UGC-style video with AI: where the line is

The interesting question is not whether that works. Often it does. The question is whether the person watching should be able to tell, and that is a question about honesty before it is one about quality. This article takes it in that order, because the legal position in Europe is moving and because we would rather state where we stand than quietly find out later.

Note upfront: this is not legal advice. We refer to one piece of legislation, and where it affects you, your own counsel is the right person to read it with.

What people mean by AI UGC

The format is familiar: someone holding a product at arm's length, talking to a phone, in a room that looks like a room rather than a studio. It works because it reads as a person's experience rather than a brand's claim.

Generated versions of it come in roughly three shapes. A synthetic presenter speaking a script you wrote. A real voice over existing footage. And a fully generated scene where a figure handles something that resembles your product.

They are not equivalent, and lumping them together is where most of the trouble starts.

Our position, in three rules

One: a generated person may never present as a real customer. The moment a synthetic figure says "I bought this and it changed my morning", you are not producing content, you are manufacturing a testimonial. That is the line for us, and it does not move when a campaign is underperforming.

Two: a real person's face or voice only with explicit permission, in writing, for the specific use. Cloning a voice from a clip someone posted is not a grey area to us, whatever the tool's terms of service permit.

Three: say what it is. If the person in the clip does not exist, the viewer should be able to find that out without detective work. In practice that means the label the platform provides, plus not building the creative on the pretence.

Those three cost you a certain kind of ad, and they are cheap compared to the alternative. Trust is the asset the entire UGC format borrows against. Spend it once and the format stops working for you, not just for that campaign.

The rule that is arriving

The EU AI Act sets transparency obligations for AI-generated content. Article 50 covers this, and it applies from 2 August 2026. The short version for a marketing team: content that is artificially generated or manipulated has to be identifiable as such, and that includes material you put out commercially. If you sell into the European Union, this reaches you regardless of where your agency sits. We referred to the same article when writing about using AI images commercially, because the obligation is not specific to video.

Alongside that, the large social platforms have their own labelling requirements for significantly altered or synthetic content, and they change more often than legislation does. Check the current rules of the platform you are advertising on before a campaign goes out rather than after.

One sentence of practical advice: decide your labelling convention once, write it into your brand guidelines, and apply it everywhere. Deciding it per campaign is how inconsistency gets in, and inconsistent labelling looks worse than no policy at all.

Where generated UGC-style content genuinely helps

Being careful is not the same as being against it. Three uses we think are sound.

Testing scripts before you commit. Produce four versions of a message and find out which one holds attention, before you book a person for a day. This is the strongest case by far, because at that point you are testing copy, not borrowing credibility.

Language variants. The same demonstration in five languages for five markets, for an audience that would otherwise get nothing in their own language. If you sell through dealers abroad, this is often the difference between a dealer using your material and making their own, which is the heart of when you do not own the shelf.

Product in context, without claims. Hands, use, scale, environment. No testimony, no "I love it", just the product doing what it does. The same limits apply as to any generated video, which we set out in what generated video does well.

Where it does not work

Anything that trades on experience. Taste, smell, comfort, fit, how something feels on skin. A synthetic figure saying it works is a sentence with nothing behind it, and audiences read that faster than marketers expect.

High-trust categories. Health, finance, safety, anything for children. The regulatory exposure is higher and so is the reputational cost.

Community. Real customers posting about your product is not a content format you can replace, it is a relationship. Generated clips do not build it and can undermine it.

A checklist before a generated clip goes out

Five questions. If any answer is uncomfortable, the clip is not ready.

1. Does anything in this clip claim an experience? Taste, comfort, result, "it changed my routine". If yes, a real person has to say it or nobody does.

2. Would a viewer who found out it was generated feel misled? Not "would they mind", would they feel misled. Those are different questions and only the second one matters.

3. Is every face and voice in it either synthetic or cleared in writing? No third option.

4. Is it labelled the way we label everything else? Same convention, every channel, including the ones where nobody is checking.

5. Does the product in it match the real product? Colour, marks, proportions. A clip can be honest about being generated and still be wrong about your product.

The fifth one catches more clips than the first four combined.

How we work with this

We produce video the way we produce stills: generated, with a creator from our own team reviewing and correcting before anything is delivered. Those creators are colleagues, not a marketplace, and a shoot with a creator is something you can book as a separate product when the material needs to be real. That combination is the subject of a person in the loop, and the wider choice between routes is in the production routes compared.

What we will not do, for any client, is produce a synthetic customer giving a testimonial about a product. If that is the brief, we will say no and offer the version we think actually works.

Want to see how we handle this in practice? See how it works, or try LensLab free with five credits.