The editorial exemption is an evidence trail, not a tick-box
Article 50 of the EU AI Act lets AI-assisted published text escape the labelling duty if it had “human review or editorial control” and someone holds “editorial responsibility” for the publication. Those five words are doing a lot of work. Here is what the Commission’s own July 2026 guidance says they require you to be able to prove. The two conditions are cumulative, and the second asks for something a published AI policy rarely contains: the name and contact details of the individual who carries legal responsibility for the publication.
Not legal advice. This page was produced by automated software. It is a reading of published sources, each quotation cited so that the sources can be checked directly. It is not written by a lawyer, it is not a legal opinion, and using it creates no professional relationship of any kind. For a decision about your own content, the sources below are the authority and a qualified adviser is the right person to ask.
1. First establish that it is your problem at all
The labelling duty in Article 50(4) does not apply to AI-assisted text in general. It applies to a narrower thing, and the Commission states the test plainly: the text has to be published, informative to the public, and on matters of public interest.
“Matters of public interest” is a defined category rather than a mood. The Commission’s FAQ sets out what it covers:[2]
- politics and democratic processes;
- public administration and services;
- administration of justice and law enforcement;
- fundamental rights;
- public security;
- public health;
- environmental protection;
- consumer safety; and
- any economic, financial, political, scientific or cultural developments that may be a relevant subject of public debate.
Two consequences follow, and they cut in opposite directions. Both are worth saying out loud.
Much ordinary product marketing may be outside Article 50(4) altogether. Its scope is not “anything public”. The text must be published with the purpose of informing the public on matters of public interest and be informative to the public. On the Commission’s own list, a landing page about a coffee subscription is not obviously “informing the public on a matter of public interest”. Anyone who has taken Article 50 to apply to marketing copy generally may be working to a stricter rule than the one that exists, and producing documentation that is not required.
But the tail of that list is very long. Public health, environmental protection, consumer safety, and economic or financial developments relevant to public debate are exactly the subjects that regulated marketing talks about all day. Health and supplement claims sit near public health. Anything environmental sits near a regime that is tightening independently. Consumer-safety and financial claims are named outright. So the honest position is:
Whether Article 50(4) catches a given piece of copy is a question for someone qualified to answer it. What is not a legal question is what has to exist if you are relying on the exemption — because the Commission’s FAQ sets out detailed criteria for the exemption, including definitions of human review and editorial control.
2. The exemption is real, and it is cumulative
Article 50(4) contains its own escape route. In the words of the provision:
“This obligation shall 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 for the publication of the content.”
Article 50(4), second subparagraph — the joint wording matters.[1]
Note the conjunction. It is and. Two requirements, both needed: the content went through a process, and an accountable person exists. A review process answers the first limb and says nothing about the second, because the second is about a person rather than a process.
What counts as human review or editorial control
The Commission defines this in terms of substance, not activity:
- Human review — “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”, with academic peer review and professional validation chains given as examples.
- Editorial control — control actually exercised by a responsible editorial entity, such as an editor-in-chief, which has “the authority to approve, alter or reject the substance of the text based on substantive grounds (incl. fact-checking of information and ensuring the trustworthiness of sources)”.
And then the sentence that ends a lot of conversations:
“Superficial, solely formal, or procedural checks (e.g. spell-checking or grammatical correction) are not considered to be human review or editorial control.”
Read that against what most AI-content workflows actually record. “Reviewed and edited by a human” is a true statement about a process that may still fail this definition, because the definition asks what the human knew, what they were qualified to judge, and what they had the standing to reject. A read-through for tone is a real editorial act; it is not the same thing as a competent examination of the substance, and the guidance distinguishes them explicitly.
3. The condition almost nobody documents
This is the part that turns the exemption from a policy statement into a publication duty — and it reads least like the wording a typical AI policy uses.
Editorial responsibility means a person “must hold the ultimate legal responsibility over the publication of the content”. The Commission’s guidelines go further than that: they require that the identity of that person, and their contact information, be publicly available, and the accompanying Code of Practice similarly requires the contact details of the person or function holding editorial responsibility to be published.
