Making Video With AI

Generated video has a reputation problem that the platform rules do not share. YouTube pays for video made with AI tools, and says in writing that declaring it “won’t limit a video’s audience or impact its eligibility to earn money”. Adobe Stock accepts AI-generated footage from contributors as long as you tick the box saying so. The advertising industry has gone further than either: 83% of ad executives say their company now uses AI in the creative process, up from 60% in 2024, and 86% of video ad buyers are using it or planning to.

What gets punished is narrower than “AI video”, and worth being precise about, because the distinction is the whole business. This page is the commercial picture: the rules, what audiences object to, where the paid work is and how to price it. The craft sits on three pages next to it — short-form for TikTok, Reels and Shorts, long-form for anything over eight minutes, and telling the difference from slop, which is the quality bar both depend on.

What audiences actually object to#

The clearest evidence comes from the two Coca-Cola Christmas campaigns. The 2026 one avoided generated people after the 2024 version unsettled everyone, and the reaction was worse. Positive social sentiment fell from 23.8% before the campaign to 10.2% after. The word that circulated was “slop”.

A study in the Journal of Retailing and Consumer Services analysed 7,822 YouTube and Reddit comments on that campaign and found the objection was not really about image quality. Viewers inferred a motive, and the motive they inferred was cost-cutting. The same paper found that disclosing AI use improves engagement and purchase intent only when the disclosure also conveys human creative investment. Disclose authorship with nothing else attached and you amplify the suspicion instead of settling it.

That pattern shows up in the survey data too. DoubleVerify’s 2026 study found 43% of North American consumers said a brand’s use of low-quality or uncanny AI ads would worsen their opinion of the brand. Note the qualifier: low-quality or uncanny. Separately, appetite for AI-generated creator content fell from about 60% in 2023 to around 26%.

So the working rule is not “avoid AI”. It is: do not use it visibly in place of the thing your audience came for. Coca-Cola’s audience came for the feeling of that 1995 advert. Replacing the craft with a render, on the one asset people had affection for, is the mistake. Using the same tools to produce a shot nobody could have filmed is not.

What the platforms permit, exactly#

YouTube’s position is the most useful because it is written down in detail. Disclosure is required when content is realistic and it makes a real person appear to say or do something they did not, alters footage of a real event or place, or generates a realistic scene that never occurred.

Disclosure is not required for any of the following, which covers most working use:

• Clearly unrealistic or fantastical scenes, and animation.
• Beauty filters, colour grading, lighting changes, blur, vintage looks and other special effects.
• Production assistance: script drafts, thumbnail generation, ideation.
• Cloning your own voice.
• Sharpening, upscaling and audio repair.

Where a label is needed, photorealistic content carries it in the player itself; everything else carries it in the expanded description. And again, in YouTube’s own words, the label does not cost you reach or revenue.

The thing that does cost you revenue is separate, and it is the Inauthentic Content policy, renamed from Repetitious Content in July 2026. That rules out content following a template with little variation, reproduced at scale, with no real author input. It applies whether or not AI was involved: a channel of slideshows with scripts read verbatim from elsewhere was already ineligible. What gets you demonetised has the full text and the format self-audit.

For stock, the two big libraries have gone opposite ways. Adobe Stock accepts generative-AI video if you mark it “Created using generative AI” at submission, hold the rights to it, and meet the same technical and legal standards as filmed footage, model and property releases included. Pond5 refuses it outright: its contributor terms require you to own what you upload, and it states that repeated submission of AI-generated content “will result in account suspension or termination”. Check the library before you build a library.

Where the money is#

In rough order of how much evidence there is that someone will pay.

Advertising variants. This is the largest and least glamorous market, and it is the one with real budgets behind it. Paid social rewards creative volume: a common brief is three hooks by two lengths by two calls to action, which is twelve versions of one concept. Producing twelve conventionally is a fortnight and a crew. Producing twelve with generated or assembled footage is an afternoon. US AI ad spending is put at $32.03 billion in 2026, close to triple the previous year.

What the client is buying is not the render. It is a brief the model can execute, and the judgement to throw away the eight versions that look uncanny. IAB found cost efficiency is now the top-cited benefit of AI in creative, at 64%, up from fifth place in 2024 — which tells you what the buyer thinks they are purchasing, and why you should price against the outcome instead.

Shots that cannot be filmed. Sold inside an ordinary production, this is the least contentious use there is. A product in a location you cannot travel to. An interior of a machine. An abstract sequence for a segment about inflation. Pre-visualisation so a client can approve a concept before anyone hires a camera. Nobody objects to any of this, because nothing was replaced.

Localisation and repurposing. Dubbing, subtitling and cutting long video into vertical clips is the best-evidenced AI service work on this site, and it has its own page: AI-assisted services. The video localisation market alone is put at around $4 billion in 2026.

Stock footage. Real, small, and slow. Adobe accepts labelled AI video, which makes this different from selling AI images on stock sites, where the largest libraries closed the door. Treat it as a use for output you generated anyway rather than as a plan.

