AI UGC ads, and where they stop working
Synthetic creator video is cheap, fast and genuinely useful for a narrow set of jobs. It also cannot hold your product, cannot testify to using it, and carries disclosure questions most brands have not thought about.
- AI is good at the parts with no product in frame: openings, captions, b-roll, translations and variant testing at volume.
- It is bad at the parts that carry the persuasion: hands on the product, texture, scale, and a person who has actually used the thing.
- A synthetic person implying a real customer experience is a claim about a customer who does not exist. That is the risk nobody prices.
- Platform rules on labelling synthetic media change, so check the current policy for each platform before the campaign, not after.
- A practical split is synthetic for volume and variants, filmed for anything that has to be believed.
Generated video reached the point where a talking face in a kitchen can be produced in minutes for the cost of a coffee, and every brand buying creator content has now been asked why it should pay a person instead. It is a fair question and it deserves a specific answer rather than a defensive one. The honest position is that synthetic footage does some of this job well, does some of it badly, and does one part of it in a way that creates risk a cheap asset does not offset. This page sets out which is which, what to settle before you run any of it, how the two approaches compare on the things buyers actually care about, and where we draw the line in our own work and why.
What AI UGC actually means in practice
The phrase covers several different things that get sold under one label, and they carry different levels of risk. Separating them is the first useful step.
| What it is | Risk level | Sensible use |
|---|---|---|
| Scripts, hooks and caption variants written by a model | Low | Volume of ideas, then a human edit before anything runs |
| Editing assistance: cuts, subtitles, reframing, dubbing | Low | Production speed on footage you already own |
| A generated presenter reading a script to camera | Medium to high | Explainer style content that makes no personal claim |
| A generated likeness of a real person | High | Only with that person's explicit written permission |
| A generated person implying they used your product | High | A claim about an experience nobody had |
Most of the cost saving people are excited about sits in the first two rows, which are uncontroversial and which almost everybody already uses. The argument is really about rows three to five.
Where synthetic video holds up
It works where the video is carrying information rather than credibility. If the viewer does not need to believe a person, a generated one is often fine.
- Opening variants. Testing twenty ways to state the same problem is cheap and tells you what to film properly afterwards.
- Explainers for software and services, where there is no physical object and nothing to hold.
- Localisation. Getting the same message into several languages without booking several shoots.
- Filler and connective footage between the parts that matter, which viewers do not scrutinise.
- Early stage testing before a product exists, to find out which message pulls before committing to production.
- Static ad variants and thumbnails, which face far less scrutiny than a face speaking.
Used this way it is a research tool. You are buying speed of iteration, then spending the filming budget on whichever message survived. That sequence gets more from both than either does alone, and it is how we test openings before a batch is commissioned.
Where it breaks
It breaks wherever the video has to prove something physical or something experienced. Four failures come up repeatedly and they are not the kind that a better model fixes soon.
The product in the hand
Scale, weight, texture, how a lid actually opens, how thick a cream is, how a fabric moves. A generated approximation of your packaging is not your packaging, and the difference is visible to the person who is deciding whether to buy the real one.
The claim of having used it
This is the important one. A face saying it changed my routine is a testimonial. If no person had that experience, the advert is describing something that did not happen. That is not a style question, it is a truthfulness question, and it lands on the advertiser rather than on the tool.
Audience recognition
Audiences are getting quicker at spotting generated faces, and the reaction in comments is rarely neutral. The format's only advantage over polished advertising is that it looks like a person. Spending the saving to remove the advantage is a poor trade.
Ownership of the output
Rights in generated material are unsettled and vary by jurisdiction and by the terms of the tool you used. Filmed work with a signed assignment gives you something you can point to. Check what your tool's terms actually grant before you build a brand asset on it.
Three things to settle before you run any of it
None of these takes long, and all three are much harder to sort out after an advert has been running for a month.