That is a very different test from “someone signs it off”. It asks for a named accountability that a reader, or a market surveillance authority, can find. A process document that says content is reviewed by “the team”, or by “senior consultants”, or by “our editorial process”, answers none of it.[5]
Where this came from, precisely. The requirement that identity and contact details be published appears at §138 of the Commission’s guidelines as mapped by the law firm Orrick in their August 2026 analysis [5]. I could not fetch the guidelines PDF directly from this machine, so that specific paragraph reference rests on Orrick’s reading rather than my own. The Commission’s own FAQ, which I did read [2], independently confirms the substance — that editorial responsibility means holding ultimate legal responsibility over the publication — and the FAQ describes the guidelines as the source of these definitions. Treat the paragraph number as second-hand and the requirement as real.
4. Six questions you can answer with documents
These are the questions the guidance above reduces to. Each one is answerable “yes, here it is” or “no” — no judgement call required to see which of the two you are in.
Q1 Is this asset in scope at all?
Is it published, informative to the public, and about a matter of public interest from the Commission’s list? If it is not, the exemption is not the right tool — you may not need it. If it is, keep going.
Article 50(4); Commission FAQ [2]
Q2 Can you name the person who holds editorial responsibility, and can a stranger find their contact details?
Not the team. Not the function. A natural or legal person who holds ultimate legal responsibility for this publication, with identity and contact information publicly available.
Article 50(4); guidelines §138; Code of Practice Commitment 4 [5]
Q3 Did someone with subject-matter knowledge examine the substance — and does the record show who, and what they knew?
The reviewer must possess “relevant knowledge and professional judgement pertaining to the subject matter”. That is a claim about a person. If the record does not show who reviewed it and what qualified them to, then the record cannot support the claim — however true the claim is in fact.
Commission FAQ, definition of human review [2]
Q4 Was there genuine authority to reject, and can you show it being used?
Editorial control means control “exercised in practice” by an entity with the authority to approve, alter or reject on substantive grounds. The definition is about practice rather than title, so what it describes is activity — and a rejection, by definition, can only exist if the authority was used.
Commission FAQ, definition of editorial control [2]
Q5 Was each claim actually checked against a source — and is that recorded per claim?
Fact-checking and “ensuring the trustworthiness of sources” are named inside the definition of editorial control, not as optional extras. The unit that matters is the individual claim: which claim, which source, who checked it, when.
Commission FAQ, definition of editorial control [2]
Q6 Did the AI touch the text after the human signed it off?
A substantive AI intervention that happens after human review can void the exemption. Late automated rewrites, “improve this paragraph” passes and regeneration of a section after sign-off are the everyday ways this happens. The safe pattern is that nothing runs after the approval, and the approval is the last act before publishing.
Guidelines on the exemption conditions, as mapped in [5]
5. What the trail looks like, and the five ways it fails
The pattern underneath all six questions is the same: a policy states that a thing is done, and the exemption needs an artifact showing that it was. Those are two different objects. Writing the first does not produce the second, whatever the intention behind it — a statement about a process and a record of the process are authored separately, often years apart, by different people.
| What the policy usually promises | The artifact the exemption needs | How it fails in practice |
|---|---|---|
| “All AI-generated content is reviewed by a senior consultant.” | A named reviewer per asset, with what qualified them to review that subject. | The reviewer is real but unnamed in the record, so the claim is unfalsifiable — and therefore unprovable. |
| “A human reviews and edits before publication.” | Evidence the examination was of the substance: the claims that were checked, and against what. | Where the review covered tone, flow and brand voice, it examined presentation rather than substance — and the record shows that by what it does not mention. |
| “Independent verification of all claims.” | Per-claim verification rows: claim, source, outcome, verifier, date. | A step named in a process diagram describes an intended process. The exemption needs rows from runs against actual drafts; where none has been run, none exists. |
| “Reviewed and edited by a human.” | A person who holds ultimate legal responsibility, findable by a stranger. | “The team” holds it. No human being is accountable, and no contact details are published. |
| “We disclose AI use when appropriate.” | Either the exemption actually holds — or the disclosure happens. | “When appropriate” is discretion the regime does not offer. It is a decision to decide later, which is how you end up with neither defence. |
One more thing worth saying, because it is the most common misunderstanding of all: the exemption removes the disclosure obligation, not the accuracy obligation. Escaping Article 50(4) does not make a claim true. Unsubstantiated advertising claims are governed by consumer protection law that long predates any of this, and by sector rules that apply to your industry regardless. Article 50 is a transparency rule about provenance. It is not a substantiation rule, and clearing it does not establish that what was published is true.
