Anthropic announced an important innovation: the application of a invisible watermark on texts generated with Claude. This watermark will be used for make recognisablethrough the use of some automatic tools, the artificial nature of a textso that anyone can know if this was somehow processed with Anthropic’s AI. The innovation will allow Anthropic to comply with article 50 of the European AI Act, which came into force on August 2nd, with the C2PA (Coalition for Content Provenance and Authenticity) system. The novelty, however, will not remain confined to the European market alone and will be extended by the company outside the EU.
How Antropic’s watermark works for Claude’s lyrics
The news announced by Anthropic on this Claude support page is part of the commitments that the company directed by the Amodei brothers has undertaken in the context ofarticle 50 of the Code of good practice on the transparency of content generated by AI provided for by the European AI Act.
Due to this, Claude models launched in the European Union since August 2, 2026 they will be the first to be affected by the news. Anthropic is also working to introduce it also in models launched before August 2nd, for which the legislation provides a deadline set at December 2, 2026. Starting from this date, therefore, previous models will also have to adapt.
The invisible watermark will be applied to content produced by AI models regardless of the service through which they are provided: Claude, the Anthropic API platform, Claude Code, Claude Cowork, Claude Tag, etc. The application of the watermark is also expected when access to the models is via AWS, Google Cloud or Microsoft Foundry, even if Anthropic specifies that «Signed provenance metadata may not be supported on all platforms, depending on the features offered by each platform».
The watermark will be embedded directly into the generated text content. In Claude’s case it will be a watermark invisible to the human eye that should not change the meaning, quality or readability of the answers provided by the AI. The signal is inserted directly by the model as it generates the text, so it does not depend on the product or interface you are going to use.
An interesting feature concerns the fact that the watermark can follow the text even when it is copied and pasted elsewhere. It can also survive some changes, albeit not too heavy ones (but we’ll come back to this shortly). This means that the watermark will not simply be a label visible next to Claude’s response, but ainformation invisible and embedded in the text itselfdesigned to be identified only through specific tools.
What about the files generated with Anthropic models? For example images in JPG, PNG or SVG format? In this case the template will attach some digitally signed provenance metadata. These will follow the open standard C2PA (Coalition for Content Provenance and Authenticity), developed to record information on the provenance of digital content.
The limits of the C2PA system
In all this it is good to take some things into account limits of technology. As Anthropic explains «the absence of a detectable mark does not mean that the content was not generated or processed by AI». This is because the content generated by Claude may not have a detectable watermark if the text has been heavily reworked, paraphrased, incorporated into other texts or translated and the same applies if the passage is short and there is not a sufficient amount of text to make an accurate analysis. As for the file metadata, however, these may have been removed via format conversion, saving, screenshots, and so on.
