Skip to content
Sign in
AI Tech

The EU’s AI marking rules are live. The tool to verify them isn’t.

02/09/2026 · 9 min read
EU AI content marketing rules

AI Governance

Article 50 of the AI Act has been in force since 2 August 2026. Anthropic’s release of Claude Fable 5.1 this week shows what that does and does not deliver for people in Europe.

By Clayton Samuel (Mr Wangdoo), QFA  |  Published 2 September 2026

Since 2 August 2026, providers of generative AI systems have had a legal duty in the European Union to mark what their models produce. Article 50 of the AI Act requires synthetic output to be marked in a machine-readable format and made detectable as artificially generated. The obligation is in force. It is not a consultation, not a phase-in, not a proposal.

What has not arrived at the same pace is the ability for anyone to check the marks.

Anthropic’s release of Claude Fable 5.1 and Mythos 5.1 this week makes the gap easy to see, because the company shipped both halves of the problem in one announcement. The models’ output carries a watermark. The tool that reads it went out in private preview.

The short version

  • Marking obligations under Article 50 applied from 2 August 2026.
  • Around 190 organisations signed the EU’s voluntary Code of Practice on transparency of AI-generated content by the end of July.
  • Anthropic signed, and its newest models now watermark their output.
  • The detection API shipped in private preview, not as the free public tool earlier reporting had indicated.

What Article 50 actually asks for

The duty sits on providers — the organisations that build and place generative systems on the EU market — under Article 50(2). A parallel duty under Article 50(4) falls on deployers, the organisations using those systems under their own authority for certain purposes, who have to label some categories of content for the people who see it.

Two things are worth separating here, because coverage tends to collapse them. Marking is machine-readable and usually invisible to a reader. Labelling is the visible disclosure. Article 50 requires both, in different circumstances and from different parties. A company can easily be both provider and deployer, with both sets of obligations running at once.

The Commission published its finalised guidelines on implementing Article 50 on 20 July 2026, alongside a Code of Practice on Transparency of AI-generated Content that had been published the previous month. The Code is voluntary. Signing it gives an organisation a single framework recognised across all 27 member states and shifts future enforcement toward monitoring adherence rather than case-by-case assessment.

Declining to sign does not remove the obligation. Non-signatories still have to satisfy their national market surveillance authority that whatever they have implemented is adequate, judged individually. One useful detail for anyone with an archive: content generated before 2 August 2026 does not have to be marked retroactively.

By the end of July, around 190 organisations had signed. The provider-side list includes Aleph Alpha, Anthropic, Black Forest Labs, Cohere, Google, Meta, Microsoft, Mistral, OpenAI and Synthesia. The deployer side pulled in names from well outside the technology sector — Bulgari, Fastweb, Getty Images, Iberdrola, Lenovo and Lufthansa among them. The Commission noted that roughly half the signatories are small and recently founded companies. Two task forces for signatories were due to launch this month.

Where the Anthropic release fits

Anthropic signed the Code in July. Its text and file outputs now carry a watermark as a result, applying to models released after 2 August 2026. Fable 5.1 and Mythos 5.1 are the first Claude models to arrive after that date. Fable 5.1 and Mythos 5.1 are the same underlying model; the difference between them is the level of safeguards applied. The company says the watermark is invisible to anyone without the detection API, has no practical effect on the quality or content of what the model produces, and contains no information about the user, their organisation or their conversations.

The mechanism is worth understanding, because it explains the limits. The technology is based on Google DeepMind’s SynthID-Text. Rather than attaching metadata or inserting hidden characters, the system adjusts the randomness the model uses when choosing between possible next tokens. Across a long enough response, those choices form a statistical pattern that can indicate the model was involved. Nothing is added to the text; the text is simply shaped as it is written. Because the watermark is part of the text itself rather than attached as metadata, simply copying a response elsewhere does not remove it; Anthropic says it can survive some editing, though rewriting enough of the text will eventually erase the signal.

Anthropic is applying the watermark to all Claude output worldwide, not just to responses served to EU users, because the company says there is no reliable way to limit it by region. Older Claude models are expected to receive the watermark over the coming months.

