Anthropic has begun embedding machine-readable marks in everything Claude generates, a step the company says is required to meet the EU AI Act’s new transparency rules — and one it is applying to every user worldwide, not only in Europe.
Anthropic describes two separate mechanisms. Generated text carries an imperceptible watermark woven directly into the output, which the company says survives copying, pasting and moderate editing without changing the text’s meaning or readability. Generated files in supported formats — including SVG, PNG and JPG — get signed provenance metadata following the C2PA open standard, which flags that a file passed through Claude.
Why now
The trigger is Article 50(2) of the EU AI Act’s Code of Practice on Transparency, whose obligations took effect across the bloc on August 2. Anthropic says any Claude model launched in the EU on or after that date supports marking from launch, and it is retrofitting older models during a transition period. Marking now covers the Claude consumer app, the Claude Platform API, Claude Code, Claude Cowork and Claude Tag, plus versions of Claude accessed through AWS, Google Cloud and Microsoft’s cloud platforms — though metadata support varies by partner.
What the marks can and can’t tell you
Anthropic is explicit about the limits. A detected mark means content “may have been processed by Claude,” not that Claude wrote it from scratch — the same watermark would appear if someone pasted their own draft and asked Claude to lightly edit it. The reverse holds too: unmarked text isn’t proof of human authorship, since heavy editing, short excerpts, screenshots or format conversion can all strip the signal. Anthropic says it is still building public detection tools and technical documentation so users and platforms can check for the marks themselves.