Imagine a 15-second product ad built from a live-action shoot. Generative AI extends the background, a human approves the English captions, and a second tool generates extra image area for the vertical cut. The brand plans to run it in the United States, the European Union and India.
“Did this ad use AI?” is no longer a useful enough question. The operational question is: which asset, after which transformation, is going to which platform and region, with what disclosure state?
On July 9, 2026, Google announced expanded AI transparency for ads. Across Search, YouTube and Discover, people can use the “How this ad was made” section in My Ad Center to see whether AI created or edited an ad. Google says ads made with its own generative advertising tools receive disclosures automatically; advertisers can use a control to identify assets created elsewhere.
Google Ads Help adds an important regional detail: ads containing assets designated as AI-created or AI-edited can receive visible overlays when campaigns target the European Union, India or New York. The same help page also says not all AI-created content needs a label, requirements vary, and the platform setting does not guarantee compliance with a specific regulation.
That changes the handoff between production and media buying. A production company cannot treat a single MP4 as the complete answer. A media team cannot infer provenance from a filename on launch day.
Disclosure is becoming a routing problem
Teams often record one project-level sentence: “AI was used in this campaign.” That sentence fails in three ways.
First, a master ad becomes horizontal, vertical, square, silent, localized and product-specific derivatives. Only one derivative may contain a generative edit.
Second, one asset can be treated differently at different destinations. Google currently places information in My Ad Center and, in some regions, visibly on the ad. An advertiser selection, automatic Google labeling and signals arriving from another platform can lead to different operational states.
Third, an AI label does not validate the creative. It does not prove product geometry is accurate, a person is properly licensed, the depicted location is real, or a claim is substantiated.
The unit a brand needs to manage is therefore:
asset version × destination × region × disclosure state.
That is ONCE’s operating inference from the public documentation. Google does not require advertisers to use the table below.
Give every asset an internal source code
A shared vocabulary lets production, legal and media teams exchange the same facts. These five codes are internal workflow categories, not legal definitions or substitutes for platform rules.
C: Captured or source-native
Footage, audio or approved 3D renders created directly by cameras, recording systems or the brand’s source pipeline. Routine cropping, compression and grading should still be logged; “captured” should not be misrepresented as “untouched.”
A: AI-assisted processing
Examples include denoising, masking, resizing, minor repair or caption assistance. The useful question is not whether the software contains AI. It is whether the process materially changed how a viewer would understand a person, product, place or event.
G: Generative alteration
The workflow creates a new background, replaces an object, outpaints a frame, synthesizes movement or introduces a situation that did not exist in the source capture. The changed area and its significance need an explicit record.
S: Primarily synthetic
The principal person, scene, action or visual subject is model-generated. The workflow also needs separate control over tool or source information, likeness and asset rights, product facts and human approval.
U: Unverified
Legacy files, agency handoffs or versions without production records. U does not mean prohibited. It means the asset should not move into every destination by default. The next action may be investigation, restricted use or rebuilding a verifiable version.
One ad may contain several codes: captured product footage (C), a generated environment extension (G) and AI-assisted captions (A). A single project-wide label can conceal that structure.
The ONCE AI Disclosure Routing Table
1. Asset and parent version
Record a stable asset ID, file hash or controlled filename and the master from which it derives. The vertical edit should not be “final_v7_new.” It should trace to a known parent and a specific transformation.
2. Source and generative change
Record the C, A, G, S or U code, the tool category and what was created or changed. Teams do not need to retain irrelevant sensitive prompts, but they need enough detail for a reviewer to understand the scope of the transformation.
3. Reality and claim impact
Did the change affect a real person, product construction, functional demonstration, location, event or comparison? Extending a sky and inventing a waterproof test are not equivalent production decisions.
4. Destination and region
Put platform, ad product, placement and target region on the same row. “Google” or “global” is too broad. Search, YouTube and Discover placements—and campaigns targeting the EU, India, New York or other regions—may present disclosure differently.
5. Label and machine-readable state
Track four things separately: advertiser designation, automatic platform labeling, expected visible overlay, and the presence of C2PA Content Credentials or another machine-readable signal.
Keep an “unknown” state. The C2PA FAQ explicitly says embedded provenance metadata can be removed intentionally or accidentally. The standard therefore supports durable credentials using mechanisms such as watermarks, fingerprints and cloud retrieval. The absence of visible Content Credentials does not prove that AI was not used.
6. Owner and launch action
Name who confirms production facts, who confirms likeness and asset rights, who selects the platform label state, who approves launch, and who can pause, replace or investigate the ad if the label or creative is challenged.
The routing table should travel with the asset into the media launch sheet. It should not remain only in the production vendor’s project folder.
Three handoffs where teams get this wrong
Google-generated, then edited elsewhere
Google says images and videos generated inside its advertising tools receive SynthID and C2PA machine-readable marking. But after download, transcoding, external editing and re-upload, the media team should inspect the current file and platform state. The origin of the first version is not proof that every downstream signal survived.
A captured master with generative outpainting only in vertical
Calling the whole campaign “live action” hides the vertical file’s generative alteration. Calling the whole campaign “fully AI-generated” is equally inaccurate. Preserve the C parent, the G derivative and a destination-specific decision for the vertical cut.
An agency delivers only the finished file
Move it into U status until the team has enough information. The buyer does not need every working file, but it should receive the tool category, material-change description, relevant rights, recommended platform designation and an accountable owner. A limited launch does not turn missing records into verified non-AI status.
A label is neither compliance nor quality assurance
Google warns that use of its AI label setting does not guarantee legal compliance. Voluntary disclosure also does not solve three separate problems:
- A generated product error remains inaccurate after it is labeled.
- Missing likeness or source rights are not repaired by transparency.
- Different destinations may require different treatment; one campaign note cannot make that decision.
The routing table therefore sits beside product truth, rights records and version approval. For the generation side of the workflow, ONCE’s article on why the brief becomes the product when Google can generate a campaign covers the inputs a brand should own. This article covers what happens after generation, when derivatives enter different media destinations.
Price provenance into the usable launch combination
Two production bids should not be compared only by cost per finished file. A more complete variable is:
Effective cost per verified launch combination = (creative and production + rights verification + version control + source/change records + destination/region disclosure review + replacement readiness) ÷ approved asset–destination–region combinations.
This is not an industry benchmark. It reveals scope. A low bid without parent-version records, change history or launch ownership pushes the verification cost into the media team’s final mile. A higher bid that supports several approved regions and placements may have a lower usable unit cost.
Six questions before upload
- Which exact file is being uploaded, and which master did it inherit from?
- What did AI create or materially change—not merely which software was used?
- Did the change affect a person, product, place, event or advertising claim?
- Which platform, ad product, placement and region will receive it?
- What are the advertiser designation, automatic platform label, visible overlay and machine-readable credential states?
- Who approves launch, and how will the team pause or replace the asset if the label or creative is wrong?
Google’s new interface does not say every AI ad is the same. It shows that AI transparency data is entering the distribution system. The scalable capability is not one blanket AI statement. It is a production-to-media workflow that routes every derivative asset to the right disclosure decision.
Sources
- Google Ads & Commerce Blog, “Expanding AI transparency in ads,” July 9, 2026.
- Google Ads Help, “Use AI content label settings and disclosures”.
- Google My Ad Center Help, “AI Transparency in ads”.
- C2PA, “C2PA FAQ”.
- Adobe Blog, “Expanding access for Content Credentials to enterprises, brands and consumers,” March 26, 2024.
This article translates public platform and standards documentation into a production-to-media handoff method. It is not legal advice for a particular jurisdiction.
