Do AI-generated videos actually work in advertising?

Short answer: images very well, physics still only partly. Modern text-to-video models produce convincing light, texture, camera movement and now audio. Where they struggle isn't beauty — it's obeying the rules of the world: objects colliding, liquid pouring, fabric falling, a hand genuinely gripping a jar.

This is a measured limit. In the VideoPhy benchmark, built by UCLA and Google researchers, models were prompted with solid–solid, solid–fluid and fluid–fluid interactions and the outputs were assessed by human evaluators. Even the best model of that period produced videos adhering to both the caption and physical law in only 39.6% of instances. The follow-up, VideoPhy-2 (March 2025), built a harder test across 200 distinct actions and reported the best model achieving only 22% joint performance — both semantic adherence and physical commonsense — on the hard subset.

Our findings reveal major shortcomings, with even the best model achieving only 22% joint performance (i.e., high semantic and physical commonsense adherence) on the hard subset of VideoPhy-2.

— Bansal et al., VideoPhy-2, 2025

Transposing those numbers directly onto today's models would be wrong — both studies measure the models available at their publication date, and models have improved noticeably since. In our own search we found no equivalent, human-evaluated published result for current models. But the *class* of failure hasn't changed: models are better at producing a convincing image than a consistent world. For advertising, that distinction is the whole point.

• Does well: atmosphere, texture, abstract visuals, backgrounds, transitions, emotional scenes without the product in them.
• Struggles with: hands and fingers, liquid behaviour, character consistency across a long scene, on-screen text.
• Riskiest at: the product itself — packaging proportions, logo typography, closure mechanisms, and what the product actually does on skin.

On generation cost, yes — dramatically. Per Google's Gemini API pricing page (updated 21 July 2026), standard video with audio on Veo 3.1 costs $0.40 per second at 720p and 1080p and $0.60 at 4K; the Fast tier is $0.10 per second at 720p and the Lite tier $0.05 at 720p. On those rates a single eight-second shot costs roughly $3.20 at standard quality, or about $0.40 on Lite (our arithmetic, a direct multiple of the listed per-second rate).

That number alone is misleading, because the cost of an ad creative isn't the cost of generation — it's the cost of a usable creative. Acceptance rates for AI video are low: outputs where the hand looks wrong, the packaging deforms or the light doesn't sit with the brand get discarded. If you use one attempt in ten, your true unit cost is ten times the listed rate, plus the hours of the person doing the discarding.

• Visible cost: the per-second generation fee — low and predictable.
• Hidden cost 1: acceptance rate. How many generations for one usable shot?
• Hidden cost 2: selection and review labour — the time of whoever watches and filters the output.
• Hidden cost 3: legal review. As the labelling and claims sections below show, this is no longer an optional step.
• Hidden cost 4: brand-consistency correction — fitting the generated image to your palette, your actual packaging and your category codes.

The right comparison isn't "is AI video cheaper than a studio shoot". They do different jobs — we set that out separately in when you genuinely need a studio photoshoot. AI video's real economic effect is less about replacing production than about making variations possible that you previously couldn't produce at all.

The real gain is in iteration; the illusion is expecting finished work in one pass. Visualising an idea, seeing whether an opening lands, putting five visual treatments of the same message side by side — these used to take days and now take minutes. Because ad systems learn through creative variation, that gain touches performance directly.

The illusion is this: generation time shrinks while decision time doesn't. The approval loop — brand, legal, and where relevant regulatory review — moves at the same pace, and with AI output it often lengthens, because new questions get added: does this face resemble a real person, is this our packaging, is this scene making an efficacy claim? Most teams speed up generation tenfold without speeding up approval, so total time falls far less than expected.

We can't back that with a statistic — we found no reliable independent publication measuring this cycle, and inventing an estimate would violate this article's own rule. But the mechanism is clear and testable inside your own process: measure not generation time but total brief-to-live time. If the gain is there, it's real.

In text, a hallucination is a fabricated fact; in video it's a fabricated *reality*. The scene a model produces may not match the rules of the world, or the truth of your product — and the more convincing the image, the later the error gets caught. The VideoPhy findings above are that failure, measured.

