A product film uses AI in three places: reflections are cleaned up, a fictional warehouse is extended beyond the practical set, and a realistic synthetic presenter explains a genuine product feature. If the delivery note says only “contains AI,” the buyer still cannot answer the questions that matter at release: what was materially changed, which technical marks survived the post-production chain, and which public-facing version needs what form of disclosure.
Since Article 50 of the EU AI Act became applicable on 2 August 2026, the practical procurement problem is no longer solved by adding one generic label at export. Brands selling into the EU need a delivery structure that separates content classification, technical marking and audience disclosure—then assigns an owner to each decision.
Start with roles, not with a label
Article 50 of Regulation (EU) 2024/1689 creates different transparency duties for different actors. Providers of AI systems that generate synthetic audio, image, video or text outputs must, subject to the rule’s scope and exceptions, make those outputs detectable and mark them in a machine-readable format. Deployers have separate disclosure duties in specified cases, including image, audio or video content that constitutes a deepfake.
The European Commission’s Article 50 FAQ makes the operational split clearer. Provider-side marking is a technical obligation. Deployer-side disclosure is what the audience receives when the relevant conditions are met. The legal person under whose authority an AI system is used may be the deployer; a contractor does not automatically take over that role merely because it produced the asset.
That distinction should appear in the statement of work:
| Delivery question | Evidence to retain |
|---|---|
| Did the system output carry a machine-readable mark? | System and version, original output, available provenance data, and results after import, transcode and export |
| Does the released content require audience disclosure? | Classification rationale, market and channel, approved visual or audio treatment, first-exposure placement and approver |
A contract can allocate tasks. It cannot rename the legal roles created by the regulation.
“AI was used” is not a sufficient classification
The Commission’s FAQ says that the provider marking obligation contains an exception where the AI system performs an assistive function for standard editing or does not substantially alter the input data or its semantics. Its examples also indicate that ordinary backgrounds, special effects and technical pre- or post-production work in film will not necessarily make content falsely appear authentic.
At the other end of the spectrum is the Act’s definition of a deepfake: AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear authentic or truthful. The decisive production question is therefore not simply whether a tool used AI. Teams need to record what the source was, what changed, whether the change reached the final output, and how it represents a real person, object, place or event.
Creative, artistic, satirical and fictional work is not a reason to skip classification. Article 50 provides for disclosure in an appropriate manner that does not hamper display or enjoyment of the work, which means the treatment may differ while the decision still has to be made.
This article is a production and procurement framework, not legal advice. A brand’s qualified counsel should confirm the applicable role, scope and disclosure treatment for the actual asset, market and release.
Machine-readable marking does not complete audience disclosure
Machine-readable marks support detection and provenance. Audience disclosure supports human understanding. The Commission says that, where a deployer disclosure is required, relying only on a machine-readable mark is not enough. The disclosure should be clear and distinguishable and provided at the latest at the first exposure, in a visible or audible manner.
This changes what “approved master” should mean. A technically marked source does not prove that every platform version has the correct audience treatment. A social cut with a visible disclosure does not prove that the original system output and its provenance information were retained through the finishing chain.
A production partner can inspect, preserve and report what it receives. It should not promise to fulfil an AI system provider’s statutory obligation on that provider’s behalf. Likewise, the brand should not leave the final classification decision to an editor working at the end of the schedule.
The ONCE three-layer AIGC transparency handoff
Layer 1: content classification sheet
Record the source, AI operation, resulting change and intended use at shot or asset level. “AI-assisted” is too broad to be useful. Distinguish actions such as cleanup, masking, generative extension, synthetic performance, voice cloning and fabricated events. For every line, mark the status as a production fact, an open legal question or an approved decision.
The sheet should also state whether the changed element appears in the final delivery. Experiments that never entered the cut belong in the production record, but they should not be confused with release content.
Layer 2: technical mark retention record
Retain the original system downloads, file information and any available provenance credentials. Check what survives at defined points: source download, edit import, intermediate render, final master and pre-upload distribution file.
Use honest statuses—detected, not detected, not supplied by the system, or not yet verified. This layer is not a certificate of legal compliance. It shows the buyer what technical information survived, where it may have been lost and what questions remain for the system provider.
Layer 3: audience disclosure and release matrix
Make the unit of control “market × channel × version.” Record whether audience disclosure is required, its visual or audio form, placement at first exposure, language, accessibility treatment, approving party and final filename.
A product film may be delivered to a marketplace page, paid social, connected television and a retail display. Those environments have different layouts and audience encounters. One abstract note attached to the master cannot control them all.
Combined, the three layers turn a video file into a reviewable release package.
A hypothetical three-version release
Consider a China-linked home appliance brand preparing an EU launch film. Version A uses AI only to clean reflections and room noise. Version B places the appliance in an obviously fictional future kitchen. Version C uses a realistic synthetic “engineer” to demonstrate genuine functions.
The production team should not wait until delivery day and attach the same label to all three. A stronger sequence is:
- Record the AI-touched assets, shots and meaning changes in each version.
- Retain the original system outputs and test the state of technical marks through the production chain.
- Ask the brand’s qualified counsel to classify each release against Article 50 and any other applicable rules.
- Approve the disclosure treatment for every distribution environment and bind that approval to the final file.
This scenario is intentionally hypothetical. It is not a legal conclusion about the three versions. Its purpose is to show that transparency decisions follow the actual content and release path, not a single yes-or-no field for AI use.
Six acceptance criteria for an AIGC advertising brief
When sourcing an AIGC production partner, request:
- an AI system, model version and asset register;
- a record of source inputs and generation or manipulation steps;
- retention of original outputs and available provenance credentials;
- technical mark checks at agreed production stages;
- a market-, channel-, language- and version-level disclosure matrix;
- explicit approval boundaries across brand, legal, media and production teams.
The delivery wording should remain precise. Production is responsible for accurate records and for implementing the approved treatment. The obligations of the system provider, deployer and final publisher still depend on their real roles and the facts of the release.
Change the order of work before changing the badge
If disclosure is discussed only after all versions are finished, the team may have to reopen title-safe areas, remix audio, re-export languages or reconstruct missing source history. A better order is to confirm markets and roles at kickoff, log AI assets during production, complete content classification before picture lock, and approve each version before release.
The Commission’s guidelines and voluntary Code of Practice on Transparency of AI-generated Content provide the primary regulatory context. The advertising industry is also developing operational guidance: IAB’s AI Transparency & Disclosure Framework V2, published on 18 August 2026, uses a risk- and materiality-based approach and explicitly warns against indiscriminate labelling. Industry guidance can support a workflow, but it does not replace applicable law.
For brands connected to China’s supply chain, AIGC ad compliance is now a delivery-design question spanning creative development, asset management, finishing and media release. If you need to turn an existing AI production workflow into an auditable three-layer handoff, brief ONCE on the actual release path before the first asset is generated.
