Output from newer Claude models will carry a concealed marker that survives copy-paste off the platform, and files will be tagged with C2PA provenance data — but older models have to be retrofitted before they can label anything.
A recent support-page notice sets out how Anthropic intends to mark the text, code and files that Claude generates — and to apply the EU AI Act's disclosure requirements to everyone who uses its products worldwide, rather than to European users alone.
For prose and code from the newer Claude models, the marker is to be built into the output itself, and it stays in place after someone copies the material and pastes it somewhere beyond the Claude platform.
Files are handled by a separate route. Anything Claude turns out as a file will be tagged with C2PA, a standard for recording where content came from; other companies have already adopted it to flag media made by AI.
Neither treatment reaches every model at once. August 2 is the day the EU AI Act's transparency obligations took effect, and the notice puts the dividing line there: models released after it come with the watermarking tool included from the outset, while earlier Claude models fall inside a grace period and have to be modified after the fact before they can apply the labelling.
Anthropic's position is that a watermark indicates content passed through Claude, not that Claude necessarily wrote the whole thing. The company is also planning to release its own tools for detecting the marks.
Anthropic is not the only company that has to work the EU AI Act's disclosure requirements through: OpenAI, Google and Meta are all likely to run into them as well.