Should an AI Video Ad Be Labeled? Decide Before the Final Export

A global campaign is due tomorrow. The product and performer were filmed in a studio, the location was extended with generative AI, the English voiceover is synthetic, and two seconds of a product action were generated to complete the edit. Only now does someone ask: “Do we need an AI label?”

That decision has arrived too late. Disclosure is not simply a watermark option. It depends on what was created or changed, which market will see it, what a reasonable viewer may believe, and whether the people, product claims and source assets were properly cleared. A label cannot repair a missing consent or make an inaccurate demonstration truthful.

Google changed its advertising policy in July 2026 to permit text or visual labels inside image and video ads created or modified with AI. It also began rolling out an AI-label setting across Google Ads, Display & Video 360, Campaign Manager 360, Merchant Center and Ads Editor. Google explicitly says that using this setting does not guarantee compliance with a particular regulation. The policy update is specific about both points.

The regulatory environment has also moved from a future discussion to an operating requirement. The European Commission says the transparency obligations in Article 50 of the AI Act apply from August 2, 2026. They address specific situations, including machine-readable marking at the system level and deployer disclosure for deepfakes; they do not turn every use of AI-assisted editing into the same consumer-facing obligation. The Commission's Article 50 FAQ explains the scope and distinctions.

For brands producing assets near a Chinese supply chain and distributing them internationally, the practical requirement is a version-level decision record: what happened to this asset, where it will run, and who approved its treatment.

Separate provenance, audience disclosure and compliance

Provenance records where the image, sound or performance came from and how it changed. Machine-readable approaches such as C2PA Content Credentials can preserve source and edit history. C2PA itself does not decide whether a creative is truthful, lawful or worthy of audience disclosure; it provides verifiable provenance assertions. Its 2026 implementation guidance includes an AI-disclosure assertion that can capture more than a binary label. C2PA describes that implementation approach here.

Consumer-facing disclosure answers whether a person should be told, at or before first exposure, that an apparently real person, place, voice or event was generated or materially manipulated. The Commission's guidance distinguishes this perceptible disclosure from machine-readable marking: embedded provenance alone may not satisfy a deployer's duty to inform people.

Legal and platform approval considers the country, audience, industry, placement and exact content. A platform control can implement one part of the decision. It does not issue a legal opinion for the advertiser.

These layers can work together, but they are not substitutes. Metadata does not ensure that a viewer sees a disclosure. A visible “AI-generated” label does not establish rights to a person's face or voice, or evidence for a product claim.

Route the production according to what the viewer could misunderstand

The following is an ONCE production triage, not a legal classification. A brand should still review the actual creative and destination market.

The EU definition of a deepfake centers on generated or manipulated image, audio or video that resembles an existing person, object, place, entity or event and could falsely appear authentic. Context and audience expectations matter. Clearly fictional or artistic treatment may be handled differently, but that does not mean production teams should stop recording AI involvement.

The IAB AI Transparency & Disclosure Framework V2 offers a voluntary, materiality-based industry approach. It distinguishes routine post-production and technical assistance from synthetic content that could mislead people about authenticity, identity or representation. It is useful operational guidance, not law, and stricter local requirements take precedence.

The ONCE AI Creative Disclosure Passport

ONCE recommends delivering an AI Creative Disclosure Passport with each final version. It is a compact record with eight fields:

1. Version identity: filename, language, aspect ratio, duration and release date.
2. Distribution scope: country or region, platform, and paid or brand-owned placement.
3. AI involvement: how AI was used in script, image, video, voice, translation and post-production.
4. Authenticity impact: which changes improved technical quality and which created an unfilmed person, action, object or environment.
5. People and rights: consent and limits for faces, voices, performances, music, source imagery and references.
6. Product-fact check: whether generated content changed appearance, dimensions, functions, speed, results or conditions of use.
7. Disclosure decision: machine-readable provenance, platform setting and/or visible or audible disclosure, with its rationale.
8. Approval and review: responsible brand, legal or compliance owner, plus the trigger date or change that requires a new review.

This is not a compliance certificate and a production partner should not present it as legal advice. Its value is operational: the client can see what happened to the version, which decisions are complete, and what still needs approval by the brand or local counsel.

Put the decision into four production gates

At the brief gate, identify markets and uses. “Make an international AI commercial” is not enough. A paid EU campaign, a US brand-social post, a trade-show screen and an internal concept film are different deployment contexts. Named likenesses, synthetic voices and product demonstrations should be separate approval items.

At the script and board gate, ask what the viewer will believe. An ordinary-looking shot may imply a real store, actual customer, proven product outcome or genuine endorsement. If that implication could affect a buying decision, choose among live capture, obviously stylized treatment, appropriate disclosure, or removing the claim.

At the production and post gate, preserve source-to-version relationships. The audience does not need to see every prompt or internal file, but the project team should be able to trace live-action masters, generated shots, voice models, permissions and significant edits. When one synthetic shot is reused across languages, its disclosure status should travel with it.

At the delivery and media gate, review each final export—not only the hero film. Google's disclosure may appear in My Ad Center, on an ad or in another product surface depending on the implementation and local requirements. Teams still need to check cropping, readable language, re-encoding and what happens when the same file moves to a different platform.

Does an AI label reduce delivery?

Google's published policy says permitted AI labels will not violate its restrictions on text overlays and watermarks. It does not promise that a label can never affect campaign performance, and it does not say that using the label automatically triggers an algorithmic penalty.

Those are different claims. Audience, bidding, creative quality, landing page, inventory, review status and market can all affect results. Compliance and performance testing should therefore remain separate. Where disclosure is required, satisfy it first and then test the clearest compliant execution under comparable conditions. Where it is not required, do not hide material synthetic content merely to make the work appear fully captured in camera.

Buy a controlled deliverable, not only an MP4

When sourcing AI or hybrid video production, ask the production company four questions:

• Which shots were captured, generated, composited or AI-enhanced?
• Did the work use an identifiable person, voice or protected source, and where may it run?
• Who makes the disclosure decision for each market and how will the label be implemented?
• Who rechecks the decision when the market, platform or edit changes?

A low-cost vendor may deliver a finished file while leaving the brand to reconstruct its origin, rights and disclosure status on launch day. That is not necessarily a saving; it is unpriced operational work and risk transferred to the buyer.

For brands working with ONCE remotely, the better brief includes intended territories and placements, approved product facts, any real-person or voice use, supplied-asset rights, and the owner of the final disclosure decision. ONCE can then design live action, generated scenes and post-production as one controlled chain, with real product evidence preserved where it matters and each output carrying its own production record.

The strategic principle is simple: mature AI production is not defined by making AI invisible. It is defined by being able to explain where it was used, why it was appropriate, what the viewer needs to know, and who owns the release decision.

Sources checked October 1, 2026. This article explains a production workflow and does not provide legal advice.