Does the EU AI Act require you to label AI-generated text?
AI Architect
Key takeaways
- Article 50 became applicable on 2 August 2026 under Article 113 and was not postponed by the Digital Omnibus.
- Machine-readable marking under paragraph 2 binds the provider of the AI system rather than the site publishing its output.
- Deployer disclosure for AI-written text reaches only text published to inform the public on matters of public interest.
- The editorial exemption has two limbs joined by "and", so naming an editor is not enough on its own.
- Providers of systems placed on the market before 2 August 2026 meet the marking duty only from 2 December 2026.
The claim that the EU AI Act requires you to label AI-generated content is repeated constantly and is wrong as stated. Article 50 contains four separate duties, attached to two different roles, with different triggers and different artefacts. Which one reaches you depends on what you build and what you publish.
What does Article 50 of the EU AI Act actually require?
Four things, and reading them as one rule is where most compliance work goes wrong.
Article 50 of Regulation (EU) 2024/1689 became applicable on 2 August 2026 under Article 113. Paragraph 1 makes providers design systems that interact directly with people so that the person is informed they are dealing with an AI system, unless that is obvious. Paragraph 2 makes providers of generative systems mark output as artificially generated. Paragraph 3 makes deployers of emotion recognition and biometric categorisation systems inform the people exposed to them. Paragraph 4 makes deployers disclose deep fakes, and in a second subparagraph, certain AI-generated text.
Which party carries which Article 50 duty
Show as text
| # | Layer | Note |
|---|---|---|
| 1 | Provider of a system that talks to people | Para 1. Tell the person it is an AI system. |
| 2 | Provider of a generative AI system | Para 2. Mark output machine-readably. |
| · | duties change hands here (breakpoint) | A deployer inherits none of the above. |
| 3 | Deployer publishing a deep fake | Para 4. Disclose image, audio or video. |
| 4 | Deployer publishing AI-written text | Para 4. Public-interest information only. |
Paragraph 5 governs how all four are delivered: the information goes to the person “in a clear and distinguishable manner at the latest at the time of the first interaction or exposure”, and must meet accessibility requirements.
Who has to mark AI output in a machine-readable format?
The provider of the AI system, which for most companies means the model vendor rather than the company publishing the output.
Paragraph 2 binds “providers of AI systems, including general-purpose AI systems, generating synthetic audio, image, video or text content” to ensure output is “marked in a machine-readable format and detectable as artificially generated or manipulated”. The obligation is qualified by feasibility: solutions must be effective, interoperable, robust and reliable “as far as this is technically feasible”, against the state of the art.
Two carve-outs sit in the same paragraph. The duty does not apply where systems “perform an assistive function for standard editing” or do not substantially alter the input data or its semantics. A grammar checker is out. A system that writes the paragraph is in.
If you call somebody else’s API, you are almost certainly a deployer of that system under Article 3, and this paragraph is not yours to satisfy. What you can do is check whether the provider’s marking survives your pipeline.
When does a publisher have to disclose AI-generated text?
Only in one situation, and it is much narrower than the general advice suggests.
The second subparagraph of paragraph 4 reads: 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.”
Does published AI-written text carry a duty
Show as text
| # | Layer | Note |
|---|---|---|
| 1 | Is the output text rather than image or video? | The deep fake duty covers other modalities. |
| 2 | Is it published to inform the public? | On matters of public interest, says para 4. |
| 3 | Did a human review or control the content? | First limb of the exemption in para 4. |
| 4 | Does a named person hold editorial duty? | Second limb, for the publication itself. |
| · | both limbs, or the duty stands (breakpoint) | The exemption joins them with "and". |
| 5 | Otherwise, disclose it to the reader |
Three limits are doing work in that sentence. It applies to deployers, so a provider’s duty under paragraph 2 is a different question with a different answer. It applies to text, while the first subparagraph of paragraph 4 covers deep fakes, defined in Article 3(60) as image, audio or video. And it applies to publication aimed at informing the public on matters of public interest, which is a purpose test rather than a channel test. The European Commission’s FAQ on Article 50 describes the duty in exactly those terms.
Does human review remove the Article 50 disclosure duty?
Under two conditions together, and the conjunction is the part people drop.
The exemption in paragraph 4 applies “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.” Two limbs, joined by “and”. A named responsible editor satisfies the second limb and says nothing about the first.
The Commission’s guidance puts a floor under what the first limb means, describing it as substantive examination by qualified persons together with control by a responsible editorial entity with authority over content approval. Approval implies the opportunity to withhold approval, which implies review happening before publication rather than after it.
That is a real line, and an automated publishing pipeline sits on the wrong side of it. Naming an accountable editor does not convert a system that publishes without review into one that reviews. The honest reading is that the exemption describes editorial practice, and either a publisher has that practice or it does not.
What changes for AI systems on 2 December 2026?
One narrow thing, and it is the only Article 50 date still in the future.
