Claude Now Watermarks Its Output: What UK Firms Should Do

If you use Claude to draft anything that leaves your business, that text now carries a hidden signature.

What happened

On 02/08/2026 the EU AI Act's Code of Practice on Transparency of AI-Generated Content came into force. Anthropic has signed it, as a provider of both generative AI models and generative AI systems, and confirmed its plans publicly on 11/08/2026.

There are two mechanisms. The first is an imperceptible watermark woven into generated text. You cannot see it, and Anthropic says it does not change the meaning, quality or readability of a response. Because it sits inside the words themselves, it travels with the text when you copy and paste, and it may survive a round of editing.

The second is signed provenance metadata attached to generated files such as .svg, .png and .jpg. That follows the Coalition for Content Provenance and Authenticity (C2PA) open standard, which also records whether a file has been tampered with since Claude touched it.

Any Claude model launched on or after 02/08/2026 supports marking from day one. Older models are being retrofitted during a transition period. The marking applies across Claude Platform, Claude, Claude Code, Claude Cowork and Claude Tag, and when Claude models are accessed through Amazon Web Services (AWS), Google Cloud or Microsoft Foundry.

One detail matters more than the rest for readers here: marking applies wherever Claude is offered, worldwide. This is not an EU-only feature.

Why it matters for UK businesses

Britain sits outside the EU, but not outside this. Anthropic chose to apply marking globally, so your output is marked whether you are in Manchester or Munich.

If you sell into the EU, there is a second layer. Anthropic's own guidance is blunt about it: if you deploy Claude inside your own product, you should independently assess what Article 50 requires of your products and services. The obligation does not stop at the model provider.

The bigger practical issue is what the marks actually prove, which is less than most people will assume. A detected mark tells you that content was processed by Claude. It does not tell you that Claude wrote it. Paste a colleague's draft into Claude for a proofread, a translation or a tidy-up, and the output can carry a mark even though every idea in it came from a person.

That distinction has teeth. Think about tender submissions, grant applications, regulated reporting, academic work and any client contract with an AI disclosure clause. A blunt reading of a watermark could put a business on the wrong side of a term it never actually breached.

The reverse gap is just as real. No mark does not mean no AI. Text generated before marking was supported, heavily rewritten passages, very short extracts, and files whose metadata was stripped by a format conversion or a screenshot may all come back clean.

How we see it at Adevious AI

We think this is a reasonable step, and a weak signal beats no signal. Detection tools give clients, regulators and the public something to work with, which is better than the free-for-all we have had for the last three years.

What we would push back on is treating detection as proof. It is evidence that Claude touched the text at some point, in some way. Any policy built on a binary read of a watermark, pass or fail, AI or human, will get people wrong. Expect the first few detection disputes to be messy.

For most UK small and mid sized businesses, nothing breaks this week. Claude output reads exactly as it did on Friday. What has changed is the audit trail sitting quietly underneath it.

One thing to do this week

Spend twenty minutes writing down every place Claude output ends up in your business: proposals, tenders, marketing copy, customer emails, reports, code comments. Then check each destination against three questions.

Does the client, funder or awarding body have an AI disclosure clause?

Would a detected watermark on that document breach a term or cause an awkward conversation?

Does your team understand the difference between drafting with Claude and proofreading with Claude, given both can produce a mark?

If the answer to the first question is yes anywhere, agree the disclosure wording now, before somebody else runs a detector over your work. Explaining that you used Claude to tighten a draft is a straightforward conversation to have upfront. It is a much harder one to have after a client has drawn their own conclusion.

If you want a hand mapping where AI output flows through your business, or setting a disclosure policy your team will actually follow, get in touch with Adevious AI. We are happy to talk it through.

Sources:

https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content

https://techcrunch.com/2026/08/11/anthropic-says-it-will-watermark-text-generated-by-its-ai-models/

https://www.euronews.com/next/2026/08/11/eu-compliance-delivered-globally-anthropic-to-watermark-claudes-output-worldwide

https://www.searchenginejournal.com/anthropic-claude-watermarks-eu-ai-act-code/585355/

https://www.techtimes.com/articles/323873/20260811/claude-now-watermarks-text-everywhere-mark-proves-processing-not-authorship.htm