The duty lands on the deployer
If an agency generates a video of a recognisable person and a brand publishes it, the labelling duty belongs to the agency and the brand, not to the company that built the generator. Article 50(4) binds deployers, the parties that use an AI system under their own authority in a professional capacity, and for commercial synthetic media that is the marketing team, the production partner or the publisher.
The vendor owes a different thing. Its duty under Article 50(2) is to mark output in a machine-readable format, discharged inside the file and aimed at detection tooling. Nothing about it produces a sentence the audience can read. Article 50 is where the roles reverse, as the provider and deployer page sets out: 50(1) and 50(2) are provider duties, 50(3) and 50(4) are deployer duties.
The date makes this a decision for this week. Article 50(4) applies from 2 August 2026, days away. The Digital Omnibus, Regulation (EU) 2026/1744, in force since 27 July 2026, moved several deadlines, including the high-risk regime and the 50(2) marking date for systems already on the market. It did not move 50(4), so anything scheduled to publish in August is in scope on the day it goes live. The page on what lands on 2 August 2026 sets out the rest of that date.
Where an agency operates the tool and a client publishes the output, the regulation does not allocate between them. Settle in the contract who does the labelling and who checks it, but treat that as an allocation of work and cost. An indemnity does not move a statutory duty off the party that owes it.
What counts as a deepfake, and what does not
The definition sits in Article 3(60): 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. Two conditions, both required.
- Resemblance to something that exists. Not only faces. The text names objects, places, entities and events, so a synthetic shot of a real building or a real protest is caught with nobody in frame.
- It would falsely appear authentic or truthful. The content has to be capable of being taken for the real thing by a person looking at it.
The second condition is why "we used AI, therefore we label" overshoots. A flat vector illustration, a stylised render, a plainly impossible scene: none would be mistaken for a record of something that happened, so none is a deepfake, however synthetic. The first condition is why the duty reaches further than teams expect in the other direction, because a photorealistic image of a high street that never looked like that meets both tests with nobody in the frame. Note the scope too: Article 3(60) covers image, audio and video only, and text falls to the second limb of Article 50(4).
The hard case is the fully synthetic person.A generated model corresponding to no real individual fails the resemblance limb on the face of the text. Two things still bite. The provider's Article 50(2) marking duty turns on synthetic output rather than on resemblance, so it is unaffected. And the setting can supply the resemblance the face does not: place that invented model at a real, identifiable event and the composite is arguably back inside Article 3(60).
Marking is for machines, disclosure is for people
The most common failure is a team that has done one duty and believes it has done both. The provider must mark synthetic audio, image, video or text in a machine-readable format detectable as artificially generated or manipulated, using technical solutions that are effective, interoperable, robust and reliable as far as technically feasible. Provenance manifests and invisible watermarks live here. The voluntary Code of Practice on transparency of AI-generated content, published in final form on 10 June 2026, is the likely de facto route to showing this was done properly, and it asks for at least two layers of marking where one technique cannot meet the four statutory criteria. Art 50(2)
The deployer must disclose to the people who see the content. That disclosure is human-visible by definition, and no watermark satisfies it, because a person scrolling a feed cannot run a detector. Art 50(4)
The exceptions do not cross over either. Article 50(2) excepts systems performing an assistive function for standard editing, or not substantially altering the input data or the semantics thereof, and that exception is drafted for the marking duty alone. If a tool is lawfully unmarked on that basis and the output still resembles a real person or place and would appear authentic, the deployer discloses anyway.
The gap between those dates produces a result that reads as a contradiction and is not one. Between 2 August and 2 December 2026 you can lawfully deploy a generative tool whose output carries no machine-readable mark, and still owe the full human-visible disclosure on the content you publish with it. The grace period is the provider's, not yours.
| Content | Provider marking, Art 50(2) | Deployer disclosure, Art 50(4) |
|---|---|---|
| Synthetic video of a named executive reading a script | Yes | Yes, both limbs of Art 3(60) are met |
| Cloned voice of a real person in a radio spot | Yes | Yes, audio is inside Art 3(60) |
| Photorealistic image of a real place that never looked that way | Yes | Yes, places count, no person needed |
| Stylised illustration or obviously synthetic render | Yes | Usually no, it would not appear authentic |
| AI-drafted news piece on a matter of public interest | Yes, text is inside 50(2) | Yes, unless the editorial exception is met |
| AI-drafted product description or internal memo | Yes, text is inside 50(2) | No, not published to inform the public |
| Crop, colour correction or noise reduction on a real photograph | Excepted as assistive standard editing | No, the photograph is authentic |
The carve-outs, and how far they actually go
Article 50(4) carries a carve-out for creative work and an exception for text. Both are read too generously.
