A final file called “final_v8.mp4” no longer tells a brand enough. It does not show which shots were captured, which were generated, whether machine-readable provenance survived export, or whether an audience-facing disclosure is needed in the European Union. Those decisions used to live in a production chat. They now belong in the delivery package.
Article 50 of the EU AI Act has applied since August 2, 2026. It creates different transparency duties for providers of generative AI systems and for deployers that use those systems in professional activity. For a brand publishing commercial content in the EU, the practical question is not simply “Was AI used?” It is which role the company occupies, what the output represents, how an audience may understand it, and which technical and human disclosures are required.
Machine-readable marking and audience disclosure are separate jobs
Under Article 50(2), providers of AI systems that generate synthetic audio, images, video or text must make outputs machine-readable and detectable as AI-generated or manipulated. The rule includes limits and exceptions, including assistive standard editing that does not substantially alter the input or its meaning.
Deployers have a different obligation. The European Commission's Article 50 guidance explains that a legal entity such as an advertising company can remain the deployer even when employees, contractors or freelancers operate the tools on its behalf. If the published image, audio or video constitutes a deepfake, the deployer must disclose that artificial generation or manipulation clearly and distinguishably, no later than first exposure.
An embedded provenance signal is therefore not a substitute for a visible or audible disclosure. A buyer should ask who supplies the system, who controls publication, what existing reality is being simulated, whether the audience may take it as authentic, and who signs off each layer.
Not every generated background or visual effect is automatically a deepfake
The Commission's guidance describes three cumulative criteria: the content must closely resemble an existing or plausibly existing person, object, place, entity or event; it must have that real-world reference; and it must be capable of falsely appearing authentic or truthful in the relevant context.
Context matters. The guidance says generated backgrounds, visual effects and technical pre- or post-production used in standard filmmaking are unlikely, by themselves, to make content appear falsely authentic. Evidently artistic, creative, satirical or fictional works still have a disclosure route when they contain deepfakes, but the disclosure may be made in a way that does not impair the experience.
This is why a universal “AI used” sticker is a poor production policy. Brands need an asset-level assessment.
The ONCE Three-Gate Test: referent, realism and context
| Gate | Production question | Evidence to retain |
|---|---|---|
| Real-world referent | Does the content simulate a real or plausibly existing person, product, place, organisation or event? | Reference list, factual sources, rights and consent |
| Perceived realism | Could the audience read the important image, voice, action or performance as genuinely captured or true? | Capture/generation boundary and altered elements |
| Deployment context | Where will the asset appear, and does that audience expect documentation or obvious creative fiction? | Market, channel, placement, version and first-exposure disclosure plan |
A clearly fantastical product world is different from a synthetic spokesperson who looks and sounds like a real endorser. A real product composited into a generated kitchen cannot be classified from the background alone. If that kitchen implies false dimensions, installation conditions or product performance, the asset may create an advertising-truth problem in addition to an AI-transparency question.
Stop accepting one flattened master as the entire delivery
For EU-facing AI content, ONCE separates delivery into two layers that must survive independently.
Technical provenance layer
Record the tool and version, generation date, source assets, captured and generated regions, machine-readable marking status, output hash, and a verification after export, transcoding and platform upload. C2PA Content Credentials is an open technical standard for recording provenance and edit history. It is not the only compliance route specified by Article 50, and it does not replace an audience-facing disclosure where one is required.
Audience disclosure layer
Prepare the actual disclosure for each market and placement: where a visual label appears, how long it remains, what an audio statement says, whether a vertical crop preserves it, whether silent playback still communicates it, and how the destination remains consistent. Metadata that a person cannot perceive at first exposure does not do this job.
ONCE's earlier Creative Passport establishes whether an asset can be deployed. The new transparency fields add why disclosure is needed, which layer carries it and who approves the final decision.
Add an AI Transparency Delivery Pack to the contract
A useful package contains at least eight fields:
- Asset identity: unique ID, version, market, channel and expiry.
- Reality references: people, products, places, organisations or events being simulated.
- Generation boundary: what was captured, conventionally edited, generated, cloned or composited.
- Facts and rights: product claims, likeness, voice, source-media and reference permissions.
- Machine-readable marking: what the tool writes and whether it remains detectable after final delivery.
- Disclosure variants: visual, audio, copy and destination treatments.
- Responsible parties: what the brand, agency, producer and media owner each review.
- Approval record: decision, rationale, approver, date and conditions for reassessment.
Acceptance should test more than the presence of a generic label. Which element does the disclosure describe? Does it survive cropping, compression and re-editing? Can an audience perceive it at first exposure? Does a new market or placement change the assessment?
Start with the assets most likely to be mistaken for reality
Audit four groups first: synthetic spokespeople and cloned voices; UGC-style testimonials that look like real customer evidence; generated demonstrations of product performance or use; and advertising narratives that recreate real places or events. Run each through the Three-Gate Test, then keep it, add disclosure, export a provenance-preserving version, or replace the risky element with captured production.
The Commission says content generated before August 2, 2026 does not need to be labelled retroactively under one blanket rule, although voluntary labelling is encouraged where possible. A brand therefore needs an updateable decision record, not a one-time screenshot saying “AI checked.”
The durable production advantage is shifting from access to models toward the ability to explain where every important frame came from, why it can run and what changes when the same asset moves to another market. Structured provenance reduces rework today and becomes a reusable content-governance asset for future campaigns, partners and channels.
Sources: European Commission, Guidelines on transparency obligations for providers and deployers of AI systems (July 20, 2026); Article 50 Q&A (updated July 24, 2026); Code of Practice on Transparency of AI-generated Content (updated July 31, 2026); EU AI Act Article 50; and C2PA Content Credentials 2.4. The Three-Gate Test and AI Transparency Delivery Pack are ONCE production and procurement methods. Specific deployments should be assessed for their market, content and legal context.
