Anthropic has announced the addition of machine-readable markers to the output content of Claude. This mechanism consists of two layers: invisible watermarks embedded in the text, and digital signature metadata attached to generated files. Starting from August 2nd, all newly released Claude models have begun implementing this marking scheme.

This is part of Anthropic's transparency code of conduct signed to comply with Article 50 of the EU AI Act, but the implementation is not limited to the EU - it will take effect wherever Claude is used globally.

Specifically, the first layer is a text watermark. When Claude generates text, it weaves in an imperceptible watermark that does not affect the reading experience or content quality. The characteristic of this watermark is that it "follows the text" - if you copy and paste a paragraph written by Claude into an email, document, or blog, the watermark remains, and may still be present after some editing.

The second layer is file metadata. When Claude generates SVG, PNG, JPG, and other types of files, it adds signature metadata that complies with the C2PA standard. C2PA is a content traceability open protocol promoted by companies such as Adobe, Microsoft, and Google. OpenAI and Google's image generation tools have already been using this standard.

The coverage is extensive. All product lines, including API, Claude website, Claude Code, Claude Cowork, and Claude Tag, are applicable. When calling Claude through AWS, Google Cloud, or Microsoft Foundry, the text watermark also takes effect, but the file metadata depends on the support of each cloud platform.

Previously, Claude had never publicly deployed text watermarks. OpenAI had previously developed a text watermark solution, but due to various considerations, it was never launched. This time, Anthropic directly implemented the watermark at the text level, which is a relatively advanced step among large model manufacturers.

For people who use Claude to write things, a direct impact is that emails, reports, and articles drafted with Claude will contain watermarks that can be detected by machines. Anthropic states that it is developing accompanying detection tools, and in the future, third parties will also be able to perform detection.

However, there are also practical limitations. Detecting a watermark can only indicate that the content "may have been processed by Claude," but cannot confirm that Claude is the original author, because many people use it for polishing, translating, and summarizing existing content. Conversely, the absence of a watermark does not mean the content was not written by AI - the watermark may disappear after the text is heavily rewritten, and short texts may not have reliable signals. Files may also lose metadata after format conversion or screenshots.

Existing Claude models released before August 2nd are currently being supplemented with marking functions, and the specific timetable has not been announced yet. If you build products based on the Claude API, Anthropic advises you to independently assess your compliance requirements with Article 50 of the EU AI Act.