Anthropic will watermark Claude's output. Here's what *Article 50 actually requires*.

What Anthropic is doing

Anthropic has signed the EU's code of practice on transparency of AI-generated content, in both roles: as a provider of the Claude models, and as a provider of Claude.ai ^[1]. The code is voluntary, and Anthropic is one of the first to join. The commitment covers machine-readable watermarking, disclosure when people talk to Claude and labelling of AI-generated content.

That sounds straightforward. But the law itself doesn't mention watermarks, and it doesn't say how you show where a piece of content came from when the question is asked months later.

What Article 50 actually says

Article 50 has four duties ^[2]:

1. Providers must make sure people know when they are interacting with an AI system, unless it is obvious.
2. Providers of generative AI must make sure content is marked in a machine-readable format and detectable as artificially generated or manipulated.
3. Deployers must tell people when images, audio or video published to inform the public are deepfakes, unless it is obvious.
4. Deployers must tell people when AI-generated text is published to inform the public and no person has reviewed it and takes responsibility for it.

The rule applies from 2 August 2026. For generative AI systems already on the market before that date, the marking duty in point 2 applies from 2 December 2026 ^[3].

That is it. The article does not say watermark. It does not say blockchain. It says the content must be marked in a format machines can read, and it must be detectable as AI-made. How you do that is left to the code of practice and to whoever implements it.

Why watermarks are not the whole answer

Think of a watermark like the sticker on a piece of fruit. It tells you something about where it came from, as long as nobody peels it off.

Watermarks work when they stay on. They can be stripped, either on purpose or by accident. An image saved as a different format, compressed or edited might lose its mark. A screenshot certainly does. So if someone turns up with an unmarked image and asks whether your system made it, the watermark won't help if it is gone.

That is not a criticism of Anthropic. Watermarking is a sensible way to meet the marking duty. But it is not the same as proving later that your system generated a particular piece of content, or that it didn't.

The question is provenance

The hard question is not whether the content is marked. It is whether you can show where it came from, who requested it and when.

Say a generated image appears somewhere without its watermark, and the question is whether it came from your system. Or say someone claims your chatbot generated a piece of text that breaks the law, and you need to show what your system actually produced on that day. The watermark doesn't answer that if it is missing or disputed.

What does answer it is a record, kept at generation, that nobody can quietly edit afterwards.

A record of what the AI generated

Here is where a digital witness fits. The thing to seal is a record of what was generated: the output, the prompt or request, the model version, the user reference and the time. Sealing takes a fingerprint of that record, not the content itself, so the content never leaves your systems.

Anyone you share the record with can later check that it matches the sealed fingerprint. Change a single character in the record and the fingerprint won't match. That check can be done by an auditor, a supervisor, a platform or the person who disputes where the content came from.

The witness doesn't replace the marking duty in Article 50. It gives you a second layer: a checkable log of what your system generated, whether the watermark survives or not.

Your system sends the fingerprint to WitniumChain through the REST API, the SDK or the MCP server, and gets a witness ID back. To check, anyone you share the record with calculates the fingerprint again and compares it with the sealed one. The first 2,000 witnesses a month are free. Read the docs for how to build it in.

How it would work

1. At generation, your system creates a record: prompt, output, model version, user reference and timestamp.
2. The fingerprint of that record is sealed, and you keep the witness ID with the record.
3. When someone asks whether your system made a piece of content, you share the record and the witness ID, and they check the fingerprint.

What this means for you

If you run a generative AI system that will still be live on 2 August 2026, or on 2 December 2026 for systems already on the market, here are the questions worth asking now:

• How will we mark content so it meets Article 50(2)?
• What happens if the mark is stripped or lost?
• Can we show, six months later, what our system generated on a given day?
• Who keeps the record, and could they change it?
• If someone disputes whether we generated a piece of content, what evidence do we have?

Marking is the law. Keeping a provable record is the practical step that makes the marking hold up when it matters.

What to watch

2 August 2026: Article 50's transparency duties apply to AI systems.

2 December 2026: The marking duty in Article 50(2) applies to generative AI already on the market before 2 August 2026.