- Disclosure. Decide how the advert indicates that the person is not real, and check the current labelling policy of every platform you will run it on. Those policies have changed repeatedly and are worth reading rather than remembering.
- Likeness. If anything resembles a real person, get written permission from that person for that use. A creator's face is theirs, and a licence to run their video is not a licence to generate more of them.
- Claims. Every statement in the script still has to be supportable by evidence you hold. A synthetic speaker does not lower the bar, and in a dispute it makes the advertiser the only party in the room.
Canadian rules treat advertising that creates a false impression as a problem regardless of how it was produced, and content reaching United States audiences sits under the FTC endorsement guides, which are concerned with whether a testimonial reflects a real experience. Neither framework cares which software was used.
A sensible split of the budget
Treat them as different tools rather than rival answers. The question is which job each is doing in your funnel.
| Job | Use | Why |
|---|---|---|
| Finding the message | Generated variants at volume | Cheap enough to be wrong twenty times |
| Proving the product works | Filmed, with a real person | Hands, scale and texture cannot be faked convincingly |
| Testimonial and review content | Filmed only | An experience nobody had is not a testimonial |
| Explaining a service with nothing to hold | Either | No physical proof is involved |
| Localising a proven message | Generated dubbing of filmed footage | Keeps the real person, changes the language |
| Scaling what already works | Filmed, in batches | The winner deserves real production, not more variants |
In money terms that usually means a small experimental budget for generated variants and the bulk of the spend on filmed assets once you know what to say. Our fixed-price packages are built for that second stage, and the budget calculator will size the first.
What we do, and what we will not do
Every asset we deliver is filmed by a person we commissioned, offered a fee in writing and paid. We use software for the ordinary production jobs: subtitles, reframing, cutting and drafting hook variants that a person then rewrites.
We will not generate a person who implies they used your product, and we will not generate a likeness of any creator beyond the footage they actually filmed for that job. Our creator agreement takes an assignment of specific delivered work, and it is not a licence to manufacture more of somebody. That limit exists because both the creators and the brands we sell to are better served by content whose provenance we can state plainly.
If your budget is genuinely too small for filmed content, we will say so and point you at product seeding or at hiring one or two creators directly, both of which are written up here for nothing. That is a better answer than selling you synthetic testimonials at a price that looks like a bargain.
Questions
Is AI UGC cheaper than hiring creators?
Per asset, obviously yes, and that is the whole argument for it. The comparison that matters is cost per usable asset in your ad account. A cheap video that nobody watches past three seconds costs more than an expensive one that sells, and neither approach gets a pass on that test.
Do I have to disclose that an ad was generated?
Platform labelling requirements for synthetic media exist and have changed more than once, so check the current policy for each platform rather than relying on a summary. Separately from platform rules, advertising that creates a false impression of a real customer experience is a problem in both Canada and the United States regardless of production method.
Can AI copy a creator's face if we already licensed their video?
Not on the basis of that licence alone. Rights in the footage and rights in a person's likeness are separate things. Generating more of somebody needs their specific written permission for that use, and our agreement with creators does not grant it.
Does generated content perform worse?
It depends entirely on the job. For openings, captions and explainers the difference is often small. For anything that has to be believed, the performance gap tends to be wide, because the format's advantage is that it looks like a person rather than an advert. Test it in your own account rather than trusting anyone's published figure, including ours.
Who owns a generated video?
Read the terms of the tool that made it, because what you get varies and the law in this area is still moving. Filmed work with a signed assignment is the clearer position, which matters if the asset becomes a significant part of your advertising.
Will you make AI ads for us?
We supply filmed content made by people we commissioned and paid. We use software for subtitles, reframing and drafting hook options that a person rewrites, and we say so. We do not produce synthetic people presented as customers, and we will tell you plainly when your budget suits seeding or direct hiring rather than a package.
General information, not legal advice. Platform policies on synthetic media change, so confirm the current rules before running a campaign.
Get a shortlist.
Tell us the product and the budget. Names, rates and recent work come back at no cost.