Copy a text produced by Claude and paste it into an email, a document or a website. With the new models from Anthropic, an invisible signal will be able to travel together with the wordseven surviving some changes. No writing “generated by AI” at the bottom of the page, no stamp appearing on the screen: the watermark is incorporated by the model while it generates the text and can be identified through automatic tools.
The novelty arises from the transparency obligations of theEuropean AI Actapplicable from 2 August 2026. Anthropic has signed up to the European Code of Practice on Transparency of Content Generated by Artificial Intelligence and has explained how it intends to brand Claude’s outputs. However, there is an important distinction: the watermark inserted in the texts and the C2PA standard used for some files are two different tools. Putting them in the same container risks making the system seem much simpler than it is.
The watermark is written directly inside the text
For the lyrics Anthropic talks about a watermark embedded directly in the output. It is imperceptible to the reader and, according to the company, does not change the meaning, quality or readability of the response. Since it is inserted at the model level, it does not depend on the interface used: it affects the outputs of models supported on Claude, Claude Platform via API, Claude Code, Claude Cowork and Claude Tag. The watermark should also be present when the same templates are used through AWS, Google Cloud, or Microsoft Foundry.
Verb choice matters: Anthropic says the signal can survive copy-paste and some editing. It doesn’t promise an immortal digital tattoo. Heavy rewriting, paraphrasing, translating, or merging with other material can make the watermark no longer detectable. Even a very short passage may contain too little text to get a reliable result. And the absence of the signal, therefore, .
There is also the opposite problem. Finding the watermark does not necessarily mean you have discovered who wrote those sentences. Claude may have been used to correct, summarize or translate a text originating elsewhere. Anthropic itself specifies that a detected mark indicates that the content it may have been drawn up by Claude; alone it does not reconstruct its entire history. We therefore do not yet have a universal author detector. We have a lead.
The C2PA enters the files generated by Claude
For images and other supported file types the mechanism changes. When Claude generates, for example, a JPG, PNG or SVGAnthropic plans to attach provenance metadata digitally signed according to the open standard C2PA, Coalition for Content Provenance and Authenticity.
Here we are not looking for a particular imprint in the sequence of words. A sort of cryptographic identity card is associated with the file which can contain information on its origin and allow you to verify whether the content or associated information has been modified. The C2PA standard uses digital signatures and cryptographic hashes to make alterations to source data detectable.
C2PA also provides more resistant systems, the so-called soft bindingwhich can exploit invisible watermarks or fingerprints of the content to find the credentials of origin if these are separated from the file. The standard, however, openly recognizes a rather practical limitation: metadata can be removed.
In Claude’s case you just need to think about everyday operations such as a format conversion, a new save or a screenshot. Anthropic lists these cases among those that can make detectable source data disappear. Here too, therefore, no indelible mark.
And C2PA certifies provenance, not truth. The same organization that develops the standard specifies that knowing the origin and history of content can help evaluate it, but . An AI-generated photograph can have perfectly valid provenance credentials and still depict something that never happened.
The AI Act calls for machine-recognizable content
The change of direction comes from article 50 of the AI Act. From 2 August 2026 the suppliers concerned must adopt systems that allow mark in machine-readable format and detect content generated or manipulated by artificial intelligence. The European Code of Good Practice defines common technical measures for text, images, audio and video. Adherence to the Code is voluntary; the obligations established by law, however, are not.
Anthropic has chosen to apply the markings of the models also supported outside the European Union. For i new Claude models launched in the EU from 2 August 2026 the system should be available from launch; for the previous ones the company is still working on the implementation.
On models already on the market the regulatory situation deserves caution. The European Commission reports that the proposal AI Omnibusagreed by the co-legislators, provides – if definitively adopted – a transitional period until 2 December 2026 to adapt some systems placed on the market before 2 August. It is therefore incorrect to treat that date as a general extension already set in stone.
The regulation also distinguishes the obligations of suppliers from those of those who use these systems professionally. Transparency rules are foreseen for texts generated or manipulated by AI and published to inform the public on issues of general interest, with an important exception when the content is subjected to human review and remains under editorial responsibility. A rather relevant difference for newspapers, publishers and other entities that use generative tools within a normal editorial process.
For now there is still one piece missing: the control tool
Anthropic promises that users and third parties will be able to verify the presence of its watermarks and provenance metadata. At the moment, however: the technical documentation will arrive later.
This is perhaps the most useful limitation to keep in mind. Claude’s watermark will be able to provide a signal of provenance, C2PA will be able to tell something about the history of a file, and the AI Act tries to build common rules around these tools. However, none of the three transform a suspicious text into laboratory evidence.
A watermark found can tell us that Claude passed by there. One that is missing may simply have been deleted along the way. AI begins to leave footprints; identifying every step with certainty is yet another job.
