What Article 50 of the EU AI Act Actually Requires Since 2 August 2026: Reading Article 50 Correctly

Since 2 August 2026, Article 50 of the EU AI Act (Regulation (EU) 2024/1689) has applied. Alongside it, one summary has circulated widely: that every piece of AI-generated text now has to be labelled. The provision is more specific than that. Its duties attach to defined uses and defined roles, and for a large share of business content the visible label does not become due. This piece sets out what Article 50 requires, who it binds, and where the text-labelling duty begins and ends.

The timing has added to the confusion. The Article 50 duties took effect on the 2 August 2026 date that had been in the calendar for months, and they arrived shortly after the Digital Omnibus had moved other parts of the AI Act back, which left the impression that everything landing on that date was new and broad. What actually applies is narrower, and reading it precisely is the difference between a proportionate compliance response and a labelling policy wider than the law requires.

What took effect on 2 August 2026, and what the Digital Omnibus left in place

The Digital Omnibus, now Regulation (EU) 2026/1744, entered into force on 27 July 2026 and moved several deadlines. It postponed the obligations for stand-alone high-risk systems under Annex III of the AI Act to 2 December 2027, and the obligations for AI embedded in already-regulated products under Annex I to 2 August 2028. Those were the parts of the Act that were hardest to operationalise on time, and those are the parts that moved.

The transparency duties in Article 50 were left in place. They are treated as comparatively low-burden and citizen-facing, and the legislator did not delay them. The position is therefore straightforward: since 2 August 2026, the Article 50 obligations apply in full, and the only concession inside Article 50 is a short grace period. Providers of generative AI systems that were already on the market before 2 August 2026 have until 2 December 2026 to bring their machine-readable marking into line. Everything else was due on the August date.

The exposure is material. Non-compliance with the Article 50 transparency obligations sits in the penalty tier of up to 15 million euros or 3 percent of total worldwide annual turnover, whichever is higher (Art. 99(4) AI Act). Enforcement runs through the national market surveillance authorities, not the AI Office.

Two roles, and they carry different duties

The first step in reading Article 50 is to fix your role before you ask what you owe. The Act draws a firm line between the provider (Anbieter) of an AI system, meaning the party that develops it or has it developed and places it on the market or puts it into service under its own name (Art. 3(3) AI Act), and the deployer (Betreiber), meaning the party that uses an AI system under its own authority in the course of a professional activity (Art. 3(4) AI Act).

For most companies reading this, and for the agencies that produce content on their behalf, the role is deployer. You take a general-purpose AI model built by someone else, you use it to help produce content, and you publish. Using a third-party model does not by itself turn you into its provider. That distinction decides which paragraphs of Article 50 point at you, because the marking duty and the labelling duty rest on different parties.

What content is caught

Article 50 contains four duties, and only some of them concern text. The first is a provider duty. Systems designed to interact directly with people, such as chatbots and voice agents, must inform the person that they are dealing with an AI system, unless that is obvious to a reasonably well-informed user (Art. 50(1) AI Act). The second is also a provider duty, and it is the marking obligation discussed below (Art. 50(2) AI Act). The third is a deployer duty attached to emotion recognition and biometric categorisation systems, which requires informing the people exposed to them (Art. 50(3) AI Act). The fourth is a deployer duty with two limbs that work differently: image, audio or video content that is a deep fake, and text published to inform the public on matters of public interest (Art. 50(4) AI Act). The sections below take the text limb first and then the deepfake limb.

What the structure does not contain is a free-standing rule that all AI-generated text must be visibly labelled. The text obligation lives inside Article 50(4), qualified by a subject-matter condition and by exemptions that a professionally produced piece will usually satisfy.

