The line through the frame

On July 20, the European Commission published its final Guidelines on Article 50 of the AI Act. Fifty-one pages. Thirteen days before enforcement begins on August 2. The regulation that has been counting down across the last forty articles of this series finally produced its instruction manual, and the instruction manual drew a line through the middle of the frame.

Backgrounds on one side. Performers on the other.

The Guidelines state, in plain language, that "AI-generated background scenes, special effects and technical pre- and post-processing are unlikely to make content falsely appear authentic." Those are not deep fakes under the regulation. No mandatory label. No disclosure obligation beyond what the editorial exemption already provides.

Then the other half: "essential elements that impact audience perception such as fully AI-generated actors, digital replicas of real or deceased actors, de-aging and simulated performances are likely to" make content appear authentic. Those are deep fakes. Full labeling. Full disclosure. Penalties up to fifteen million euros or three percent of global turnover.

The regulation just separated the set from the performer. The world around the actor from the actor in the world.

The hybrid got paperwork

Jon Erwin's Young Washington grossed $40 million this month. Real actors on a soundstage. AI-generated environments projected on LED walls behind them. Ben Kingsley performed. The model built ancient cities around him. Under the Guidelines, the backgrounds are not deep fakes. The performance is real. The production sits cleanly on the non-deep-fake side of the line because the person in the frame was in the room.

Rahi Anil Barve shot two actors on an iPhone for $360 and generated everything around them. Same principle. Same side of the line. Barve kept the people and gave the model the world. The regulation agrees with his instinct.

Soderbergh used ten minutes of AI imagery in a ninety-minute documentary. The AI portions contain no human performers. No faces, no likenesses, no simulated performances. Thematic surrealism. Under the creative works exception, those sequences require reduced disclosure that does not "hamper the display or enjoyment of the work." Credits, companion materials, not a watermark stamped across the dream sequence.

Three filmmakers. Three different projects. All on the simple side of the paperwork. Not because they planned for the regulation. Because they made the same creative decision for creative reasons: the person stays real, the world gets generated.

The other side

Val Kilmer's posthumous AI performance in As Deep as the Grave lands squarely on the deep fake side. Digital replica of a real person. Simulated performance. Full labeling obligation.

Tilly Norwood, the fully AI-generated performer, meets the definition through a clause the final Guidelines expanded from the draft: "persons" explicitly covers "digital replicas of real people, realistic AI-generated avatars or personas, and personal characteristics or expressions such as a person's image, voice, behaviour and performances." A photorealistic portrait of an invented person still qualifies. Tilly is invented. Tilly is photorealistic. Tilly is a deep fake under the regulation, with or without a body to trace her back to.

Gene Wilder's reconstructed voice narrating a Netflix game show. A deceased performer's vocal characteristics reproduced and deployed. Deep fake.

The Guidelines closed a gap many expected to remain open. The draft version left room to argue that a fully fictional AI performer with no real-world source might escape the definition. The final version sealed it. If the output "can plausibly exist," it qualifies. Plausibility, not provenance, is the trigger.

Thirteen days

Fifty-one pages of guidance, published thirteen days before enforcement. The Guidelines are non-binding, technically. Only the Court of Justice of the European Union can give an authoritative interpretation. But national market surveillance authorities are expected to follow them. The Code of Practice on Transparency of AI-Generated Content has been formally assessed as adequate. It is the only EU-wide recognized compliance framework.

Thirteen days is not enough time to redesign a production pipeline. It is barely enough time to read the document twice. The companies that have been preparing since the regulation was adopted in March 2024 are ready. The companies that assumed enforcement would be pushed back, the way the watermarking deadline was pushed to December 2, are discovering that transparency obligations stayed on schedule.

One provision will catch producers by surprise: posting content on the "globally accessible internet" can trigger EU obligations even for deployers outside Europe. The Guidelines do not require targeting the EU market specifically. Global accessibility is enough. A filmmaker in Los Angeles who posts an AI film on YouTube with a deep fake element is, under this reading, subject to Article 50(4) labeling requirements because the content is accessible in the EU.

