TerradsTerrads
← All posts

Meta can tell your ad was made with AI, and documents no penalty for it

Meta names C2PA as one way it detects third-party AI tools, labels its own output automatically, and documents no delivery or eligibility consequence.

Meta detects third-party AI-generated ad content with industry-standard methods, "such as C2PA", an open list rather than a complete one. When it detects that metadata it attaches an "AI info" label inside the ad's "About this ad" screen. Content made with Meta's own generative features is labeled automatically without any detection step. Meta's documentation describes what gets labeled and where, and describes no consequence for delivery, performance or eligibility. The only hard requirement is for ads about social issues, elections or politics.

The question underneath this, asked constantly and answered mostly by guessing, is whether Meta can tell and whether it will punish you for it. The answer to the first is yes, under specific conditions. The answer to the second, as far as anything Meta has published goes, is no.

What Meta actually detects

Meta's help center splits this into two cases, and they work completely differently.

Its own tools label without detecting. Use Background Generation, Image Generation or Add Animation "to create or significantly edit an image or video", and an AI info label is applied automatically. Nothing is recognized here; Meta knows because Meta did it.

Third-party tools are detected from the file. In Meta's words:

Meta uses industry-standard detection methods, such as C2PA to identify when ad content has been created or edited using third-party generative AI tools. When we detect this metadata, we label the content accordingly.

Read that sentence carefully, in both directions. The detection Meta documents is metadata detection: provenance written into the file by the tool that made it, under the C2PA standard. But "such as" is an open list, and the newsroom announcement describes the same capability only as "industry-standard signals" without naming C2PA at all. Meta names one method it uses and doesn't say it is the only one. Treating the documented example as an exhaustive list is an assumption Meta hasn't made.

Keep this separate from what Meta does with organic content, which the same page flags as a different system: "Labeling and identifying AI-generated content is different for organic content." Claims about the organic label do not transfer to ads, and most of the advice circulating about this mixes the two freely.

Where the label shows up

In most cases the AI info label is not on the ad. You reach it through the three-dot menu → About this ad → the AI info label, alongside the other transparency information Meta puts there: payer and beneficiary details, social-issue ad details, partnership ads info.

Two related changes: Meta has renamed the Sponsored label to Ad on Facebook and Instagram, though it notes the change may not be available in every region. And you are told in advance: "Before publishing, you will see a message during the ad creation process letting you know when an AI info label will appear on your ad." You should never be surprised by one.

Boosted posts follow a different path again. Organic content keeps the organic labeling approach, and Meta says organic labels "may remain present when a piece of organic content is directly boosted to an ad via the boosting flow".

Five regions where the label moves to the front

Meta lists regions where local law or guidance imposes AI transparency requirements, and where advertisers may self-disclose that generative AI was used. Where they do, "the label will appear on the ad next to the Sponsored label" rather than inside the menu. Meta names: the European Region, California, New York, India and Taiwan. It adds that in Taiwan and South Korea, detecting AI-generated photorealistic humans in images or videos made with Meta's own tools may itself trigger an automatic label. (In the European Region, EU law adds a rule of its own for deep fakes: what Article 50 of the AI Act asks of advertisers.)

Region is not the only thing that moves the label to the front, though. Meta's newsroom post sets one global trigger: for its own generative tools, "when these tools result in the inclusion of an AI-generated photorealistic human, the label will appear next to the Sponsored label (not behind the three-dot menu)." No region qualifier, beyond Meta's general note that the experience may vary where law requires.

A synthetic person, in other words, is the one case Meta has committed to showing the audience by default. Keep that in mind before briefing one. It also matches the usual audience objection, which is rarely "AI touched this" and usually "that person isn't real".

So the label's visibility is a function of where the ad delivers and of what is in the frame, not only of which tool made it.

What a label costs you

Nothing that Meta documents.

Meta's pages describe at length what is labeled, where the label appears, which regions differ, and how the policy will keep evolving. They contain no statement that labeled ads are delivered less, cost more, are reviewed differently or become ineligible for anything. The one genuine requirement is the one that has existed all along: advertisers must disclose when an ad about social issues, elections or politics contains a photorealistic image, video or realistic-sounding audio created or edited with AI. For ordinary commercial ads there is nothing you are required to declare.

Absence of a documented penalty is not proof that no effect exists. An auction is not a published rulebook, and if you have measured a difference on your own account with a controlled test, trust that result over any documentation. But it does mean the widely shared belief that a label throttles delivery is currently sourced to nobody.

So should you strip the metadata?

No, and the reasoning isn't moral.

Removing provenance is a trade. What you buy is the absence of a label that lives two taps deep in a menu and carries no documented cost. What you spend is this:

  • In five named regions, self-disclosure is the compliant route, and Meta builds an affordance for it. Regulatory pressure on AI transparency is moving in exactly one direction, and arriving with your provenance already intact is the cheap version of that future.
  • For social-issue, election or political ads, disclosure is mandatory. Suppressing the signal doesn't remove the obligation; it just means you're the one who has to remember.
  • Provenance is becoming an asset. The same metadata that says "a model made this" is the mechanism that will let you prove a piece of footage is real when that matters to you.

If your creative only works while the audience doesn't know how it was made, the problem is the creative. The audience objection isn't "an AI touched it". It is the uncanny result, the invented product detail, the synthetic person presented as a real customer. Editing a file header fixes none of those, so fix the creative instead.

Sources

Related: what Meta actually requires you to disclose, and the enhancements Meta may apply to your creative without being asked — several of which trigger the label themselves.