Three weeks ago the rules changed. On 2 August the EU's transparency obligations for synthetic media became applicable, and California's provenance requirement was deliberately timed to land the same day. Most coverage of AI video is still about which tool makes the prettiest clip.
The two facts that define the year
Put these side by side and the whole strategic picture falls out.
Nobody can tell. Peer-reviewed work on human detection of synthetic video puts performance at roughly coin-flip level, and separate research suggests only a small single-digit percentage of people can reliably distinguish AI-generated video from real footage. Your audience is not spotting your AI content. They're watching it.
They mind when they find out. Survey work through 2026 consistently finds a substantial minority reacting negatively to AI in branding, a large majority wanting brands to disclose it, and academic research showing that identifying content as AI-generated makes consumers measurably more sceptical and less engaged.
So the risk was never that viewers would catch you. It's that the label now arrives by law, by platform policy, or by someone else's discovery — and the reaction happens then, at a moment you don't control.
The 2026 position Detection failed, so disclosure became mandatory. The question stopped being "can we get away with it" and became "what does our work look like once it's labelled?"
That reframing matters because it changes what you optimise for. If nobody can tell, realism is no longer the goal — it's table stakes and it buys you nothing. What the label does is put your creative judgement on public display alongside your efficiency saving.
What the rules actually require
Broad strokes only, and you should take proper advice for your own situation — this is a fast-moving area and the detail matters more than a blog post can carry.
The EU AI Act, Article 50. Applicable from 2 August 2026. It requires machine-readable marking of synthetic audio, image and video content, with additional disclosure duties for deepfake material — defined broadly as AI-generated or manipulated content resembling real people, places, objects or events in a way that could falsely appear authentic. It reaches beyond the EU: content targeting EU consumers is covered regardless of where the company sits. Penalties are substantial.
California. The state's provenance requirement was re-timed to align with the same August date, meaning two major markets began enforcing on one day.
Sector rules. Political advertising already carries AI disclosure requirements in several jurisdictions, and the direction of travel toward commercial advertising is clear rather than speculative. Industry bodies published voluntary guidance through early 2026 ahead of the statutory deadlines.
Content Credentials, briefly
The technical answer to machine-readable marking is provenance metadata, and the standard consolidating around it is C2PA — with Content Credentials as its practical implementation.
Rather than stamping a watermark on your video, it embeds a tamper-evident, cryptographically signed record of how the file was made and edited. Platforms and devices can read it; viewers can inspect it if they choose. Major AI video tools and platforms have been adopting it, and standardisation efforts have been consolidating through 2026.
The practical implication for a marketing team is narrow but real: check whether your video tools write Content Credentials, and whether your editing and export pipeline preserves them. Provenance metadata that gets stripped during a routine export is worse than useless, because you'll believe you're compliant when you aren't. That's a fifteen-minute check with whoever owns your production stack, and worth doing before it's a problem.
Synthesia
Creates professional AI videos with avatars without cameras or video editing skills
Best for: AI Video Creation
Where the volume actually is
Adoption is not in question. Industry estimates put AI-generated material at a meaningful and rapidly growing share of digital video, with the large majority of marketing teams using it in at least one campaign per quarter and generation volumes rising several-fold over two years. The cost floor has fallen far enough that per-second generation costs are now a rounding error against traditional production.
But the aggregate numbers hide the important pattern, which is where teams use it.
| Use case | Authenticity stakes | Verdict |
|---|---|---|
| Ad creative variants for testing | Low | Strong fit — volume is the whole point, and no one expects a variant to be documentary |
| Localisation and language versioning | Low | Strong fit — arguably the single best application in marketing |
| Product explainers and demos | Low–medium | Good fit — provided the product behaviour shown is accurate |
| Motion graphics, B-roll, backgrounds | Low | Strong fit — nobody was ever meant to believe the stock footage |
| Founder or leadership communication | High | Poor fit — the person being real is the entire message |
| Customer testimonials | Very high | Don't — synthetic testimonial is a deception problem, not a disclosure one |
| Creator and influencer content | Very high | Poor fit — everyday creators outperform precisely because they're real |
There's a length dimension worth layering on top. The split between short clips for discovery and long video for trust maps almost exactly onto where AI helps: the discovery layer is volume work where synthesis pays, and the trust layer is where a real person on camera is doing the actual job. Teams that generate at the top of the funnel and stay human at the bottom tend to get the efficiency without the credibility cost.
