EU AI Video Transparency: Production Handoff: ONCE

A commercial is finished in Shanghai. The product is live action, the background was extended with AI, the model’s lip movements were adapted for German, and the final cut was graded and exported into multiple platform versions. Does the campaign need an “AI-made” label in Europe?

There is no universal yes-or-no answer. The costly mistake is waiting until export to ask legal whether a badge should be added. Article 50 of the EU AI Act separates two jobs: providers of generative AI systems must make synthetic outputs machine-readable and detectable, while brands, agencies, or other deployers have audience-facing disclosure duties in cases such as deepfakes. One layer does not automatically satisfy the other.

The European Commission published its implementation guidelines on July 20, 2026, and the Article 50 transparency obligations have applied since August 2, 2026. The rules are not a blanket demand to label every AI-assisted ad. Standard editing assistance or changes that do not substantially alter the input or its semantics can fall outside the provider marking obligation. The Commission also says AI-generated backgrounds, special effects, or routine film post-production are not automatically deepfakes. Context, resemblance, message, audience expectations, and deployment all matter.

This is not legal advice. The procurement lesson is more practical: preserve the facts needed to make the legal and platform decision before the final render. ONCE calls that delivery the T-5 Transparency Handoff.

1. Applicability record

Do not stop at “AI was used.” For every relevant shot, record:

whether AI merely assisted an edit or generated/replaced image, audio, or text;
whether the simulated subject is fictional or resembles an existing or plausibly existing person, object, place, entity, or event;
whether the intended audience could mistake the result for authentic footage, performance, or product capability;
whether the work is evidently artistic or fictional, or relies on perceived reality as proof; and
the market, platform, and viewing context in which people will first encounter it.

The Commission’s Article 50 Q&A describes three cumulative elements for a deepfake: sufficient resemblance, a subject that exists or could plausibly exist, and a false appearance of authenticity or truthfulness. The same technique can therefore create a different risk in a fantasy brand film and in a product demonstration. Procurement should require a shot-level record, not a single AI/non-AI flag for the whole master.

2. Generation and edit record

For each AI-touched shot, retain the asset ID, tool and model, generation or edit date, source inputs, material changes, human reviewer, and approved output. The purpose is not to expose every prompt or place confidential data in public metadata. It is to answer: where did this version come from, what changed, and who approved it?

Localization branches belong in the same record. Face replacement, lip synchronization, cloned voice, digital humans, multilingual dubbing, and background replacement can create a new authenticity decision even when the parent master was already approved. “DE-final-v7” is not enough. The name must connect to the master, transformation, market, and approver.

3. People, rights, and product-truth record

Transparency does not replace consent, and provenance does not prove truth. Likeness, voice, performance, third-party footage, product appearance, and performance claims still require separate clearance. If a model generates a character closely resembling a real founder, engineer, or customer, a visible AI notice does not automatically solve authorization or deception questions.

This record can connect to ONCE’s rights operating pack: which permission covers the shot, where and how long it may run, whether it may be retrained or adapted, and who removes it at expiry. The disclosure and rights decisions share an asset ID without being collapsed into one checkbox.

4. Machine-readable provenance custody

Article 50 requires providers of generative AI systems, as technically feasible, to mark synthetic outputs in a machine-readable and detectable form. C2PA Content Credentials are one open-standard mechanism for recording origin and edit history. They are provenance signals—not a verdict that the content is accurate, lawful, or truthful.

Brand acceptance should therefore require three items:

the original tool export and a checksum;
verification results after each material edit, transcode, and platform upload; and
a record of where credentials were lost, plus any watermark, soft binding, or internal ledger used as a fallback.

This is not theoretical. C2PA notes that provenance may be incomplete when an asset passes through a tool that does not understand Content Credentials. Its specification supports workflows where credentials can be preserved or reattached, but no production team should assume every format conversion or platform upload will keep them intact.

5. Channel disclosure and release responsibility

Machine-readable marking is not the same as a disclosure that people can perceive. The Commission says a deployer publishing deepfake content must disclose it no later than first exposure in a clear, distinguishable, and perceivable form. It cannot rely only on embedded technical marking.

Each distribution version should state:

market, platform, ad account, and asset ID;
whether visible or audible disclosure is required, including language, placement, and duration;
who activates the platform’s AI label setting and where the confirmation is stored;
whether crops, responsive layouts, subtitles, or other adaptations can hide the notice; and
who has final authority to release that version.

Google’s July 9, 2026 ads transparency update demonstrates the two-layer reality. Google adds disclosures for assets made with its own tools and gives advertisers a control for externally generated assets. Its Campaign Manager 360 guidance also warns that using the setting does not guarantee compliance with a particular regulation. A platform toggle is an execution aid, not outsourced accountability.

Compare vendors by transparency delivery, not generation price

AI can reduce the cost of producing some shots while adding work for classification, records, verification, and multi-market delivery. A more useful procurement model is:

Transparency delivery cost = applicability review + generation/version records + rights verification + provenance preservation tests + channel disclosure/localization + exception rework

This is not an industry benchmark. It makes proposals comparable. Two vendors may both quote a “30-second AI ad,” while one delivers only an MP4 and the other supplies an auditable master, shot ledger, rights status, verification evidence, and market-ready disclosure variants.

Use three release gates:

Proceed: AI is routine assistance and does not materially alter the input or the audience’s understanding of authenticity; keep a basic record.
Specialist review: generated environments, people, product behavior, or voices could affect the authenticity assessment, even within a clearly creative or fictional context.
Pause and escalate: the work closely simulates a real person, product, place, or event and could appear authentic, or the team cannot connect source materials, permissions, model use, and the final version.

The procurement decision

EU AI video transparency is a supply-chain deliverable, not a sticker added in post. Brands do not need to treat every AI-assisted edit as the same risk. They also cannot treat a provider watermark as the complete answer.

The practical purchase is the T-5 handoff from pre-production onward. Brand, agency, and local compliance owners can then decide whether and how to disclose based on the actual use. That record remains valuable after the first EU campaign: it supports localization, platform declarations, rights renewals, and dispute handling without sacrificing version control for generation speed.