This article is general information, not legal advice, and it contains no cost or performance figures — those depend entirely on your market, format and brief. Platform policies and AI rules differ by country and change often.
You need a person on screen saying something about your product. There are three realistic ways to get one, and they are not simply cheap, medium and expensive versions of the same thing. They differ in what you own afterwards, what you are allowed to claim, what you have to disclose, and how long the ad stays usable.
- A generic AI avatar — a synthetic presenter from a tool’s library, resembling no particular real person.
- A licensed AI likeness — an AI version of a real, identifiable person who has agreed to it for your brand.
- A conventional shoot — a human in front of a camera, with a signed release.
Here is how they actually compare.
What you hold at the end
This is the question most briefs skip.
With a generic avatar, you hold rendered video and whatever your tool’s terms grant. Read them: some licences restrict advertising use, resale, or using the same avatar in categories like health and finance, and some reserve the right to retire an avatar from the library. You do not hold anything exclusive — a competitor can use the same face.
With a licensed likeness, you hold video plus a licence: a named person who agreed to a named brand, in named countries, for a stated period. The value is in the record. If anyone asks whether the person on screen consented, you can answer with a document rather than an assurance.
With a shoot, you hold footage plus a release. The limit is usually in the release, not the footage — a usage period that expires, or territories you did not buy. Plenty of brands discover years later that an ad they still consider theirs is out of licence.
Speed, volume and iteration
Synthetic routes win here, and the win is not mainly about the first video.
A generic avatar gives you same-day turnaround and effectively unlimited variants: re-record a line, swap a price, test six hooks. A licensed likeness sits in the middle — generation is fast, but a well-run licence puts script approval and brand approval in front of it, which is friction by design. A shoot concentrates everything into one day: anything you did not capture needs another one.
If your campaign depends on weekly iteration, that difference matters more than any other item in this article. If you are making one brand film a year, it barely matters at all.
Credibility, and the line you cannot cross
All three can present information. Only one can give a genuine testimonial.
The FTC’s endorsement guidance is direct: “an endorsement must reflect the honest opinion of the endorser and can’t be used to make a claim the marketer of the product couldn’t legally make”, and “you can’t talk about your experience with a product if you haven’t tried it” (FTC). The Commission’s rule on consumer reviews and testimonials, published in the Federal Register in August 2024, addresses testimonials that misrepresent that they come from someone who does not exist or who lacks actual experience with the product (FTC).
Read that against the three options:
- A generic avatar can explain a feature, demonstrate a process or read a factual claim. It cannot have an opinion, because there is nobody there to have one. “I switched and I love it” from a presenter who does not exist is the category the rule is aimed at.
- A licensed likeness can carry a genuine endorsement if the person has actually used the product and has separately agreed to that specific testimonial. Licence plus script approval is not enough on its own; a real experience and an explicit yes are both required.
- A shoot has the same requirement. Being physically present does not make an untrue endorsement true.
The line sits in the same place when the ad is styled as user content rather than a straight presenter read, and the implied authenticity of that format makes it easier to cross — see AI UGC ads and where the testimonial rules bite.
So the decision is partly a script decision. Write the ad first, then ask which route can legally deliver those words.
Disclosure: two of the three need a label
Both AI routes attract transparency duties; the shoot generally does not.
Platforms ask for it. YouTube requires creators to disclose content that “makes a real person appear to say or do something they didn’t do”, or that “generates a realistic scene that didn’t actually occur” (YouTube Help). Meta requires people to “disclose, using our AI-disclosure tool, whenever they post organic content with photorealistic video or realistic-sounding audio that was digitally created or altered” (Meta). TikTok’s advertising rules prohibit using “the likeness (audio or visual) of a public figure without their permission” in misleading contexts (TikTok Ads Policy). Our platform-by-platform guide sets out where each switch lives.
Law adds to it. In the EU, deployers of a deep fake must disclose that content has been artificially generated or manipulated, and providers must mark synthetic outputs machine-readably (Article 50) — see our Article 50 guide. Note that a photorealistic generic avatar can fall inside a deepfake test even though it depicts nobody real, because the test is about realism and plausibility rather than identity.
