AI Video and UGC: What Can Be Produced Today, and What Cannot
AI video is footage generated from a text or image prompt by a diffusion model, with no camera involved. UGC, user generated content, is the opposite: real footage shot by real people, usually on a phone, deliberately rough. The two get grouped together because both promise cheap volume, and that is where most briefs go wrong. This article sets out what each format actually delivers in 2026, what the law has required since 2 August 2026, and how to build production that survives the next tool change. To be transparent: Collective Agency produces AI content creation for Swiss brands, so we have a commercial interest here. The figures below all trace back to primary sources.
What is the actual difference between AI video and UGC?
The distinction matters because the two solve different problems, and a brief that confuses them produces expensive disappointment.
AI video gives you control over things a camera cannot easily deliver: a product in a location you never travelled to, a scene with no permit, twenty variants of the same shot with a different background each time. What it does not give you is credibility. Viewers increasingly recognise the look, and in a feed full of it, polish reads as advertising.
UGC gives you exactly the opposite. It is believable because it is imperfect, and it works in placements where a polished spot is skipped. What it does not give you is control or speed. You depend on creators, on their calendars and on their interpretation of your brief.
A practical way to decide:
The message depends on trust: a testimonial, a use case, a recommendation. Use UGC. AI cannot manufacture credibility, and pretending otherwise is where the legal problems start.
The message depends on the product looking right: materials, colourways, a location, a scenario. AI video is often faster and cheaper than a shoot, and nobody is being deceived.
You need thirty variants of one idea for testing. AI video, because the marginal cost of variant thirty is close to zero.
You need one hero asset that carries a campaign for a year. Shoot it. The cost difference disappears across twelve months, and the result does not date the way generated footage does.
The most common mistake we see is treating AI video as a cheaper camera. It is not a cheaper camera. It is a different instrument with different failure modes.
What can AI video actually do in 2026, and where does it still fail?
The honest answer is that the gap between demo and deliverable is still wide, and it sits in predictable places.
What works reliably now: short shots of five to ten seconds, product and object motion, camera moves through a scene, atmosphere and texture, and stylised worlds that do not claim to be real. Synchronised speech has improved to the point where a generated voice no longer immediately gives itself away in a short cut.
What still fails, and what you should plan around:
Consistency across shots. The same character in three cuts is still work, not a setting, even with reference images.
Hands, text and logos. Anything with fine structure and a correct answer is where generation breaks visibly.
Physical accuracy of your own product. A model that has never seen your packaging will invent it, confidently.
Long takes. Beyond roughly ten seconds, coherence degrades and you are editing around artefacts rather than telling a story.
There is a cheap test before you commit a budget. Take your actual product, give a model three references and ask for five seconds of it turning on a plain surface. If the packaging comes back wrong, the shape drifts or the logo dissolves, no amount of prompting will fix it in that campaign, and you have learned it for the price of a few generations rather than halfway through production. We run this on every project before we quote generation as an option.
The consequence for planning is that AI video is a shot-level tool, not a film-level one. Briefs that assume a finished thirty second spot from a prompt are the ones that overrun. Briefs that treat generation as one department in an edit, alongside stock, shot footage and motion design, tend to land.
"A photorealistic AI image that depicts people, products or situations true to life can fall under the labelling obligation." Read the analysis.
Tobias Voßberg, IP Lawyer for Copyright, AI and Media, 2026, KI-Kennzeichnung in der Werbung nach der KI-Verordnung
What has been legally required since 2 August 2026?
This is the part most production conversations skip, and it changed recently enough that a lot of published advice is out of date. The transparency obligations in Article 50 of the EU AI Act have applied since 2 August 2026.
The structure matters more than the headline. The obligation to mark content technically falls on the providers, meaning the companies that build the generation tools. Advertisers count as deployers, and their duty under Article 50 is narrower: they must disclose when people are exposed to deep fakes, and to AI-generated content on matters of public interest published without human review.
So the common fear, that every AI-assisted frame now needs a visible label, is wrong. But the opposite conclusion is also wrong. A photorealistic depiction of a person, product or situation that a viewer would take for a recording is exactly the case the rules were written for.
In Switzerland the position is different and worth stating precisely, because it is frequently misreported:
There is no general Swiss labelling obligation for AI-generated advertising, and none has been passed. Parliament rejected a specific deepfake regulation in May 2025.
The Federal Council is pursuing its own approach, with a consultation draft expected by the end of 2026. Transparency is explicitly named as a topic.
