026, presents the technique as a version of the SynthID-Text approach published by Google DeepMind in a 2024 Nature paper. The company’s account explains the design, but is not an independent evaluation of the detector’s effectiveness.

Coverage depends on the model and the surface where it is used. Anthropic says that Claude models launched in the European Union from August 2, 2026, will support watermarking from launch, and that it is working to add it to earlier models. Its support documentation lists specific models and conditions, so the announcement should not be read as automatic coverage for every version and feature.

Anthropic says it applies the watermark to outputs from compatible models across its products and certain cloud services, and that text watermarking is deployed globally. Availability of each type of marking depends on the configuration: for example, the documentation specifies conditions for some Sonnet 4.6 outputs and for C2PA credentials on files. There is no basis for claiming that every response, file, or integration always carries every signal.

Text detection is in private preview, available to eligible organizations and certain businesses with verification obligations. Anthropic plans to expand access, but does not describe the API as a general-purpose tool available to everyone.

What is marked and what can be checked

The method and its scope vary by content type and platform compatibility.

Content or featureSignal describedMain limitation
Text generated by a compatible modelA statistical pattern embedded in word-generation choicesDetection estimates possible Claude involvement; it does not prove exclusive authorship.
A compatible file generated by ClaudeA C2PA Content Credential in the metadata, where the feature and platform support itThe credential indicates that Claude processed the file; it does not by itself establish who created the original content.
Text or file with no detected signalNo compatible signal recognized by the detectorFailure to detect a signal does not prove that Claude or another AI was not involved.
02

How the text signal works

Language models generate text by choosing successively among possible words or word fragments. At many points, several options can sound equally natural without changing the meaning. Anthropic says its method uses some of these low-stakes choices to create a statistical pattern.

According to the company’s description, a key and the context of preceding words influence the selection among options the model might already choose. A detector can then check whether a sequence is consistent with the choices associated with that key and estimate the likelihood that Claude was involved.

The technique does not insert a visible label or hidden characters into the text. Anthropic says the signal changes the source of randomness used for certain choices, not the meaning the model is trying to express. That describes the mechanism as the company presents it; it does not guarantee that every response will retain identical quality or replace performance testing.

There is less room to apply a watermark when a response requires an exact choice. In a factual statement with only one correct answer, or in code that must use a precise term, favoring another option could introduce an error. Anthropic therefore says the signal is generally sparser in factual text and code than in prose with more valid alternatives.

The watermark also needs enough usable decisions to form a detectable pattern. A brief greeting, a fixed phrase, or a very short sentence offers fewer opportunities than a longer passage. Length alone does not guarantee detection: the content, available choices, and extent of later changes all matter.

03

What detection shows—and what it does not

A positive detection means the text contains a signal compatible with Claude’s watermark. Anthropic describes this as an indication that Claude may have been involved at some point, not conclusive proof of the content’s full provenance.

The signal cannot distinguish between text written from scratch by Claude and text that Claude reviewed, translated, summarized, or edited. The underlying material may come from a person or another source, while words added or changed by the model may contribute to the pattern.

The watermark does not identify a user, organization, or specific conversation either. Anthropic says the signal and its key contain no data that would allow such identities to be recovered. Detection concerns possible system involvement, not who requested or authorized the work.

A negative result does not rule out the use of Claude. The detector may fail to recognize a signal if the model did not support watermarking, if the text was changed, if the passage is too short, or if the platform did not apply that type of watermark. Paraphrasing, translating, or mixing the content with other writing can weaken the signal; light editing may not remove it completely.

A detected watermark should therefore not be used on its own to decide who the author is, whether a work violates a rule, or whether content is entirely artificial. The signal offers limited information about the process, not a complete account of how the text was created or transformed.

How to interpret results cautiously

Read the result as a signal of possible involvement, not as an automatic decision about authorship.

ResultReasonable conclusionConclusion it does not support
A compatible watermark is detectedClaude may have generated or processed some of the content.Claude wrote the entire text or a specific person commissioned it.
No watermark is detectedThe detector found no compatible signal in that material.The text must be human-written or never passed through Claude.
The signal appears in a translation or editClaude may have selected some of the words present.The original content belongs to Claude or its intervention was substantial.
04

Editing, translation, short text, and code

The strength of the signal depends on how many useful choices the model could make and how much of the marked text remains. Anthropic says short passages are difficult to assess because they contain fewer generation decisions; a short sentence may not provide enough information for a reliable estimate.