The marking duty, and how digital watermarking works

Article 50(2) requires providers of generative AI systems, including general-purpose models, to mark their synthetic output, whether audio, image, video or text, in a machine-readable format so that it can be detected as artificially generated or manipulated. The statute asks that the technical solution be “effective, interoperable, robust and reliable” as far as this is technically feasible. Two features of this duty are easy to miss. It sits with the provider, so the model vendor does the marking, and the mark is machine-readable, so it is aimed at detection systems and stays invisible to a reader glancing at a page. Article 50(5) governs the manner of any Article 50 disclosure that is presented to a natural person, requiring it to be clear and distinguishable and to meet the applicable accessibility requirements (Art. 50(5) AI Act), while the paragraph 2 mark itself remains machine-readable.

The word watermark suggests something more durable than the technology delivers. Two families are in use. A statistical or embedded watermark, of which Google’s SynthID is the best-known example, biases a model’s word choices or an image’s pixel data at generation so that a detector can later recognise the pattern, invisibly to a person. A cryptographic provenance standard, of which C2PA, published as Content Credentials and standardised as ISO/IEC 22144, is the reference, attaches signed metadata recording how a file was created and edited. Both carry limits a compliance function should understand. Text watermarks fall below the detection threshold in short passages and are degraded by paraphrasing, and cryptographic metadata is stripped the moment someone screenshots the content or a platform re-encodes an upload. Independent testing has shown invisible image watermarks removed at high rates under adversarial conditions. The Commission has channelled the detail into a voluntary Code of Practice on Transparency of AI-generated Content, adopted on 10 June 2026 with roughly 190 signatories by July; a provider that signs it can point to it as evidence of compliance, and one that does not must show equivalent means. For a deployer, the vendor’s signature status signals whether the marking being relied on will hold up.

Separately from the provider’s machine-readable mark, stylometric detectors have emerged that infer from the writing itself how likely a passage has been AI-produced, without reading any watermark. Such a tool can indicate that a text may be AI-assisted, but it does not establish whether the content underwent human review or who holds editorial responsibility, which is, however, what the disclosure duty under Article 50(4) turns on.

When AI-generated text does not have to be labelled

The text duty sits in Article 50(4) AI Act. It applies where a deployer uses an AI system to generate or manipulate text, that text is then published for the purpose of informing the public on matters of public interest, and it requires the deployer to disclose that the text was artificially generated or manipulated. Read closely, the condition has two elements that both have to be present. The text has to be published to inform the public, which points to material addressed to a general audience, and it has to concern a matter of public interest. The Commission’s guidance describes matters of public interest by reference to areas such as democratic and political debate, public administration, justice and fundamental rights, public health, safety of consumers, protection of the environment, and significant economic, financial or cultural developments. Classification is fact-specific and the list is indicative, so the workable approach is to look at the subject and the audience of each piece.

A set of worked classifications gives the practical feel of where the line runs. Several common formats fall inside the subject-matter condition. A law blog that explains a new regulation, a court decision or a reader’s rights addresses justice and fundamental rights, and is within scope. A health blog that gives information on conditions, treatments or public-health questions addresses public health, and is within scope. News posts and current-affairs reporting are the paradigm case. Published political opinion and commentary concern democratic debate, and are within scope. Essays on law, policy, finance or politics on a platform such as Substack sit within the enumerated areas. A LinkedIn post that analyses a legal or financial development for a general professional audience can fall inside the condition, because it informs the public on a legal or economic matter, though a personal career update or a networking note does not. A white paper on a legal or regulatory question of general concern can fall inside it as well, where it is published to inform and not solely to promote a product. The present article is itself an example: it informs the public about a regulatory development, so it meets the subject-matter condition, and it reaches the labelling question only through the exemption discussed in the next section.

Other formats sit outside the condition, either because the subject is not a matter of public interest or because the text is not published to inform the public. A lifestyle blog on travel, food, and similar topics does not address a matter of public interest in the sense the article uses, so it stays outside, unless a given piece crosses into health or consumer-safety advice. Product descriptions and other marketing copy are outside, because their purpose is commercial. Handouts, training materials and client memos that are circulated internally or to a defined group, without being published to the public, are outside, because the element of informing the public is not met. One-to-one correspondence and customer-service replies are outside for the same reason. Borderline items such as a handout, a white paper or a LinkedIn post fall either way on their facts, and in practice resolve through the editorial exemption discussed in the next section.

