TL;DR
Anthropic has committed to embedding watermarks in text produced by future Claude models and attaching signed provenance metadata to generated files, to satisfy transparency obligations under the EU AI Act. Critically, the marking will apply everywhere Claude is sold, not only inside the EU. Existing models will be retrofitted during the transition period European law allows.
What gets marked
The commitment appears in a Monday help document. It covers text and files across the API and the consumer apps, plus Claude Tag, Claude Cowork and Claude Code — and reaches models served by third parties, Microsoft Foundry and Google Cloud among them, as well as AWS.
“When a supported Claude model generates text, it weaves an imperceptible watermark directly into the text itself,” the company says, adding that it survives copying and pasting and can outlast a degree of editing. Because the marking happens at model level, it does not matter which product the text emerged from. File provenance will use the C2PA standard.
Detection tooling has not been published yet, though EU law requires Anthropic to explain how the marks can be identified.
The obvious objections
The Register notes that open source tools already exist to strip C2PA metadata, and that researchers have shown image watermarking can be defeated. For text, an optical character recognition pass would plausibly discard obscure markers.
Anthropic’s own caveats are notable: a detected mark is not conclusive proof Claude produced something, and no mark does not mean AI was uninvolved. That is a fair description of the technology’s limits, and also an acknowledgement that the scheme’s value may be compliance rather than detection.
Looking forward
For UK organisations, the worldwide scope is the operative detail. Britain has no equivalent transparency duty, but British users of Claude will get marked output anyway, because the vendor has chosen one global behaviour over separate regional builds.
This is the Brussels Effect arriving through a supplier rather than a statute — the same route by which UK firms inherited GDPR-shaped tooling. It follows the EU opening reporting channels for AI Act breaches last week. Firms whose clients care about provenance may find this useful; firms passing off generated copy as human work should assume it now carries a signature they cannot see.