AI Ad Labels Are Now an Account Setting

Google expanded its AI-content disclosure rule to California, South Korea and Vietnam on 5 November. Meta shipped a Graph API field one day earlier that adds or removes an AI label. Read as one story, they look like AI labelling getting real. They are not one story. Google's rule governs ad assets and depends on someone in the account designating them. Meta's field governs Page posts and has nothing to do with ad creative. Two different switches, two different surfaces, and one conclusion: the AI label stopped being a policy you read once and became an attribute somebody has to own.

That distinction is not pedantry. An advertiser who reads both announcements and concludes "we're covered" is wrong in both directions — their ad assets may still be undesignated inside the six jurisdictions, and their organic posts carry a label that a publishing tool can now edit.

What Google actually changed

The change is a jurisdiction list, not a new mechanic. Google's help doc now reads, verbatim:

AI regulations in the European Union, India, New York, California, South Korea, and Vietnam require that ads with certain AI-generated or edited assets include disclosures and/or labels that inform consumers that the ads were made with AI.

And on where that shows up:

For campaigns that target the following locations, ads with assets that you designate as AI created or edited will also include visible overlays on the ads.

The operative word is designate. Google is not scanning your creative library and labelling what it finds. The disclosure depends on a human — or an upload template — marking a specific asset as AI created or edited while the ad is built or reviewed. Google does state that assets produced through its own fully automated creative features may be labelled on your behalf to help with compliance. That covers Google's output. It does not cover the background you generated elsewhere and uploaded.

Because the label is a per-asset designation rather than an account toggle, it travels with the asset through bulk uploads, editor workflows and agency handoffs. Whatever carries it in can also fail to.

The last row is the one people miss. About this Ad and the "How this ad was made" section of My Ad Center are visible globally, not just in the six markets. A designated creative shows its provenance in Honduras too; an undesignated one shows nothing, anywhere.

Meta's field is on the other surface entirely

The v27.0 changelog entry reads:

Apps using Graph API v27.0 can add or remove a post's AI-generated self-disclosure while editing the post.

The mechanics: POST /{post-id} and POST /{page-post-id} accept an optional boolean is_ai_generated. Set it to true to add the disclosure, false to clear it, or omit it to preserve the existing value. Older API versions are unchanged — the capability lands with v27.0 only.

Three things follow, and the news coverage of this got two of them wrong.

First, the scope. The parameter edits Page posts. Meta's documentation does not extend it to ad creatives, and ad labelling on Meta is a separate system driven by the platform's own detection. So the natural headline — "Meta now lets advertisers remove AI labels from ads" — describes a capability that does not exist. It lets a publishing tool remove the self-disclosure from an organic post.

Second, omission is safe; a false is not. Because Meta preserves the existing value when the field is omitted, a scheduler that simply does not know about is_ai_generated will not damage anything. The exposure is narrower and nastier: a tool that explicitly writes the field — a bulk attribute reset, a post-cloning flow, a template that fills every optional boolean with a default — can send false and clear a correct disclosure. Nothing in the post UI announces that the label changed. The self-disclosure becomes a value that can silently drift out of sync with the asset it describes.

Third, the label is now editable, which makes it data. Before v27, the AI disclosure on a post was effectively write-once through the API. Now it is a mutable attribute with no audit trail in the product. That is a governance problem, and it is the same governance problem the Google designation has.

The evidence that voluntary labelling does not happen

Resemble AI published its 2026 Synthetic Advertising Report on 22 October. The scan covered 37,768 image and video files across 15,433 ads from 509 companies in 26 industries, pulled from Meta's public Ad Library in September 2026. It flagged 3,379 files — 8.9% — as AI-generated or manipulated.

Of those 3,379 flagged files, two mentioned AI anywhere in the ad copy. A rate of 0.06%. None of the 971 files checked for C2PA provenance credentials carried any, including every file the report classified as a confirmed fake. Seventy percent of the labelled fakes were photoreal; 46% depicted a person.

Read that honestly, including its limits. The detector is the vendor's own product and the report publishes no error rates or score thresholds, so the flag rate is a vendor number, not a census. The copy search is also blind to a label burned into the video, so 0.06% measures written disclosure only — how many of the rest carried an on-screen badge is unknown. But the direction is not ambiguous. When disclosure is voluntary and unenforced, roughly nobody does it. Which is precisely the gap the six-jurisdiction rule was written to close, and precisely why an account-level designation field is the wrong place to leave unattended.

The decision nobody writes down: which assets count

Google's language is "certain AI-generated or edited assets," which means the operator, not the platform, draws the line. In account work we find the same absence repeatedly: the person who builds the ad and the person who owns compliance are different people, and the designation field sits in the gap between them. Nobody decided; the default answered.

So decide once, in writing, per asset class:

Two rules make that table survive contact with a team. First, the ambiguity goes in writing with a date and a name, because in a market with a regulator the documented decision is the defence — not the label. Second, when the call is genuinely 50/50, designate. A disclosure costs you a badge. The other side of the coin is a finding you cannot undo, because the ad already ran.

A twenty-minute audit you can actually run

1. Pull the list of campaigns whose geo targeting includes the EU, India, New York, California, South Korea or Vietnam. If that list is empty, you have no exposure under this rule today and the rest is housekeeping.
2. For each campaign, export asset-level data and find assets with no AI designation and no documented source.
3. Ask one question per asset: could a reasonable customer believe this depicts something that existed? If no, designate.
4. Check who built the ad. If Google's automated creative features produced it, verify the label is present instead of assuming compliance handled it.
5. Map your organic publishing chain. Which app holds v27 permissions on the Page, and does it write post attributes? Confirm it omits is_ai_generated rather than sending false.
6. Write the decisions down per asset, with a date and an owner. Six jurisdictions with six regulators means the paper trail is the control.

What the rule does not say

It does not ban AI creative. It does not require a label on every asset — "certain" is doing real work in that sentence. It does not reach organic posts, which is Meta's surface and a separate regime. It does not resolve upscaled photography for you. And it does not come with a penalty schedule you can look up on the platform, because the penalty lives in the local law, not in Google's help centre.

What it does do is convert a policy you read once at onboarding into a per-asset attribute that two different platforms manage on two different scopes in overlapping geographies. That is an operations problem before it is a legal one: one field, one owner, one date, one place where the decision is written down.

If nobody in your account owns the designation field, the default answer is unlabeled. In six markets, unlabeled is now the non-compliant answer.

If you would rather walk the account with someone who does this weekly, that is where our Digital Ads & PPC work begins — with an asset-level export, not a slide. What Meta's MCP server defaults to is the same lesson from the tooling side, and the product browsing opt-out is the clearest case of a platform default answering an operator question silently. Still mapping control between Smart+ and Advantage+? The purchase-data comparison sorts the automation layer first.