Anthropic Says Claude Will Carry Hidden Watermarks to Meet EU AI Act Rules

If you’ve ever wondered whether a piece of writing was produced by a human or an AI, you’re not alone. From job applications to school essays to marketing copy, the question of who — or what — actually wrote something has become harder to answer. Now, under new EU rules, AI companies are being asked to make that question easier to resolve. And Anthropic, the company behind the Claude family of AI models, has announced exactly how it plans to do that.

The company posted details of its approach on its official channels, explaining that Claude models launched in the EU on or after 2 August 2026 will carry embedded watermarks in text and digitally signed provenance metadata in supported files. The move is tied directly to the EU AI Act’s transparency obligations, which come into force on that same date.

What the EU AI Act Actually Requires

The EU AI Act is Europe’s wide-ranging legal framework for artificial intelligence. Among its many requirements are transparency rules for AI-generated content — meaning that, under certain conditions, users should be able to tell when content has been produced by a machine rather than a person.

To give companies a practical path to meeting those obligations, the European Commission published a Code of Practice on the marking and labelling of AI-generated content on 10 June 2026. Around 190 organisations signed up to that code in July 2026, including Anthropic and, according to the company, several other major AI model developers. That’s a broad slice of the industry moving in the same direction at the same time.

The code is a compliance framework. It sets out how providers can demonstrate they’re meeting the Act’s requirements. It doesn’t guarantee that watermarks will be impossible to spoof or that every piece of AI-generated text will be caught — but it does create a baseline standard that providers are expected to meet.

How Anthropic’s Watermarking Actually Works

This is where it gets interesting, because Anthropic’s approach is a bit different from what many people might picture when they hear “watermark.”

There are no hidden characters inserted into the text. No extra tokens are used. And readers won’t be able to spot anything unusual just by reading the output. The watermark is embedded at a technical level — machine-readable, not human-readable — meaning it’s designed to be picked up by detection tools rather than spotted by the human eye.

Anthropic has also been clear that the watermarking approach isn’t specific to Claude alone, and that it cannot be traced back to a specific person, organisation, or chat session. So if you’re worried about your AI-assisted draft report being linked back to your account, that’s not how this works. The marking identifies the content as AI-generated; it doesn’t identify you.

A Wider Industry Move

It would be easy to read this as Anthropic going it alone. But the picture that emerges from the company’s own statements is one of broader industry alignment.

Anthropic says other major AI providers have signed the same EU transparency code and will implement watermarking of their own. The specifics of how each company will handle this — which models, which output types, which timelines — will vary. But the direction of travel across the sector appears consistent.

That matters, because a watermarking system only really helps if it’s widespread. A single company marking its outputs while others don’t is a bit like one supermarket labelling its food allergens while the rest of the high street doesn’t bother. The EU’s approach of getting multiple providers to sign a common code is an attempt to avoid exactly that problem.

The Limits of What Watermarking Can Do

There’s a genuine debate about whether watermarking is a complete answer to the problem of AI-generated content. Critics point out that watermarks can potentially be stripped or spoofed, and that detection tools aren’t always reliable. A compliance framework can set expectations, but it can’t guarantee enforcement in every case.

The Commission hasn’t claimed otherwise. The code is framed as a way to help users identify AI-generated content — not as a silver bullet. And Anthropic’s own framing is measured: this is a compliance measure, one that the company says shouldn’t affect output quality or the experience of using Claude.

Luca Bertuzzi, a technology policy journalist who has followed the EU AI Act closely, has previously noted that the Act’s transparency provisions represent “a minimum floor, not a ceiling” — meaning companies can go further, but must at least meet the baseline. That context is useful here. Watermarking is the floor being set, not the limit of what’s possible.

What This Means for Kent Residents

For anyone in Kent using AI tools — whether for work, study, or running a small business — the practical effect is that some of the text and files produced by Claude and other major AI platforms may start carrying machine-readable labels identifying them as AI-generated, especially if you’re working with EU-facing services or regulated workflows. If your business produces marketing content, reports, or customer communications with AI assistance and you deal with EU clients, it’s worth understanding that provenance labelling is becoming a standard part of how these tools operate. More broadly, as a UK consumer, you may find that AI transparency features developed for EU compliance roll out across platforms you already use, even if the UK’s own AI regulatory framework is still taking shape.