TL;DR
- "Can I generate it" and "can I run it as an ad" are two different questions, and only the first one has an easy answer.
- Ad platforms set their own disclosure rules for synthetic media, and those rules change. Check the platform's current policy at campaign time — not the one you read about last year.
- Cloning a real person needs written, specific, time-bounded permission. "They said it was fine" is not a permission structure, and consent given for one campaign is not consent for the next one.
- A stock AI avatar removes the consent problem and introduces a different one: you don't control who else is using that same face, including your competitors.
- The liability doesn't move. If an AI presenter states a spec, a price or a claim that turns out to be wrong, the advertiser answers for it — the generator does not.
The generation question stopped being interesting some time ago. You can produce a presenter who speaks four languages in an afternoon, and the output is good enough that most viewers won't stop scrolling to interrogate it.
The questions that actually stop campaigns are elsewhere: who agreed to this face existing, does the platform need you to say it's synthetic, and who answers when it states something that isn't true. None of those are solved by better generation.
Do I have to say the ad is AI-generated?
Sometimes, and the answer is a moving target rather than a fact you can memorise.
Ad platforms have been adding disclosure requirements for synthetic and digitally altered media, particularly where a real-looking person appears to say or do something. The direction of travel has been consistently toward more disclosure, not less. What varies is which content is covered, how the disclosure must appear, and whether it is a self-declaration in the ad manager or something visible in the creative itself.
The practical rule is procedural, not factual: check the current policy of the platform you're actually buying on, at the time you launch. Not the version summarised in an article — including this one. A policy you read about six months ago may have been replaced, and the rejection notice won't explain which clause you missed.
Build this into the campaign checklist next to the budget approval. It takes ten minutes and it is the cheapest step in the entire process.
What do I need if we're cloning a real person?
Written permission, specific about what gets generated, and bounded in time.
The gap we see most often isn't malice — it's informality. Someone senior says "yes, use my voice," everybody proceeds, and eighteen months later there is a synthetic version of that person in a campaign nobody discussed with them, possibly after they've left the company.
Three things the permission should actually name:
What will be generated. A voice model is a different thing from a full presenter, and a presenter who can be given new scripts is different again from a fixed set of renders. People consent to what they picture. Make sure the picture is accurate.
How long it lasts, and what happens at the end. Including whether the trained model is deleted or retained. This is the clause that is almost always missing, and it is the one that matters most after someone leaves.
Which campaigns and platforms. Consent for a recruitment video is not consent for a paid social campaign. Broad permission is easier to get once and much harder to defend later.
Two things are true at once here, and it's worth separating them. A production partner should require that basic permission exists before building a model — that's a floor, and you should expect it to be asked for. But the responsibility for that permission being real, adequate and still valid remains yours. The studio checking that something was signed is not the same as the studio having obtained informed consent on your behalf, and it will not be treated as such if it's ever disputed.
If nobody has asked you for a release before a model was built, that is a signal about the process, not a convenience.
Is a stock AI avatar safer?
It removes the consent problem cleanly. There's no real person whose permission can expire or be withdrawn, which is the sharpest risk in the whole category.
It introduces a commercial one instead: the same face is available to everyone. Including your competitors, including businesses in unrelated categories doing things you'd rather not be visually associated with, and including whoever is generating low-quality content in your market this month.
For a one-off explainer, that's a non-issue. For a brand intending to build recognition around a consistent presenter — where the whole point is that people start to know the face — it's a structural limitation. You're building recognition on a shared asset.
There's a middle path some brands take: commission a custom avatar that isn't a clone of a real employee and isn't off the shelf. It costs more and it solves both problems at once.
Who's responsible if the AI says something wrong?
The advertiser. This one is not ambiguous, and it's the point people most often assume has changed.
A generated presenter is a production method. If the ad states a price that isn't current, a specification that's wrong, or a guarantee the business can't honour, the responsibility sits exactly where it sat when a human actor read the same line off a card. Nothing about the generation method transfers liability to the tool.
The practical consequence is worth stating plainly: AI output needs the same factual sign-off as a scripted shoot, not a lighter one. In practice it often gets less, because generation is fast and the review process wasn't designed for something that can produce forty variants before lunch. Speed is exactly why the sign-off matters more, not less.
If you're producing many variants, the check that scales is on the source — approve the facts once, in a form the variants are generated from, rather than reviewing forty finished videos and hoping.
In practice that means script approval, before generation. Not sign-off on the finished renders. The order matters more than it looks: approving renders means you are checking forty outputs for a mistake that was made once, upstream, and the odds of catching it in the fortieth are not good.
