AI Act Article 50: what actually has to be labelled

Since 2 August 2026 the transparency duties in the AI Act's Article 50 have applied. The common reading is that anything AI was involved in now has to be labelled, and that reading produces two errors at once: work nobody needs to do, and three duties that genuinely are yours drowning in that work.

Article 50 has seven paragraphs. Four of them carry disclosure duties, a fifth governs how and when the disclosure is given, and the exemptions are written into the text. The line does not run where most people guess.

Who bears what in Article 50

Paragraphs 1 and 2 address the provider. Paragraph 1 requires whoever builds a system intended to interact directly with people to make sure those people are told, unless it is already obvious to a reasonably well-informed, observant and circumspect person. Paragraph 2 requires whoever builds a system generating synthetic audio, image, video or text to make sure the output is marked in a machine-readable format and detectable as artificially generated.

Paragraphs 3 and 4 address the deployer, meaning the organisation putting the system to work. Paragraph 3 requires whoever uses a system for emotion recognition or biometric categorisation to inform the people exposed to it. Paragraph 4 requires two things: that whoever uses a system to generate or manipulate a deepfake discloses that the content is artificially generated or manipulated, and that whoever publishes AI-generated text to inform the public on matters of public interest labels it.

Paragraph 5 binds both ends. The disclosure under paragraphs 1 to 4 has to be clear and distinguishable, given at the latest at the time of the first interaction or exposure, and it has to meet applicable accessibility requirements. The Commission's code of practice counts it among the obligations it covers, alongside paragraphs 2 and 4.

The split has a practical consequence that is easy to miss. An organisation writing in an AI tool is not the provider of that tool.

The marking duty everyone talks about sits with whoever built the tool, not with you writing in it.

That does not make you free. Paragraph 4 carries two duties, so yours are three of them out of two paragraphs, and they are of a different kind: they concern what you put out and who you expose to what, not whether technical marking sits in the file.

The exemption that takes most of the labelling

Even for the provider the marking duty is narrower than it looks. Article 50(2) does not apply to the extent the system performs an assistive function for standard editing, nor to the extent it does not substantially alter the input data provided by the deployer or the semantics of that data. All four paragraphs additionally carry an exemption for systems authorised by law to detect, prevent, investigate or prosecute criminal offences.

The Commission's own question page on Article 50 puts the same thing plainly: the obligation to mark does not apply when the AI system performs an assistive function for standard editing.

This is the exemption that settles most everyday cases. A tool that proofreads, shortens, adjusts the tone or suggests a headline for a text you wrote yourself alters neither the content substantially nor its meaning, and so triggers no marking duty under paragraph 2. The exemption stops there: if the text is published on a matter of public interest, paragraph 4 applies anyway. A tool that produces a finished text from a prompt does alter it substantially, and there the marking sits with whoever built the tool.

The dividing line is not whether AI was involved. It is whether the AI helped you edit, or produced the content.

Where the line runs in what you publish

The duty most likely to be yours is the second limb of paragraph 4. It reaches AI-generated or manipulated text published in order to inform the public on matters of public interest. A customer magazine, a press release on a societal question, a public authority informing residents.

The exemption is drafted so that the text need not be labelled where it has undergone human review or editorial control and someone holds responsibility for the publication. In its guidance the Commission sharpens that last part: a person has to hold the ultimate legal responsibility over the publication of the content, the human review included.

That is the sentence that makes the exemption usable, and at the same time harder to meet than it sounds. Somebody having read the text is not enough. The responsibility has to sit with someone, a natural or legal person, who answers in law for the fact that it was published.

An organisation with a responsible editor, or a communications director who formally approves publications, is outside. An organisation where the tool writes and somebody clicks publish without the responsibility being assigned is inside, and labels.

Images follow a different rule from text, and that is where most people guess wrong. The first limb of paragraph 4 reaches only image, audio and video content that constitutes a deepfake. The definition sits in Article 3(60): content resembling existing persons, objects, places, entities or events that would falsely appear to a person to be authentic or truthful. Two conditions, and both have to be met.

