The most repeated claim about the EU AI Act since the spring is also the least accurate, that any AI-assisted article published in the EU now needs a visible label. It does not. This piece walks through how AI text disclosure actually works under Article 50 of the EU AI Act, in the order a practitioner would test a page, and shows why the compliant answer for most agencies is a documented editorial process rather than a disclaimer. The sources are the Article 50 text, the European Commission's 2026 transparency guidelines, and law-firm analysis from Stibbe and Herbert Smith Freehills Kramer. Teksyte sells SEO services and operates properties that publish under this regime, which is disclosed here.
Where the text obligation actually applies
The disclosure duty on written content is narrow by design. Under Article 50(4), a deployer must disclose AI-generated text only where that text is published to inform the public on matters of public interest, and even then only where no human editorial review with named responsibility has taken place (Confirmed, Article 50(4) and Commission guidelines, 2026). Two gates, not one. A service page, a product description or an internal document clears the first gate by never being public-interest text at all. A reviewed news or analysis piece clears the second gate through the review itself. Only unreviewed, public-interest writing sits inside the obligation.
The exemption is a process, not a paragraph
The common reading is that the exemption is a form of words you add to a page. It is not. The Commission guidelines tie the exemption to a human editorial review process with a natural or legal person holding editorial responsibility for publication (Confirmed, Commission guidelines, 2026). That is an accountability record, not a sentence in the footer. The correction matters because a disclaimer widget satisfies nobody, while a named editor and a dated sign-off satisfy the actual text of the law. An agency that adds a label but cannot say who reviewed a piece has done the cosmetic thing and skipped the substantive one.
A decision tree you can run on any page
Run each page through four questions, in order, and stop at the first that resolves it.
Is the page public-interest text? If it is a commercial, product or internal page, stop. No text obligation (Confirmed, Article 50 scope).
Did a human review it before publication? If yes, and a named person holds editorial responsibility, record that owner and the date. Exemption applies (Confirmed, Article 50(4)).
If no human reviewed it, add review, or add a clear and distinguishable disclosure the reader meets at first interaction, presented accessibly (Confirmed, Commission guidelines, 2026).
Log the decision against the page, so the accountability trail exists if a market-surveillance authority ever asks.
The tree resolves most catalogues at question 1 and most editorial work at question 2. The residue that reaches question 3 is small, and for that residue disclosure is cheap.
What this costs an agency in practice
The honest figure is time, not licence fees (Observed). Assigning a named editor to each public property and recording a dated sign-off adds minutes per article, not a budget line, and it produces a governance artefact worth having on its own terms. Teksyte runs this across its own network, which is disclosed here as a commercial operator of those sites. The firms finding this expensive are the ones that were publishing public-interest content at volume with no human in the loop, and for them the cost is the review they had skipped, surfaced by the law rather than created by it.
What to set up this week
Name an editor for every public-facing property and give them a standing record, a simple log of piece, reviewer, date. Sort your library once against the public-interest test so you know which pages the obligation can even reach. For the unreviewed public-interest residue, add review rather than a label, because review is what the exemption rewards. Synthetic images and video sit under a separate machine-readable marking duty, worth handling in the same review. Done this way, AI text disclosure stops being a compliance scramble and becomes what it should have been, a record of who stands behind what you publish.