The organising principle is authenticity stakes, not production complexity. AI video is excellent where nobody was claiming the footage was real in the first place, and dangerous wherever the content's value derives from a real person standing behind it.
The one-question test: would a viewer feel deceived on learning this was synthetic? If yes, no efficiency saving justifies it. If no, generate freely and mark it properly.
The disclosure paradox, and how to work inside it
Here's the tension nobody has resolved. Consumers overwhelmingly say they want disclosure. Research also shows disclosure reduces engagement and increases scepticism. Both findings are robust. Marketers are being asked to do something audiences demand and then penalise.
Three things reconcile it in practice.
The penalty is quality-dependent. Industry research suggests that clear communication paired with high creative standards can improve attitudes toward AI advertising, and even lift attention. Read alongside the scepticism findings, the plausible reading is that disclosure doesn't punish AI — it punishes bad AI, by removing the ambiguity that previously let mediocre work pass unexamined. The label puts your creative judgement on show.
Blanket labelling dilutes the signal. Tagging a video as AI-generated because the subtitles were auto-generated is both unhelpful and, arguably, misleading in the other direction. Proportionate disclosure — specific about what was synthesised — is more useful to viewers and less costly to you. Some industry guidance explicitly notes that obviously fictional or impossible content doesn't typically warrant AI-specific labelling, though ordinary advertising rules about misleading consumers still apply.
Some brands are leaning in. A visible minority now treat disclosure as a positioning move rather than a compliance chore — stating openly which tools they use and where, on the theory that the audience will find out anyway and volunteering is cheaper than being caught. Whether that reads as confidence or as excuse-making depends entirely on the quality of what's being disclosed, which is the same conclusion as the first point.
None of this is separable from the broader question of brand trust in an era of AI-generated everything. Video is simply where the tension is sharpest, because video was the format audiences trusted most.
The saturation problem
A second-order effect that gets less attention than the legal one and may matter more commercially.
When generation costs approach zero, volume explodes. Analysis of social feeds through 2026 suggests a substantial fraction of short-form content is now low-effort AI output, and audiences have developed a vocabulary of contempt for it. The problem for a legitimate brand is that generic AI video is now a recognisable genre — the smooth camera moves, the too-perfect lighting, the uncanny middle distance — and being fluent in that genre reads as low effort even when your intent was good.
This is the content sameness problem arriving in video, and it's worse here than in text because the visual signature is more distinctive. Default outputs from the same handful of models share a look. If your ad looks like the slop, it will be sorted with the slop, regardless of your budget or intentions.
The defences are unglamorous: art-direct rather than accept defaults, mix synthetic and real footage rather than going fully generated, keep a distinctive visual system across everything you publish, and treat AI video as a production method inside a creative direction rather than a substitute for having one.
Practically, most of the distinctiveness gap is a specification gap. The default output is what you get when the brief is thin, and video models reward detailed direction — shot type, lens character, lighting, pace, colour treatment — in the same way careful prompting improves any AI output. A team with a written visual specification produces recognisably different work from the same model as a team typing a sentence, which is worth more than switching tools.
Where the performance data points
Some genuinely encouraging findings sit alongside the caution, and they cluster in one place: personalisation at scale.
Large-scale academic work has found AI-generated personalised video ads outperforming both personalised image ads and generic video ads on click-through, by margins in the single-digit percentages. That's a real effect and it points at the strongest commercial case for the technology — not replacing your hero film, but producing the hundred variants of it that were previously uneconomic.
That maps directly onto creative testing discipline, where the binding constraint has always been how many variants you can afford to produce. Remove that constraint and the testing programme is the beneficiary, not the brand film. It also feeds naturally into shoppable video on connected TV, where variant volume and localisation both matter.
Two cautions on the numbers, though. The uplift is in personalisation, not in synthesis — the gain comes from relevance, and AI is the mechanism that makes relevance affordable. And a single-digit CTR improvement is meaningful at scale and invisible on a small campaign, so don't build a business case on it below serious volume.
A note on the statistics in this space
Worth saying plainly, because this topic has a data-quality problem that's unusual even by marketing-statistics standards.