Either way, plan the label into the creative. A disclosure retrofitted onto a finished cut always looks like an apology.
Where the legal exposure sits
Each route has a different main risk.
Generic avatar: that the face is not as generic as the vendor thinks. Ask how the avatar was created, whether the person it was built from consented to advertising use, and what indemnity you have if a claim arrives. “It’s synthetic” is not an answer to “who is that?”.
Licensed likeness: that the consent is thinner than the paperwork suggests. The newer US statutes make vague grants legally weaker, not just ethically worse — California and New York have made certain replica contract terms unenforceable where the intended uses are not described with reasonable specificity and the person was not represented (AB 2602; S7676B). More in digital replica laws in the US.
Shoot: that your rights lapse while the ad is still running, or that you reuse footage outside the territory or period the release covered.
All three share the ordinary rule that an ad must not mislead. No production method fixes a claim you cannot support — see how to write a script for an AI spokesperson ad.
Localisation, changes and shelf-life
Languages and markets. Synthetic routes make many-language versions straightforward; a shoot generally means a performer per language, or subtitles. With a licensed likeness, check the licence actually covers the languages and countries you are planning — this is a common overreach.
Changes. Price moves, offers end, products get renamed. Avatars and licensed likenesses can be re-rendered. Filmed footage cannot be re-spoken.
Shelf-life. All three expire, for different reasons. Avatar terms and library availability change. A likeness licence has an end date and a renewal fee. A model release has a usage period. Put whichever applies in a calendar at the point of launch, not when someone notices.
Cost shape, not cost
We will not invent numbers, but the shape of the spend differs predictably.
A shoot front-loads: crew, talent, location, post, and a real cost for every change afterwards. An avatar tool is usually a subscription, so the marginal cost of the tenth variant is close to the first, and the real cost is the licence terms you accepted. A licensed likeness is typically per licence and per renewal — the ongoing payment is the point, because it is what keeps the person’s consent live and current rather than bought out once.
A short decision guide
- Generic AI avatar — explainer content, feature walkthroughs, high-volume testing, factual scripts with no personal endorsement, and audiences you are confident will not read “real person” into it.
- Licensed AI likeness — you want a specific, recognisable human with a genuine relationship to the product, you need frequent updates or many markets, and you need to be able to prove consent.
- Conventional shoot — flagship brand work, physical demonstration, performance nuance, documentary feel, or any format where audiences would react badly to synthetic presentation.
And mixing is normal: a shoot for the hero film, a licensed likeness for the always-on social cut, avatars for internal and explainer video. Decide per format rather than per brand.
Questions to ask any vendor
- Who is the person this presenter was built from, and what did they consent to?
- Does your licence permit paid advertising, in my categories, in my countries?
- Is the presenter exclusive to me, or shared with anyone, including competitors?
- How are outputs marked as AI, machine-readably and visibly (content credentials)?
- What happens if the person withdraws consent, or the avatar is retired?
- What indemnity do I have if a likeness claim arrives?
- When does my right to keep running this ad expire?
Where Likeness fits in
Likeness is being built for the middle option: real, verified adults who have agreed to specific brands and uses, with a licence record a business can point to, labelled output, and a payment each time a licence is issued or renewed. It is in development and not yet open — see for businesses and trust and consent, or email [email protected].
General information only, not legal advice, and no claim that any production method is compliant with any rule. Platform policies and AI transparency laws change often — check the current versions for your markets before you commit to a format.
Sources
- Federal Trade Commission, FTC’s Endorsement Guides: What People Are Asking
- Federal Trade Commission, 16 CFR Part 465: Rule on the Use of Consumer Reviews and Testimonials (final rule)
- YouTube Help, Disclosing use of altered or synthetic content
- Meta Transparency Center, Community Standards: Misinformation
- TikTok Ads Policy, Misleading and false content
- European Commission AI Act Service Desk, Article 50: Transparency obligations
- California Legislature, AB 2602 (2024)
- New York State Senate, S7676B
- C2PA