An AI disclosure sits safely in the lower-right corner of a landscape ad. Then the media team crops the film to 9:16, a platform button lands on the same corner, and a partner downloads and recompresses the asset for another market. The disclosure disappears even though everyone believed the job was finished.
That failure is not mainly about forgetting compliance. It happens because the disclosure was treated as a graphic burned into one export, rather than as a delivery layer that must survive every approved release.
Article 50 of the EU AI Act has applied since 2 August 2026. The practical buying question for an international brand is not “Did anyone use AI?” It is: when a particular release requires audience-facing disclosure, who makes that determination, how will the disclosure render in each placement, and which exact file or platform configuration is approved to go live?
Three different obligations are often collapsed into one
The European Commission’s current Article 50 FAQ separates responsibilities across the value chain.
Providers of systems that generate synthetic audio, images, video or text must support machine-readable marking and detectability, subject to the law’s qualifications. Deployers must clearly disclose image, audio or video that constitutes a deepfake when the relevant conditions apply. The Commission also states that a machine-readable mark does not, by itself, satisfy a deployer’s duty to make a disclosure understandable and perceivable to people.
So provenance inside a file is not the same thing as a viewer-facing label. A platform information panel is not the same thing as a disclosure that survives a download. Neither is a substitute for knowing which party approved the public release.
Outsourcing does not automatically move the role. The Commission uses an advertising company as an example: a legal entity can remain the deployer when employees, contractors or freelancers operate the system on its behalf and under its control. The actual workflow and control relationship matter more than the word “vendor” in a purchase order.
Not every AI-assisted ad needs the same visible label
Article 50 is not a universal instruction to stamp every asset that touched an AI tool.
The provider-side marking duty includes qualifications for assistive standard editing and changes that do not substantially alter the input or its meaning. The deployer-side visible disclosure test for images, audio and video concerns deepfakes: content that resembles existing or plausibly existing people, objects, places, entities or events and could falsely appear authentic or truthful.
Context matters. The Commission’s guidance points to the audience’s expectations, the deployment setting and the message conveyed. Clearly artistic, creative, satirical or fictional works have an adapted disclosure regime: disclosure should exist where required, but it may be presented in a way that does not interfere with enjoyment of the work.
Calling an advertisement “creative” is not an exemption by itself. A photorealistic product test that never happened, a cloned endorsement or a fabricated real-world event can create a different authenticity problem from an obviously fantastical background.
This article is a production procurement framework, not legal advice. Campaigns involving identifiable digital replicas, politics, health, finance or high-risk product claims should be escalated to qualified local counsel and the relevant platform policy team.
Three disclosure mechanisms fail in different ways
1. Burn the disclosure into the video
A burned-in disclosure survives downloading and reposting. But it can be cropped out when a 16:9 master becomes a 9:16 cut, obscured by subtitles or calls to action, or rendered unusable when the language changes. If the only master contains a market-specific label, the team also loses a clean source for markets or placements that require a different treatment.
2. Let the platform render the disclosure
Platforms can surface information based on advertiser settings or their own generation tools. Google’s 9 July 2026 announcement, Expanding AI transparency in ads, describes automatic disclosures for assets created with Google’s tools and an advertiser control for assets created elsewhere.
That overlay belongs to a particular platform and delivery context. It may not survive downloading, embedding, syndication or a move to another media owner. “The platform handles it” is therefore not a cross-channel acceptance standard.
3. Render the disclosure through the ad-delivery layer
Programmatic video can use the VAST <icons> element so the player renders an appropriate disclosure. Google Ad Manager’s current AI labelling guidance also warns that support depends on the player and integration. Third-party demand is not automatically labelled by Google publisher products.
This method is easier to update centrally, but reliability now depends on the publisher, ad server, player and SDK. The production team can no longer test only the MP4; someone must test the real delivery path.
A three-layer handoff: mezzanine, disclosure track and release branches
For AI or hybrid campaigns that need to cross markets and placements, ONCE recommends separating the handoff into three layers.