In cosmetics there are two extra layers. First, product accuracy: a model doesn't generate your bottle, it generates the idea of a bottle. Proportions drift, logo typography degrades, a pump becomes a mechanism that doesn't exist. Even if the consumer doesn't consciously register it, the gap between the advertised product and the one that arrives comes back as returns and lost trust.

Second, and more serious: showing a result. AI can "show" how skin changes over eight hours — but that depiction rests on no measurement at all; it's the model's statistical guess. Under cosmetics regulation it is a visual claim, and claims must be substantiated. We treat that separately below, because it is AI video's one genuine red line in this category.

• Check every generated frame against a real reference of the product (packaging, proportion, logo, closure).
• Never let a model generate a scene showing what the product does on skin or hair — only real footage plus real evidence can carry that.
• Scrutinise hands, nails, lashes and on-screen text specifically; these are where models are least consistent.
• Physics-heavy scenes — crowds, mirrors, water, hair — raise the error rate; keep those in real production.

The research isn't one-directional, but the weight sits on the negative side. Koning and Voorveld's 2025 experiment (Journal of Interactive Advertising, N = 304) measured the effect of an AI disclosure on an ad created with generative AI. The result cut both ways: the disclosure raised participants' conceptual AI knowledge and attitudinal persuasion knowledge, which decreased trust in both the advertisement and the organisation. The same study also found an opposing path, where the disclosure's effect on attitudinal persuasion knowledge increased trust.

The results showed that AI disclosures increased conceptual AI knowledge and attitudinal persuasion knowledge, resulting in a decrease in trust towards the advertisement and organization. However, a positive effect of AI disclosures was also found…

— Koning & Voorveld, Journal of Interactive Advertising, 2025

A second study points the same way. In Baek, Kim and Kim's research in the International Journal of Advertising, disclosing AI-generated content was associated with unfavourable attitudes toward the ad, with perceived ad credibility mediating the relationship. Their second study found the negative effect weakened among participants who perceived AI as more human-like than machine-like. The limit is worth stating: this work was conducted in a prosocial (nonprofit/donation) advertising context and doesn't transfer one-to-one to beauty e-commerce.

What's striking is that both studies use the same theoretical frame: the Persuasion Knowledge Model. An AI disclosure functions as a signal that triggers the consumer's "this is trying to persuade me" reflex — the exact inverse of the mechanism in our piece on why UGC outsells studio content. If user content wins because it doesn't wake that reflex, an AI label loses because it does.

The correct conclusion isn't "so don't disclose" — as the next section shows, disclosure is now a legal obligation in many cases. The correct conclusion is: don't use AI where disclosing it destroys the value. Saying you used AI for an emotional background scene costs the ad almost nothing; saying it about a "customer experience" narrative collapses the narrative entirely.

In Türkiye and the EU, in many cases yes — and both regimes start within days of each other. Per the Turkish Ministry of Trade's announcement, amendments to the Regulation on Commercial Advertising and Unfair Commercial Practices, published in the Official Gazette of 1 July 2026 (No. 33297), take effect on 1 August 2026. In the Ministry's own terms, where an ad features AI-generated digital characters indistinguishable from humans, this must be stated clearly, understandably and distinguishably; and using an AI-generated digital copy of a real person in a way that suggests they actually tried or recommend a product is prohibited.

On the EU side, Article 50 of the AI Act, Regulation (EU) 2024/1689, applies from 2 August 2026. It imposes two distinct duties: providers of systems generating synthetic audio, image, video or text must ensure outputs are marked in a machine-readable format and detectable as artificially generated; deployers must disclose content constituting a deep fake.

Deployers of an AI system that generates or manipulates image, audio or video content constituting a deep fake, shall disclose that the content has been artificially generated or manipulated.

— Regulation (EU) 2024/1689, Article 50(4)

On top of the law sits a platform layer that is already live. Per Meta's own documentation, "AI info" normally appears in the three-dot menu at the top right of an ad — but where the image includes an AI-generated photorealistic human, the label appears next to the "Sponsored" label instead. Meta also notes that ads created or modified with third-party AI tools don't receive this automatic labelling, and that for ads about social issues, elections or politics disclosure is already the advertiser's own obligation.