Article 50 applied on schedule on 2 August 2026. The Digital Omnibus postponed several AI Act deadlines, and the transparency obligations were not among them. The single concession is transitional. The Commission’s FAQ states it directly: “A limited grace period is envisaged only for AI systems placed on the market before 2 August 2026 and only as regards the marking and detection obligation for AI-generated content (Article 50(2) of the AI Act). Providers of such systems must comply with those obligations only as from 2 December 2026.”
So the deployer duties in paragraphs 3 and 4 bind now, for every system. The provider marking duty binds now for systems placed on the market from 2 August 2026, and from 2 December 2026 for everything older. If you are a provider with a system that predates the cutover, the date on your market-placement record is the fact that decides which applies.
How does a Code of Practice fit into this?
It supplies a documented route to showing compliance, without changing what is owed.
The Commission’s Code of Practice on marking and labelling AI-generated content was published on 10 June 2026 after three drafting rounds, and covers Article 50(2), (4) and (5). By 31 July 2026 roughly 190 organisations had signed it. The Commission and the AI Board confirmed it as an adequate voluntary tool for demonstrating compliance with those obligations.
The Commission is plain about what that does and does not mean: “Even though adherence to the code is voluntary, the transparency requirements under article 50 of the AI Act are legal obligations.” Signing buys a recognised method. Declining leaves the obligation intact and the method for demonstrating it up to you, in front of a market surveillance authority.
Article 50(7) also holds a reserve power. If the Commission judges a code inadequate, it may adopt an implementing act specifying common rules under the Article 98(2) procedure. The voluntary phase is the current state rather than a permanent settlement.
How does this site handle Article 50?
Directly, because the question is not hypothetical here.
This site publishes on governance and compliance through an automated pipeline. Articles are drafted, checked against automated gates and published without a human reading them first. Editorial responsibility is named and the correction record is public, so the second limb of the exemption is met. The first limb is not, and pretending otherwise would be the exact error this article is about.
Whether these articles are “published with the purpose of informing the public on matters of public interest” is genuinely arguable, and we are not going to resolve it in our own favour for tidiness. The safer and cheaper choice is to disclose regardless, which is why the methodology page states how articles are produced and the automated process is described rather than implied.
The general point for any team reading this. The exemption rewards a practice. If review happens after publication, say so, and label.
Do this
Work out which Article 50 duty lands on your product
The order matters. Each step removes a class of content from scope, so running them out of order produces a labelling policy far wider than the text supports.
Decide whether you are the provider or the deployer of each system
Article 3 turns on who develops and places a system on the market against who uses it under their own authority. A company can be both, for different systems, on the same page.
Separate machine-readable marking from human-visible disclosure
Paragraph 2 is metadata attached to output by the provider. Paragraph 4 is a statement to the person reading. They are different artefacts and different parties, and conflating them produces work nobody owes.
Split your output by modality before applying paragraph 4
The deep fake duty reaches image, audio and video. The text duty is a separate sentence with a narrower trigger, so a text pipeline and a media pipeline get different answers.
For text, test the public-interest purpose honestly
The trigger is text published with the purpose of informing the public on matters of public interest. Write down which of your surfaces meet that description and which plainly do not.
Check both limbs of the editorial exemption together
The text asks for human review or editorial control AND a natural or legal person holding editorial responsibility. An automated pipeline that names an editor satisfies only the second.
Record the date each system was placed on the market
Systems on the market before 2 August 2026 pick up the paragraph 2 marking duty from 2 December 2026. That distinction is worth having in writing before an authority asks.
Decide on the Code of Practice deliberately
Signing gives a documented route to showing adequacy. Declining is allowed, and means you carry the argument yourself. Either way, write down which you chose and why.
Frequently asked questions
- Does Article 50 mean every AI-written blog post needs a label?
- The paragraph 4 duty on deployers reaches text published with the purpose of informing the public on matters of public interest. A product description or a support macro is not that. The Commission's own guidance describes the duty in those terms rather than as a blanket labelling rule.
- Who is a provider and who is a deployer?
- Article 3 defines a provider as the party that develops an AI system and places it on the market, and a deployer as a party using an AI system under its own authority. A company that publishes output from a third-party model is usually a deployer of that system.
- Is a visible "written with AI" note the same as machine-readable marking?
- No. Paragraph 2 asks for marking in a machine-readable format that makes output detectable as artificially generated. Paragraph 4 asks for disclosure to the person reading it. One is metadata or a watermark, the other is a human-visible statement.
- Does signing the Code of Practice make you compliant?
- The Code is voluntary and the Commission describes it as an adequate tool for demonstrating compliance. The obligations in Article 50 are legal ones either way, so a non-signatory has to show adequacy to market surveillance authorities on its own terms.
- What about a satirical or fictional deep fake?
- Paragraph 4 limits the duty for evidently artistic, creative, satirical or fictional work to disclosing the existence of generated content in a manner that does not hamper the display or enjoyment of the work.