Not yet mature enough

Article 50 of the EU AI Act goes live on August 2. Synthetic content marking, deepfake labeling, chatbot disclosure. Fines of up to fifteen million euros or three percent of global annual turnover. Four hundred and fifty million people in the enforcement zone. The deadline has been on the calendar since 2024. Nobody can say they were surprised.

The Code of Practice was supposed to provide the operational how. Which marking techniques satisfy the law. Which detection mechanisms qualify. Which disclosure formats meet the standard. The European AI Office spent months drafting it, circulated a first version in December 2025, collected feedback, and finalized it this week. Three weeks before enforcement.

Buried in the finalized Code: forensic detection mechanisms were demoted from mandatory to optional, "explicitly acknowledged as not yet mature enough to meet Article 50(2) EU AI Act quality requirements."

The regulation's own compliance framework just admitted that detection is not ready.

What moved

Three shifts between the December draft and the final Code tell the story. First, forensic detection went from mandatory to optional. A provider can offer it, but the Code no longer requires it, because the technology cannot reliably distinguish AI-generated content from photographed content at the quality threshold the regulation demands. Second, watermark interoperability was pushed to February 2, 2027, six months after the August enforcement date. Providers must implement an interoperability solution by then, choosing from four approved pathways. Until then, each provider's watermarking system is its own island. Third, the previously strict multilayer marking requirement was loosened. A single marking layer now suffices for systems in physically controlled environments and for free-form text that cannot transport metadata.

The marking side works. SynthID embeds imperceptible watermarks. C2PA signs metadata with provenance chains. The EU icon, a capitalized "AI" displayed in a standardized visual format, is finalized and freely available. If a provider marks their output, the mark survives distribution. The infrastructure for saying "I made this with AI" is functional.

The detection side is something else. The tools that would identify an unlabeled deepfake in circulation, catch a factory stripping watermarks before posting, or verify whether something flagged as suspicious was actually AI-generated: optional. The Code says so.

Zero retention

The finalized Code introduced a new privacy protection absent from the December draft. Content submitted for detection must be stored only for the duration of the detection process and permanently deleted immediately afterward. The term is "zero retention." Only minimal traffic logs survive.

The detection system sees each frame once and forgets it. The generation system remembers everything it was trained on. That is not a criticism of the privacy protection, which is sensible and necessary. It is an observation about the structural asymmetry. One side of the equation accumulates knowledge continuously. The other side is prohibited from accumulating anything at all.

The asymmetry runs deep. Earlier this year, Grok labeled Netanyahu's authentic video as "100% deepfake" and the man had to film a second video holding up his fingers. Generation has one well-defined objective: produce convincing output. Detection has an unbounded one: distinguish every form of generated content from every form of captured content, across every model version, every compression format, every distribution channel. The regulation assumed those two sides would be closer to parity by now. The Code just documented that they are not.

Catching the willing

On August 2, the regulation catches providers who comply. Providers who embed watermarks will embed watermarks. Providers who attach C2PA metadata will attach C2PA metadata. Providers who display the EU icon will display the EU icon. These are the organizations that would have labeled their output regardless, because the marking infrastructure is ready and the business incentive to comply is real. Fifteen million euros is a clarifying number.

The regulation has no reliable mechanism for catching providers who strip watermarks, skip metadata, and distribute unmarked AI-generated content into the same channels as photographed footage. Not because the regulation doesn't want to catch them. Because the detection technology the regulation was designed to rely on has been formally acknowledged as not ready.

This is a speed limit with no radar guns. The posted limit applies. The fine is real. The drivers who obey, obey. The ones who strip their plates drive past nothing capable of catching them.

The editorial exemption holds

The Code reinforces what the regulation established. No label is required if AI-generated content "has undergone a process of human review or editorial control" and a natural or legal person holds editorial responsibility for its publication. The editorial exemption survived every draft, every round of feedback, every amendment. It is the regulation's quiet recognition that the interesting question was never "did AI touch this?" It was "did a human decide what this looks like?"

A filmmaker who builds structured prompts specifying lens behavior, lighting direction, compositional placement, and atmospheric texture, then iterates through takes, reviews every frame, and takes editorial responsibility for the result, exercises editorial control. The EU icon appears nowhere on their work. Every framework that has weighed in on this question, the Academy, the DGA, the Golden Globes, SAG-AFTRA, Chinese regulators, YouTube, the courts, the EU itself, converges on the same answer: a human made the creative decisions.

The filmmaker exercising vocabulary satisfies every test. The factory that strips watermarks satisfies none but faces no reliable detection. The regulation was built for a world where marking and detection were symmetric. They are not. The Code is the first official document to say so.

Twenty-one days

The detection infrastructure will improve. Forensic tools will mature. The February 2027 interoperability deadline will produce something. But on August 2, when Article 50 goes live for 450 million people, the enforcement architecture arrives with its center missing. The regulation can say what must be labeled. It cannot yet verify what wasn't.

The distance between what a filmmaker knows and what a model comprehends has always been the interesting territory. The regulation just mapped the same terrain from the enforcement side. Marking works. Detection doesn't. The word that keeps appearing in every draft, every framework, every technical assessment is the same: not yet.

In a regulatory landscape where everything is oversold and underbuilt, a compliance framework admitting that its own detection mechanisms aren't ready is a rare act of honesty. The filmmakers who exercised vocabulary, iterated through takes, and documented their creative decisions have already built the editorial record the exemption requires. The factories that stripped watermarks and posted first-output content into unmarked channels just learned they have twenty-one more days of no consequence.

After that, the consequence is real and the detection is optional. The space between those two facts is where the next twenty-one days live.

Bruce Belafonte is an AI filmmaker at Light Owl. He reads compliance frameworks for a living now and blames no one but himself.