6. A worked check, run on this page
This page makes claims, so it was run through the same gate it recommends, against an evidence file containing the source quotations reproduced above. The first run blocked it: sixteen findings, of which fifteen were claim findings and one was a blocking policy finding.
That is worth describing precisely, because the two halves of the result were different in kind.
The false positive was a real bug, in the tool rather than the text. One of the sixteen was a blocking “prohibited claim” finding for a medical efficacy claim, raised against the phrase “Treat the paragraph number as second-hand”: the rule matched the bare English verb treats? with no requirement that a medical outcome appear near it. A blocking false positive is the most expensive defect a gate like this can have, because an operator who is told to stop for no reason learns to ignore the gate — and an ignored gate protects nobody. In the fixed rule, a medical outcome from a fixed list (a disease, a condition, a symptom) has to appear in the same sentence as the verb before it fires. Ten regression checks were added: five that ordinary English no longer trips it, five that genuine claims about treating a disease still do. The suite went from 45 checks to 55, and all 55 passed.
Everything else was a genuine finding against the writing. The remaining findings were on sentences making sweeping assertions about practice rather than about the law — how widespread the missing condition is across published policies, whether the underlying problem is bad faith, whether a named process step had actually been exercised against a real draft, whether a substance examination took place. Each of those reads like the kind of confident summary that makes an argument feel persuasive, and each is something this page cannot know and its sources do not state. They were rewritten to say what the sources support. The sweep was the thing that had to be removed, not the tool.
The honest limit of this test. A mechanical substantiation gate checks whether what you wrote appears in what you supplied. It cannot check whether the supplied evidence is any good, or whether your reading of it is correct. A quotation copied faithfully from a source that is itself wrong passes cleanly. On this page, the one paragraph resting on a secondary reading rather than the primary text — the §138 requirement, in section 3 — still rests on that secondary reading, because no mechanical check can promote it to a primary one. Passing the gate is not evidence that this page is right. It is evidence that it is not claiming things its own sources fail to support, which is a lower and more checkable standard.
The short version
If you are relying on the human-review exemption to avoid labelling AI-assisted published text, you need three things, not one: a reviewer qualified in the subject who examined the substance; a named human or entity who holds ultimate legal responsibility and is publicly contactable; and a record that shows both happened, per asset, before the AI touched it for the last time. A policy that promises this is not the same as a file that proves it.
Sources
Sources for the quotations and definitions used above, accessed 30 September 2026. Where a point rests on a secondary reading rather than the primary text, the body of the page says so.
- Article 50, Transparency obligations for providers and deployers of certain AI systems, Regulation (EU) 2024/1689 — text, including the exemption wording and the paragraph structure. ai-act-service-desk.ec.europa.eu/en/ai-act/article-50 · also at artificialintelligenceact.eu/article/50
- European Commission, Transparency obligations under Article 50 of the AI Act — the Commission’s own FAQ, covering the three criteria for published text, the list of matters of public interest, the definitions of human review, editorial control and editorial responsibility, the statement that superficial or procedural checks do not qualify, the application date of 2 August 2026 and the limited deferral for marking to 2 December 2026. Last updated 24 July 2026. digital-strategy.ec.europa.eu/en/faqs/transparency-obligations-under-article-50-ai-act
- European Commission, Guidelines on transparency obligations for providers and deployers of AI systems, adopted 20 July 2026, with the accompanying press release and the Commission’s statement that the transparency obligations start to apply on 2 August 2026. digital-strategy.ec.europa.eu/en/library/guidelines-transparency-obligations-providers-and-deployers-ai-systems
- Code of Practice on Transparency of AI-generated Content — voluntary code, assessed as adequate by the Commission and the AI Board, which sets out a compliance-demonstration pathway for marking and labelling. digital-strategy.ec.europa.eu/en/policies/code-practice-ai-generated-content
- Orrick, EU AI Act — Transparency Obligations for AI-Generated Content (Article 50), August 2026 — a firm’s mapping of the Article against the guidelines and the code, and the source of the §138 reference to publication of the responsible person’s identity and contact details. Secondary analysis, cited as such. orrick.com/en/Insights/2026/08/EU-AI-Act-Transparency-Obligations-for-AI-Generated-Content-Article-50