How to Tell If an Ad Is AI-Generated: 9 Signals (2026)

What metadata reveals an AI-generated ad?

Metadata can reveal the tool chain, the export path, and whether the file carried a provenance record when it left the editor. If the asset still has C2PA or IPTC-style fields, you may see creation data, edit history, or the software that wrote the file. If those fields are gone, that is not proof the ad is human-made. It only means the easiest evidence has been stripped.

If you are auditing a static image, inspect EXIF and IPTC first. A generator name, an export timestamp, or an editing app can tell you whether the asset came straight from an AI workflow or passed through human cleanup before publication. That does not prove the ad was deceptive. It only tells you which tools touched the file.

C2PA Content Credentials are built for that trail. The spec treats a manifest as a cryptographically bound provenance record, so a clean file can carry more than a caption or a screenshot ever will. When a file gets screen-recorded, recompressed, or re-exported through a platform that strips metadata, the trail weakens fast.

• Look for creator software names in the file properties.
• Look for C2PA or Content Credentials claims.
• Compare timestamps against the public post time.

No metadata is a signal too.

How do Meta and TikTok label AI creative?

Meta and TikTok both label some AI creative, but the labels mean different things. Meta's Expanding GenAI Transparency for Meta's Ads Products says its 'AI info' labels appear on ads made or significantly edited with Meta's own generative tools, and Meta said it would also surface third-party AI signals inside 'About this ad'. TikTok's Misleading and False Content policy requires realistic AI-generated content to be labeled or disclosed, and TikTok Ads Manager can place an AI-generated content disclaimer on the ad itself. Those labels confirm disclosure or detection; they do not prove authorship.

Use the label to sort, not to settle.

Meta's labels are also more visible now because they sit inside 'About this ad', not only on the asset itself. TikTok splits the job between in-feed labels and its ad-disclaimer system. That means you should treat the label as a platform note, not as a universal badge of AI authorship.

That matters because labels are often policy objects, not forensic ones. The FTC's Endorsement Guides still care about truthfulness and disclosure when a synthetic voice or face is used to persuade. A clean label is useful. It is not a verdict.

Which visual glitches still betray AI video in 2026?

AI video still leaks through motion, hands, and scene logic. The fastest tells are still fingers that merge or multiply, jewelry that clips through skin, shadows that point the wrong way, product edges that wobble between frames, text that mutates, and reflections that do not match the room. The more the clip depends on realism, the easier it is to catch. Stylized ads hide more, so you have to inspect continuity, not just one frame.

Hands still fail first.

• Hands: extra fingers, fused knuckles, nails that change shape, wrists that bend too far.
• Physics: liquid that moves uphill, fabric that floats, object weight that never quite lands.
• Text: package copy, on-screen CTAs, and logos that change letter by letter across cuts.
• Reflections: mirrors, sunglasses, chrome, and phone screens that show the wrong scene.
• Continuity: earrings, watches, hairlines, and product placement that jump between frames.

You open a competitor reel. The lip sync is clean, but the right hand changes from four fingers to five across two frames, the watch face flips from analog to blank, and the jar label bends when the camera pans. That is enough to mark the creative for manual review, even if the ad is still live.

Text remains one of the easiest failures to catch. Generative models can make a pretty package shot, then lose the brand name on the next frame or mutate the CTA on the lower third. If the logo is wrong, the ad is either synthetic or heavily manipulated, and either way it deserves review.

How can you check an ad's C2PA credentials?

Check the original file in a viewer that can read Content Credentials, not a repost or a screenshot. Look for a manifest, then compare the claim chain against the visible edit history and the software that exported the file. If the asset was re-encoded, screen-captured, or stripped by a platform, the manifest may be gone. That loss is not a clean bill of human authorship. It is just a broken chain.

The C2PA Content Credentials specification is useful because it does not ask you to trust a caption alone. It is built around a content binding and a signed claim, which is why provenance survives best when the asset stays close to the source file. If the ad went from source export to social post with no intermediate manipulation, you have a better chance of reading the trail.

