AI avatars, UGC-style ads and the rules: what to make and disclose

An AI avatar can show a product in use; it cannot turn an invented customer into a real one. If you make UGC-style ads with actors, generated faces or cloned voices, this guide helps you decide what the viewer is being asked to believe, which disclosures belong in the video, and which platform controls to check before you ship.

Start with the representation, not the tool#

UGC-style describes a way of filming: a person speaking plainly to a camera, often in the setting where the product is used. It does not tell you whether the speaker bought the product, was paid to appear, or even exists. Those are different propositions. Write them down before you script the first line. “Here is how this works” can be a demonstration; “I bought this and it changed my mornings” presents a consumer experience. An AI label answers how an image or voice was made, not whether the experience happened.

In the US, the FTC's Endorsement Guides define an endorsement by what consumers are likely to believe about the speaker's opinions or experience. The speaker can appear to be a person without being one. The guides require an honest account and say a claimed user must actually have used the product at the time of the endorsement.^[1] The FTC's separate consumer-reviews rule also reaches a testimonial conveyed by a likeness, not just a written review; it prohibits creating testimonials that materially misrepresent the speaker's existence, use or experience.^[2] You cannot repair a fabricated experience by changing the face or adding “AI-generated” in small print.

That distinction is useful at the brief stage. Make a column for speaker, basis of claim, relationship to brand, production method, and destination. A founder describing their own product, a customer describing something they actually did, an actor reading a script, and a synthetic presenter can all have different truthful scripts. Do not paste a customer sentence onto an interchangeable face until you have checked whose sentence it is, whether the customer agreed to its use, and what the face implies about the speaker.

The rules changed; the questions did not#

The FTC revised its Endorsement Guides in 2023 to address, among other things, virtual influencers and fake reviews. Its consumer-reviews rule took effect on October 21, 2024. TikTok's current advertising policy addresses undisclosed generated or significantly edited ads. The European Commission says the AI Act's Article 50 transparency obligations apply from August 2, 2026; it describes a later December 2, 2026 date for the machine-marking duty of systems already on the market.^[3][4][5][6] Dates matter because a screenshot of a platform's old help page is not a durable permission slip.

These are parallel checks. The FTC asks whether your ad deceives about an endorsement or claim. An ad platform asks whether your upload needs an AI label and whether the underlying ad satisfies its policies. EU rules can ask both the system provider to mark certain outputs in a machine-readable way and the deployer to disclose a deepfake to people on first exposure. The Commission specifically says metadata alone is not enough for that viewer-facing disclosure.^[6] A US ad placed on TikTok can therefore need a truthful script, a material-connection disclosure, and a platform AI disclosure for different reasons. The presence of one does not discharge the others.

The UK provides a further cross-check. Its current CAP Code says marketing must be recognizable as marketing, prohibits fake consumer reviews, requires incentivized reviews to be identified, and requires permission for testimonials. It also expects documentary evidence that a testimonial is genuine unless it is obviously fictitious.^[7][8] Some older explainers quote earlier CAP rule numbers; use the live rule text rather than carrying their numbering into your approval sheet. Treat the jurisdiction question as part of distribution planning, not a last-minute caption edit.

A real experience cannot be supplied by a synthetic face#

The FTC says an ad presenting performers as actual consumers should use actual consumers in audio and video or clearly and conspicuously disclose that they are not actual consumers. It separately warns that pairing an endorsement with somebody else's image or likeness can deceive if it misrepresents a material attribute of the real endorser.^[1] This is why permission to quote a real review does not automatically grant permission to portray its author with an actor who appears to be that customer. You may use a performer, but the presentation needs to tell the viewer what the performer is doing.

The consumer-reviews rule is more direct about fabricated testimony. It covers the business creating or selling a review or testimonial that materially misrepresents that the speaker exists, used the product, or had the stated experience. Dissemination can also be prohibited when the business knew or should have known about the misrepresentation.^[2] FTC staff explicitly say there is no blanket prohibition on AI stock avatars in marketing; a testimonial given by an avatar runs into the rule when the underlying testimonial is fake or false. The same staff guidance warns that avatars can still be deceptive under the broader FTC Act and Endorsement Guides.^[9] “Avatars are allowed” is therefore an incomplete creative brief.

Trace each first-person line to a source. “The lid fits under my cabinet” is not a property of the rendered pixels; it is a claim that a person used an object in a place. If that person is an actor demonstrating a measured fit, make it a demonstration: show the object and cabinet, use accurate dimensions, and do not add a fictional purchase history. If a customer really said the line, keep the original record and permission to reuse it. If neither exists, rewrite the scene around the product rather than manufacturing a customer.

Payment and gifts create a separate issue. The FTC requires disclosure of material connections that a significant minority of the audience would not expect, including free or discounted products even when the brand did not demand a review.^[1] A paid creator can tell the truth and still need to say they were paid. Conversely, a conspicuous “Paid partnership” label cannot turn an untrue experience into a true one. For help timing the product proof rather than making the speaker carry every claim, see the short-form ad beat sheet.

