AI transparency is no longer a note for legal teams to add just before media launch. Between June and August 2026, Meta expanded AI information for ads made with third-party tools, Google introduced AI-label controls across its advertising products, the EU AI Act's Article 50 transparency duties began to apply, and C2PA published deployment guidance for Content Credentials. Together, these changes create a production question: when a platform, regulator or client asks how an asset was made, a final MP4 cannot provide the full answer.
For overseas brands working with Chinese supply chains and production partners, the practical response is not to watermark every AI-assisted frame. It is to separate three decisions: what the audience must be told, what the platform must be told, and what the brand must be able to prove later.
What has changed—and what has not
The European Commission's July 24, 2026 Article 50 guidance says the obligations apply from August 2, 2026. Providers of systems that generate synthetic audio, image, video or text must support effective machine-readable marking. Deployers must provide clear, perceivable disclosure for deepfakes and certain other covered uses. The guidance also distinguishes standard editing assistance and technical film-production processes from outputs that substantially alter meaning or falsely appear authentic.
Google's July 9, 2026 advertising policy update permits text or visual labels inside AI-generated or modified image and video ads and introduces AI-label settings across Google Ads, Display & Video 360, Campaign Manager 360, Merchant Center and Ads Editor. Google explicitly warns that using the setting does not guarantee compliance with every local rule.
Meta's June 1, 2026 update to GenAI transparency for ads extends beyond Meta's own tools. Meta says it will use industry-standard signals to detect ads created or edited with third-party AI tools and show an “AI info” label in “About this ad.”
C2PA's July 2026 deployment resources describe Content Credentials as tamper-evident, machine-readable records that can show where an asset came from and which creation or editing actions occurred. They do not decide whether content is true. Machine-readable provenance also does not replace a visible or audible disclosure where one is required.
The confirmed trend is that regulation, platforms and production tools are building separate transparency mechanisms. The ONCE inference is that brand handoffs need to preserve those mechanisms across agencies, exports and markets.
The missing production line behind the final file
Traditional delivery tracks resolution, aspect ratio, captions, audio and duration. Hybrid live-action and AIGC production adds an invisible line: which shots were captured, which were generated or replaced, whose likeness or voice was used, which tools touched the asset, and who approved its final context.
If that history exists only in a freelancer's chat thread, a market expansion or agency change turns into reconstruction. The costly part is rarely adding a label. It is discovering that the source render, consent record, model version or approval decision cannot be found, forcing the campaign to remake work it already paid for.
A brand should therefore buy a film plus a provenance passport. This extends the logic of making AI product demonstrations verifiable: one system supports product claims; the other supports the origin and permitted use of the image and sound.
The ONCE three-layer provenance passport
The passport does not crowd every frame with technical detail. It separates transparency into three layers that can be scoped and accepted independently.
Layer 1: audience-facing disclosure
This layer answers when a viewer needs to know. The decision depends on whether content resembles a real person, product, place or event; whether it could reasonably appear authentic; and which market and medium will receive it. When disclosure is required, it must be perceptible at the appropriate point, not buried only in metadata.
Layer 2: platform declaration
This layer records what the delivery system knows. Google, Meta, YouTube and other channel fields should be set for the actual deployed version. One upload does not automatically describe every derivative. A platform-applied label does not eliminate the brand's need for an approval record.
Layer 3: machine credentials and the human ledger
This layer answers what the brand can prove later. Where the toolchain supports C2PA or equivalent provenance, production should test whether credentials survive export, transcode and handoff. Where it does not, a human ledger must fill the gap. Both should point to the same asset identifier and approved version.
The layers are not substitutes. Machine credentials support provenance, platform settings support distribution, and audience disclosure supports understanding. A project may not activate every layer, but that should be a documented judgment rather than an omission.
Use three risk routes instead of one blanket rule
To avoid labeling everything merely because AI touched it, route the brief before production.
Route A: assistive editing. Noise reduction, routine colour work, resizing or caption cleanup that does not materially change a person, product or event may need a production record more than a prominent on-screen label. Applicable platform and local rules still control the final decision.
Route B: synthetic but not pretending to document reality. Stylised environments, abstract transitions, fictional spaces and concept imagery need a shot-level generation record, a platform-declaration decision and context-appropriate disclosure where audiences could otherwise be misled.
Route C: high resemblance or deception risk. A recognisable person's digital double, synthetic voice, a product performance that looks like a real test, or a fabricated news-like event requires deeper control: audience disclosure, platform status, likeness or voice permissions, generation scope, human approval and version evidence.
This routing is a production framework, not a legal determination. Public figures, children, political content, health or finance claims and high-risk markets should trigger specialist legal and platform review.
Eight fields every deployable asset should retain
Each approved version should lead back to:
- a unique asset ID, hash or unambiguous version reference;
- a classification: captured, AI-assisted, partly generated or fully generated;
- the principal tools, models or services and the creation date;
- rights status for source media, likeness, voice, trademarks and training use;
- material actions such as extension, replacement, face work, dubbing or compositing;
- target channels, territories, and organic or paid status;
- completion state for visible disclosure, platform fields and machine credentials;
- reviewer, approval date, restrictions, expiry and the next review trigger.
The consumer does not need to see the complete ledger. The brand, agency and production partner need to be able to transfer it. A vendor that hands over only “final.mov” without source, rights and disclosure status has not completed a durable delivery.
Quote and accept the work as production
Provenance work should be a scoped production task, not free administration. A quote can separate source registration, rights and consent packaging, market or platform disclosure configuration, and final credential verification. A new territory, changed performer, new media use or regenerated shot should update the passport instead of being treated as a free export.
The better procurement question is not “Do you use AI?” It is: How do you distinguish assistive work from material generation? How does each shot trace back to its source? What survives transcode? Who sets platform labels? Who can produce the approval record if a distributor challenges the asset?
A useful cost equation is:
Total AI transparency cost = recording and verification + disclosure-version production + rights review + expected rework from missing provenance.
The last term is the probability of a source or permission gap multiplied by remake, reapproval, missed-media-window and redeployment costs. Brands do not need an invented industry average. They can use their own number of variants, approval chain and launch schedule.
The responsibility shift
None of this makes AIGC advertising less valuable. It makes accountable production more valuable. A partner that can explain origin, permission and approval creates a more reusable asset than a supplier that can only generate a polished frame.
Set the risk route and disclosure strategy in the brief, maintain the provenance passport during production, and accept the audience, platform and evidence layers at handoff. The MP4 remains the film. It is no longer enough to prove where, how and by whom the film can be used with confidence.