Your own channel. Fully monetisable if you are the author of it. Which length you pick changes everything about the economics: short-form earns roughly a twentieth to a thirtieth per view of long-form, and only long-form carries mid-roll advertising. The failure case is well documented: creators who adopted AI video tools early tripled their output on average, and only 34% held or improved their engagement rates. Tripling output and losing engagement is a worse position than before, because you now have three times the work and the same income. Faceless AI channels covers where that ends.

Pricing, and the trap in it#

Two ways to buy generation, and they lead to very different conclusions. These are the providers’ own published prices, not aggregator compilations, because the compilations disagree with each other by a factor of five.

Per second, paying by API:

Provider and platform list prices, September 2026: Google, OpenAI, and the fal.ai listings for Wan, Kling and Seedance. Veo includes audio by default. OpenAI’s batch tier is half price, which suits variant work where you are not waiting on any single render. This is not the whole field — Hailuo, LTX, Grok Video and Higgsfield’s own models price similarly, and MiniMax H3 Max bills a streamed directing session rather than clips.

So a thirty-second spot is $1.50 to $21 of compute. Now apply a real keeper rate: practitioners generally discard several takes for each one they use, so multiply by four or five. A finished thirty seconds is therefore somewhere between about $8 and $105. Still small, and still no basis for a price.

Per month, paying by subscription, which is what most people actually do. Runway is worth working through because it publishes both halves: Gen-4.5 costs 60 credits per five seconds, and the tiers include a fixed monthly allowance.

Runway’s published prices and credit costs, September 2026. The output column is arithmetic from those two figures; the keeper-rate column assumes you bin three takes in four, which is our assumption and not their claim.

That second table is the one that changes plans. Thirteen seconds of usable footage a month is not a production pipeline, it is a trial. Anyone selling regular client work is on the top tier or on API billing, and should know which before quoting a retainer.

Price the deliverable. A set of twelve tested ad variants with hooks written and the duds discarded is worth what twelve variants are worth to that advertiser’s media spend, and has nothing to do with your render bill. The same logic as everywhere else on this site: see AI-assisted services for the version of this argument with published rate comparisons, and the rate calculator for pricing usage rights, which matter more here than usual because generated footage is easy for a client to assume they own outright.

One clause to watch. Contracts increasingly ask for the right to use delivered material as training data, and for likeness rights over anyone appearing in it. If you generated a presenter, whose likeness is it? If you used your own face or voice, you are granting something with a long tail. Contract clauses that cost you money has the language.

Common mistakes#

Pitching AI as the discount. The client hears “cheaper” and reprices you permanently, and the audience infers exactly the motive the Coca-Cola research documented. Sell the speed and the volume of variants; keep the production method out of the sales pitch.

Shipping a take you have not checked against the known failure modes. Garbled text, morphing hands, a face that changes between shots. The artefact checklist takes two minutes and is the difference between assisted and slop.

Generated people in emotional work. Both Coca-Cola campaigns are the case study, and the 2026 one shows that removing the human faces was not enough. Anywhere the point of the asset is warmth or trust, generated humans are the highest-risk choice available.

Skipping the disclosure toggle because the platform might penalise it. It does not, and failing to disclose runs a three-strike escalation ending in removal from the Partner Programme. The toggle is free; the omission is not.

Uploading to a library that bans it. Pond5 terminates accounts for repeat submissions. Read the contributor terms first.

Building a workflow on one endpoint. Models and prices in this category change every few months and providers retire APIs with short notice. Keep the process portable and do not promise a client something only one model can do.

Volume as the strategy. The Inauthentic Content policy exists precisely to stop this, and the engagement data says it does not work even where it is permitted.

Who this suits#

People who already sell video work, and want a higher margin on the parts clients do not value: variants, inserts, pre-vis, localisation.

People with a commercial eye who can look at nine generated takes and identify the one that does not look wrong. That judgement is the scarce input now, and it is not evenly distributed.

Anyone selling to advertisers, where the demand is documented and the budget is somebody else’s media spend.

Who it does not#

Anyone hoping to generate video and have the video itself be the product. The output is close to free and getting closer.

Anyone building a channel on volume. The policy blocks it and the engagement numbers say it fails first.

Anyone unwilling to disclose. The requirements are narrow and cost nothing to meet, and undisclosed realistic content is the one version of this with a real penalty attached.

Anyone whose audience relationship rests on it being them on camera. That is an asset AI cannot manufacture, and it is worth more now than it was.

Sources: YouTube, disclosing altered or synthetic content and YouTube Help; Adobe Stock AI video submission guidelines; Pond5 contributor legal guidelines; IAB, The AI Gap Widens, 104 ad executives, October 2025 to January 2026; When AI ads backfire, Journal of Retailing and Consumer Services; Marketing Dive on the Coca-Cola sentiment figures; eMarketer on AI ad spend. Model prices are published lists from the providers and from the platforms hosting the models: Google, Gemini API pricing, OpenAI pricing and Runway pricing. Third-party aggregator compilations of these figures disagree with the price lists and with each other, so we have not used them. The output-per-month and keeper-rate columns are our arithmetic and our assumption respectively. Checked 8 September 2026.