A finished master is no longer a complete handoff when generative AI has materially shaped an ad. A brand also needs to know where the source material came from, what AI changed, who approved each decision, which disclosure route applies in each market, and whether machine-readable provenance survived the final export.
This is becoming an operational issue, not a theoretical ethics debate. A platform may add a label automatically. It may detect signals left by a third-party tool. A visible disclosure may move from a menu into the ad itself because of local rules. If the media team is still messaging the editor on launch day to ask, “Does this count as AI?”, the production was not actually finished.
“How it was made” is becoming part of the ad surface
On July 9, 2026, Google announced a “How this ad was made” panel in My Ad Center across Search, YouTube and Discover. Ads created with Google’s generative advertising tools receive an automatic disclosure; advertisers can use a control for work created elsewhere. Depending on local requirements, a label may also appear directly on the ad. Google separately warns that using its label setting does not by itself guarantee compliance with every local rule.
Meta’s June 1, 2026 update to GenAI transparency for ads places AI information inside “About this ad.” Meta says it is also beginning to detect ads created or edited with third-party AI tools through industry-standard signals. The presentation varies: significant edits may be disclosed inside the menu, while an AI-generated photorealistic person can trigger a label next to “Sponsored.”
TikTok’s April 2026 advertising policy requires significantly edited media and AIGC to use its AIGC label or a clear advertiser-provided disclosure. Fully generated media, a subject shown doing something that did not happen, and AI voice cloning are listed as significant examples. Lighting, color, background changes and denoising are treated as minor examples. Undisclosed AIGC can be rejected or restricted.
These sources establish that the platforms are building and expanding disclosure systems. They do not create one universal global rule. Account controls, media type, market law and the risk of the depiction still matter.
Platform labels, provenance and claim approval do different jobs
The European Commission’s July 2026 Article 50 guidance says the transparency obligations apply from August 2, 2026. Providers of generative AI systems must support machine-readable detection of synthetic output. Deployers may also need clear, perceivable disclosure for deepfakes. The guidance distinguishes assistive standard editing from changes that materially alter an input or its meaning.
This creates three separate production questions:
- Provenance asks whether the file can carry a detectable account of its origin and changes.
- Platform disclosure asks what the audience sees in a particular ad environment.
- Brand approval asks whether the people, product facts, rights and claims are actually valid.
The C2PA’s July 31, 2026 Content Credentials implementation guide shows how machine-readable records can distinguish generated, AI-enhanced, camera-captured and human-edited content. It also describes affected regions, source ingredients, model disclosure and human oversight. But a credential is not a truth certificate. It can record that a claim was generated; it cannot prove the claim is accurate.
One global master can create four unanswered questions
Consider a product film that combines a live-action pack shot, an AI-extended environment, a synthetic voice and a digital double. Delivering one 30-second master leaves the next team to discover:
- Which shots only received cleanup, and which depict something that never happened?
- Can a platform detect third-party AI signals in the exported file?
- Which visible disclosure applies in the EU, the United States and other target markets?
- Do vertical, horizontal, subtitled and dubbed versions retain the same provenance record?
A platform version therefore needs more than an aspect-ratio suffix. It needs a disclosure route. The useful buying unit is a deployable version, not an exported file.
The ONCE STAMP handoff
ONCE proposes a five-part STAMP record for every commercial version materially involving AI. It does not replace legal review. It gives production, brand, media and counsel the same underlying facts.
S — Source
Identify the live footage, product images, human likenesses, voices, fonts, music, references and third-party assets used, together with their permitted uses. AI transformation does not cure an unclear license.
T — Transformation
Describe what AI actually did: full generation, local replacement, outpainting, background substitution, digital double, voice cloning, dubbing, denoising or conventional grade. For video, record changes by shot or time range instead of applying one vague label to the whole film.
A — Audience
Record the country, language, platform, ad type and foreseeable audience. Whether a depiction could be mistaken for a real person, event, place or product result depends partly on context. Photorealistic people, children, regulated claims and public-interest subjects deserve a higher human-review threshold.
M — Marking
Specify the disclosure for each platform version: an in-platform control, visible text, audible disclosure or a combination. Record language, position, duration and safe area. When the platform adds its own label, preview the ad to ensure that label does not cover captions, product proof or the call to action.
P — Proof
Retain the approved export, a reliable identifier or hash, Content Credentials validation, platform preview, named approver, date and scope. Provenance signals can change during transcoding or re-export, so validate the actual delivery file rather than relying on a tool’s general claim of support.
Route production by the nature of the change
Use a light-assistance route for denoising, grading and routine cleanup that does not change the meaning of the scene. Keep a basic version record and still confirm the target platform and market requirements.
Use a material-generation route for synthetic scenes, altered human actions, cloned voice or imagery that could change how a viewer understands reality. Prepare source and rights records, then choose the visible disclosure and account setting for every market-platform combination.
Use a high-risk human route for photorealistic digital doubles, health or financial claims, children, political and public-interest content, and imagery likely to be treated as evidence of a real event or product result. Production supplies the record; the responsible brand or qualified adviser makes the final determination. Multiple market versions may be safer than forcing one global master to carry every obligation.
Price the deployable version
An internal comparison can use:
Full cost per deployable version = (creative and production + source and rights review + AI change record + market/platform disclosure design + versioning + credential validation + approval archive) ÷ versions approved for the intended market and platform.
This is not an industry benchmark or legal formula. It is a procurement control. A supplier that generates 40 files but delivers only three traceable, correctly marked versions has not delivered 40 usable assets.
Five questions to settle before the quote
- Which countries, platforms and ad types will receive the work?
- Which shots, voices or people will be fully generated or materially altered?
- Who decides the disclosure, and who approves claims and likeness rights?
- Will the handoff include version-level sources, platform previews and the approved final files?
- Who rechecks disclosure and provenance after transcoding, subtitle replacement or recutting?
If those questions have no owner, “exported” does not mean “ready to publish.” ONCE’s AIGC commercial production begins by assigning clear roles to live action, generation and compositing, then carries the STAMP record into each intended delivery version.
Sources and limits
- Google Ads, “Expanding AI transparency in ads,” July 9, 2026.
- Meta, “Expanding GenAI Transparency for Meta’s Ads Products,” updated June 1, 2026.
- TikTok Advertising Policies, “Misleading and false content,” updated April 2026.
- European Commission, “Transparency obligations under Article 50 of the AI Act,” updated July 24, 2026.
- C2PA, “A New Implementation Guide for Content Credentials,” July 31, 2026.
This article translates public sources into a production and procurement method. It is not legal advice and does not assume identical controls in every market or account. Review the actual market, platform, asset and contractual responsibility before release.