The creative carve-out limits the form. Where the content is evidently artistic, creative, satirical or fictional, or analogous, the disclosure is limited to a form that does not hamper the display or enjoyment of the work. The obligation is narrowed in its manner, not switched off. A film or a piece of satire still discloses, it simply does not have to burn a banner across the face of the work: end credits, a caption in the description or an unobtrusive corner mark all respect the limit.
The word carrying the weight is "evidently", judged from the side of the audience rather than from the brief. A commercial with a comic tone is not evidently satire. If a reasonable viewer could take the content as a record of something that happened, the carve-out is the wrong thing to rely on.
The text limb is narrower than it looks, and its exception is stricter. The second limb of Article 50(4) reaches AI-generated or manipulated text published with the purpose of informing the public on matters of public interest. Both filters have to be passed before the duty exists at all, and most commercial copy passes neither: a landing page sells rather than informs, and a product range is not a matter of public interest. Coverage of an election, a public health question or a planning decision is.
Where the limb does bite, the disclosure falls away only if the content underwent a process of human review or editorial control and a natural or legal person holds editorial responsibility for the publication. Both are required. A rubber stamp does not satisfy the first, because the text asks for a process of review or control rather than an approval event, and a workflow where nobody can be named as answering for the output does not satisfy the second. If you rely on this exception, record who holds editorial responsibility, in advance, by name or by role.
Where the label goes, surface by surface
Article 50(5) sets the manner for both 50(1) and 50(4): the information must be provided clearly and distinguishably at the latest at the time of the first interaction or exposure, and must conform to applicable accessibility requirements. First exposure is the whole test, so the question on every surface is where exposure actually begins for the person.
- Social video.Exposure begins on autoplay in the feed, muted, with the caption truncated after a line, so a caption-only disclosure arrives after first exposure and often never. Put a legible label in the first frames and keep it on screen. The platform's AI-content flag is worth setting too, but it renders in the platform's own interface and generally does not survive a download and a reshare, so treat it as a supplement.
- Display advertising. Exposure is the impression, not the click, so a note on the landing page is too late. The disclosure has to sit inside the creative and stay legible at the smallest placement it gets trafficked to. A label readable in the master file and illegible at 300 by 250 has not been provided clearly and distinguishably.
- Synthetic voiceover. On an audio-only surface it has to be audible, at the start, before the content it qualifies. On a video with a cloned voice, on-screen text does not reach a listener with the screen off and a spoken line does not reach a deaf viewer, so carry it in both channels.
- Articles and long-form pages. Put it with the byline or beside the asset it describes, not in a page footer. Where a synthetic image or clip is embedded, the caption and the alt text should carry the same words, because accessibility is part of the Article 50(5) duty and a label that exists only as pixels reaches no screen-reader user.
The regulation prescribes no wording. It fixes only what has to be conveyed, that the content has been artificially generated or manipulated, and short and specific works: "AI-generated video", "This voice is synthetic". "Enhanced with AI" is weaker, because a viewer can read it as describing a tool used on real footage rather than as saying the footage is not real.
Penalties, and the harder line above Article 50
Article 50 breaches sit in the middle Article 99 tier: 15,000,000 euro or 3 percent of total worldwide annual turnover, whichever is higher. Under Art 99(6) SMEs and start-ups pay the lower of the two instead. The penalty provisions have been enforceable since 2 August 2025, so the tier is in place before the duty attaches, and the penalties page has all three.
The AI Office begins enforcing from 2 August 2026, and national market surveillance authorities designated under Article 70 handle most of the rest. Designation has been uneven: member states had to name their national competent authorities by 2 August 2025 and only 8 of the 27 did so on time, which the national authorities tracker keeps current. Expect inconsistent early enforcement, not an absence of it.
One category is not a labelling problem at all. From 2 December 2026 the Digital Omnibus adds a ninth prohibited practice under Article 5, covering AI used to generate child sexual abuse material or non-consensual intimate imagery, binding providers and deployers alike. That is the top tier, 35,000,000 euro or 7 percent of worldwide turnover, and no disclosure cures it. Consent and subject-matter checks belong upstream, in the production process.
Four things worth doing before 2 August 2026:
- Inventory the synthetic assets already scheduled or in market. An asset built in June and still running in August is in scope on the application date.
- Run each one through the two limbs of Article 3(60) and record the answer. "No existing person, place or event is resembled" is a defensible finding when it is written down at the time.
- Ask each generative vendor in writing what it marks, in what format and from what date, and whether it treats itself as inside the 2 December 2026 transitional period. The answer changes your evidence position, not your duty.
- Fix the placement per surface against the first-exposure test, then check it at the smallest rendered size, with sound off and with a screen reader. The remaining limbs are mapped on the Article 50 transparency page.