The clean mezzanine master contains the approved picture, sound, subtitle-safe composition and any retained provenance information, without permanently welding one market’s visible label into the only source. It is not a “no-label version” intended to bypass disclosure. It is the controlled source from which approved releases are made.
The disclosure track contains the replaceable audience-facing elements: icon, wording, spoken notice, first-exposure timing, duration, language, contrast, accessibility treatment and safe-area coordinates. It needs a version number and an owner, just like a subtitle track.
The release branches are the actual files or platform configurations cleared to run. Each branch binds together territory, language, channel, aspect ratio, disclosure mechanism, approver and file hash.
The same discipline applies to product-video URLs. If a brand already knows that overwriting one URL can confuse automated commerce systems, it should not quietly replace a disclosure-bearing creative under the same identity either. See ONCE’s guide, Stop Overwriting Product Video URLs: AI Commerce Needs Versioned Creative.
Use the ONCE TRACK card for every release
TRACK is not a legal acronym. It is a production acceptance card.
T — Territory
Record the actual market, launch date and whether the release is organic, paid or syndicated. “Europe version” is not a territory list. A limited transition that applies to some provider-side marking obligations should not be mistaken for a general delay for campaign disclosure.
R — Role
Name who selects the system, controls generation, approves the content, traffics the asset and publishes it. Contracts can allocate tasks, but a line saying “supplier handles compliance” is not a substitute for mapping the real control relationship.
A — Authenticity
Assess whether the release resembles a real person, product, place or event, and whether it implies a real test, endorsement or occurrence. The key question is not how much AI was used; it is what the intended audience may reasonably understand as authentic evidence.
C — Channel
For every aspect ratio and placement, specify whether disclosure is burned in, rendered by the platform or delivered through VAST. Test first-exposure visibility, collision with subtitles and interface overlays, muted playback, downloading and resharing.
K — Keep
Retain the clean mezzanine, disclosure track, final release branch, file hash, approval record and recheck date. Screenshots of platform settings can support the record, but they do not replace the released file and version manifest.
Change the denominator to approved market-placement releases
A quote based only on the number of exported files hides the work that actually expands with transparency requirements: classification, disclosure design, localisation, channel testing and controlled re-export.
A more useful formula is:
Cost per compliance-ready release = (base creative and production + role/authenticity classification + disclosure-track design + localisation + channel QA + rework) ÷ market-placement releases that pass TRACK.
This is a decision model, not an industry benchmark. A brand can enter its own territories, languages, aspect ratios, media owners and approval rounds to compare two approaches: one permanently labelled master versus a clean mezzanine plus controlled disclosure track and approved release branches.
For a short-lived campaign in one market and one placement, a correctly burned-in disclosure may be enough. For campaigns spanning multiple platforms, frequent localisation, downloadable assets or ongoing iteration, the disclosure track is more likely to become a reusable production asset.
What is confirmed, what is analysis, and what is prediction
Confirmed: Article 50 has applied since 2 August 2026. Provider-side machine-readable marking and deployer-side audience disclosure are distinct. The EU provides optional labelling icons, while platforms including Google are adding disclosure controls. The Commission’s EU icon guidance says labels should be perceivable by first exposure and should not be hidden by intervening overlays; using an icon does not, by itself, establish compliance.
ONCE analysis: The three-layer handoff, TRACK card and cost formula are production controls derived from those requirements. They are not mandated templates in the AI Act.
Prediction: AI disclosure is likely to become a versioned production layer, much like subtitles, music rights and product-claim approvals. Markets and platforms will not converge immediately. The durable brand asset is therefore not one permanent badge in one corner; it is a release system that can place, verify and replace disclosure without losing control of the creative master.
Source boundary
This article was checked against Article 50 of Regulation (EU) 2024/1689, the European Commission’s Article 50 FAQ updated 24 July 2026, its transparency guidelines published 20 July 2026, the EU labelling-icon guidance, and current Google Ads and Google Ad Manager documentation. It does not replace legal advice for a specific campaign.
