Article 50 touches a very large number of organisations, and its deadline has already passed: it falls under no postponement and has applied since 2 August 2026. It is not the first: the prohibitions in Article 5 and the AI literacy obligation have applied since 2 February 2025, apart from the two prohibitions added by the omnibus and the two paragraphs construing them, and the penalty regime since 2 August 2025.
This page deals with the law: which obligation, on whom, since when, with which exemptions. Choosing a technical marking solution is a separate subject, covered on the page about marking AI content.
Paragraphs 1 and 2 fall on the provider. Paragraphs 3 and 4 fall on the deployer. That is the key to the whole article, and it is where the most frequent error sits: a company using a tool under its own authority does not have to mark the system’s outputs; it has to disclose certain uses.
What does Article 50 of the EU AI Act require?
Four transparency obligations, on four different situations — and the count is not in the text. Article 50 numbers nothing: it is the Commission’s draft guidelines, not yet adopted, that read it as carrying four. The grid is convenient, provided you know that paragraph 4 carries two of them, on different regimes.
| Obligation | Who carries it | Basis |
|---|---|---|
| Inform the person that they are interacting with an AI system | Provider | §1 |
| Mark synthetic content in a machine-readable format and make it detectable as artificially generated or manipulated | Provider | §2 |
| Inform people exposed to an emotion recognition or biometric categorisation system of the operation of the system and process their personal data in accordance with Regulations (EU) 2016/679 and (EU) 2018/1725 and Directive (EU) 2016/680, as applicable | Deployer | §3 |
| Disclose that a deep fake has been artificially generated or manipulated | Deployer | §4, 1st subpara. |
| Disclose that text published to inform the public on matters of public interest has been artificially generated or manipulated | Deployer | §4, 2nd subpara. |
A deep fake is defined at Article 3(60) as “AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful”.
What exemptions exist, and to which obligations?
Nine exemptions across all four paragraphs, and they do not overlap. Paragraph 1 carries two, contextual obviousness and authorised law-enforcement use. Paragraph 2 carries three. Paragraph 3 carries one, narrower. Paragraph 4 carries three, plus a reduction of the duty for artistic works. Being exempt from one obligation releases you from no other — this is where most real answers are decided.
Paragraph 1, two exemptions. Contextual obviousness: the information is not due where this “is obvious from the point of view of a natural person who is reasonably well-informed, observant and circumspect, taking into account the circumstances and the context of use”. And law-enforcement use authorised by law, “subject to appropriate safeguards for the rights and freedoms of third parties”, itself subject to a counter-exception where the systems “are available for the public to report a criminal offence”.
Paragraph 2, three exemptions. An assistive function for standard editing; not substantially altering “the input data provided by the deployer or the semantics thereof”; and authorisation by law for law-enforcement purposes.
Paragraph 3, a narrower exception, and it is a trap. Paragraph 3 covers use permitted by law “to detect, prevent or investigate criminal offences”. Paragraphs 1, 2 and 4 add “or prosecute”. Prosecution is therefore not covered at paragraph 3. Never line the four law-enforcement exceptions up in one sentence.
Paragraph 4, two distinct regimes. For deep fakes, a law-enforcement exception applies first. Then, where the content “forms part of an evidently artistic, creative, satirical, fictional or analogous work or programme”, the obligation is not lifted: it 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 lighter modality, not an exemption.
For text of public interest, the second subparagraph carries two alternative exceptions. The first is law-enforcement use authorised by law. The second is editorial, and it is cumulative: the content must have “undergone a process of human review or editorial control” and “a natural or legal person holds editorial responsibility for the publication of the content”. One without the other is not enough.
Is a model provider caught by Article 50?
No, and this is the costliest misreading of the article. Paragraphs 1 to 4 all address AI systems. A provider of a model, open-source or not, is not made subject to Article 50 merely because its model generates synthetic content: its own obligations sit in Chapter V. The reservation bears on systems released under a free licence, which remain subject to Article 50.
An important reservation on open-source licences. Article 2(12) excludes from the Regulation systems released under free and open-source licences, unless “they are placed on the market or put into service as high-risk AI systems or as an AI system that falls under Article 5 or 50”. An open system that falls within Article 50 therefore remains fully subject to it.
When and in what form must you disclose?
The moment. The information must be provided “in a clear and distinguishable manner at the latest at the time of the first interaction or exposure”, and must “conform to the applicable accessibility requirements”.
No substitution. Paragraph 6 says it expressly: paragraphs 1 to 4 “shall not affect the requirements and obligations set out in Chapter III, and shall be without prejudice to other transparency obligations laid down in Union or national law for deployers of AI systems”. Marking your content settles neither the GDPR, nor high risk, nor sectoral obligations.
And a deployer cannot lean on the provider. The draft guidelines on Article 50 (not adopted, and in any event non-binding) state that “deployers cannot rely on the machine-readable marking embedded in the content by the provider under Article 50(2) AI Act, since those markings are not immediately clear and distinguishable for the natural persons exposed to the deep fake content”. The two obligations are distinct, and so are their audiences: the machine on one side, the human on the other.
What happens on 2 December 2026?
