The EU AI Act and AI-Generated Video: What Changes on August 2, 2026 for Brands and Agencies

Updated July 24, 2026

On August 2, 2026, the transparency obligations of Article 50 of the AI Act — the European regulation on artificial intelligence (Regulation (EU) 2024/1689) — become applicable. For anyone producing or commissioning video, the core point is this: AI-generated or AI-manipulated content that falls under the regulation's definition of a deepfake must carry a perceivable disclosure.

Read that carefully, because the wording is where everything happens: not everything the AI touches needs a label. But what does need one, needs to be done properly — and figuring out where the line sits is an assessment worth making before you go live, not after.

On July 20 the European Commission published its final guidelines on how the rules apply, so the operational picture is now clear. No need to panic — you need a method. And method happens to be my trade.

What Article 50 actually says

Article 50 sets out two distinct obligations, for two different players.

The first concerns the people who build the tools. Providers of generative AI systems must mark synthetic outputs in a machine-readable format. Solutions already in use across the market include invisible watermarks (such as Google's SynthID), provenance metadata and cryptographic methods based on content provenance standards. The regulation, however, does not mandate any single technology: it requires solutions that are effective, interoperable, robust and reliable, as far as technically feasible. In a standard production workflow, this obligation sits with the system provider — not with the production that simply uses the tool.

The second concerns the people who use those tools. When a piece of content qualifies as a deepfake under the regulation, whoever deploys it must disclose its artificial nature in a way people can perceive. (There is also a separate obligation covering AI-generated text on matters of public interest published without adequate editorial control — but that is not our field, and I won't cover it here.)

And here the definition matters. The guidelines set out three cumulative criteria for content to qualify as a deepfake: it must bear a high degree of resemblance to the subject it simulates; it must depict something that exists, could plausibly exist or could have existed; and it must be capable of falsely appearing authentic or truthful to the viewer. That third criterion is not abstract: it also depends on the context of publication, the message of the content and the expectations of the audience.

So the correct chain is not "AI video, therefore label." It is: AI-generated or manipulated video → assessment of the three criteria and the context → possible qualification as a deepfake → proportionate form of disclosure.

"So is my ad a deepfake?"

It's the question everyone is asking right now, and it's the right one. The answer is not automatic.

If your video contains photorealistic people, places or situations that a viewer could reasonably interpret as authentic footage, the content may fall under the deepfake definition. But realism alone is not enough: the subject depicted, the message, the context of distribution and the audience's expectations all count. That's why the assessment has to be made project by project.

The guidelines offer an example anyone in post production will appreciate: generated backdrops, special effects and technical work embedded in a normal film production may not falsely appear authentic, precisely because the audience doesn't read them as documentation of reality. Twenty years of VFX taught me the same thing from the other side: there's a pact between the viewer and the work, and that pact is part of the context.

There is also a specific regime for deepfakes that form part of evidently artistic, creative, satirical or fictional works: the disclosure is still required, but it can take a form that doesn't get in the way of experiencing the work — typically the end credits, the description, the project page.

And advertising? It sits in the middle: a commercial is a creative work, but it's also commercial communication aimed at the public. There is no single graphic solution that fits every project. The prudent approach is to plan a clear but proportionate disclosure, weighing the format, the context of distribution and the creative nature of the work. With one fixed point: as a general rule, the disclosure must be perceivable at the latest at the viewer's first exposure. Less intrusive options — such as a note in the end credits — are reserved for cases where the creative-works regime genuinely applies.

Who is responsible

The regulation distinguishes between providers — those who develop and supply the system — and deployers, meaning the person or organisation using it under their own authority. In a video production, the deployer might be the agency, the production company, the professional or the client, depending on who actually controls how the tools are used. That's why responsibility, traceability and the form of disclosure should also be defined contractually — before the project starts, not at delivery.

One misconception worth clearing up straight away: the tool's automatic watermark does not cover you. Machine-readable marking and perceivable disclosure are two separate obligations. And there's a detail few have noticed: for systems already placed on the market before August 2, 2026, Regulation (EU) 2026/1744 gives providers until December 2, 2026 to bring their technical marking into line — but the disclosure obligation for deepfakes remains applicable from August 2. Even if your tool doesn't mark its outputs yet, the assessment is still yours to make.

Violations of Article 50 can carry fines of up to €15 million or, for companies, up to 3% of total worldwide annual turnover, whichever is higher. For SMEs, proportionate criteria and caps apply, including the lower of the absolute amount and the percentage. But for anyone working seriously, the fine is not the point: platforms already have their AI labelling systems live, and the more structured clients will start asking for compliance guarantees directly in the brief.

