Anthropic to watermark Claude outputs globally under EU AI transparency push

Anthropic has signed the EU AI Act Code of Practice on transparency for AI-generated content and says new Claude models will ship with labeling built in from August 2026. The policy applies worldwide, not only in the EU. The rollout covers Claude products including the API, Claude, Claude Code, Claude Cowork, and Claude Tag.

Claude-generated text will carry invisible watermarks that Anthropic says do not affect meaning, quality, or readability. The marks survive copying and pasting and “may persist through some editing.” Supported files, including .svg, .png, and .jpg images, will receive signed provenance metadata based on the C2PA standard to indicate Claude processed the file and to help reveal later tampering.

Anthropic says a detected watermark does not prove Claude wrote the content, because people may use Claude to proofread, translate, summarize, or edit human-authored material. A missing watermark also does not settle authorship: older models, heavy editing, translation, short passages, format conversion, or screenshots can remove or weaken signals. The practical value will depend on how well the labels survive real-world editing and reformatting.

Anthropic plans to provide verification tools for users and third parties, but the article does not give a release date. Developers integrating Claude into products must assess which Article 50 requirements apply to their own services. For schools, businesses, and platforms, the key takeaway is that Claude content may become easier to flag, but watermarks should be treated as evidence to interpret carefully rather than a standalone verdict.