The "AI-generated" label: the two routes that lift it, and what each one demands

Article 50 of the EU AI Act covers a narrow category of publications, and the obligation falls away as soon as the content has undergone human review or editorial control under editorial responsibility that someone actually holds. The guidelines the Commission published on July 20, 2026 set out what each of those two routes demands, what counts for neither, and the one move that voids the exemption once it is earned.

The obligation covers a narrow category of text

Before looking for an exemption, establish whether the obligation exists at all. The second subparagraph of Article 50(4) does not cover AI-generated text in general: it covers text "published with the purpose of informing the public on matters of public interest," and the guidelines set out the three elements, which must all be met (point 131). Whether you have to label your articles as AI-generated walks through the content types covered, chatbots included. This one assumes your text is inside the scope.

Whether the regulation reaches you is a separate question. A deployer established in the Union is bound. A deployer established in a third country is bound as well where the output of the AI system is used in the Union and the deployer itself foresees that dissemination, by directing or authorizing it. It is not bound where its content reaches audiences in the Union through channels that are unforeseeable and outside its control (point 13).

One example in the document is worth reading to the end, because its parenthesis carries the information. The text places outside the scope "AI-manipulated text that is part of a company's advertisement or product descriptions," and then qualifies: "not including any claims related to e.g. health, consumer safety or sustainability." An ordinary product page therefore carries nothing. The same page, once it adds an environmental argument, moves into an area where two regimes meet: Article 50, and the law on environmental claims, where twelve new prohibitions become applicable on September 27, 2026.

The "or" in Article 50 opens a second route

Here is the sentence that decides, in the official English text of the regulation. The obligation "shall not apply where the use is authorised by law to detect, prevent, investigate or prosecute criminal offences or 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."

The structure is an "or" followed by an "and." Either form of control is enough on its own, but named editorial responsibility is added in both cases. The guidelines put it as two cumulative conditions (point 133), the first of which can be met in two ways. What each of them covers is in that document, not in the regulation, which does no more than name them.

The two routes describe two different acts, and it is the Commission's document that defines them. The guidelines are published in English only, so everything quoted below is the source itself, verbatim, with the point number for checking.

• Human review is "the deliberate examination of the substance of the content by one or more natural persons possessing relevant knowledge and professional judgement pertaining to the subject matter under scrutiny," the Commission citing academic peer review and professional validation chains. "Fact-checking the accuracy of the content is a minimum requirement that should be part of that review" (point 134).
• Editorial control is "the control exercised in practice by a responsible editorial entity (e.g. an editor-in-chief) over the content having the authority to approve, alter or reject the substance of the text based on substantive grounds," which includes fact-checking and ensuring the trustworthiness of sources (point 134).

The difference lies in the competence required. The first route asks for someone who knows the subject; the second asks for someone who holds authority over publication and exercises it on substance. A publisher who cannot put a specialist on every article therefore has a workable path, provided the control bears on substance and is real.

What counts as neither review nor editorial control

Point 135 closes the door on four common practices, and it does so in a single sentence: "Superficial, solely formal or procedural checks (e.g. spell-checking or grammatical correction), the mere existence of an editorial policy, automated review processes or cursory editorial approval without substantive engagement by the human reviewer or the editorial entity, cannot fulfil the conditions."

Each of those targets a common way of believing yourself compliant. Publishing an AI usage policy exempts nothing if nobody reads the copy. A sign-off given in one click, unread, does not exist as far as the provision is concerned. Two of the counter-examples in the document are blunt: a website posting AI-generated articles on Union policy without any deliberate human review, and AI-generated articles "reviewed and edited by another AI system and where a human editor performs a mere superficial, grammatical check before publication."

That last case is ours. We have our articles checked by several automated agents before publication, and that step falls squarely within the "automated review process" that point 135 excludes. It serves quality. The exemption rests on the human review or editorial control exercised by the person who signs and takes responsibility.

An AI edit after sign-off voids the exemption

The most operational point in the document is also the least reported: of fifteen French-language pages about Article 50, returned by two searches on September 16, 2026, one states it. Where an AI system is used to modify, supplement or reformulate content after editorial sign-off, the result must be treated as AI-generated or manipulated: "Any substantive AI intervention occurring after the human review or editorial control process has taken place will therefore cause the exception to become void" (point 136).

The exemption is lost through the order of operations, with nothing visible changing in the text. The concrete cases are ordinary: an automated optimization pass triggered after review, a summary or standfirst produced by the model at publishing time, a headline rewritten last to fit a display constraint. The rule to hold on to is simple: human sign-off must be the last substantive step, and anything that comes after it has to go back through it.

The consequence is heavy for anyone publishing in several languages. A version translated by a model after the source text was signed off is an AI intervention subsequent to review. The Commission confirms it from the other side, listing among the exempted cases "AI-supported translation of a human-written article whereby the translation has undergone human review." Eight languages therefore call for eight reviews, not one review followed by seven machine translations.

The person who takes responsibility has to be findable

The second condition is sometimes treated as a formality. It requires a natural or legal person to "hold the ultimate legal responsibility over the publication of the content, including the human review or editorial control": an individual, an editorial board, or the publishing company (point 138).

