What Anthropic actually turned on
On August 2, 2026, the transparency commitments Anthropic made under Article 50(2) of the EU AI Act's Code of Practice became legally enforceable. Anthropic's Claude Help Center describes two separate marking mechanisms it has now switched on. Generated text carries what the company calls an imperceptible watermark woven directly into the text itself, a statistical signal in word choice that Anthropic says does not change the meaning, quality, or readability of a response, and that can survive copying, pasting, and some light editing. Generated or Claude-processed image files in SVG, PNG, and JPG format instead get signed metadata following the Coalition for Content Provenance and Authenticity, or C2PA, open standard, the same provenance format used elsewhere across the industry to record how a file was produced and whether it was later altered. Claude models launched in the EU on or after August 2, 2026 support the marking from launch; older models are still being updated to add it.
Why the company chose global over geo-fenced
The more interesting decision is not that Anthropic added a watermark, but where it applied it. Article 50(2) is an EU instrument, and a narrower rollout could have shipped the marking only to EU-registered accounts or EU data regions, treating it as a cost to contain at the border. Anthropic instead says the marks apply wherever Claude is offered worldwide, across Claude.ai, the API platform, Claude Code, Claude Cowork, Claude Tag, and cloud deployments through AWS, Google Cloud, and Microsoft Azure. Practically, a business in London, Zurich, or Toronto drafting content with Claude is now marked the same way as one in Berlin or Paris, even though only the EU rule forced the change. Building the compliance behavior into the product's default architecture, rather than into a regional switch, is cheaper to maintain and harder for any single market to route around, but it also means the EU's transparency requirement has quietly become the whole product's baseline.
What this changes for a business already using Claude
For a communications team in London, or any European company already drafting marketing copy, client emails, or code with Claude, the mark now exists regardless of whether the company itself discloses AI involvement, because Article 50 places its own disclosure duties on the deployers of certain AI systems, not on model vendors alone. A vendor-side watermark does not satisfy a business's own Article 50 duties, and it is not proof of authorship either way: Anthropic's documentation says a detected mark only means content may have been processed by Claude, not that Claude wrote it start to finish, and the absence of a mark is not evidence AI was not involved, since heavy editing, paraphrasing, or translation can remove the signal entirely. The realistic use for a compliance or comms team is narrower than it sounds, one weak, defeatable signal among several, not a substitute for an actual AI-use disclosure policy, and not something to lean on in a dispute over authorship.
The skepticism Anthropic left out of its own docs
Anthropic's own framing throughout is neutral, procedural language about a code-of-practice commitment. The Register's account of the same rollout is pointedly more cynical, headlining it a sop to the EU and warning that overly aggressive marking could alienate Claude customers or push some toward open-weight models that carry no watermark at all. The outlet also notes that researchers have already demonstrated tools capable of stripping watermarks from AI-generated images, undercutting confidence that the text-side signal will hold up any better once someone specifically targets it. That gap between Anthropic's confident, procedural framing and outside reporting's skepticism about durability is worth noting for anyone deciding how much weight to put on Claude marked it as a compliance answer: the mark is real, but its resilience is unproven, and Anthropic effectively says so itself, just in more careful language than the headline writers used.
Read next: Two AI Rules Landed Today, One Lets You Check | Brussels Was Briefed Before the Blog Post