The extraterritorial reach matches GDPR's pattern. The enforcement lag will also match GDPR's pattern. The first cases will take years. The compliance anxiety will start next Saturday.

What the line validates

The regulation did not invent the distinction between background and performer. Filmmakers arrived at it independently through craft.

Erwin arrived at it because LED walls with real cameras preserve the relationship between actor, director, and lens. Barve arrived at it because AI lip sync was unconvincing, so he removed dialogue and kept real faces. Soderbergh arrived at it because he asked what AI was good at and the answer was hallucinatory imagery, not human performance. Mendiboure arrived at it because calm dialogue scenes broke the 180-degree rule while action sequences hid the seams.

None of them read the regulation first. All of them, through different experiments in different countries, concluded that the person belongs to the filmmaker and the world belongs to the model. The regulation now agrees. Not because it studied their work. Because both the craft and the law converge on the same observable truth: the audience treats a generated background as a special effect and a generated person as a deception.

That asymmetry lives in the viewer, not the technology. A generated sunset behind an actor does not change the audience's relationship with the actor. A generated actor changes the audience's relationship with everything.

The creative works exception has an asterisk

The draft Guidelines offered filmmakers a relatively generous creative works exception: reduced disclosure for content forming part of "evidently artistic, creative, satirical, fictional or analogous works." The final version kept this but added a counter-exception that narrows it considerably. When content combines informative and creative characteristics, "the informative character always prevails." Standard labeling applies.

A documentary that uses AI-generated imagery to illustrate historical events is informative first, creative second. A fictional film is creative. An advertisement is neither. The regulation leaves no room for the blurred categories that most commercial content occupies.

For filmmakers working in pure fiction, the exception holds. For documentary filmmakers, the exception is gone before it arrives. Soderbergh's Lennon documentary will need to navigate this carefully. The AI sequences are thematic, not informative. They illustrate philosophical conversation, not historical events. That distinction may be clear to a filmmaker. Whether it is clear to a market surveillance authority in Brussels is a question that will be answered after August 2.

The holistic test

The final Guidelines replaced the draft's simpler deep fake assessment with a "holistic" test. Whether content "would falsely appear to a person to be authentic" now depends on resemblance, substantive message, deployment context, the environment in which the content appears, and the composition and expectations of the foreseeable audience. Where the audience does not expect content to be authentic, the content may fall outside the definition entirely.

A filmmaker screening AI-generated work at a festival with full disclosure is presenting to an audience that expects synthetic content. A filmmaker posting the same clip on social media without context is presenting to an audience that expects documentary reality. Same content. Same pixels. Different legal obligations based on the room it enters.

The holistic test is an invitation to argue. It rewards filmmakers who control their distribution context and punishes those who let content circulate without framing. This is, accidentally, an argument for structured creative processes that include documentation, disclosure, and editorial control. The vocabulary this series has documented for a hundred and forty-three articles is simultaneously better filmmaking, stronger copyright protection, clearer awards eligibility, and now legally simpler compliance.

Seven days

Seven days remain before the EU AI Act's Article 50 transparency obligations begin. Sixteen institutional frameworks sit on the gradient. The EU's line through the frame is the most specific any of them have drawn. Not "how much AI is acceptable" or "was a human involved" or "did you have consent." This one says: this half of the image is a special effect, and this half is a deception, and the line runs between the world and the person in it.

The filmmaker who generates the environment and keeps the performance has always been making the stronger work. Now that filmmaker is also making the legally simpler work, the more copyrightable work, the more awards-eligible work, and the less paperwork-intensive work.

That convergence did not happen because any institution read this series. It happened because the craft and the law are both looking at the same thing: the audience's relationship with the person in the frame. When that person was never in the room, the audience needs to know. When the sunset behind them was generated by a model on a soundstage, nobody needs to be told. They already know. They have always known. The Guidelines just wrote it down.

Seven days. The line is drawn. The question is which side of it your production stands on.

Bruce Belafonte is an AI filmmaker at Light Owl. He has read more regulatory guidelines this month than screenplays and considers the ratio a sign of the times.