Providers of systems generating synthetic content placed on the market before 2 August 2026 must comply with the whole of paragraph 2 — machine-readable marking and detectability, two obligations, not one. This is the only short deadline still ahead for the companies concerned, and it is routinely halved in the telling.
The basis is Article 111(4). Saying “the marking catch-up” amounts to asserting that detectability was already due on 2 August 2026 for those systems. It was not.
Three limits not to lose: the catch-up binds neither deployers, nor systems that do not generate synthetic content, nor the obligations in paragraphs 1, 3 and 4, which have been enforceable since 2 August 2026 with no extra time.
One gap worth knowing. Article 111(4) covers systems placed on the market before 2 August 2026; it says nothing about putting into service. The draft guidelines extend the benefit to systems “placed on the market or put into service before 2 August 2026”. That document is not adopted and cannot widen the text: a system developed and put into service internally, never placed on the market, does not fall within the letter of Article 111(4).
On earlier content. The draft guidelines consider that content generated or manipulated before 2 August 2026 does not have to be marked or labelled retroactively, but that text generated before that date and published on or after it must be labelled. That is the nuance that will catch the most people out.
What are the Article 50 guidelines worth?
The Commission has prepared guidelines on Article 50. Mind their status: communication C(2026) 5054 final of 20 July 2026 approved only the content of a draft. Formal adoption is expressly conditional on all language versions being available, and the document exists in English only. It is therefore not yet applicable.
And adoption will not make it binding. The document says so itself: “These Guidelines are non-binding. Any authoritative interpretation of the AI Act may ultimately only be given by the Court of Justice of the European Union (‘CJEU’).”
One point worth knowing for smaller providers: technical feasibility is described there as “an objective notion that is not dependent on the specific resources and capabilities of individual providers”. In other words, limited means are not an argument for insufficient marking.
The Regulation, for its part, brings three objective factors into the balance, and not the provider’s own resources: the technical solution must be as effective, interoperable, robust and reliable as is technically feasible, “taking into account the specificities and limitations of various types of content, the costs of implementation and the generally acknowledged state of the art” (Article 50(2)).
What is the penalty for breaching Article 50?
Article 99(4)(g) expressly covers “transparency obligations for providers and deployers pursuant to Article 50”. Ceiling: EUR 15 000 000 or 3 % of total worldwide annual turnover, whichever is higher (with the reverse rule, the lower figure, for SMEs and for small mid-cap companies).
Three questions to place yourself
Are we provider or deployer, system by system? If you place a system on the market under your own name, you carry paragraphs 1 and 2. If you use it under your own authority, you carry paragraphs 3 and 4. Many organisations are both, on different systems.
Do our systems generate synthetic content, and when were they placed on the market? That date decides whether the 2 December 2026 catch-up applies.
Do we publish AI-generated or AI-assisted informational text? If so, two exceptions are open: law-enforcement use authorised by law, and the editorial exception, the latter being cumulative, so both branches have to be demonstrable.
What this page rests on
Article 50, Article 111(4), Article 2(12), Article 3(60) and Article 99(4) of Regulation (EU) 2024/1689 in the consolidated version of 27 July 2026, as amended by Regulation (EU) 2026/1744, the authentic text remaining the one published in the Official Journal. The Commission’s guidelines on Article 50, whose draft content was approved by communication C(2026) 5054 final of 20 July 2026: not formally adopted, English only, non-binding, and quoted here as they stand.
Sources. Regulation (EU) 2024/1689, consolidated text of 27 July 2026 and Regulation (EU) 2026/1744, both consulted on 27 August 2026. The Commission’s page on the guidelines on the transparency obligations in Article 50, consulted on the same date.
Frequently asked questions
Since when has Article 50 applied?
Since 2 August 2026, and it was not postponed. It sits in Chapter IV, which the third paragraph of Article 113 does not list among the exceptions, so it falls under the general date of application. Regulation (EU) 2026/1744 amended only its paragraph 7, on codes of practice.
Who has to mark AI-generated content?
The provider of the system, not the deployer. Paragraph 2 requires providers of systems generating synthetic content to ensure the outputs are marked in a machine-readable format and detectable as artificially generated or manipulated: two distinct obligations, and meeting one without the other is not enough. The deployer’s duties are elsewhere: disclosing a deep fake, and disclosing AI-generated text published to inform the public on matters of public interest. Those are the two subparagraphs of paragraph 4. Paragraph 3 carries a different deployer duty: informing the persons exposed to an emotion recognition or biometric categorisation system.
Is a model provider subject to Article 50?
No, and this is the costliest misreading on the subject. Paragraphs 1 to 4 all refer to AI systems, never to models. A model provider, open-source or not, is not made subject to Article 50 merely because its model generates synthetic content. Its own obligations sit in Chapter V.
What happens on 2 December 2026?
It is a catch-up deadline set by Article 111(4). It binds only providers of systems generating synthetic content placed on the market before 2 August 2026, and it covers Article 50(2) in full: machine-readable marking and detectability. Saying it concerns only marking amounts to saying detectability was already due for those systems: it was not.
This content is a technical and regulatory information resource. It does not constitute legal advice.