Does the EU AI Act Require You to Label AI-Written Coverage?

What Article 50 actually requires

Article 50 is the transparency article of the EU AI Act, Regulation (EU) 2024/1689. It places duties on two distinct parties who are rarely the same organisation: the providers who build and supply AI systems, and the deployers who put those systems to use and publish what comes out of them. Reading it as a single obligation on publishers is the most common way to get it wrong.

On providers, paragraph 2 requires that systems generating synthetic audio, image, video or text mark their outputs “in a machine-readable format and detectable as artificially generated or manipulated”. This is a technical marking duty attached to the tool rather than to the publication.

On deployers, paragraph 4 carries the obligation that matters most to publishing: text generated by AI and published “to inform the public on matters of public interest” must be disclosed as artificially generated. The same paragraph carries the deepfake disclosure duty for image, audio and video.

Paragraph 1 adds a separate duty for systems that interact directly with people: they must make that clear, unless it is obvious to a reasonably well-informed person.

Paragraph 5 governs how the information must be given — clearly and distinguishably, and at the latest at the time of the first interaction — so a disclosure added after publication, or buried in terms a reader never opens, does not discharge the duty. The timing requirement is as operative as the disclosure itself.

The editorial carve-out, and how narrow it is

The exemption is written into paragraph 4 itself. The disclosure duty does not apply “where the AI-generated content has undergone a process of human review or editorial control and where a natural or legal person holds editorial responsibility”.

Both limbs must hold. A human must have reviewed or exercised editorial control, and an identifiable person or legal entity must hold responsibility for the publication.

Automated review does not satisfy the first limb, and an unattributed page satisfies neither. A publication that cannot name who is answerable for a given piece has no carve-out available to it, which means the exemption is unavailable precisely where AI-generated text is most likely to appear without anyone having read it.

This makes the exemption a test of editorial process rather than of effort or of how much text a human changed. It aligns closely with what we found when examining bylines as credibility signals: the recurring question, from regulators and platforms alike, is who is answerable for the page.

It is also a carve-out that a publisher can engineer deliberately. Editorial review and named responsibility are both process decisions, and both are documentable in advance rather than argued about afterwards.

What it means for placed and contributed content

A sponsored article drafted with AI assistance, edited by a named editor at the publication and published under that publication's editorial responsibility falls inside the exemption on the face of the text.

A contributed post generated with AI and published without human editorial review, on a matter of public interest, does not. Contributor networks are precisely where editorial review is thinnest, which makes them the format most exposed by this provision.

The phrase “matters of public interest” is not defined in the article, and a great deal turns on it. Commentary on policy, health, finance or technology plainly can qualify, while purely promotional commercial copy is less obviously caught — but much placed content sits precisely between those poles, which is where the uncertainty bites hardest.

The marking duty on providers is unaffected by the editorial exemption. Paragraph 4 relieves the deployer's disclosure obligation; it does not remove the provider's obligation under paragraph 2 to mark outputs in a machine-readable way.

Nor does the exemption apply to the assistive-editing question, which paragraph 2 handles separately: the marking duty does not attach to assistive functions that do not substantially alter the input data.

How this sits alongside disclosure rules you already have

This is a distinct obligation from advertising disclosure. The FTC's requirements concern whether a material connection between advertiser and publisher is disclosed, which we cover in our report on paid placement. Article 50 concerns whether content was machine-generated.

A placed article can therefore carry two independent duties at once: to disclose that it was paid for, and to disclose that it was AI-generated, unless the editorial carve-out applies to the second.

Satisfying one says nothing about the other. A clearly labelled sponsored post generated without human review still engages paragraph 4; a human-edited article that hides its commercial relationship still engages the FTC.

It is also distinct from search-engine policy. Google's position on machine-generated content is a ranking and spam matter, covered in our report on AI-generated coverage, and carries no legal force whatever — a page can rank perfectly well and still breach a disclosure duty, and the reverse is equally true.