The principle underneath it is worth stating flatly, because it is the one thing about AI production that should not be negotiable: the facts stay human-gated. Generation can be automated. Whether a claim is true cannot be, and no volume of output is a reason to relax that.
Where these actually earn their place
The honest version, which is narrower than most of the marketing around this category:
Volume and repetition. Several language versions of the same content. Material that must be reissued whenever a specification or a price changes. Training and explainer content where the presenter is delivering information rather than vouching for it. In that territory the economics are genuinely different, because re-recording becomes a text edit and re-recording is the expensive part.
In practice the work that actually lands here is narrower than the category's marketing suggests: educational and informative talking-head pieces, AI UGC, and product demos. That's the real usage pattern, and what those three share is that the presenter is conveying information rather than lending it credibility.
Where they don't work is anywhere the point is that a real person is putting their name to something. A founder's conviction. A customer's relief. A technician who has visibly done this a thousand times. Those aren't delivery problems and no amount of render quality solves them.
A note on the absence of horror stories. We haven't yet had a case where an AI version was produced and then abandoned for a real shoot. That is worth saying and worth not over-reading: it reflects a small number of projects, deliberately kept inside the narrow band above. It is not evidence that the approach generalises past that band, and it shouldn't be used as one. The honest summary is that within a well-chosen scope it has held up, and the scope was chosen conservatively.
Before you launch
Four checks, none of which take long:
- The platform's current synthetic-media disclosure policy, read this week
- Written permission for any real person whose likeness or voice is used — specific, time-bounded, naming the platforms
- Factual sign-off on the claims, at the source rather than per variant
- A decision, written down, about what happens to a trained model when the person leaves or the campaign ends
None of these are about whether the technology is good enough. It is. They're about the fact that an ad is a statement your business is making, and the method of production doesn't change who made it.
Related reading on this site: AI-Generated Content for Brands covers where AI content fits overall, and When Not to Use AI Video covers the cases where a real shoot is the cheaper answer.
Thinking about a project like this?
Tell us what it needs to do — we'll take it from there.
Frequently asked questions
- Do I have to disclose that an ad uses AI-generated people?
- Ad platforms set their own disclosure requirements for synthetic and digitally altered media, and these have been tightening rather than loosening. Because the specific rules differ by platform and change over time, the reliable practice is to check the current policy of the platform you're buying on at the time you launch, rather than relying on what was true for a previous campaign.
- Can I clone a real person's face or voice for an advert?
- Only with that person's clear permission, and the permission should be written, specific about what will be generated, and bounded in time. A verbal "sure, go ahead" from an employee or a client's staff member is not a permission structure — it doesn't survive their departure, a dispute, or a second campaign nobody discussed.
- Does the studio or the client handle permission for voice and face cloning?
- Both, at different levels. A production partner should require that basic permission exists before building any model, and you should expect to be asked for it — if nobody asks, treat that as a signal about the process. But responsibility for that permission being real, adequate and still valid stays with the client. A studio confirming something was signed is not the same as a studio having obtained informed consent on your behalf.
- Is a stock AI avatar safer than cloning someone?
- It removes the consent question, which is the sharpest risk. It introduces a commercial one instead: the same avatar is available to everyone, including competitors in your category, and you have no control over what else that face appears in. For a brand that intends to build recognition around a presenter, that's a real limitation.
- Who is responsible if an AI presenter says something inaccurate?
- The advertiser. A generated presenter is a production method, not a party to the claim — if the ad states an incorrect price, specification or guarantee, the responsibility sits where it always did. This is why AI output needs the same factual sign-off as a script read by a human actor, not a lighter one.
- Should we use AI presenters at all?
- They earn their place in high-volume, repeatable work — multiple language versions, content that gets reissued whenever a spec changes, training and explainer material. They're a poor fit where the point of the video is that a real person is vouching for something. Match the tool to the job rather than to the saving.
Keep reading
- When NOT to Use AI Video (A Production Studio's Take)A studio that runs an AI line tells you where to keep it off — the trust layer, real emotion and testimonials — and the moment high-realism AI quietly stops being cheaper than just shooting it.
- Festive Campaigns at Speed: CNY, Raya and Deepavali in DaysHow AI lets brands turn festive campaigns around in days — and the one rule that decides whether audiences love it or turn on you: frame it as extra creative ambition, never as "we saved money."
- Always-On Social Content: Why a System Beats One-Off ShootsOne-off shoots leave you with a spike and then silence. Here's why a content system — real production and AI working together — keeps your social presence alive between shoots, without pretending AI does it for free.