Anthropic to Watermark Claude Output Under EU AI Act Transparency Rules

Anthropic has signed the EU AI Act’s Article 50(2) Code of Practice on transparency, committing to mark content its models generate. Claude models launched in the EU on or after August 2, 2026 will carry machine-readable marks from day one, and Anthropic says it is retrofitting the capability onto models released before that cutoff during a legal transition period. The marking applies across every surface where Claude runs — the API, the Claude apps, Claude Code, Cowork, and Tag — and extends to Claude accessed through AWS, Google Cloud, and Microsoft Foundry, worldwide.

Two techniques do the work. Generated text gets an imperceptible watermark applied at the model level, so it persists across products, survives copy-paste, and can withstand some editing without altering the text’s meaning or readability. Generated files such as SVG, PNG, and JPG receive signed provenance metadata following the C2PA open standard, which both identifies Claude as the processor and reveals whether a file was later tampered with. Anthropic plans to release tools and documentation letting third parties detect both mark types.

The company is candid about the limits. A detected mark only signals that content ‘may have been processed by Claude’ — it doesn’t establish authorship, since users often run Claude to proofread, translate, or summarize material written by someone else. Conversely, the absence of a mark proves nothing: short passages, heavy paraphrasing, format conversion, or screenshots can all strip the signal. For anyone building products on Claude, Anthropic notes that meeting Article 50’s requirements remains the developer’s own responsibility.