One point before leaving this section. The subject-matter condition is only the first gate. Even where it is met, a label is required only if the text was in fact generated or materially altered by an AI system and the editorial exemption does not apply. Text a person writes without AI assistance is outside Article 50(4) regardless of subject. AI-assisted text that passes through a genuine editorial process is exempt, which is the subject of the next section.

What human review and editorial control have to look like

Even for text that is published to inform the public on a matter of public interest, Article 50(4) lifts the disclosure duty where the AI-generated content has undergone human review or editorial control and a natural or legal person holds editorial responsibility for the publication. The Commission’s Guidelines on the transparency obligations, published on 6 August 2026, put content behind those words, and the content is demanding enough that a formality will not meet it.

Human review means a deliberate examination of the substance of the content by one or more natural persons with the relevant knowledge and professional judgement. Editorial control means control by a responsible editorial entity, an editor-in-chief being the model the guidelines use, with the authority to approve, alter or reject the substance of the text on substantive grounds, including fact-checking the information and verifying the reliability of the sources. Editorial responsibility means that an identifiable person or organisation carries the ultimate legal responsibility for the act of publication. The guidelines are explicit that a superficial pass does not qualify. Running a draft through a spell-checker or tidying its grammar is not human review and is not editorial control.

Organising the editorial process comes down to three fixed points. There has to be a named person with competence in the subject who examines the substance of each piece before it is published. That person has to hold real authority, meaning they can change the text or stop it going out on the merits, including on the accuracy of statements and the reliability of the sources. And the review has to be distinct in substance from the drafting, so that where an AI tool produced the draft, a person still engages with what it says. In a small setting the author and the reviewer can be one person, provided the review is a separate, deliberate step and not the act of writing itself. In an organisation, the workable form is to name an editorial owner for each channel, so responsibility for what is published sits with a person and is not diffused across a team.

Documentation is what lets you show, if a market surveillance authority asks, that the review was substantive. It does not need to be elaborate. A short record kept with each published item will do: the date, who carried out the review, that the substance and the sources were checked, and the decision to publish. Retaining the AI-assisted draft alongside the reviewed and published version gives a visible change history that evidences the examination. The purpose of the record is narrow but real. The exemption depends on the review having happened in substance, and a contemporaneous note is the ordinary way to demonstrate that afterwards.

Many of the people asking about this are not publishers in the traditional sense. A lawyer writing a blog, a physician posting on a health topic, a politician publishing commentary, or an entrepreneur writing on a policy question is both the author and the person who decides to publish. The AI Act does not require a media organisation or an editorial department. It requires that a natural or legal person hold editorial responsibility, and an individual author who publishes under their own name holds exactly that. So a lawyer who uses an AI tool to produce a first draft, then reads it properly, corrects it on the merits, and publishes it under their name meets the exemption. The exposure for this group is narrow and specific: publishing AI-drafted text under their name after no more than a glance, which is the one case the guidelines say does not qualify.

For content providers who are not publishers, a short standard operating procedure keeps the position clean.

1. Before publishing, ask two questions. Does the piece concern a matter of public interest, and was any part of the text generated or materially changed by an AI tool? If the answer to either is no, no Article 50(4) label is due, and you can publish.
2. If the answer to both is yes, read the full text yourself, or have a competent person read it, and check the facts and the sources. Change or remove anything that does not hold.
3. Record that you did so. One line kept with the piece is enough: the date, who reviewed it, that the substance and sources were checked, and that you approve publication under your name.
4. Keep the AI-assisted draft together with the published version, so the change history is visible.
5. Publish under a named person or organisation that takes responsibility for the content.
6. If all you have done is run a spell-check or skim the draft, either carry out a proper review or add the AI-generated disclosure before publishing.

Followed as a habit, this procedure resolves the labelling question for an individual professional without a label in almost every case, because the review that good practice already calls for is the review the exemption requires.