A full translation offers more opportunities for a watermark than light proofreading, because the model chooses many words again. By contrast, when Claude reviews human-written text without changing much of the wording, it may make too few changes for a detectable signal. The outcome depends on the length and extent of the model’s intervention.

Paraphrasing and editing can weaken detection, but the documentation does not quantify how well the system withstands each transformation or provide a general success rate. Anthropic says a complete rewrite can remove the watermark, while light editing may not. There is no public threshold for predicting the result for every text.

In code, many choices must meet exact syntax or output requirements, leaving fewer safe alternatives where a pattern can be applied. Anthropic says comments may allow some flexibility, but that does not mean the whole program will be detectable or that the watermark validates its functionality.

These limitations also matter when assessing an alert. Before attributing a text to a tool, consider whether it is short, highly factual, translated, edited, or generated using a version or surface without compatible marking. Such conditions can explain detection failures; they cannot establish in advance whether a result is true or false.

How to assess a signal without overinterpreting it

A useful check separates the detector’s result from the conclusion someone wants to draw.

  1. 01Confirm that the model, feature, and content type could carry a watermark in that configuration.
  2. 02Check whether the passage is long enough and allows flexible wording choices.
  3. 03Account for translation, editing, paraphrasing, format conversion, or mixing with text from other sources.
  4. 04Describe the result as a signal compatible with possible involvement, not proof of exclusive authorship.
  5. 05Use other relevant information for decisions about authorship, rules, or provenance.
05

Text detection and file credentials

Anthropic describes two distinct mechanisms. Text can carry a statistical pattern embedded in word choices; compatible files may include C2PA Content Credentials, signed metadata that records that Claude processed the file.

A C2PA credential is not the same as a text watermark. Anthropic’s documentation says one may be attached when Claude generates a compatible file and the platform supports the feature. C2PA-compatible tools can read the credentials, but their presence does not by itself establish who originally authored the content.

Metadata can disappear when a file is converted to another format, saved again, or captured in a screenshot. So the absence of a credential does not show that Claude was not involved. Coverage also depends on the specific feature and the platform used to create the file.

The text watermark detection API remains in private preview for the categories of organizations Anthropic describes. The company says it plans to expand access, without setting a general availability date in this documentation. The existence of the detector should not be confused with universal public access.

Anthropic presents watermarking as part of its commitments associated with the European Union’s AI Act Code of Practice. The company says it applies the watermark globally at launch because it does not have a durable way to limit it by region. This is Anthropic’s stated approach; watermarking alone does not resolve questions of ownership, liability, or compliance in a specific case.

06

What remains unknown about effectiveness

Anthropic’s pages explain the design, intended uses, and limitations the company recognizes. Those materials do not publish a complete independent evaluation of accuracy, false positives, and false negatives across different passage lengths, languages, models, and transformations. An explanation of the system is not the same as external validation.

Anthropic refers to Google DeepMind’s work on SynthID-Text and summarizes quality tests conducted by that company. Those results belong to the evaluation described in the paper Anthropic cites; they should not be presented as an independent measurement of Claude’s specific deployment or as evidence that every edit preserves the watermark.

Anthropic’s corporate page says its internal tests found no impact on content, creativity, or readability. Because this is Anthropic’s account of internal testing, it does not by itself establish that the effect is zero across all tasks or configurations.

A rigorous evaluation would need reproducible results specifying model version, language, passage length, configuration, and type of edit, as well as detection thresholds and error rates. Without those details, the responsible conclusion is limited to the scope the company describes: a probabilistic signal of possible Claude involvement, with important limits for short passages, transformed text, and unsupported outputs.

What the documentation supports

A public description of the mechanism and demonstrated performance are different things.

ClaimStatus in the documentation
Anthropic describes a technique based on statistical patterns in generation.The company’s primary account of its system design.
The detector can estimate whether Claude was involved.A stated probabilistic scope; it does not confirm full authorship.
The watermark works equally well across every language, length, or edit.Not established by the materials described.
Watermarking guarantees correct content or a signal that never disappears.Unsupported; the watermark does not validate accuracy and is not infallible.
07

Keep exploring

Sources consulted

03

Corrections and transparency

If you spot incorrect or outdated information, send us a correction with the page and source we should review.

Submit a correction