On August 2, 2026, the EU AI Act stopped being a calendar item and started being enforced. The European Commission's AI Office and national market surveillance authorities began applying the transparency rules in Article 50 that same day: chatbots have to say they are chatbots, synthetic media has to carry machine-readable marking, and anyone deploying a deepfake has to disclose it. The penalty ceiling for breaching those duties is 15 million euros or 3 percent of worldwide annual turnover, whichever is higher.
I read the coverage the way most working directors did, which is to say I looked for the sentence that tells me what changes on Monday. Most of it was written for compliance officers. Almost none of it was written for the person who actually assembles the shot.
So I went and read Article 50 itself. There is something in there that the legal briefings mention in passing and the tech coverage skips entirely, and it is the part that matters most to anyone making films.
What the EU AI Act actually requires from August 2
Two obligations sit at the center of this, and they land on two different people.
Article 50(2) is about the tool. It requires providers of AI systems that generate synthetic audio, image, video, or text to mark their outputs in a machine-readable format, detectable as artificially generated or manipulated. That is a watermarking duty, and it belongs to the company shipping the model. For generative systems already on the market before August 2, the Commission gave until December 2, 2026 to comply.
Article 50(4) is about you. It requires deployers, meaning the person or company putting the content in front of an audience, to disclose when they have generated or manipulated a deepfake. The Act defines that broadly: image, audio, or video content resembling existing persons, objects, places, entities, or events, that would falsely appear authentic. And the obligation does not depend on intent. You can have no intention of deceiving anybody and still owe the disclosure.
There is also a limit on 50(2) worth knowing, because it draws a line a lot of editors are already standing on. The marking duty does not apply where the AI system performs an assistive function for standard editing, or where it does not substantially alter the input data or its semantics. A cleanup pass is not the same object as a generated performance. The law says so explicitly.
Does the EU AI Act apply if you are not in Europe?
This is the question I got asked three times in two days, and the honest answer is that geography is the wrong axis. The Act attaches to where the output lands, not to where the edit suite sits. If the film, the ad, or the social cut reaches an audience in the European Union, you are inside the scope. I work out of São Paulo. Commercial work I have delivered over fourteen years has run in more than one market, for brands like Nestlé and Carrefour, and nobody on those jobs ever asked me which continent the color pass happened on.
The other thing worth knowing: enforcement is not only top down. Any person, including a competitor, can file a complaint with a national market surveillance authority. That is a different pressure than an audit. That is the market policing itself.
The carve-out nobody is reading
Here is the sentence. Article 50(4) says that where the content forms part of an evidently artistic, creative, satirical, fictional or analogous work or programme, the transparency obligations are "limited to disclosure of the existence of such generated or manipulated content in an appropriate manner that does not hamper the display or enjoyment of the work."
Read that again, because it is doing something unusual for a regulation.
The law is saying that a film is not a document. It is saying that when the frame is evidently fiction, the audience is not being deceived, and so the disclosure shrinks to an acknowledgment that does not damage the work. No burned-in badge across your final shot. No banner over the second act.
Brussels wrote a distinction into statute that our industry has understood since the first dissolve: the audience knows it is watching something made. A matte painting was never a lie. A miniature was never a lie. Neither is a generated element, provided the work is evidently a work.
The catch is that the exemption is anchored on the word "evidently." It protects the artistic frame, not the artist. Something presented as a document, a testimonial, a piece of reportage, or a real person saying real words does not get the lighter treatment, and it should not. That is the boundary, and it is a boundary about honesty, not about technology.
Who actually carries the burden
The anxiety I keep hearing is that labelling will mark AI work as second class, that a disclosure line is a scarlet letter, that audiences will discount anything that carries it.
I do not believe that, and not because I think audiences are generous. Because the label is not a verdict on quality. It is a statement about method, and our industry has published statements about method for a century without anyone thinking less of the film. Nobody walked out of a theater disappointed to learn there was a visual effects unit.
What the rules do change is subtler and more demanding. When the watermark duty sits with the model provider and the disclosure duty sits with you, the law has quietly named you the author of the delivery. Higgsfield, Kling, Seedance, whoever you generate with: their obligation ends at the file. Yours begins when you decide what the file is for and who sees it. You cannot outsource that to the tool, and after August 2 you cannot pretend the question is not yours.
That is the same argument I made about YouTube's monetization rules being an authorship test, and the same one underneath the Academy's disclosure debate. Three institutions, three different mechanisms, one convergent conclusion: the accountable human is the unit the system is built around.
What I am changing in my own delivery
Very little, and that is the point of saying it out loud.
I already log what was generated, with which model, at which stage, because when a client asks a question six months later I want an answer instead of a guess. That log is now most of a compliance record. The one thing I have added is a delivery note that states plainly where synthetic elements appear in the piece, written for a producer rather than a lawyer.
If you do not keep that log, start this week. Not because a regulator is coming for a São Paulo edit suite. Because the first time a client asks you to certify what is in a cut and you have to reconstruct it from memory, you will discover it is unreconstructable, and the loss will be yours before it is ever the law's.
None of this is legal advice, and a piece running to an EU audience with money attached deserves an actual lawyer. What I can tell you is what the text says and where the interesting part is.
Where this lands
The EU AI Act did not make AI filmmaking harder on August 2. It made the author visible.
Every regulation of a new medium eventually resolves into the same question, which is who is answerable for what reaches the audience. Cinema has answered that question for a hundred years with a credit block. The Act is asking for a version of the same thing, in machine-readable form, with a deadline.
The people who will struggle are the ones who were using these tools to avoid the question of authorship in the first place, who wanted the output without owning the decision. The people who will not notice are the ones who were already keeping track, already deciding, already treating the generated element as one more thing they are responsible for on a long list of things they are responsible for.
I would rather be in the second group. The tools amplify what you are willing to stand behind. Sign the work.
