Anthropic plans to make future Claude models produce invisibly watermarked text as part of its compliance with the European Union’s AI Act. The company will initially apply the system worldwide rather than limiting it to Europe and expects to provide an API that can assess whether a passage was likely written, at least in part, by Claude.
The watermark will be based on SynthID-Text, a technique introduced by Google DeepMind and described in a 2024 Nature paper. It creates a statistical signature by altering how a language model makes low-stakes choices among suitable words. Anthropic said the implementation should have negligible effects on generation speed and will not add tokens or increase the cost of serving the model.
Language models generate text by repeatedly selecting the next token from a set of candidates. In many cases, several candidates can express substantially the same idea without changing the answer’s quality or accuracy. A conventional model may use arbitrary randomness to resolve those choices. With watermarking enabled, Claude will instead derive that randomness from a secret key and the preceding context.
Over a sufficiently long response, those key-guided decisions form a pattern that a compatible detector can test. The mark is not visible to readers, and it does not require extra characters, metadata or an attached label. Detection produces a probability rather than definitive proof that Claude generated the text.
Anthropic said its internal evaluations found no effect on the content, creativity or readability of Claude’s responses. It also pointed to Google DeepMind testing in which users and human evaluators did not register a statistically significant quality difference between watermarked and unwatermarked output. The technique is designed to influence selections only when multiple plausible options are available, rather than forcing the model to choose an unsuitable or obscure term.
The system has important technical limits. Short passages contain relatively few word-selection decisions, leaving less evidence for a detector to analyze. Confidence should generally improve with longer samples because they provide more opportunities for the keyed pattern to appear.
Watermarks may also be weak or absent in material where exactness sharply limits the model’s options. Factual completions, equations and executable code often have one clearly correct next token, so applying a statistical nudge could damage accuracy or functionality. Code comments and other flexible language within a programming response may still carry the pattern, but the code itself is expected to contain less watermarking than ordinary prose.
Light editing creates a similar challenge. If a user asks Claude to correct only grammar and punctuation in human-written material, most of the returned words originated with the user. The relatively small number of model-selected changes may not supply enough evidence for reliable detection. More extensive rewriting gives the watermark more room to register.
A successful match will not establish that an entire document came from Claude, nor will a failed match prove that a person wrote it. Anthropic’s detector will be tied to the company’s key and cannot identify text made by another provider using a separate key or a different watermarking method. Rephrasing and mixed human-AI authorship also complicate any attempt to treat watermark detection as a universal authorship test.
Anthropic said the watermark itself will not encode information about a user, organization or conversation. Its purpose is to identify a statistical relationship between the output and Claude’s generation process, not to trace text to an account.
The company is adopting the measure after joining the EU Code of Practice on Transparency of AI-Generated Content in July 2026 alongside other major model providers and roughly 190 signatories in total. The code calls for providers to mark AI-generated material. Anthropic said it lacks a durable way to confine the feature by region, leading it to choose a global rollout at launch while it continues evaluating implementation options.
A forthcoming detection API will give outside users a way to check text for Claude’s watermark. Anthropic has not provided launch details for that service, leaving access arrangements and the practical thresholds for interpreting results to be specified later.
Sources: Anthropic