Anthropic is now watermarking Claude’s writing

Claude models launched in the EU on or after August 2, 2026 embed a machine-readable watermark in every piece of generated text, which is applied automatically and invisible to the reader. Files such as images carry signed provenance metadata using the C2PA standard where the format supports it.

Anthropic is applying the change everywhere Claude is used, not only inside the EU. Article 50 of the EU AI Act requires deployers of an AI system that generates public-interest text to disclose that AI generated it, with the duty falling on the deployer, generally the publisher or content producer putting the system to use. Anthropic has said the global rollout comes down to a technical limit, as it doesn’t yet have a reliable way to apply the watermark only within the EU, so every Claude model gets it. Clearly the requirement to comply, as a user, only exists if you are inside the EU or sharing your content inside the EU. This matters for GEO too. As brands increasingly use AI to help produce content intended to appear in search and AI-generated answers, knowing how that content was created becomes part of the wider conversation around transparency.

For models already in use before August 2, Anthropic is adding watermarking support in stages over the coming months.

Anthropic has said the change carries no cost to output quality, and that a watermarked response reads identically to an unwatermarked one. The method statistically biases which words Claude selects during generation, in a pattern too subtle for a reader to notice but detectable by software built to check for it.

A detected watermark shows that Claude was involved in producing a piece of text. It does not show how much of that text is still Claude’s own drafting by the time it reaches a reader.

Anthropic’s own guidance is very clear. Light editing will likely leave a watermark in place. A complete rewrite, replacing every word, removes it.

Between those two extremes sits most PR and marketing output: an AI-drafted paragraph added to, restructured and tightened before it reaches a client, but very rarely retyped from scratch.

Content that has been through that kind of editorial pass may still carry Anthropic’s mark, even where a writer would reasonably describe the piece as their own. After all, the AI has produced its words from your prompt, which you probably wrote using the ask Traci prompt writing structure, and sweated blood over, right?

Ryan Law, a content marketing director at Ahrefs who has written about the rollout, explains the real risk plainly: it only becomes a problem when “you’re passing off AI work as human-written when you’ve promised otherwise.”

Most marketing teams and organisations have treated a human editorial pass as sufficient grounds to call a piece human authored. That assumption gets harder to defend once a watermark can survive exactly that kind of pass.

Article 50’s disclosure duty sits with the organisation publishing the content, not the tool provider. Agencies are not the ones legally required to label copy under EU rules. Content only avoids the disclosure duty once a named person has actually read it and taken responsibility for what’s published, not just glanced at it. The Commission’s guidance rules out spell-checking as a stand-in for real review.

There’s a separate reputational risk. A client who is told a piece was written by a human being in their agency and then discovers a heavily edited AI draft can still carry Anthropic’s mark, may feel less likely to trust the agency.

A watermark that survives editing requires us to ask whether someone actually took editorial ownership of a draft before a byline went out. Bylined thought leadership puts that question to its sharpest test: the named individual’s expertise is the entire premise of the piece.

None of this demands a new process. Agencies that already run genuine editorial oversight, checking accuracy and brand fit before a byline goes out, are already answering that question in practice. A real human editor reading for sense and context catches inconsistencies a pattern-matching checker can miss.

You should also treat detection tools with some caution. AI checker benchmarks regularly turn up false positives, flagging work as AI-written even when no AI touched it at any stage, from brainstorming to final edit. A single tool’s verdict is too shaky a foundation for policy, given how unreliable AI checkers can be.

Detection tools exist on the other side too. Independent verification services are already testing ways to check for Anthropic’s mark, and the company has said it will publish its own detection API so third parties can check content directly.

A quick Google also shows up several APIs that claim to be able to remove the watermark entirely, no matter how much human interaction has happened or not happened.

So, overall, what should you do? I would advise that you ensure you are happy how much AI you have used in your editorial process and be confident that you believe you’ve used it as a tool, rather than as a replacement for a human writer and editor. Then, be open about how much AI you used; this article for instance was drafted by an AI with a long and complex prompt, rewritten by one human and the heavily edited by another.

Would it have a watermark, if we had used the relevant version of Claude? Who knows, but we would be happy to defend the use of AI in its creation if required.

Just like the lemon juice writing held up to the heat, a watermark tells you Claude was involved somewhere in a piece of text. It doesn’t tell you how involved, although the detection tools are only going to get more sophisticated from here.

We get that it’s hard to keep up with the latest thinking in PR and marketing. Subscribe to our twice weekly blog, direct to your inbox, for the answers to all the burning questions your board ask – from using AI in marketing to how to spend your budget against a backdrop of constantly multiplying channels.