Anthropic is marking Claude's output as machine-generated worldwide, according to a support document updated on 11 August and picked up by reporters the same day. Text gets an imperceptible embedded watermark; .svg, .png and .jpg files get signed provenance metadata under the C2PA standard.
The legal hook
The driver is Article 50(2) of the EU AI Act and the accompanying Code of Practice on Transparency of AI-Generated Content, which took effect on 2 August 2026 and which Anthropic has signed. The Act's ceiling for transparency-type breaches is €15m or 3% of worldwide turnover — a statutory maximum, not a penalty anyone has imposed here.
Scope is the story
The obligation is European; the implementation is not. Marking applies across the API, Claude, Claude Code, Claude Cowork and Claude Tag, and through AWS, Google Cloud and Microsoft Foundry, for users everywhere. Models shipped on or after 2 August support it at launch; older ones are still being backfilled.
What the mark does not prove
Anthropic is explicit that a detected mark is not evidence Claude authored the text — Claude may only have edited or translated it — and that the absence of a mark is not evidence of human authorship. The watermark survives copy-paste and "may persist through some editing"; C2PA metadata can be stripped with existing open-source tools.
Not an announcement
There was no blog post. This is an updated help-centre article that journalists noticed, which is a materially different act from a lab launching a provenance product.
Who moves next
Every lab serving European users faces the same Article 50 obligation. Anthropic is the first to publish a concrete implementation, which makes its choices — an invisible text mark plus C2PA on files, applied globally rather than by region — the reference the others will be measured against.