Existing law applies regardless: data protection, personality rights, copyright and unfair competition law. Claiming a production is something it is not is already actionable without any AI-specific rule.
If you advertise into the EU, the EU rules reach you, wherever your company sits.
Our practical position: label when a reasonable viewer could mistake generated footage for a recording of something that happened. That threshold is easier to defend than any rule text, and it holds in both jurisdictions. We wrote about the image side of this separately in AI images in advertising and the legal position in Switzerland.
Why you should not build production on a single tool
There is a concrete example running out this month. On 24 March 2026 OpenAI notified developers that the Videos API and the Sora 2 models would be removed from the API on 24 September 2026. Anyone who built a production pipeline on that endpoint has had six months to move, and the deadline is now days away.
This is not a criticism of OpenAI. Their published notice periods are clear and reasonable: at least six months for generally available models, at least three months for specialised variants, and as little as two weeks for anything labelled preview. The company states plainly that preview models are not recommended for business-critical production unless you can migrate on short notice.
The lesson is about how you organise, not which vendor you pick. A production setup that survives this looks like:
Prompts and references stored as documented assets, not in a chat history that dies with the account.
A defined output specification, meaning resolution, aspect ratios, colour handling and delivery format, so a new model can be measured against the old one instead of guessed at.
At least two tools qualified for each job, so a shutdown is a switch rather than a rebuild.
Rights and provenance recorded per asset, including which model produced it. When labelling questions come up later, this is the only thing that answers them.
Teams that treat generation as a subscription tend to rebuild every time the market moves. Teams that treat it as a documented process change one line in a specification.
What a production setup that holds up looks like
The question we get asked most often is what to do first. The order below reflects what actually causes projects to fail, rather than what is most interesting to build.
Fix the brief before the tool. Most disappointing AI video traces back to a brief that described a look instead of a message. Generation amplifies whatever the brief contains, including its vagueness.
Build the reference set first. Your product, your colours, your typography, your people, captured properly once. Every generated asset afterwards is measured against it, and without it you get plausible footage of a product that is not yours.
Decide the labelling threshold in advance and write it down. Deciding case by case under deadline pressure is how inconsistent disclosure happens, and inconsistency is worse than either choice.
Separate hero from volume. One or two assets carry the campaign and deserve a shoot or serious post. The variants around them are where generation pays for itself.
Measure creative fatigue, not production cost. The saving from generating instead of shooting is real but small. The saving from replacing a fatigued creative in days instead of weeks is where the budget actually moves.
On UGC specifically, one thing is worth saying plainly: generated footage styled to look like UGC is the single riskiest thing in this space. It borrows the credibility of a real person without a real person, which is precisely the deception the transparency rules are aimed at, and it is the case where a viewer would reasonably feel misled. If you want the UGC effect, work with creators. If you want volume and control, use generation and let it look like what it is.
If you want the wider context on how AI-assisted content behaves in search and in AI answers, we covered that in Google indexing of AI content and in AI content creation dominates the feed.
Planning AI video or UGC for a campaign?
We build AI-assisted content production for Swiss brands from Zurich: reference systems, generation, edit and the labelling decisions that come with it. If you want a straight assessment of whether your next campaign is a generation case, a UGC case or a shoot, talk to us. If the honest answer is to hire a camera, we will say so.
"AI providers will have to design AI systems to inform users when they are directly interacting with an AI and they will have to add machine-readable marks to enable the detection of AI-generated or manipulated content." Read the guidelines.
European Commission, 2026, Guidelines on transparency obligations for providers and deployers of certain AI systems, 20 July 2026
Summary
AI video and UGC are routinely grouped together because both promise cheap volume, but they solve opposite problems: generation gives control, UGC gives credibility. This article sets out what AI video reliably delivers in 2026 and where it still fails, what the EU AI Act has required since 2 August 2026, and why Switzerland has no general labelling obligation while existing law still applies. It also explains why building production on a single tool is fragile, using the shutdown of OpenAI's Videos API on 24 September 2026 as a live example, and describes a production setup that survives the next model change.
Related Articles
ChatGPT Ads Are Live in Switzerland: What Changed on 24 August
Published
By
Dino Reichmuth
How ChatGPT Ads Work: Formats, Costs and Where They Run
Published
By
Dino Reichmuth
AI Images in Advertising: The Legal Position in Switzerland
Published
By
Lars Kienle
AI Content Creation Dominates the Feed
Published
By
Dino Reichmuth