Anthropic Watermarks Claude Text Worldwide

Anthropic did not build a Europe-only Claude. After signing the EU AI Act’s Article 50(2) Code of Practice on transparency of AI-generated content, the company began weaving imperceptible watermarks into Claude’s generated text and attaching signed provenance metadata to supported files, then applied those marks wherever Claude is offered. Brussels wrote a regional rule. Anthropic shipped a planetary default.

The mechanics are split by medium. Text carries an invisible watermark Anthropic says leaves meaning, quality, and readability alone, and that can survive copy-paste and some light editing because it lives inside the string itself. Supported image files such as PNG, JPG, and SVG get digitally signed C2PA metadata that records Claude processed the file and can show whether that file was later tampered with. The company says marking covers Claude apps, the API, Claude Code, Claude Cowork, Claude Tag, and access through AWS, Google Cloud, and Microsoft Foundry, with the usual platform caveats.

7 min read

Timing tracks the statute. Article 50’s transparency duties became applicable on August 2, 2026. Models launched in the EU on or after that date are supposed to mark at launch; older models are being updated through a transition window. Anthropic’s support page is unusually candid about what a hit means: a detected mark shows content "may have been processed by Claude," including proofreading, translation, or summarization of someone else’s draft. Missing marks do not clear the content. Heavy edits, short passages, stripped metadata, pre-marking models, and unsupported surfaces can all erase the signal.

The label nobody sees

A machine-readable mark is not a warning badge for a human reader. Under the Act, providers owe detectability; deployers often owe disclosure when they publish synthetic or heavily assisted public-interest text, deepfakes, and similar material. Anthropic’s watermark can help automate the first half of that chain. It does not finish the second. Readers still meet unmarked pages until a platform, newsroom, or regulator turns the detector on.

That gap is why the global choice matters. Europe already forced authentic-looking deepfakes toward visible labeling duties. Anthropic’s text marks extend the same philosophy into ordinary prose, then refuse to keep a cleaner unmarked product for users outside the bloc. Compliance becomes product design. A San Francisco assistant writing a cover letter in Lagos now inherits a Brussels theory of what machine speech owes the public record.

The honesty in Anthropic’s limitations section is the best part of the rollout, and the most damning. If a mark can appear on human text Claude merely touched, and vanish from Claude text someone paraphrased, then watermarking is a provenance hint, not an authenticity oracle. Treating it as either a perfect detector or a useless gesture misses the point. It is a new default assumption about synthetic language: the utterance should carry a machine-readable history even when that history is incomplete.

Sonar’s judgment is blunt. Invisible worldwide marking is the right direction for a lab that wants to keep shipping powerful text models into messy public life. It is also a quiet admission that "AI-generated" will not be a clean binary, and that Europe’s Article 50 is already writing the house style for how the rest of the planet talks about machine authorship. The watermark will not settle the philosophy of originality. It will make the argument harder to dodge.