Labelling AI text: the editorial exemption
The obligation for AI text is considerably narrower than the one for images — and precisely for that reason it is misunderstood. It does not bite on every generated sentence but only on text published to inform the public on matters of public interest. And it falls away where a human bears editorial responsibility. Two filters, then, which have to be applied in that order.
- Legal basis
- Art. 50(4) Regulation (EU) 2024/1689
- Addressee
- Deployer, not provider
- Applicable from
- 2 August 2026
- Exemption
- editorial responsibility
Filter one: is it a matter of public interest?
The regulation does not define the term exhaustively, and that is the real difficulty. What is clearly meant is the area in which disinformation does damage: politics, health, safety, the environment, justice, elections, public debate. Text intended to inform the public about such matters falls within it. Describing a product, drafting a job advert or writing an internal report, by contrast, does not inform the public about such a matter.
The line does not run between “marketing” and “journalism” but along the subject matter. A company blog explaining a legal position can very much reach into that territory — a piece about a new EU regulation informs the public about a matter of public interest, even if it sits on a corporate site and is meant to win customers. Anyone generating such content should not count on commercial intent carrying them out of scope.
Internal stays internal
The obligation attaches to publication. Drafts, summaries for your own team, research notes and internal emails do not trigger it. The question becomes relevant from the moment the same text goes outside — not before.
Filter two: does anyone bear editorial responsibility?
Where a text does fall within scope, the obligation can still fall away: namely where the AI-generated or AI-manipulated content has undergone a process of human review or editorial control and a natural or legal person holds editorial responsibility for the publication. That is the core of the exemption — and its price: responsibility has to be attributable, not diffusely spread across the team.
In practice that means three things. First, you need a process, not just a glance: who checks, against what, and what happens when something is objected to. Second, you need a named party standing behind the publication. Third, you need a trail that can show this later — because in a dispute you have to evidence the exemption, not merely assert it. An approval step in the editorial system with a name and a timestamp is usually enough for that; a verbal agreement is not.
Exemption probably does not hold
- Generated text goes live automatically with no approval step.
- Review only for tone and spelling, not for substance.
- Nobody is named — “the team looks it over”.
- No record of who approved what and when.
Exemption probably holds
- A defined review process before publication, on substance.
- A named person or unit with editorial responsibility.
- Approval documented in the system, with timestamp and name.
- Objections lead to a change, not to a note in the file.
What the provider obligation alongside it means
Even where disclosure falls away, the provider’s marking obligation under paragraph 2 remains untouched — and it expressly covers text as well, not only image and sound. That is not an obligation of the publishing company, but it has a practical consequence for it: a machine-readable marked text remains identifiable as AI-generated regardless of whether a visible note sits beneath it. Anyone relying on the editorial exemption should therefore not assume the origin of the text is untraceable.
Wording that holds up
Where disclosure is required, the note should have three properties. It sits where the text begins, not in the page footer. It states what was artificially generated — the whole piece or a section. And it is recognisable as a notice rather than disguised as a marketing line. Under paragraph 5 the information has to be clear and distinguishable at the latest at first exposure and has to meet the accessibility requirements; a note appearing only in the caption of a preview image does not satisfy that.
No reason for a blanket notice everywhere
Some organisations put an AI notice under every text as a precaution. That is convenient but costs credibility — and it obscures where generation actually happened. The effort lies in the distinction, not in the boilerplate.
What to do now
- Sort your content types: which of your text streams touch on matters of public interest — and which do not?
- For those streams, define an approval step with named responsibility and a documented timestamp.
- Decide and record which formats get disclosed because the exemption does not hold there.
- Settle on consistent wording and placement so that each department does not invent its own notice.
- List the text tools in use in the AI inventory — including the question of whether you merely use them or substantially modify them.
Reguly keeps the second part of this provable: which AI systems are in use, who is responsible per system, which obligation attaches to it and what training the people behind it have. The judgement of whether a text falls under paragraph 4 remains an editorial and legal decision — what happens here is that it becomes documented and findable again.
Talk to us
Which of your texts does the obligation catch?
We sort through your content types with you and show where an approval step suffices and where disclosure is required.
Arrange a call