The second condition settles most cases, and it rules most things out. An illustration, a chart or a stylised image in a house style announces itself as an image and does not purport to show something that actually happened. It triggers no disclosure duty for you, whatever tool made it. The duty arises once the image presents something as real and could reasonably be taken for it: a person who never said it, a place that does not look like that, an event that did not happen.

Paragraph 2 remains, covering synthetic audio, image, video and text whether or not it resembles anything real. It sits with the provider, and it is about something other than a visible line under the picture: the text requires the output to be marked in a machine-readable format and detectable as artificially generated. Your disclosure under paragraph 4, by contrast, has to be given clearly and distinguishably under paragraph 5 to whoever meets the content. Two different things, both called labelling in everyday speech.

Your duty under paragraph 4 also has no transition period. It has applied since 2 August 2026. The four-month window running to 2 December that gets mentioned sits in Article 111(4), and it covers providers only, the marking duty in paragraph 2 only, and only systems that were on the market before 2 August.

There is a relaxation for deepfakes in art and satire. Where the material forms part of an evidently artistic, creative, satirical, fictional or analogous work or programme, the obligation is limited to disclosing the existence of such content in an appropriate manner that does not hamper the display or enjoyment of the work. That is a relaxation of form, not an exemption from the duty.

Two instruments in six weeks, and what they are worth

The requirements are hard to turn into technology, and the Commission has put two instruments beside them.

The code of practice on transparency of AI-generated content was finalised on 10 June 2026 and covers paragraphs 2, 4 and 5. It was drawn up in two working groups with providers of generative systems, developers of marking and detection technology, associations of the organisations that deploy them, civil society and academic experts. On 20 July the Commission additionally published guidelines on the transparency obligations for providers and deployers.

What the code is worth is written out on the Commission's own page, and it repays a careful reading. Adherence is voluntary, while the transparency requirements in Article 50 are legal obligations. The Commission and the AI Board have confirmed that the code is an adequate voluntary tool for demonstrating compliance, and signatories can rely on its measures.

It is a weaker instrument than it sounds. Being able to demonstrate compliance is not the same thing as the presumption of conformity a harmonised standard confers, and no harmonised standards under the regulation exist yet.

The objection that carries weight

Here is the pull against our own reading, and it is that the generosity of the exemptions cuts both ways.

If you rely on the provider's marking to tell what is AI-generated, you will find gaps. Article 111(4), added by the omnibus Regulation (EU) 2026/1744, gives providers whose generative systems were on the market before 2 August 2026 until 2 December that year to meet the marking duty in Article 50(2). Until then the absence of a mark tells you nothing about those systems. If the system reached the market later, the duty bites now.

And the code had its legal framing changed after it was written. The omnibus entered into force on 27 July 2026 and replaced Article 50(7) in full. The Commission's power to approve a code of practice by implementing act was removed and replaced with an assessment of whether applying the code is sufficient. The instrument meant to tell you how to meet the requirement had its terms rewritten barely seven weeks after it was finalised.

The three questions that settle it

Two organisations using the same tool can get different answers, and the difference does not lie in the tool.

Are you the provider or the deployer of this particular system? If you built it, or put your name on it and placed it on the market, paragraphs 1 and 2 are yours. If you use somebody else's tool they are not. Most organisations are deployers in everything they run and providers in nothing.

Is the AI assisting your editing, or producing the content? Proofreading, shortening and tone adjustment of something you wrote alter neither the content substantially nor its meaning. A finished text from a prompt does. The answer settles whether a marking duty arises at all and where it sits, which is what you need when you check what your vendor owes you.

Does someone hold the ultimate responsibility for what you publish? That is the question that settles whether your public-interest text needs labelling, and it is answered inside your own organisation rather than in the regulation.

An organisation that can answer those three knows which of its systems carry a duty. One that cannot does not know which systems exist, and then the inventory is the work rather than the labelling.