AI UGC video: when it sells and when it looks cheap

Last Tuesday I had two fifteen-second cuts side by side on the monitor. Same product, a frosted jar of face cream. Same script, same length, same bathroom. One looked like a woman who had just found something that worked. The other looked like an advert pretending to be that woman.

On the big monitor nobody in the room could say which was which. Then we played them on a phone at arm's length and it took about two seconds. The second cut had a face nobody has ever met, lit by a light nobody owns, saying a sentence nobody says out loud.

That gap is the whole subject of this post. AI UGC is the format brands ask us about most. It is also the one that goes wrong most publicly, because it fails in front of an audience that has watched ten thousand real phone videos and knows exactly what one looks like.

What is AI UGC video?

AI UGC video is advertising made to look like a customer filmed it on a phone, except the person, the setting or both are generated rather than shot. The script is still written, the product is still real, and someone still directs every take. It borrows the grammar of user content, handheld framing, a bathroom shelf, a direct address to camera, without a creator in the loop.

Brands come to it for three reasons: cost, speed and volume. A creator-sourced UGC round typically runs somewhere between 150 and 600 euros per video once you add usage rights, and takes two to three weeks from brief to files. A pack of ten AI variants from a studio like ours lands in about a week, and the per-video cost falls as the pack grows, because the expensive part is the set-up, not the tenth take.

None of that matters if the result looks cheap. So the first question is not "can we", it is "should we, for this product".

When does AI UGC actually sell?

It sells when the product does the talking. Four conditions: the product is visually clear, the pitch can be shown rather than asserted, the claims are low-stakes, and you need volume to test against. If three of the four hold I would recommend it. If one holds, I would not.

A visually clear product. A frosted jar, a drinks can, a phone case, a pair of socks. Something a hand can hold up to a window. A dashboard, a subscription or a feeling does not work here, because the format has nothing to point at.

A demo, not a verdict. Open the lid, swatch it on the wrist, pour it, click it shut. The hand-and-product shot is the most reliable thing we make, because the product is real and the hand only has to hold it. A generated person demonstrating is honest; a generated person vouching is not, and I come back to that below.

Low-stakes claims. "Smells like grapefruit" is fine. "Cleared my eczema in a week" is not a sentence to put in a generated mouth, whatever the product does. Health, money and safety claims should come from real people with real contracts, and in most markets they also carry disclosure obligations. I wrote up the rules as I understand them in the disclosure post.

Volume for testing. The honest economics of UGC are that most variants lose. Of a batch of twenty, two or three usually earn their spend and the rest exist to find those two. AI is good at the rest. It lets you test twelve hooks, three settings and two voices before you pay a creator to make the winner properly.

When does AI UGC look cheap?

It looks cheap when it pretends to be something it is not. The four failures I see most, in order of damage: fake testimonials, generic faces, over-polish and a voice that does not match the body. Each is a small lie the viewer catches before they can name it.

Fake testimonials

A generated person saying "this changed my skin" is not a testimonial. It is an actor without an actor's contract, and audiences treat it as one the moment they sense it. Keep generated people to demonstration and description. If you want a verdict, get a real one from a real customer and film them on their own phone.

This is also where the legal exposure sits. A fabricated endorsement is a problem in the US under truth-in-advertising rules and a problem in the EU under consumer law before anyone gets to the AI-specific disclosure obligations. Check with counsel, but my working rule is simple: no opinions from people who do not exist.

Generic faces

Generators drift toward a median face. Symmetrical, about twenty-eight, lit from the front, no pores, no history. Audiences read that face as "advert" within a frame or two, which is the one thing UGC is meant to avoid.

The fix is direction. We write people, not demographics: a gap in the front teeth, a jumper with a worn cuff, reading glasses pushed up into the hair, a kitchen with crumbs on the table. Ask for someone your audience could sit next to on a bus. Then keep that person consistent across the pack, which is its own discipline and the subject of the consistency post.

Over-polish

UGC's entire value is that it looks like nobody paid for it. Generated footage defaults to the opposite: shallow depth of field, a frame locked on a tripod, a key light that costs more than the bathroom. The result is beautiful and dead.

We spend real time making footage worse on purpose. Handheld drift, a tap in the corner of frame, a slightly blown window, a grade pushed toward a phone sensor rather than a cinema camera. The full list of what gives a shot away is in the uncanny checklist; for UGC the rule is to run that list backwards and put some of the flaws back in.

Mismatched voice

The voice is too clean for the room, or its rhythm does not match the shoulders, or the lips land a quarter-second behind. Lip-sync is the loudest tell in the whole format, and it gets louder the longer the on-camera line runs.

Fixes, in order of cost: cut to the product while the voice continues, keep on-camera lines under three seconds, and record the voice with a real person in a real room so at least one half of the pairing is true.

UGC's entire value is that it looks like nobody paid for it, so most of the work is making expensive footage look cheap on purpose.

How does a studio build an AI UGC pack?

Like a shoot, with a shot list instead of a prompt box. We brief, write the hooks, pick two or three settings, photograph the real product for plates, generate takes against that plan, then select, fix, grade and cut. The generation step is the shortest part of the week.

For a pack of ten finished variants we typically generate somewhere between eighty and a hundred and twenty takes. Most are rejected in the first pass on a phone screen. The survivors get the product plate composited in where the geometry matters, a grade that matches them to one another, and a cut that keeps every on-camera line short.

The thing clients underestimate is the product. A face can be approximately right and pass; a jar cannot. The lid is the wrong height, the frosting is too even, the logo has become a shape. So for anything with a label or a specific silhouette we shoot the real object once, under the light the scene will use, and bring that in. It is slower, and it is the difference between a pack the brand will run and one they will quietly bury.

Should your brand use AI UGC at all?

Use it for hooks, demos and volume, and use real people for verdicts. That split holds up for nearly every brand I have talked to, and it is cheaper than either extreme.

A short decision list, the one I run on calls:

• The product fits in a hand held up to a window. If it does not, this is probably not the format.
• The strongest claim is something a stranger could verify by watching the video. If it is not, the strongest claim needs a real person.
• You need more than five variants. Under five, a creator is often the better buy.
• You are willing to disclose that the people are synthetic where the platform or the law asks. If that disclosure would kill the ad, the ad was leaning on a lie.
• The product has text, a logo or a precise shape. Then budget for product plates, not just generation.

The hybrid most of our clients land on: AI UGC for the top of the funnel, where you are testing which angle stops a thumb, and creator UGC for the proof, where a real person says what they think. The AI pack finds the message; the creator delivers it.

What I'd tell you on a call

Bring the product and the one claim you most want to make, and I will tell you in about five minutes whether this format can carry it. Most of the time the answer is "yes for the demo, no for the verdict", and the pack we scope reflects that.

If you want to see what a finished pack looks like before we talk, the work on the front page has a few. If you want to know how we catch the cheap-looking shots before they reach you, start with the uncanny checklist, then how we keep one face and one jar consistent across forty shots.