Anthropic Watermarks Claude’s Text — Worldwide, Not Just EU

Anthropic confirmed on 11 August that text produced by its newest Claude models now carries an imperceptible watermark, and that files Claude generates ship with signed provenance metadata. The change covers every model released after 2 August 2026 — and, unusually for a compliance measure drafted in Brussels, Anthropic is applying it wherever Claude is sold, worldwide, rather than only to users covered by European law.

Two separate mechanisms do the work. For prose, the company says the watermark is embedded at the model level, directly into the text itself, so that it survives being lifted out of the chat window: it “will travel with the text when it’s copied and pasted elsewhere, and may persist through some editing.” For generated files — currently SVG, PNG and JPG — Claude attaches signed metadata following the Coalition for Content Provenance and Authenticity (C2PA) standard, which also makes later tampering detectable.

The marking is not confined to the consumer app. Anthropic lists the Claude Platform (API), Claude, Claude Code, Claude Cowork and Claude Tag, plus Claude as served through Amazon Web Services, Google Cloud and Microsoft Foundry. Models released before 2 August are not yet covered; the company says it is working to add support before 2 December 2026, when the EU’s grace period for systems already on the market expires.

The legal driver is Article 50 of the EU AI Act, whose transparency obligations took effect on 2 August 2026 and require synthetic content to be machine-readable as such. Anthropic has signed the bloc’s accompanying code of practice on transparency, alongside Black Forest Labs, Google, Meta, Microsoft, OpenAI and Synthesia. Non-compliance is expensive: penalties under the transparency regime run up to 15 million euros or 3 percent of global annual revenue.

The most consequential detail is what a mark does not mean. Anthropic’s help centre is blunt that detection proves the text passed through Claude, not that Claude wrote it: “Claude may not be the original author. People often use Claude to proofread, translate, summarize, or convert files.” That distinction matters for anyone hoping to treat a watermark as a plagiarism test. An essay a student wrote themselves and then ran through Claude for a grammar pass will carry the same mark as one produced from a single prompt.

The reverse inference is weaker still. The absence of a detected mark, the company notes, does not mean content was not AI-generated — it may come from an older model, have been edited heavily, or arrive in a file type that carries no marking at all. Very short passages may not hold enough signal to be read, and file metadata is stripped by routine acts like taking a screenshot or converting between formats.

What is still missing is the other half of the system. Anthropic says detection details will arrive in forthcoming technical documentation, which leaves open the question of who can actually read a mark — platforms, publishers, schools, or only Anthropic. Google DeepMind is further along, having extended SynthID across text, image and audio, while OpenAI has moved more slowly on text since adopting SynthID and C2PA for images in May.

Until that tooling is public, the watermark is a claim rather than a check. A signal that survives copy-paste is genuinely useful provenance infrastructure — but only once someone other than the company that embedded it can read it.