Anthropic plans to equip future Claude models with a text-watermarking mechanism that allows anyone with the appropriate key to estimate the likelihood that Claude participated in creating a text. In a post dated August 14, 2026, the company explained that the watermark would not be visible to readers, would not add hidden characters or metadata, and would not require additional tokens or increase the cost of running the model.
The move comes as part of Anthropic’s compliance with the European Union’s AI Act. In July 2026, the company, along with other major model providers and approximately 190 signatories in total, signed the European code of practice on transparency for AI-generated content. According to Anthropic, the European Union has required AI providers serving its market to label AI-generated content since August 2.
How does the watermark work?
Language models generate text one word at a time, choosing at each step from several possible words. When multiple options are close in meaning and quality, the watermark uses a key and a set of preceding words to determine the source of randomness that resolves the choice. The choice therefore remains effectively random, but the sequence of words leaves a pattern that can later be detected using the key.
Anthropic says this mechanism does not push Claude to choose words it would not normally use, nor does it permanently bias it toward a particular word. The company compares it to using a sequence of digits from the value of pi instead of rolling dice: the result appears random, but analyzing the sequence of results may reveal the source of the randomness.
No practical effect on text quality
According to internal tests, Anthropic observed no effect on the content, creativity, or readability of Claude’s text. The company also points to the SynthID-Text study, which tested the technology on part of Gemini’s traffic and found no statistically significant differences between evaluations of watermarked and unwatermarked text. In a controlled study, human evaluators were unable to notice a difference in quality when comparing answers side by side.
The company stresses that the watermark slows the models only negligibly and does not make the service more expensive because it does not generate additional tokens. It also contains no information that could reveal the identity of the user or their organization or conversations.
Detection limits and editing and programming cases
Detection effectiveness depends on the length of the text and the number of decisions in which the model has more than one suitable option. Detection is therefore weaker in short samples and in factual passages that allow only one precise answer, such as stating the correct name of a work by Isaac Newton. The same applies to light proofreading, since most of the words remain the user’s own and the watermark has too little on which to rely.
In programming, the watermark is generally not applied when exact output is necessary, because changing a word or symbol could break the code. However, it may appear in places with arbitrary choices, such as code comments, with little effect on the code itself. Translations generated by Claude, meanwhile, carry the watermark because the model chooses all of their words.
Availability and what the watermark proves
Anthropic is working to provide an API for detecting the watermark, but it has not yet specified implementation details. The watermark will be applied globally at launch because the company currently has no sustainable way to limit it by region. The company is also working to add it to Claude models released before August 2, 2026, taking advantage of the transition period set out in European law, with this to be completed over the coming months.
The watermark does not prove that Claude wrote the text in its entirety; it only indicates the likelihood that Claude participated in producing or processing it, and does not distinguish between original writing and extensive editing. Rewriting the text completely may remove the watermark, while light editing may not. The watermark does not change the user’s rights, ownership of the content, or legal responsibility for it.
For supported files such as PNG, JPG, and SVG, Claude will add content credentials in the form of a cryptographically signed note in the file’s metadata, in accordance with the open C2PA standard. Anthropic says this data does not change the file or include identifying information, and can be read by tools compatible with C2PA.