1. Open the source file, not the screenshot.
2. Check for a manifest or Content Credentials panel.
3. Compare the stated creator and edit history to the public ad record.

Some viewers surface credentials directly, but the workflow still starts with the original asset. If you only have a downloaded repost, you can at best ask whether the file once had provenance. You cannot recreate a lost signature from a crop.

Missing credentials are common. They do not end the inquiry.

Do AI voiceovers have detectable patterns?

Yes, but not as a standalone verdict. Listen for even pacing, missing breaths, vowels that stay too perfect, consonants that smear at the edges, and emphasis that lands on the wrong syllable in names or prices. Cloned voices also tend to flatten room tone, so the background sound shifts less than a real recording usually does. A good human voice actor can sound synthetic, and a tuned model can sound human. Audio is a clue, not proof.

Compare the cadence to the brand's earlier ads if you have them. A sudden switch from a rough, local voice to a flat, clean read is often the first sign of a cloned or heavily cleaned track.

If the ad uses a cloned voice to imply a testimonial or endorsement, the FTC's Endorsement Guides still care about truthfulness and disclosure. TikTok's policy is even more direct: realistic AI-generated audio has to be labeled. So if the voice is the hook, verify the disclosure as carefully as you listen for artifact noise.

Listen in headphones.

Why does it matter for research if a winner is AI-made?

It matters because the production method changes how you read the result. A synthetic ad can be cheap to iterate, fast to localize, and easy to A/B at volume, which means a winner may owe more to editing speed than to raw concept quality. If you are deciding whether to model the creative, do not skip a live synthetic ad just because it was made with AI; the market has already told you something about the offer, the hook, or the angle. What you need to reproduce is the mechanism, not the tool stack.

Live beats archived.

That is why the current file matters more than the archive copy. Meta's ads transparency updates in 2025 and 2026 show that labels can move into new surfaces, but the ad itself is still the signal you are buying against. A live ad that keeps spending tells you more than a clever post that died last quarter. Count the response, then inspect the origin.

If the ad is AI-made but converting, note that in the model sheet. Binary labels are less useful than a simple note on what the creative was doing. That keeps the analysis tied to response.

Which detection tools actually work?

The tools that actually work are boring. Start with the platform label, move to a provenance viewer, then use manual frame and audio review as the final check. Commercial AI detectors can help you prioritize a stack of ads, but they are not verdict machines. The best first pass is the one that shows you who published the asset, what the file claims, and whether the creative still exists in market.

The Meta Ad Library is useful here for a narrower job: it helps you see whether an ad ran, which page ran it, and how the copy or creative changed over time. It does not tell you whether AI made the asset. That is a different question.

Missing is not proof.

When provenance exists, use it before any classifier score. The file can tell you more than a probability can, because it ties the creative to a source chain rather than to a guess. If provenance is missing, keep the ad trail, the visual audit, and the audio check. That is enough to keep the research honest.

Frequently asked questions

Can I tell from a screenshot alone?

A screenshot is weak evidence. Use the file, the platform record, and any Content Credentials before you call something AI-generated. If all you have is a repost or a crop, you can suspect synthetic work, but you cannot prove it.

Are AI detectors reliable enough for research?

They are useful for triage, not proof. A detector can help you rank a large ad set, but it can miss edited content and overcall normal compression or stylized design. Use it after provenance and platform checks, not before.

What does the Meta Ad Library tell me?

It tells you what ran, not who authored it. You can use it to check the page, the copy, the creative history, and whether a variation persisted long enough to matter. It does not prove AI generation, and it does not prove human authorship.

Sources

Named rather than linked — verify before relying on any figure below.

• Meta's Expanding GenAI Transparency for Meta's Ads Products
• TikTok's Misleading and False Content policy
• C2PA Content Credentials specification
• FTC Endorsement Guides