The document then adds a publicity requirement of its own. To ensure public accountability and trust, "the identity and contact details of the legal person, the natural person or the function with editorial responsibility should be made publicly available on an easily findable location (if not yet otherwise available)," and the Commission points to a website's terms and conditions or other user-facing legal information, and to a publication's colophon or edition notice in print (point 138).

One clarification avoids a misreading here. The same document holds that information is not provided in a clear and distinguishable manner where it is buried in terms of use that readers do not read (point 142). These are two different pieces of information: the "AI-generated" label, where it is due, sits on the content and must be perceived at first exposure; the identity of the editorial responsible belongs in the site's legal information. The first accompanies the text, the second qualifies the publisher.

On our side, that line points to our legal notices, which name the publication director and give a contact address, and which every article page links to. Point 138 is satisfied by a findable identity and contact at site level. The byline at the foot of an article does another job, the one we describe in one article, one named expert.

Who is the deployer when a contractor produces the content

The obligation falls on the deployer, the term the regulation uses for whoever uses an AI system under their own authority. That authority is defined by the decision to deploy the system and by control over the manner of its actual use, outputs included. No technical control is required (point 12).

The guidelines draw the practical consequences without hedging. Employees acting on instructions are not deployers separate from their employer, journalists and content creators included. A legal person "remains a deployer even if it involves third parties (e.g. contractors, freelancers) in the operation of the system on its behalf and under its responsibility and control" (point 14). Hosting providers and platforms whose role is limited to disseminating content created by third parties are not deployers (point 16).

The counter-example is what settles a real case: "a company that merely commissions an advertising agency to produce an advertisement, without taking decisions and exercising control over whether and how the advertising agency uses AI in the production process, is not a deployer." In other words, the qualification follows whoever decides on the use and on publication, whichever hand typed the prompt. A publisher who picks a tool, configures it, signs off and publishes is the deployer, whichever contractor operates it. A contract leaves that qualification intact; what it can do, and should do, is organize the record.

Three dates, three recipients

Article 50 has applied since August 2, 2026, and it addresses providers and deployers alike, depending on the paragraph. December 2, 2026 concerns only the provider's marking and detection obligations: the AI Act calendar for a publisher separates the two regimes, and our article on C2PA and SynthID gives the real state of the technical signals.

A third date is worth knowing because it catches content inventories. Outputs generated before August 2, 2026 do not need to be labeled retroactively, public-interest text included. But, point 154 adds, "conversely, if texts that have been generated or manipulated before 2 August 2026 are published on or after that date, they need to be labelled." A draft produced in the spring and posted today therefore falls under the full regime, and under the exemption if it has been reviewed.

Frequently asked questions

Is a company blog public-interest text?

Usually not. A product page, a range comparison, a how-to guide, a customer story do not inform the public on a matter of public interest. What tips it is the subject: as soon as a page carries a health, consumer-safety or sustainability claim, it may fall inside the scope, whatever the format. The qualification is made page by page.

Is review by another model enough?

No. Point 135 places automated review processes among the things that do not meet the conditions, and one of the counter-examples in the document is precisely an AI-generated article, reviewed by another AI, then approved by a human on form alone. An automated check serves the quality of the text; the exemption stays attached to human review or editorial control.

Where should the label go when it is due?

The regulation imposes a moment rather than a location: the information is provided in a clear and distinguishable manner at the latest at the time of first exposure (Article 50(5)). It must be perceivable without any specific tool or dedicated action, and it does not meet that condition if it is relegated to terms of use or hidden under layers of menu options.

Do we need an author page to qualify?

No. What the guidelines ask is that the identity and contact details of the person or the function holding editorial responsibility be public and easy to find. The site's legal information does that job. A page per byline serves other purposes, editorial and credibility ones.

Are these guidelines binding?

They are not, and they say so: only the Court of Justice of the European Union can give an authoritative interpretation (point 5). They are adopted under Article 96(1)(d) of the regulation and addressed to competent authorities as much as to companies, which makes them the reference reading in the absence of case law. The other recognized way to demonstrate compliance is adhering to the Code of Practice on Transparency of AI-Generated Content, which the Commission's FAQ states has been assessed as adequate by both the Commission and the AI Board.

What does a company risk by not labeling?

Failure to comply with the transparency obligations of Article 50 carries an administrative fine of up to EUR 15 000 000 or 3 % of total worldwide annual turnover for the preceding financial year, whichever is higher, and whichever is lower for an SME. Supervision lies with the market surveillance authorities designated by the Member States, which can act on their own initiative or following a complaint, any affected person having the right to lodge one under Article 85.

Sources

• Regulation (EU) 2024/1689 (AI Act), consolidated version, Article 50 and Article 111(4). The consolidated version is the one that incorporates Regulation (EU) 2026/1744 of July 8, 2026
• Commission guidelines on the transparency obligations of Article 50, C(2026) 5054 final of July 20, 2026: the document itself, 51 pages. The point numbers cited in this article are those of that document
• Commission FAQ on the transparency obligations of Article 50
• Code of Practice on Transparency of AI-Generated Content, voluntary adherence

This article describes the state of the law and of administrative guidance as recorded on September 16, 2026 from the official texts. It is not legal advice.