• YouTube: realistically altered or synthetic content — making a real person appear to say or do something they didn't, altering footage of a real event, or generating a realistic scene that never occurred — must be disclosed at upload; beauty filters, colour adjustment and production assistance like scripts are exempt.
• TikTok: per the platform's own announcement, labelling realistic AIGC has been required for over a year; TikTok also began reading C2PA Content Credentials to auto-label content made elsewhere, and attaching those credentials to TikTok content.
• Meta: the AI info label, surfaced visibly on ads containing photorealistic AI humans.
• None of these substitute for the law — platform compliance and legal compliance are two separate obligations.

The technical infrastructure is maturing too. C2PA, the Coalition for Content Provenance and Authenticity, offers an open technical standard recording a piece of content's origin and edit history; in its own words Content Credentials work "like a nutrition label for digital content", with Adobe, Amazon, BBC, Google, Meta, Microsoft and OpenAI on its steering committee. On Google's side, SynthID embeds an imperceptible watermark into images, audio, text and video at the moment of generation, described as designed to survive cropping, filters, frame-rate changes and lossy compression.

The moment it shows the product working. Cosmetics regulation doesn't limit claims to text: Regulation (EU) No 655/2013, which sets the common criteria for justifying cosmetic claims, covers pictures and figurative signs alongside text, whatever the medium. Its honesty criterion is unambiguous: presentations of a product's performance shall not go beyond the available supporting evidence. In Türkiye the counterpart is TİTCK's Guide on Claims for Cosmetic Products, applying the same set of common criteria.

Translated to AI video, that's sharp: a model can generate footage of "even skin tone in four weeks" with no measurement behind it whatsoever. The more convincing the image, the stronger the claim — and the evidence still doesn't exist. That's not a technical flaw but a direct regulatory exposure, and it doesn't shrink as models improve. It grows, because the unsubstantiated claim becomes more believable.

The second red line is fabricated testimony. The US Federal Trade Commission's final rule, announced 14 August 2024, prohibits reviews and testimonials that misrepresent being by someone who does not exist — explicitly naming AI-generated fake reviews — and opens the door to civil penalties against knowing violators. In Türkiye, as noted above, presenting an AI-generated digital copy of a real person as having used a product is prohibited from 1 August 2026.

So "let's generate a customer testimonial video with AI" sits squarely in prohibited territory under two separate legal regimes. It's the most tempting and most dangerous AI use for beauty brands: tempting because it skips the hardest part of producing UGC, dangerous because it is a precise description of the prohibited conduct.

The line runs here: use AI where imagination belongs, and don't use it where evidence belongs. If you're producing a visual world, an atmosphere, a feeling, an abstraction, the model is a good tool. If you're showing what the product is or what it does, that frame has to be real.

Where AI genuinely adds value

• Concept visualisation and storyboards: see the idea before the shoot, align the team, kill expensive wrong decisions early.
• Backgrounds, atmosphere and abstract scenes: visual texture where the product isn't central and no claim is carried.
• Variation generation: testing different visual treatments of the same message — ad systems learn through diversity.
• Reformatting and adaptation: fitting existing footage to other aspect ratios and placements (platform-native tools already do this).
• Localisation: subtitles, voice-over and text variants — without touching the footage.

Where you shouldn't

• Frames showing the product itself: packaging, texture, colour, scale — that's controlled photography's job.
• Efficacy or result demonstrations: before-and-after, skin change, added volume, reduced shedding — no visual claim without evidence.
• Customer testimonials and user-experience narratives — both prohibited and destructive of the very trust advantage you're after.
• Anything generating a real person's face, voice or likeness (permissions and regulation each create separate problems).
• The visible hero creative for a brand whose positioning rests on transparency — don't pick a tool that contradicts your own claim.

One closing caution: AI video's real cost isn't a budget line, it's a trust line. When a consumer notices that a scene in your ad is fake, they don't re-evaluate the scene — they re-evaluate the brand. That is precisely what the research above measures: what's lost isn't only the credibility of the advertisement, but the credibility of the organisation.