Anthropic Prepares Statistical Watermarking for Text Generated by Claude

A Global System Driven by European Requirements

Anthropic is preparing a statistical watermarking system to identify text produced by Claude. The initiative, described on August 14, 2026, comes in the context of the transparency requirements under the European Union’s AI Act and the company’s participation in the related Code of Practice alongside other AI providers.

The watermark will initially be applied globally, not just to users in the European Union. Anthropic does not yet have a sufficiently reliable mechanism for enabling or disabling it based on region.

Future Claude models will generate text covered by the new system. Models introduced before August 2, 2026, fall under the transitional period established by the European Union. Anthropic is working to extend the watermark to those models as well, but has not disclosed an exact list of affected versions or a detailed timeline.

For users, the change should remain invisible. The system will not add hidden characters, extra tokens, or visible markers. According to internal assessments, it should not materially affect the creativity, readability, content, or quality of responses. No additional costs are expected, and the impact on generation times should be negligible.

The Signature Is Built During Generation

The chosen technology is based on the SynthID-Text approach developed by Google DeepMind. The watermark is not added after a response has been generated; it is built as Claude selects each subsequent token.

When multiple linguistically plausible continuations are available, the model uses a secret key and part of the preceding context to influence the random component of the selection. Each individual choice remains normal and does not exhibit anomalies visible to the naked eye. Taken together, however, the sequence tends to produce a statistical signature associated with the model and the key.

A detector with access to the key can examine the text and calculate how closely its patterns match those expected from Claude when operating with the watermark. It does not necessarily require access to the proprietary model or particularly intensive processing: the analysis focuses on the distribution of linguistic choices already present in the text.

Effectiveness depends primarily on the amount and variety of available material. A longer text provides more decisions to analyze and therefore more statistical evidence. The system also performs better when the model has many plausible alternatives to choose from.

Conversely, a short or highly predictable response contains less useful information. In a mathematical formula, a rigid definition, or an answer with only one correct continuation, there is very little room to embed a signature without changing the meaning.

Factual Responses and Code Will Contain Fewer Signals

Anthropic will not necessarily apply the watermark when choosing an alternative could make the response incorrect. For example, after “2 + 2 =”, the correct token is “4”: selecting an alternative to strengthen the signature would compromise the reliability of the result.

The same limitation applies to factual responses with only one correct continuation, such as a numerical answer referring to George Orwell’s Nineteen Eighty-Four. In such situations, the model must prioritize accuracy over detectability.

Code presents a similar challenge. Changing a token can break syntax or alter a program’s behavior. As a result, code blocks will generally contain fewer watermark signals than prose.

The system may still operate where arbitrary choices are available, such as in code comments. Anthropic believes the impact on the actual functioning of programs will be negligible, but it has not specified uniform coverage across all programming languages or generation scenarios.

Translations produced by Claude should also retain the watermark. In this case, the model selects each word in the output language, creating new linguistic decisions that the detector can analyze.

The API Will Not Certify Who Wrote the Text

Anthropic is developing an API capable of estimating the likelihood that Claude contributed to the production of a piece of content. The result will not constitute definitive proof of the author’s identity.

A positive detection may indicate that Claude was used at some stage of the work, but it will not necessarily distinguish between text generated almost entirely by the model and human-written text that was substantially edited. Nor will the watermark demonstrate that Claude was the only AI system involved.

The signature may survive limited edits, such as grammatical corrections or punctuation changes, because many original words remain unchanged. A complete rewrite replacing the entire text should eliminate the signal. Very short samples also reduce the reliability of the assessment.

The detector must therefore be treated as a probabilistic tool for attributing possible involvement to Claude. It cannot serve as forensic certification of the author, exclusive origin of the text, or absence of human intervention.

Anthropic has not yet disclosed the API’s operational details. Availability, authentication methods, response formats, confidence thresholds, and specific procedures for handling false positives and false negatives remain unknown.

C2PA Provenance Is Coming for Images and Graphic Files

For PNG, JPG, and SVG files, Anthropic will use a different mechanism. Claude will add cryptographically signed provenance metadata compliant with the C2PA standard, indicating that the file was created or processed with Claude.

In this case, the image’s binary content will not be altered to embed a statistical signature. Instead, the information will be associated with the file through verifiable metadata whose validity can be checked using the digital signature.

This introduces an important technical distinction: for text, provenance is estimated by examining linguistic choices; for graphic files, it is declared through provenance credentials linked to the content.

The practical protection of the metadata will depend on sharing workflows and the tools used to modify or convert files. However, the announced system does not yet explain how C2PA metadata will be handled when an image is recompressed, cropped, or exported from incompatible software.

What Changes for Users and Organizations

For Claude users, the watermark should be imperceptible and require no specific configuration or action. Organizations may use the future API to conduct internal checks, analyze suspicious content, or support transparency processes.

Results should be assessed in context. A positive score does not prove that the text was written entirely by Claude, while an inconclusive result may be caused by the text’s brevity, predictability, use of code, or subsequent modifications.

The source does not describe a vulnerability or assign a cybersecurity severity rating. The impact primarily concerns regulatory compliance, transparency, and the probabilistic attribution of AI-generated content.