Walk the claim through a decision path#

Start with the viewer's likely inference, not the prompt used to generate the asset. The figure branches on the implied customer experience first. A “stop” leaf means remove or replace that claim; a “disclose” leaf means check applicable wording and placement in addition to the truth of the script. It is an editorial triage path, not a legal safe harbor.

Apply the path to both picture and sound. A genuine customer's quote over an unrelated synthetic face may give a false impression about identity; an actor in a real kitchen saying “I tested this for a month” may give a false impression about use. A cloned voice that sounds like an identifiable person raises permission and platform questions even if the text is otherwise accurate. TikTok specifically treats making a primary subject say something they did not say with voice cloning as a significant modification and restricts some unauthorized likenesses.^[5][10]

Before approving a cut, read the ad aloud without looking at its captions, then watch it muted without reading its description. Write down what a new viewer would infer in either mode. The FTC's clear-and-conspicuous definition accounts for size, contrast, location, time on screen and the medium of the original claim; a claim made visually and audibly calls for disclosure in both channels. Its examples explicitly note that a smartphone presentation can make a disclosure inadequate.^[1] The test is comprehension at the point of the claim, not whether a disclosure string exists somewhere in the upload form.

Platform labels solve a narrower problem#

Each destination defines “AI” and “disclosure” differently. On TikTok, the ads policy says to apply the AIGC label or add your own clear disclaimer, caption, watermark or sticker for applicable generated or significantly edited creative; undisclosed AI may be rejected or restricted. Its advertiser self-disclosure toggle is documented for non-Spark Ads, while Spark Ads use the creator-side label. Minor lighting, color and denoising edits are listed separately from significant edits.^[5][11] The organic-post policy requires labels for realistic AI images, audio and video, while it only encourages labeling some other significant edits. Do not merge these two policies into “all edits require a label.”^[10]

Meta labels some AI media using disclosure or industry signals, but its ad-specific treatment depends on the product and context. Political and social-issue ads have a disclosure rule when realistic synthetic depictions meet specified triggers. Meta says a photorealistic human made with its own ad-generation tools receives an AI label next to the ad indicator. Its June 2026 update says it is beginning to identify third-party AI ads using industry-standard signals.^[12][13] A regional Meta help page still says third-party created or modified ads will not receive AI info; Meta notes regional variation. Do not promise a particular automated badge. Inspect the delivered placement and follow the current ad flow for the account's region.^[14]

YouTube asks creators to disclose altered or synthetic material that realistically depicts a real person doing something they did not do, changes footage of a real event or place, or creates a realistic scene that did not occur. Its help page exempts production assistance such as scripts and captions and cloning your own voice for voiceovers or dubs from that disclosure requirement. It can attach its own label from C2PA metadata, and repeated failure to disclose can carry platform penalties.^[15] Those exemptions concern YouTube's altered-content disclosure, not an exemption from truthful advertising or someone else's voice rights.

Think of release as a series of gates rather than one “compliant” toggle. First, verify the depicted experience. Second, identify material ties and any performer who could be mistaken for a consumer. Third, decide whether the destination asks for an AI label. Finally, inspect the actual exported ad and the live placement. A failed first gate is not something a later gate can cure. This order also makes revision cheaper: rewriting a misleading testimonial before animating its avatar costs less than replacing every rendition after approval. The sequence is an illustrative production workflow, not a new legal test.

Keep an export note with the destination, the question answered in its upload flow, the creative disclosure visible in the cut, and a screenshot of the live placement. Platform badges can sit near a handle or behind a menu; the advertiser's own material-connection disclosure still has to be noticeable in the ad. The FTC staff guidance says a disclosure in a TikTok text description alone is very unlikely to be enough, and one placed at the beginning of a video is better than one saved for the end.^[16]

For a paid creator, put the relationship words in the shot list rather than asking the editor to add a tiny badge at export. For a product-only demonstration, specify which frames genuinely show the claimed result. For a synthetic speaker, write the statement of role next to the lines whose delivery could imply a real customer. Give each reviewer the same cut and the same claim ledger; otherwise one reviewer may approve the script, another the avatar, and neither the combined impression. If you cannot express the role and claim plainly in the first viewing, revise the creative before debating the exact disclosure wording.^[1]

Prove the product claim before you render it#

A truthful actor disclaimer does not prove a performance claim. The FTC requires a reasonable basis before an objective advertising claim runs and considers both express and implied interpretations. Claims saying “tests prove” demand the evidence they promise. Satisfied-customer statements generally cannot substitute for objective support for claims needing evaluation.^[17][18] Put the substantiation next to the script line in the brief, not in a folder somebody might assemble after a review notice.

