Best AI Disclosure Approaches for Ads, Ranked

The default stack is: tell Ads Manager the truth, keep C2PA or equivalent provenance on the export when the tool supports it, and add a human-readable flag on the creative for photoreal people. Meta can auto-label some of its own generative features and is expanding detection; Google has been rolling advertiser-facing AI transparency panels. Ranked first because fighting the platform label is how ads get rejected, and because relying on it alone is how third-party avatar spots ship unlabelled.

Operationalise it. Whoever uploads the ad ticks the third-party AI control when the file was made outside the ads manager. Do not strip metadata on export 'to keep the file clean'. If your generator writes Content Credentials, leave them on. Preview the live label on mobile; a tiny 'AI info' in a corner the crop removes is not a preview pass. For political and social-issue ads the rules are stricter in many regions — treat those as a different workflow, not as a UGC template. Make the tick-box part of the upload checklist next to the URL and the pixel. If agencies traffic for you, put the requirement in the insertion order. A missed third-party AI flag is not a creative choice; it is an ops miss the advertiser still owns.

The hedge: auto-labels lag. A new avatar tool will not be in the detector on day one. That is why rank 1 is 'plus your own flag' rather than 'trust the sparkle icon'. You still decide whether the person on screen could be mistaken for a real endorser. If yes, you disclose, even if the dashboard did not prompt you. The advertiser owns that call. Re-preview after Meta or Google restyles the label. UI chrome moves. A label that was visible in March can sit under a new CTA in August. Compliance is recurring, which is why 'we ticked it once' is not rank 1 by itself.

When the ad shows a photoreal person who does not exist, or a likeness doing something they did not do, a spoken or burned-in line is the disclosure that actually meets a viewer. IAB-style guidance and several 2026 rules converge here: identity is the issue, not the fact that a language model wrote the hook. Ranked second because it is visible in the creative you control, independent of whether the platform's detector fired.

Write it like a caption, not like a terms footer. 'Virtual creator — not a real person' or a local-language equivalent, on screen in the opening beats, contrast-safe, away from the crop edge. Sound-off is most impressions; a spoken-only disclosure fails mute. Do not hide it at second 27. New York-style synthetic-performer rules (in force in 2026 for in-scope ads) care about conspicuousness; a 6-point caption is a dare. Translate the line; do not leave English on a Spanish ad and call it conspicuous. If the avatar speaks Hindi, the burned-in disclosure should too. Local counsel picks the noun ('virtual', 'generated', 'not a real person') — media does not.

Hook-rate cost is real and usually smaller than people fear if the rest of the ad is a demo. Test the line as a cell if you must; do not test 'no line' in a jurisdiction that requires one. For stylised or obviously cartoon avatars, many frameworks treat identity as already obvious — still keep rank 1's platform tick. For photoreal, this is the grown-up default. Legal reviews the exact wording; media does not freelance a joke disclosure. Keep a still of the first frame with the line visible in the campaign folder. That still is what you show in a review. If you cannot screenshot it, the disclosure was never in the creative you launched.

FTC endorsement logic still applies when the 'creator' is generated. A material-connection disclosure (paid, gifted, branded) does not tell the viewer the endorser is not a person. Ranked third because AI UGC that imitates a testimonial is the highest-risk shape in this list: you can satisfy a platform AI label and still have a fake-review problem. Ranked third because testimonial-shaped AI is where platform AI labels are not enough. A sparkle that means 'made with AI' does not mean 'this endorser is not a person' and does not mean 'this result is a real customer'. Two short, separate lines are the operational answer; inventing a reviewer is never a labelling problem.

Keep them separate. 'Ad / paid' in the usual place, and 'virtual creator' (or 'this person is generated') where the face is. Do not merge into 'AI ad' — that tells nobody whether the SKU, the review or the mouth is synthetic. Never generate a testimonial that implies a real user result you do not have. The 2024 FTC fake-review rule is older than your avatar tool; it still bites. If you use a real review with an avatar, say so without inventing a face for the reviewer. 'A customer wrote this; the person on screen is virtual' is clunky and honest. Clunky beats a fake neighbour in a fake city.

Influencer-style UGC with a real person who used AI only to edit captions is a different case: paid disclosure yes, synthetic-identity maybe no. Photoreal avatars reading 'I bought this' are not that case. If your prompt includes a fake name and city, stop. That is not a disclosure problem you can label your way out of; that is a deception problem. Train whoever writes comments and landing pages on the same dual rule. An unlabelled avatar ad plus a PDP full of unnamed five-stars is how a file-level disclosure gets undone by the click. The advertiser owns that stack too.

Invisible provenance is not a viewer disclosure. It is how platforms, journalists and tomorrow's ad-review jobs decide what you made. Tools increasingly sign outputs (C2PA, SynthID-class watermarks). Ranked fourth because it should be on by default and because it does not replace a human-readable line. Think of it as the shipping label on the crate, not the warning on the toy. Ranked fourth as infrastructure: sign the export, store the original, do not flatten credentials in a random editor. It will not tell a scroller anything. It will tell a platform, a journalist or a regulator what you shipped, which is why it sits in the stack under the visible line rather than instead of it.