Article 50 is law with a stated application date, which makes it the first of the three where the consequence of getting it wrong is not a ranking outcome.

What a publisher should have in place

The article rewards documentation rather than intention. A publisher relying on the carve-out needs a record that human review actually occurred, and a named natural or legal person holding editorial responsibility for the piece — neither of which can be reconstructed convincingly after a challenge has been made.

For agencies commissioning content, the corresponding question to ask a publication is simple: who edits this, and who is named as responsible for it. Both answers exist at reputable outlets and are frequently absent at contributor platforms.

Where AI assistance is used in drafting, the safest position is to treat the disclosure as a default and the exemption as something claimed deliberately, rather than assuming the carve-out applies because an editor glanced at the file.

None of this is legal advice, and the reach of the Regulation over a non-EU publisher is a question of how the system and its outputs touch the EU market.

What can be said plainly is the date and the test. From 2 August 2026 the question is whether a human genuinely reviewed the text and whether someone is named as responsible for it.

The marking duty does not disappear with the carve-out

The most common misreading of Article 50 is that a human edit removes every obligation. It does not. Paragraph 4's exemption is addressed to the deployer's disclosure duty, and paragraph 2's marking duty sits with the provider of the generating system.

So a publisher relying on editorial review may still be handling output that carries, or should carry, a machine-readable mark identifying it as artificially generated. The two provisions operate on different parties and different objects.

Paragraph 2 contains its own limit rather than an editorial one: the marking duty does not apply to assistive editing functions that do not substantially alter the input data. Ordinary correction, reformatting and copy-editing assistance sits outside it on the text's own wording, which is the provision most newsroom use will actually turn on.

The practical consequence is that a compliance answer has two halves: what the tool marked, and what the publisher disclosed. An organisation can satisfy the second in good faith while having no visibility at all into the first, since the marking happens inside software it neither built nor controls.

For content commissioned across borders this is the harder half, because the provider obligations attach to systems placed on the EU market while the publisher may be using a tool procured entirely elsewhere, under terms that say nothing about marking. The publisher owes the disclosure regardless of what its supplier did or failed to do.

Methodology and limitations

Article 50 is summarised and quoted from the consolidated text of Regulation (EU) 2024/1689 as published at artificialintelligenceact.eu, retrieved 19 September 2026, which reproduces the Official Journal text. We attempted to retrieve the Official Journal version directly from EUR-Lex; that document was truncated before Article 50 in our retrieval and we therefore relied on the reproduction. Quotations are of the article's operative wording; paragraph numbering follows the Regulation. This report covers the transparency obligations in Article 50 only — not the risk classification regime, the general-purpose AI provisions, or the penalty structure — and it is a description of the text, not legal advice. Application to any specific publication or campaign is a matter for qualified counsel.

Conclusion

From 2 August 2026, AI-generated text published to inform the public on matters of public interest must be disclosed unless it passed through human editorial review and someone holds editorial responsibility for it. For anyone commissioning coverage in the EU that turns a question of taste into a question of compliance: the difference between a draft an editor genuinely reviewed and one that was published as generated is now the difference between an exemption and an obligation. The machine-readable marking duty on the tools themselves applies regardless, and sits alongside — not instead of — the advertising disclosure rules that already govern paid placement.

Frequently Asked Questions

Sources

1. official Regulation (EU) 2024/1689 (the EU Artificial Intelligence Act), Article 50 — Transparency obligations for providers and deployers of certain AI systems. Text as reproduced at artificialintelligenceact.eu/article/50, retrieved 19 September 2026.
2. official Official Journal of the European Union, Regulation (EU) 2024/1689, EUR-Lex. Consulted 19 September 2026; our retrieval was truncated before Article 50.
3. trade press Commentary and implementation guidance on Article 50 published during 2025 and 2026 was consulted for context on the editorial-responsibility carve-out; all quoted wording in this report is taken from the Regulation itself.