The detection side is where expectations and delivery diverge. Earlier reporting had indicated Anthropic would ship a freely available detection API — anyone could paste in a passage and get an answer. What launched is considerably narrower: a private preview open to eligible organisations, described as including regulators, law enforcement, media outlets, fact-checkers, independent researchers, educational bodies and EU civil society groups. Enterprises with their own compliance obligations can apply. The company says it intends to widen access but has not given a date.

There is a defensible reason for gating it. A detection tool that anyone can query is also a tool an adversary can query repeatedly to find out what defeats it. Restricting access buys time. It also means that for the moment, a European citizen who wants to know whether a piece of text was machine-generated cannot simply find out. The verification capability exists institutionally rather than publicly.

Video: an explainer on the AI Act’s transparency obligations. Wangdoo has no affiliation with the publisher of this video.

The structural problem underneath

Marking is only as useful as detection. A watermark nobody can read is a compliance artefact rather than a transparency measure, and the AI Act’s stated purpose is that people should know when they are dealing with AI.

There is also a blind spot that follows directly from how the marking works. The watermark is applied token by token: when a particular token is required for accuracy — meaning choosing differently would break the code or produce a wrong answer — Anthropic does not apply it. Natural language gives the model constant room to choose between equivalent phrasings. Code does not; variable names, syntax and operators are far more constrained. The result, as The New Stack reports, is that the watermark signal is weaker in code output than in prose. It can still appear in less constrained parts of a code response, such as comments, but short responses may not contain enough signal to be reliably detected at all.

A further complication sits alongside it. Analysts working through the Code have pointed out that no single marking technology satisfies every requirement Article 50 sets, since different content types and distribution paths need different approaches.

The watermark survives copying and light editing, but tools claiming to strip AI watermarks through heavier rewriting have been reported. If deliberate removal is achievable and detection is restricted to vetted institutions, the practical result is a system where compliant actors are marked and identifiable while anyone determined to avoid it may not be. The obligation binds those already inclined to follow it.

None of which makes Article 50 pointless. It establishes an expectation, creates an audit trail for regulators, and puts a legal foundation under something that had previously been a matter of vendor goodwill. But it is a foundation rather than a finished structure, and the year ahead will be about whether detection infrastructure catches up with the marking obligation or stays a step behind it.

A second European gap

One further detail in the Anthropic release deserves noting for a European audience. Mythos 5.1, which shares the same underlying model as Fable 5.1 but with more permissive safeguards for vetted users, reaches those users through two routes: a Cyber Verification Program and a Life Sciences Verification Program, the latter developed in partnership with the United States government. Anthropic’s announcement says access is currently limited to a set of US organisations and gives no date for widening it, though open enrolment for scientists is described as coming soon.

The practical position today is that European cyberdefenders and life scientists are outside those programmes. For an EU organisation planning around frontier model access, that is worth tracking — particularly given that the marking obligation driving the watermark is European law, while the loosened-safeguard version of the same model is currently an American arrangement.

My Take — Mr Wangdoo

I am reading through the regulation, the Commission’s guidance and the launch documentation here rather than having implemented any of this myself or tested a detection tool.

What strikes me is a pattern that keeps recurring in EU technology law. The obligation lands on a fixed date and is enforceable from that date. The infrastructure that makes the obligation meaningful arrives on a vendor’s timeline, shaped by commercial and security considerations that the legislation never addressed. We saw a version of this with the Cyber Resilience Act’s reporting duties, and we are seeing it again.

I do not think that is a reason to be cynical about the AI Act. Marking obligations are better than no marking obligations, and the Code of Practice giving smaller companies a single compliance route across 27 states is a real practical benefit. But the honest position is that “AI content must be detectable in the EU” and “you can detect AI content in the EU” are two different statements, and only the first one is currently true.

Transparency

Wangdoo did not test any detection tool, watermarking implementation or AI model described here. This article is based on the text of the EU AI Act, published European Commission guidance, the Code of Practice documentation, Anthropic’s public announcement and reporting from The Next Web and The New Stack. Benchmark and pricing figures relating to the models are vendor-published and self-reported.

Sources

Clayton Samuel (Mr Wangdoo)

Clayton Samuel (Mr Wangdoo), QFA

Founder & editor, Wangdoo.com. Qualified Financial Adviser with a background in electronics, web development, and cloud infrastructure.