Do AI-written finance, health, legal or news articles need a label?

Article 50(4) contains the only sentence in the EU AI Act that reaches directly into an editorial workflow:

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 shall disclose that the text has been artificially generated or manipulated.

And then, immediately, the escape hatch: this does not apply where the AI-generated content has undergone human review or editorial control and where a natural or legal person holds editorial responsibility for the publication.

Two questions decide whether a label is required on any given article: is it public-interest text, and does your process clear the carve-out. Most publishers guess at both. This page works through each.

Question 1: is it "public interest"?#

The Act does not define the phrase, and the honest position is that its edges are unsettled. What is reasonably clear from the text and the Commission's July 2026 guidelines is that the duty targets content published to inform the public on matters of public concern — the informational function, not the topic label.

Reasonably clearly in scope:

• News reporting, current affairs, politics, public health.
• Consumer-facing explainers on health, medicine, finance, tax and law, where a reader takes them as information about the world.
• Coverage of public bodies, elections, courts, safety and the environment.

Reasonably clearly outside:

• Product descriptions, pricing pages, marketing copy, release notes.
• Internal documentation, support macros, transactional email.
• Fiction and entertainment presented as such.

The contested middle is where most commercial publishing sits: the finance blog that exists to rank for "best savings accounts", the health content marketing on a supplement retailer's site, the legal-adjacent explainer on a SaaS company's blog. These are commercially motivated and informational, and being commercially motivated does not remove the informational function. The "it's marketing" argument is weaker than most content teams assume, because a reader who arrives from search does not experience it as marketing — they take it as information, on a matter where being wrong has consequences.

The practical test: would a reader treat this as information they might act on, in a domain where bad information hurts? If yes, plan for the article to be in scope and rely on the carve-out rather than on the scoping argument.

Question 2: does your process clear the carve-out?#

This is the load-bearing question, and it has two limbs. Both must hold.

Human review or editorial control. Someone competent read it before publication, with the authority and the practical ability to change or spike it. A spellcheck pass will not do it, and neither will a rubber stamp on a queue of forty drafts. The word doing the work is control.

A natural or legal person holds editorial responsibility. Someone — a named editor, or the publishing company itself — is answerable for the piece. Note that a legal person suffices: your company can hold editorial responsibility. What does not suffice is nobody holding it.

Where teams fall out of the carve-out, it is almost always the same three ways:

Volume outruns review

A pipeline generating fifty articles a week with one part-time editor is not exercising editorial control over fifty articles, whatever the workflow diagram says.

Auto-publish paths exist

A CMS that publishes on a schedule with review as an optional step means some articles ship unreviewed. The carve-out is assessed per publication rather than per policy.

Nobody is named

"The team reviews everything" identifies no person and no legal person accountable for a specific piece.

The label is not the expensive part#

Notice what the structure of Article 50(4) actually incentivises. If your process clears the carve-out, no label is required. If it does not, you must disclose. Either way you need to know which is true, and that knowledge is a property of your workflow, not of your published pages.

So the work is:

• Write down who holds editorial responsibility, by name or by legal entity, per publication or per section. Publish it — an editorial or AI-use policy page is the natural home, and it is the single most useful artefact you can produce here. It is corroboration if the question is ever asked, and it costs an afternoon.
• Make review a gate, not a step. If an article can reach production without a reviewer's action, the carve-out does not hold for the ones that slip through.
• Keep the record per article. Who reviewed it, when. Your CMS probably already stores this; the question is whether you could produce it a year later.
• Decide your default. Some publishers label AI-assisted public-interest content regardless, because a consistent rule is cheaper to run than a per-article legal judgement. That is a defensible choice and it moots the whole analysis.

Where the label goes, if you use one#

Article 50(5) applies: clear and distinguishable, at the latest at the time of first exposure. For text, first exposure is where reading starts — at the top of the article, by the headline or the byline. A disclosure appended after the body discloses at last exposure, which is not what the provision asks for. A tag on a category page the reader never visits is not disclosure either.

The Commission's Code of Practice includes uniform EU text labels and icons for this; the official icons and their placement rules covers each format.

The other duty on the same article#

Article 50(4) is the visible duty. Article 50(2) — machine-readable marking of synthetic output — is a separate obligation on the provider of the generative system, and it applies to AI-generated text regardless of whether it is public-interest content and regardless of the editorial carve-out. Human review does not un-generate the text.

In practice, machine-readable marking for text is far less mature than for images, and text marking rarely survives a copy-paste into a CMS at all. That does not transfer the duty to you, but it does mean the marking a reader's tooling would look for is usually not there. See C2PA vs watermarking vs metadata for what the marking duty actually asks for and why text is the hard case.

What we will not do, and why it matters here#

We do not run statistical AI-text detection, and we will not add it. Those classifiers are unreliable in both directions — they flag human writing and miss edited machine writing — and an external tool guessing whether your article was AI-written would be inventing the exact fact Article 50(4) puts in your hands. The provision turns on your declaration and your editorial process.

What an external check can establish, and what our scanner does: whether a visible AI-content label is present on article-like pages, whether it sits where first exposure happens, whether an AI-use or editorial policy page exists and is linked, and what all of that looked like on a given date. That is the observable half, and it is the half you will be asked to produce. The free scan checks article-like pages for visible labels and returns the findings as detected, not detected, or could not verify — never as a compliance verdict, because the carve-out you may well qualify for is invisible from outside.

Common questions

Sources and further reading

• Article 50 — Transparency obligations (EU AI Act)
• Commission guidelines on transparency obligations, 20 July 2026

Last updated September 2026. Informational only, not legal advice: this page describes what the text of the EU AI Act says and what an external check can observe, not whether any particular site complies. Corrections welcome at hello@disclosureproof.com.