Anthropic announced that, starting August 2, it will add machine-readable markers to content generated by Claude, employing a two-layer mechanism: text watermarking and file metadata. The text watermark is embedded within the generated content and may persist even after copying and pasting; file metadata is appended to files such as SVG, PNG, and JPG, complying with the C2PA standard. This initiative is designed to comply with Article 50 of the EU AI Act but applies globally across all use cases. Anthropic stated that detecting a watermark indicates that content may have been AI-generated but does not confirm the original author’s identity, and the watermark may disappear after extensive rewriting. This represents a leading practice among large model providers in AI content labeling.Author and source: AIBase
Anthropic has announced the addition of machine-readable markers to Claude's outputs, a system comprising two layers: invisible watermarks embedded within the text and digital signature metadata appended to generated files. Starting August 2, all newly released Claude models have begun implementing this marking scheme.
This is the transparency code of conduct signed by Anthropic to comply with Article 50 of the EU AI Act, but its enforcement is not limited to the EU—it applies globally wherever Claude is used.
Specifically, the first layer is a text watermark. When Claude generates text, it weaves in a watermark invisible to the human eye, without affecting readability or content quality. This watermark is designed to "travel with the text"—if you copy and paste a passage written by Claude into an email, document, or blog, the watermark remains intact and may even persist after some degree of editing or modification.
The second layer is file metadata. When Claude generates files such as SVG, PNG, or JPG, it attaches signature metadata compliant with the C2PA standard. C2PA is an open content provenance protocol jointly promoted by companies including Adobe, Microsoft, and Google, and is already in use by image generation tools from OpenAI and Google.
The coverage is extensive and applies to all product lines: API, Claude’s official website, Claude Code, Claude Cowork, and Claude Tag. When invoking Claude through AWS, Google Cloud, or Microsoft Foundry, text watermarks still apply, but file metadata depends on the feature support of each cloud platform.
Previously, Claude had never publicly deployed text watermarking. OpenAI had previously developed a text watermarking solution but never launched it due to various considerations. Anthropic has now directly implemented watermarking at the text level, taking a relatively leading step among large model manufacturers.
For those using Claude to generate content, a direct impact is that emails, reports, and articles drafted with Claude will carry machine-detectable watermarks. Anthropic states that it is developing accompanying detection tools, and third parties will also be able to detect them in the future.
However, the limitations are practical. Detecting a watermark only indicates that the content "may have been processed by Claude," but it cannot confirm that Claude is the original author, as many people use it to edit, translate, or summarize existing content. Conversely, the absence of a watermark does not mean the content wasn’t AI-generated—watermarks can disappear after extensive rewriting, are unreliable in very short texts, and can be lost when files are reformatted or screenshot.
For existing Claude models released before August 2, Anthropic is adding tokenization features, though the specific timeline has not yet been announced. If you are building a product based on the Claude API, Anthropic recommends independently assessing your compliance obligations under Article 50 of the EU AI Act.