A great many circulating figures fail on inspection. Numbers get invented or garbled by aggregator sites, then repeated across a dozen roundups until repetition confers false authority. Market-size reports with completely different scopes get quoted interchangeably. Registration counts get reported as active usage. If a statistic about AI video seems remarkably convenient, check whether it traces to a primary source — a company disclosure, a peer-reviewed study, a named survey with a stated method — or only to other blog posts.
The figures in this piece are drawn from research and disclosures that name their methods, and I've written them as approximations where the underlying sources vary. That's not false modesty; it's the honest precision available.
What to do about it this quarter
- Audit what you've already published. Find every piece of customer-facing video containing synthetic elements. Most teams underestimate this because generation crept in through editing tools rather than through a decision.
- Check your provenance pipeline. Do your tools write Content Credentials? Does your export process preserve them? Fifteen minutes with whoever owns production.
- Write a one-page policy. What you'll generate, what you won't, what gets disclosed and how. One page, agreed once, prevents the case-by-case argument every time.
- Draw the authenticity line explicitly. Testimonials, founder communication and customer stories stay real. Put it in writing before someone's deadline makes the decision for them.
- Take legal advice if you sell into the EU or California. Not blog-post advice. The obligations are live now and the extraterritorial reach catches more organisations than expect it.
If creative production is where this bites for you — needing the volume without the generic look — a social and video partner working inside a defined art direction is usually a better answer than another generation tool.
The short version
Viewers can't detect AI video, so the law now requires you to tell them. Disclosure penalises weak work and is survivable for strong work, which makes creative judgement more important than it was, not less. Generate freely where nobody was claiming the footage was real; don't go near testimonials or founder communication. And check whether your export pipeline is quietly stripping the metadata you're now obliged to embed.
Need the video volume without looking like everyone else's AI?
We build creative direction first, then use AI to scale it — variants, localisation and testing included.
Explore Branding & Creative →Frequently asked questions
Do you legally have to label AI-generated video in marketing?
In some jurisdictions, yes, as of August 2026. Article 50 of the EU AI Act became applicable on 2 August 2026 and requires machine-readable marking of synthetic audio, image and video content, with specific disclosure duties where content depicts real people, places or events in a way that could appear authentic. It applies extraterritorially — content targeting EU consumers is covered regardless of where the company is based. California's equivalent provenance requirement was re-timed to the same date. This is a rapidly developing area, so treat this as orientation and take proper legal advice for your own situation.
Does labeling content as AI-generated hurt engagement?
Research points both ways, which is the central difficulty. Studies have found that consumers become more sceptical and less engaged when content is identified as AI-generated, and surveys consistently show a substantial minority react negatively to AI in branding. At the same time, industry research suggests that clear communication, high creative standards and consistent disclosure can improve attitudes and even lift attention. The reconciling view is that disclosure penalises low-quality AI content and is survivable for high-quality work, so the label amplifies the underlying creative judgement rather than replacing it.
Can people tell if a video is AI-generated?
Largely no, and that is the reason disclosure rules exist. Peer-reviewed work has found human detection of synthetic video performing at close to chance level, and separate research reports only a small percentage of people can reliably identify AI-generated video from real footage. This has an uncomfortable implication for marketers: audiences are not distinguishing your AI content by eye, so any reputational effect comes from what you disclose, what platforms label automatically, and what someone later discovers rather than from viewers spotting it themselves.
What are Content Credentials and C2PA?
C2PA is a technical standard for attaching tamper-evident provenance metadata to media files, and Content Credentials is the consumer-facing implementation of it. Rather than a visible watermark, it embeds a cryptographically signed record of how a file was created and edited that platforms and devices can read. It matters commercially because the machine-readable marking requirements arriving in 2026 are most straightforwardly satisfied by this kind of embedded provenance, and because major AI video tools and platforms have been adopting it.
Where does AI-generated video actually work in marketing?
Best in high-volume, low-authenticity-stakes work: ad creative variants for testing, localisation and language versioning, product explainers and demos, motion graphics, and background or B-roll elements. It works poorly wherever the value of the content is that a real person is behind it — testimonials, founder communication, customer stories, and anything trading on lived experience. The reliable test is whether a viewer would feel deceived on learning it was synthetic. If yes, the efficiency saving is not worth the trust cost.
