An ad that looks compliant in the edit suite may stop being compliant after delivery.
The landscape version becomes vertical. A platform recompresses the file. An agency changes the container. Captions cover a disclosure. A downloaded copy moves into a creator's repost. If the brand's plan was simply to place an icon on the master, it has no assurance that machine-readable provenance survived or that the viewer saw a clear disclosure at first exposure.
That procurement problem became concrete in 2026. Transparency obligations under Article 50 of the EU AI Act became applicable on August 2. The European Commission's July 31 implementation page separates two roles: providers of generative AI systems address machine-readable marking and detection, while deployers disclose deepfakes and certain AI-generated public-interest text when the relevant conditions apply. On September 24, the Commission published a set of optional EU icons and stressed that using an icon does not establish legal compliance by itself.
The implication is not that every AI-assisted ad needs the same badge. The first task is to determine whether the content is in scope. When disclosure is required, however, the production contract must deliver a tested marking path—not merely a decorated master file.
Separate the three jobs that teams often collapse
| Control | What it can establish | What it cannot establish alone |
|---|---|---|
| Machine-readable marking | Whether a file carries detectable origin or edit information | Accuracy, ownership, permission or context |
| Viewer-facing disclosure | Whether a person receives a clear notice at the relevant time | Whether provenance survived transcoding or rights are valid |
| Brand approval record | Who approved a version, on what basis and for which market | Whether the audience actually saw the disclosure |
C2PA Content Credentials are a provenance technology, not a certificate of legality. C2PA describes signed, tamper-evident structures that can record source and edit history. OpenAI makes the limitation equally explicit in its current provenance guidance: detecting a signal does not prove that content is accurate, unedited, legally owned or presented in the right context.
A positive C2PA result therefore cannot replace talent permission, music licensing, product-claim substantiation or legal review. It solves one portion of the transparency problem.
AI use does not create one universal treatment
The Commission's September icon guidance includes an important boundary. Viewer-facing disclosure addresses deepfake image, audio or video that resembles existing people, objects, places, entities or events and could falsely appear authentic, plus specific public-interest text. Evidently artistic, creative, satirical, fictional and analogous works may use an appropriate form of disclosure that does not hamper enjoyment. AI-generated public-interest text can also fall under an exception when it has undergone human review or editorial control and a person or legal entity assumes editorial responsibility.
The next brief therefore needs more than an “AI used: yes/no” field. It should identify:
- which people, objects, places, entities or events were generated or materially altered;
- whether an ordinary viewer could mistake the result for a real record;
- the countries, platforms, placements and languages in which it will run;
- who is the system provider, who is the deployer and who controls final publication;
- whether a documented human review and editorial owner exist; and
- whether disclosure uses text, an icon, platform UI or a combination.
Local counsel must make the legal determination. Production's job is to preserve the facts that counsel needs and turn the approved disclosure into a deliverable that survives versioning.
The ONCE three-pass test: Verifiable source, visible first exposure, retained after delivery
Every version that requires disclosure should pass three separate checks.
1. Verifiable source
Record the generation service or tool, date, material edit steps, source assets and responsible owner. Where the workflow supports Content Credentials, digital watermarking or another machine-readable signal, validate the final master and retain the validation tool, timestamp and result.
This does not mean publishing every prompt or placing personal data and trade secrets in a credential. The goal is a defensible internal record of where the version came from, what material transformations occurred and who approved release.
2. Visible first exposure
The Commission says an applicable disclosure should be clear and distinguishable no later than a person's first exposure. It should not sit under an overlay; accompanying language should be plain; and accessibility should be considered. The Commission also reports that the tested icon performed better when paired with a text label.
That is an editing and layout constraint. A disclosure hidden at the end of a video is not the same as a notice at first exposure. Nor can a single master be scaled mechanically across landscape, vertical and square formats when platform usernames, buttons and captions occupy different safe areas.
3. Retained after delivery
The real acceptance file is the output at the end of the distribution path, not the master on the workstation. Upload the asset to the target platform or test environment, then inspect the preview, transcoded copy, permitted download and representative re-share. Can a verifier still detect the machine signal? Is the disclosure legible? Did automated muting or remixing remove an audio notice?
If a channel strips metadata, the brand, counsel, platform and technical supplier need an approved fallback or layered approach: visible disclosure, native platform controls, watermarking, fingerprinting or a retrievable publication record. C2PA itself distinguishes embedded metadata from more durable approaches using soft bindings such as watermarks and fingerprints.
Change the acceptance unit from “final file” to “validated distribution route”
The same 30-second master may enter German social video, a French ecommerce page, an Italian creator partnership and European CTV through different crops, codecs, captions and interface overlays. Approving only the master creates a false sense of completion.
Compare proposals with this denominator:
Effective cost per validated distribution route = (scope assessment record + provenance record + disclosure design and localization + version output + platform test + review and archive) / market-language-placement-format combinations that pass
The denominator is not the number of exports. A combination counts only when it passes source verification, first-exposure visibility and post-delivery retention. If the badge exists in seven aspect ratios but is obscured by the interface on two platforms, those two routes are not accepted deliverables.
This also changes vendor comparison. A cheaper bid that omits platform testing, version records and exception rework may simply push the cost back onto the brand immediately before launch.
Add eight deliverables to the contract
- An inventory of every AI-generated or materially modified element.
- Market, language, media and role mapping.
- Separate permissions for people, voices, music, source assets and product claims.
- Generation and validation records for machine-readable marks or watermarks.
- Approved disclosure copy, icon, position, duration and accessibility treatment.
- Landscape, vertical and square safe-area versions for target interfaces.
- Spot-check results after upload, transcode, download or permitted re-share.
- Retirement and replacement rules when content, rights or regulation changes.
If the work involves a real person's likeness or simulated voice, transparency still does not replace permission. ONCE's digital-likeness ledger treats identity, generation scope, usage scope, derivative rights, supporting records and expiry as a separate control. This article addresses the downstream layer: keeping the approved disclosure visible and verifiable through the European delivery route.
The practical decision for brands
The EU icon is neither a button for every AI-assisted ad nor a one-off design task. It turns responsibility, marking, timing and distribution survival into a shared production, media and legal problem.
For low-risk AI-assisted retouching that does not cause a viewer to mistake a synthetic person, place or event for a real record, document the scope decision rather than automatically applying the heaviest process. For simulated people, cloned voices, replaced real-world scenes or visuals that could pass as documentary evidence, define the disclosure path before generation or shooting and include platform validation in the estimate and acceptance criteria.
What the brand should procure is not an icon file. It is a marking and disclosure route that remains intact from source to first exposure.
Sources and scope
Variable facts are based on the European Commission's Code of Practice on Transparency of AI-generated Content, updated July 31, 2026; its EU Icons for labelling AI-generated content, updated September 24, 2026; the C2PA 2.4 specifications and July 2026 deployment guidance; and OpenAI's current provenance-signal documentation. The three-pass method, cost denominator and contract checklist are ONCE's production-procurement framework, not legal advice from the Commission, C2PA or any platform.
