What it means that Anthropic watermarks everything Claude generates for EU AI Act compliance — August 12, 2026
· CompaniesAutomation
Radar Flash Edition: Anthropic embeds invisible watermarks in Claude's text and C2PA metadata in its files, worldwide, to comply with Article 50 of the EU AI Act. What part of the obligation remains yours.
Flash Edition. Anthropic has confirmed that everything Claude generates—text and files—now comes with built-in marking to comply with Article 50 of the European Union AI Act, in effect since August 2nd. And it is applying it worldwide, not just in the EU. It is the other half of today's Radar on Spotify's AI labels: transparency stops being a checkbox and starts traveling inside the content.
What happened
The text Claude returns now carries an imperceptible watermark embedded in the words themselves, which survives copying and pasting; files generated in .svg, .png, and .jpg formats come with provenance metadata signed according to the C2PA standard. It covers every surface: the API, claude.ai, Claude Code, Claude Cowork, Claude Tag, and Claude served from AWS, Google Cloud, and Microsoft Foundry. Models launched as of August 2, 2026, are marked from day one, and previous ones are being updated (Anthropic documentation). The trigger is regulatory: Anthropic signed the Article 50(2) code of practice, enforceable since August 2nd with fines of up to 15 million euros or 3% of worldwide turnover (Euronews). The company itself sets the limits: the mark says the content passed through Claude, not that Claude is the author nor that no one modified it afterward; and heavy editing, a format change, or a screenshot can erase it. There is no public detector yet either: Anthropic says it is working on third-party verification (The Decoder).
Why it matters
Two things change for a typical business. First: the provider's marking does not exempt you from anything. Article 50 splits obligations, and the one falling on you as deployer—flagging that a bot is talking, labeling the synthetic content you publish—stays yours; Anthropic covers its side and leaves yours in plain view. The second is both more uncomfortable and more useful: everything coming out of Claude now travels with a technical signal third parties can verify. The text delivered to you by an agency as "handwritten," a candidate's cover letter, a vendor report: the information asymmetry is closing. Not to the point of conclusive proof—that doesn't exist—but certainly to the point of a signal.
For your business
Three decisions for this week. One: inventory what you publish with AI—product sheets, support replies, blog posts, images—and set your labeling policy today; declaring it yourself is cheaper than being flagged. Two: change the clause in your contracts with agencies and freelancers from "AI use prohibited" (unenforceable) to "mandatory disclosure of AI use," which is enforceable and protects you. Three: don't build processes on top of detectors—discarding a candidate because "it has a watermark" is building on a signal that its own manufacturer calls non-conclusive—and check if your retouching workflows (converting to PDF, compressing, cropping, capturing) are erasing marks that you might need to keep. If you are not sure which part of Article 50 is yours and which is your provider's, an AI consultancy can put it in writing before a client or an inspector asks.
Frequently Asked Questions
Does this force me to label my company's content?
The watermark is the provider's responsibility, but Article 50 also obligates the person deploying the system: you must notify people when they interact with an AI and label synthetic content you publish as artificial. Claude marking its output covers Anthropic's obligation, not yours.
Can I check now if a text was written by Claude?
Not yet. Anthropic says it is working on enabling detection by users and third parties and will publish the technical documentation later. Until then, the mark exists, but there is no public tool to read it.
My company doesn't sell in the EU, does this affect me?
The marking applies worldwide, so your content comes out marked regardless. The legal obligation to label reaches you if your system or its output is used in the European Union, which includes many companies that do not consider themselves "European" but serve customers who are.