Pick generators that emit credentials and keep a pipeline that does not flatten them in a random desktop editor. If you recut in CapCut or Premiere, check whether the credentials survive; if not, re-declare in the ads manager and keep the on-screen line. Store the original signed export next to the launched file. That archive is what you show in a takedown or a regulator letter. Add a 'credentials survive?' checkbox to the edit SOP. If the recut strips them, re-export from the generator or re-declare in the ads manager the same day. A signed master that never launched is not provenance; the launched file is.

Do not treat a watermark as stealth compliance. Some watermarks are invisible on purpose. Viewers never see them. Article 50-style duties in the EU (deployer transparency for AI-generated content, in force from August 2026 for in-scope uses) still want people to be able to tell. Provenance helps machines; rank 2 helps humans. You likely need both on photoreal. Watch vendor defaults. Some tools sign only on paid plans or only on some modes. If you switch a generator, assume credentials are off until you prove otherwise. The advertiser owns that assumption, not the vendor's changelog.

A dry spoken disclosure can feel more native than a banner, and it is easy to add in an AI script prompt. Ranked fifth because most paid social plays without sound. Use it as a complement to burned-in text, especially in markets where a verbal notice is customary, not as the only approach. 'I'm a virtual creator, this is still a real product' is a line you can test; burying it under a joke is how it fails conspicuousness.

Put it after the hook, not as the hook, unless legal requires first-second placement. Pair with captions that include the same words. If the avatar is speaking another language, disclose in that language. Do not let the model improvise a cute version each take — freeze the approved sentence in the brief so ten variants do not create ten legal objects. Put the frozen sentence in the prompt's don't-rewrite list. Models will soften it into a joke or bury it in a thank-you. If legal approved twelve words, those twelve words are the take. Variants copy them, they do not riff.

This approach is weak on statics and on carousels. If the campaign mix includes stills of the same face, those stills still need a visible label. A VO disclosure on the video does not travel to the image. For mixed campaigns, add the burned-in line on the statics even if the video speaks. Viewers do not see a sequence; they see one object. A spoken-only plan on a carousel is a hole. Treat every file as if it might serve alone. Mixes are a report, not a guarantee the viewer saw the VO first, and stills of the same face travel farther than the video they came from.

A site footer, a terms paragraph, or a 'how we make ads' page is good hygiene and a terrible primary disclosure. Ranked sixth because almost nobody who sees the ad will click into the policy centre before they form a belief about the person on screen. Use it as a third layer — archive, press, regulator — not as the plan you show a media buyer. Ranked sixth as the handbook layer: useful, late, and invisible in the feed. Keep the page, link it from the ads library for staff, and never tell a buyer it is the disclosure. Feed ads are judged in the first seconds, not in a policy centre after the belief has formed.

Keep the page anyway. Say which tools you use, that paid media likenesses are licensed, that reviews on the PDP are real customers unless labelled, and how to contact you. That page will not save a photoreal avatar ad that implied a user result. Do not point to it from a 6-point caption and call the job done. Update the page when you change generators or when a new market's rule lands. A stale 'we may use AI' paragraph from last year is worse than nothing in a letter. Date it. Name the categories of content (avatars, copy, stills) separately.

Where this layer does real work: employee training, agency contracts, and intake forms so a freelancer cannot upload an unlabelled clone. Compliance is a workflow. The page is the handbook. The ad is the act. Put the same rules in creator and editor contracts so a freelancer cannot upload an unlabelled clone 'because the brand page exists'. The page does not traffic the ad. The trafficker does, and the advertiser owns that person too. A missed label on a freelance upload is still your Ads Manager, your page, and your letter.

The popular founder take is that labelling 'kills authenticity'. Sometimes a label costs a bit of hook rate. An enforcement action, a rejected account, or a story that you faked a customer costs the brand. Ranked last because 'the platform will tag it' is not a policy. Detectors miss. Statics miss. New tools miss. You shipped the file. Ranked last because authenticity arguments do not pay the fine or restore the account. Some stylised work may sit outside photoreal duties after counsel agrees; photoreal UGC avatars do not get there by wishing. Detectors miss. You shipped it. The advertiser owns the miss.

There are real carve-outs in industry frameworks: routine colour correction, obviously cartoon characters, generic synthetic voice in some contexts, text-only copy. Those are not photoreal UGC avatars holding your serum. If your legal counsel — not a Slack poll — says a specific file is out of scope, document that. Do not generalise it to the whole library. If counsel writes an out-of-scope note, keep it with the campaign and do not stretch it to 'all AI'. A cartoon mascot memo is not cover for a photoreal nurse. Scope is file-level, which is inconvenient, which is why this rank is last.

This list will date. Article 50, state synthetic-performer laws, platform detectors and IAB guidance were all moving in 2026. Rebuild the runbook twice a year. Until then, the operational rule is simple: photoreal person, you disclose; endorsement-shaped, you dual-disclose; platform tick plus on-screen line; advertiser owns it even when the vendor's marketing said 'ads-ready'. Build the twice-yearly runbook review into the calendar now. Platforms, Article 50-style duties and state synthetic-performer rules were not done moving in 2026. A static 'we don't label' stance is a bet that the law will wait for your hook rate. It will not.