What does NOT change

Three fixed points, to clear away the unnecessary fears.

No retroactive obligation for content generated before August 2, 2026. The Commission encourages voluntary disclosure where possible and appropriate, but it is not required.

Assistive AI is not the target of the rule. Corrections, cleanup, standard retouching: ordinary editing remains ordinary editing. The scope covers content that is generated, or substantially manipulated, in a way that can appear authentic.

And above all: Article 50 does not prohibit producing video with AI. It introduces transparency obligations in defined cases. The rules that already existed — privacy, image rights, intellectual property, advertising law — naturally remain in force; the AI Act does not make them disappear.

How I handle it in production

I'm speaking here as someone who produces, not as a lawyer. Twenty years in post production have taught me that new rules become manageable when the work has a method. These are the five habits that help set up an AI project in a way that is compliant, documentable and verifiable — without sacrificing a millimetre of the creative idea.

1. Transparency is decided in the brief — and in the contract. Who the deployer is, who answers for the disclosure, where it goes, what form it takes, how it lives alongside the creative direction. Decide it afterwards and you're bolting it on; decide it upfront and you're designing it in.
2. I track every shot. Which shots are generated, with which model, in which version. A production log that, at the end of the job, answers the client's legal team in five minutes: "what's AI and what isn't?". It looks like paperwork; it's precision — the same precision you need to manage revisions.
3. I choose tools that mark their outputs, and I preserve provenance information where possible. The post production pipeline matters: re-compressions, transcodes and exports can compromise metadata or other forms of technical marking. It's not always possible to preserve them fully, but verifying how your pipeline behaves is part of working responsibly — and it's the kind of attention that separates a supplier from a partner.
4. I coordinate the in-video disclosure with platform labels. The native systems on YouTube, Meta or TikTok are an additional layer of transparency, not necessarily a substitute for a disclosure embedded in the content itself. As a matter of operational prudence, I design the disclosure so it still makes sense when the video is downloaded, embedded or reshared elsewhere.
5. On hybrid projects, I map the perimeter. Real footage plus generated shots is the most common scenario in actual productions — it's the territory I work in every day, and it's where shot-by-shot assessment matters most. The disclosure has to reflect the reality of the project: neither hiding the AI nor inflating it.

It's the same approach I describe in How I choose between AI and traditional VFX for every shot: the technical decision and the transparency decision are born at the same moment, in front of the same storyboard. And on projects like Doppelganger, where AI was part of the declared concept from day one, transparency took nothing away — it added credibility.

The real point

Transparency is not a tax on creativity. Audiences react badly to AI that was hidden and discovered later — not to AI that was declared well. A carefully designed label communicates one precise thing: there is someone behind this work who knows what they're doing.

Those who work with a method were, in a sense, ready before it became law. Shot-level traceability, care for metadata, consistency between content and platforms: these don't come from the AI Act. They come from how you work when you work properly.

One necessary note: I am not a lawyer, and specific cases call for legal counsel. What I can guarantee is the production side — setting up an AI video project so it's prepared from the outset to meet its transparency obligations, without the idea paying the price.

Got an AI video project about to start, and want it set up right from day one? Let's talk. Or take a look at how I work on the AI Video Production page.

Quick answers

Is every photorealistic AI-generated video a deepfake?

No, not automatically. What has to be assessed is the resemblance to existing or plausible subjects, the risk of the content appearing authentic, and the context in which it is shown. All three criteria must hold together.

Do I need to label AI videos generated before August 2, 2026?

No. There is no retroactive obligation for content generated before that date. The Commission encourages voluntary disclosure where possible and appropriate.

Does a deliberately stylised animated ad need a label?

If no one could reasonably mistake it for authentic footage, it is unlikely to meet the three deepfake criteria, and the Article 50 obligation should not apply. If the content does constitute a deepfake embedded in a creative work, the disclosure can take a form that doesn't interfere with the work — for example in the credits or the description.

What if I only use AI for cleanup, retouching or rotoscoping?

Assistive editing is not the target of the rule. The scope covers content that is generated, or substantially manipulated, in a way that can appear authentic.

What are the penalties?

Up to €15 million or, for companies, up to 3% of total worldwide annual turnover. For SMEs, proportionate criteria and caps apply, including the lower of the absolute amount and the percentage. In practice, the pressure will come from platforms and clients' legal departments before it comes from the authorities.

For edge cases, the most immediate reference is the Commission's official FAQ, which is updated regularly.