Labelling AI-Generated Text on Matters of Public Interest

There is no general duty in the EU AI Act to label AI-written text for human readers. There is one specific duty, in the second subparagraph of Article 50(4), and it is narrow enough that most published AI-assisted writing falls outside it — including, by design, most of what a newsroom produces.

The second subparagraph

Article 50(4) of Regulation (EU) 2024/1689 has two subparagraphs. The first is the deepfake rule for image, audio and video, covered on the deepfake labelling page. The second reads on deployers of an AI system that generates or manipulates text which is published with the purpose of informing the public on matters of public interest: they must disclose that the text has been artificially generated or manipulated. It carries its own exemption, discussed below. The text is on EUR-Lex.

Everything interesting about this provision is in how much it does not reach. Marketing copy, product descriptions, internal documents, customer emails, code comments, novels, most blog posts: none of them are published with the purpose of informing the public on matters of public interest, and none of them attract this duty. Article 50(2) may still require machine-readable marking by the provider of the generating system — a different duty on a different party, covered on the marking duty page.

Three cumulative conditions

The duty bites only where all three are present.

• Text generated or manipulated by an AI system. Both wholly generated text and AI-manipulated text count. A human-written article passed through an AI system that rewrites it is manipulated; one run through a spellchecker is a much weaker case, and Article 50(2)'s parallel exception for systems performing an assistive function for standard editing or not substantially altering the input gives a strong indication of where the drafters saw that line, even though it sits in a different paragraph.
• Published. Made available to the public. Internal circulation, client deliverables and private correspondence are not publication, however important their subject.
• With the purpose of informing the public on matters of public interest. Two elements at once: an informational purpose, and a public-interest subject. Neither term is defined in the Act. “Matters of public interest” has a long history in European media law and Article 10 ECHR jurisprudence, where it covers political debate, public health, the environment, the conduct of public officials and the administration of justice, and is consistently distinguished from what merely interests the public. That body of case law is the natural interpretive resource here, and it is not the same as a commercial interest in traffic.

The editorial responsibility exemption

The provision then disapplies the duty 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 for the publication of the content.

This is the heart of the provision and it is more demanding than “somebody read it”. Two conditions, joined by “and”:

• A process of human review or editorial control. A process, not an event — something repeatable that can be described, not an assertion that a person glanced at the output.
• Editorial responsibility held by a person. Somebody, natural or legal, answerable for the publication. The concept is borrowed from audiovisual media regulation, where editorial responsibility means effective control over selection and organisation of content, and it implies accountability rather than mere involvement.

The design intent is legible: a publisher with an editorial chain is already accountable for what it publishes and is not required to attach an AI label to every article; an automated content operation with no human answerable for it is exactly what the provision targets. That is why the practical effect for a conventional newsroom is usually “no label required”, which surprises people who expect the opposite.

It also means the exemption is a governance question rather than a technical one. The thing that decides whether the duty applies is whether an accountable editorial process exists and can be evidenced — a masthead, a named responsible editor, a documented review step, a corrections policy.

Why text is treated differently

The asymmetry between the two subparagraphs is deliberate. Synthetic image, audio and video that resembles real people or events carries a particular deceptive force: a viewer's default assumption about a photograph is not their default assumption about a paragraph. Text is also the modality where AI assistance is most pervasive and least separable — almost every document produced in Europe now passes through some model-assisted step, and a general text-labelling duty would have been both unenforceable and uninformative.

So the drafters narrowed on the harm rather than on the technology: text that shapes public understanding of public matters, with nobody answerable for it. Reading the two subparagraphs together clarifies both — the first is about content that looks authentic, the second is about content that informs without accountability.

Note also what Article 50(6) preserves: the Article 50 duties do not affect requirements in Chapter III or other Union or national law. Nothing here displaces the Digital Services Act, national media law, consumer protection rules on misleading commercial practices, or professional codes. A publisher out of scope of Article 50(4) may still owe disclosure under a different instrument.

What is unresolved

Three questions have no settled answer, and it is more useful to say so than to guess.

• How much human review is a “process”. A newsroom that generates a hundred market summaries an hour with a single sub-editor nominally responsible is at the edge, and nothing in the text says where the edge is.
• What manipulation means for text. The Act does not carry the Article 50(2) assistive-editing exception across into 50(4)'s second subparagraph, so its application to heavy AI editing of human-written copy is an open reading.
• Whether “public interest” will be read against Article 10 ECHR case law. That is the most likely reading and it is not the only possible one; a CJEU reference would settle it.

Article 50 applies from 2 August 2026 under Article 113, with non-compliance in the Article 99(4) penalty tier of up to EUR 15 million or 3% of total worldwide annual turnover, whichever is higher.