The same applies to synthetic before-and-after shots. A polished simulation may imply results a product cannot deliver. TikTok's advertising policy rejects exaggerated effects and before-and-after comparisons that distort the likely outcome; it also expects the landing page offer to match the ad. Meta bars deceptive or exaggerated success claims. In the UK, the ASA says an AI-made cosmetic result that does not reflect real-world results is likely misleading even when labeled as AI.^[5][19][20] The audience sees the implied outcome before they parse the provenance badge.

Build a claim ledger for each cut: exact spoken or printed words; what the shot implies; source of proof; who checked the source; and the destination page that backs up the offer. For a demonstration, keep raw footage or a reproducible product test. For a testimonial, keep the original words, relationship and permission. For a comparison, record comparable conditions. For a number, check whether it is a typical outcome rather than a selectively chosen extreme. The FTC warns that a “results not typical” tag does not cure the impression that a featured result is representative; where typical results are known and differ, the generally expected result needs a clear disclosure.^[1]

An AI presenter can help explain a real mechanism, but the prompt should not invent data, customers, affiliations or permissions. A tool can preserve an asset's provenance; it cannot certify a customer's experience. The EU Article 50 text, cited in the research through a secondary, unofficial mirror, describes a provider duty to mark certain generated outputs in machine-readable form. The Commission's primary FAQ separately explains that people still need a first-exposure disclosure for covered deepfakes.^[21][6] Confirm the governing consolidated text and your role before treating a tool's metadata as compliance. Azelify can help assemble a cut; you own the script, rights, evidence and upload decisions.

Worked examples: four briefs, four different answers#

These are fictional production scenarios, not rulings or promises that a platform will approve a specific upload. Assume the product demonstration in each case is truthful; if it makes a separate efficacy claim, return to the claim ledger. A line can be rewritten instead of discarded when the underlying product evidence is real.

The actor case is the easiest to mishandle. In the UK, CAP's current advice says stock imagery alongside a genuine testimonial was not necessarily unreasonable in one ruling, while the CAP Code still requires evidence, permission and no misleading impression.^[22][8] In the US, the FTC's rule about another person's likeness turns on whether a material attribute is misrepresented.^[1] Neither is a blanket authorization to attach any face to any quote. Record which person is speaking, whether the shot purports to show that person, and what disclosure the audience can actually see.

If a collaboration uses a creator's own account, check both the creator's label and the advertiser's placement. A gifted item, affiliate arrangement or paid appearance changes the relationship even if the review is sincere. In the UK, the ASA's influencer guide prefers a clear “Ad” label up front; its legal references have not all caught up with the newer CAP rule numbering, so use it for placement advice and consult the current Code for rule text.^[23][8] In the US, a platform disclosure button may not be sufficient if the audience misses it on a phone.^[1] For legible, visible in-video text, use the vertical caption guide.

Common mistakes#

Writing an experience first and finding a face later. This reverses the evidence chain. Start with a real person and their permitted words, or script a demonstration that does not pretend to be somebody's purchase history. The FTC's avatar FAQ is explicit that the issue is the false underlying testimonial, not the mere availability of an avatar.^[9]

Treating one badge as a universal release form. “AI-generated,” “Ad,” and “paid partnership” describe different facts. Test the viewer's inference about who spoke, why they spoke, and whether the result happened; make the needed disclosures in the video and the platform flow. A label that arrives behind a menu in one region may appear differently in another.^[13][14][1]

Moving the disclosure to the final frame. The viewer may have left before the end, and an early claim already shaped the story. Place the explanation where the misleading inference would form; if the claim is spoken and printed, check both channels. Watch the export at phone size over both bright and dark footage.^[16][1]

Using a real review to support an unrelated statistic. One customer's enthusiasm does not establish typical results or substantiate a technical claim. Keep the review's source separate from the product test. The ad should survive even if the avatar, actor or testimonial shot is removed.^[1][17]

Assuming all destinations use the same AI threshold. TikTok distinguishes organic from paid policies; Meta's labeling changes by ad category, tool origin and region; YouTube exempts some production assistance while requiring disclosure of certain realistic altered scenes.^{[10][5][13][15]} Check the exact destination after the final edit, not when the initial brief was written.

Checklist#

• Write the likely viewer inference for every speaker, result and implied customer experience; remove any invented experience.
• Keep the original customer words, dated permission and material-connection record when using a testimonial; identify actors or synthetic presenters when their role would otherwise mislead.^[1][8]
• Put objective proof beside each express or implied product claim before approving the cut.^[17]
• Check the paid-ad and organic AI rules separately for every destination, including the current regional publishing flow.^[5][13][15]
• Put necessary disclosures where they are noticeable in the video, on a phone, with sound on and off; do not rely solely on the description or platform badge.^[1][16]
• Capture the finished export and delivered placement, then re-check any rule or policy that changed between scripting and publication.