Images, audio and music: a different test, and no editorial exemption

The deployer duty in Article 50(4) has two limbs, and they do not work the same way. The text limb, set out above, turns on the public-interest condition and can be switched off by a genuine editorial process. The other limb covers image, audio and video content, and it turns on a different question: whether the content is a deep fake. The Act defines a deep fake 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 (Art. 3(60) AI Act). Where AI-generated visual or audio material meets that description, the deployer has to disclose that it was artificially generated or manipulated (Art. 50(4) AI Act).

Two consequences follow for anyone producing visual or audio content. The first concerns what is caught. Content that is evidently synthetic or stylised, which no viewer would take for an authentic photograph or recording, is not a deep fake, so the deployer disclosure duty does not attach to it. An obviously artificial illustration or an instrumental track generated by a model does not carry a deployer label. The provider-side machine-readable marking under Article 50(2) still applies, because that duty covers audio and image output as it covers text, though that is the vendor’s machine-readable mark, which a reader never sees. The line is drawn by realism. A photorealistic image that a viewer could mistake for a genuine photograph can be a deep fake even where the person shown does not exist, because the Commission applies the test by asking whether the content would appear authentic, a question a realistic invented person can meet as readily as a copy of a real individual. In music, audio that imitates the voice of a real, identifiable performer is the clearest case, while a purely instrumental generated track is generally outside the deepfake test.

The second consequence is the sharp difference from text. The image and audio limb has no human-review or editorial-control exemption. The editorial process that removes the duty for public-interest text does nothing for a deep fake. Article 50(4) gives this limb only two forms of relief: use authorised by law to detect, prevent, investigate or prosecute criminal offences, and content forming part of an evidently artistic, creative, satirical or fictional work, where the disclosure is reduced to a form that does not spoil the display or enjoyment of the work, but is still made. Neither is an editorial exemption. A newsroom with full editorial control cannot review a deep fake into compliance the way it can with text.

A worked example

Take a mid-sized company, product-neutral, that produces its own editorial content. Someone researches a topic with a general-purpose model, has it draft an outline, writes the article, then has the model revise it and apply its machine-readable watermark. A named editor decides whether it goes live. The provider-side marking under Article 50(2) is already handled by the vendor. The deployer-side text duty under Article 50(4) turns first on subject matter: product or marketing content is outside it, so no disclosure arises. If the piece addresses a matter of public interest, the editorial exemption resolves it, because a qualified person directed the text and a named editor holds responsibility for publication. Either way, no label.

Change one fact so that the model generates the published text and a human reviews it before release. For product content nothing changes. For public-interest content the answer follows from the quality of the review: a deliberate examination of the substance by a qualified person, with a named party responsible, meets the exemption, while a quick skim does not. The line turns on whether a genuine editorial process stands behind the text, whatever produced the first draft. This reasoning holds only for text. Had the same piece carried an AI-generated image of a real person or event that appeared authentic, the editorial process would not remove the deepfake disclosure duty.

Where this leaves the compliance function

Read Article 50 as a sequence. Fix your role, provider or deployer. For text, ask whether it is published to inform the public on a matter of public interest, and if so whether a genuine editorial process stands behind it. For image, audio and video, ask instead whether the content is a deep fake, and remember that editorial review will not remove that duty. Worked in that order, the labelling question resolves for most of what an organisation publishes, and the cases where a disclosure is genuinely due become visible and few.

The practical step is contained. Map your content types, textual and visual, against these tests, set down how your editorial process compares with the Commission’s definition of human review and editorial control, and record who holds editorial responsibility for each channel. An organisation that can explain on the record why each piece does or does not need a disclosure has editorial governance it can show to regulators and clients, and it has avoided disclosures the law never required. We are glad to review that mapping against the current text of Article 50 and the August guidelines before it becomes your standard.

See also our analysis on Substack:

MAYRFELD is a Frankfurt-based law firm specialising in technology and business law, advising national and international companies throughout the innovation cycle, from R&D and regulatory affairs to market entry and international expansion.