Technology

The fine print behind your cheap AI video

Making a product photo or an ad clip now costs almost nothing. The catch is buried in the terms: whether you own it, whether it copies someone else, and whether a platform will even let it run. Cheap to make is not the same as safe to post.

By Albert Richer / Updated August 18, 2026 / 12 min read

You can make a polished product video in about the time it takes to read this sentence. Type a prompt, wait a few seconds, download. It looks great, it cost almost nothing, and that is exactly where the trouble starts.

“Commercial use allowed” is the phrase every AI tool waves at you. It tells you almost nothing about who pays when something goes wrong.

Cheap and fast hides three questions the tool never asks. Do you actually own it? Are you copying someone? Will the platform even run it? Get one wrong and the free clip turns into a takedown, a lawsuit, or a fine. So before anyone posts anything an AI made, it has to clear three gates.

Gate 1: whose clip is it, really?

The tool says you own the output. US copyright law may say nobody does.

Type a prompt, get an image, and every tool tells you the same comforting thing: it is yours, use it commercially. Legally, that sentence does a lot less than it sounds.

Here is what that means for an ad. You can use it. You just cannot stop anyone else using it. If it is not copyrightable, you cannot register it or sue a competitor who lifts your exact AI video and runs it as their own. Other rules can sometimes still help (a trademark, a contract), but the copyright shield is simply gone.

“You own the output” also comes with strings the tool does not print on the button. OpenAI assigns you its rights in what you make, to the extent the law allows.OpenAI Midjourney says you own your images too, with a catch: if your company makes over $1 million a year, you have to be on a paid Pro or Mega plan just to own them.Midjourney Free and trial tiers are thinner still, and some tools bar commercial use outright or leave a watermark baked in.

There is a way back in. The Copyright Office is clear that a human who meaningfully shapes the work keeps copyright in that human part. The raw AI frame is unprotected; the edit built around it can be protected. The person in the loop is the legally valuable part of the whole workflow.Copyright Office

Gate 2: are you copying someone?

A prompt can quietly borrow someone else’s work, logo, or face. When it does, the demand letter lands on the advertiser, not the tool.

The model learned from a huge slice of the internet, and sometimes that leaks into what it hands back: a recognizable character, a brand logo in a background nobody asked for, a face that looks a little too much like a real person. If the output is “substantially similar” to something protected, it can infringe, no matter which tool made it. And the person who published it can be on the hook, not just the company that built the model.Copyright Office

The math is the scary part. One protected frame can carry statutory damages of up to $30,000 per work, and up to $150,000 if a court finds the copying willful.17 U.S.C. 504 A $20 clip does not come with $20 of risk.

This is not hypothetical. The tools marketing teams actually use are in court.

Two lessons hide in those cases. First, if a logo or watermark shows up in a generation, kill it: in the UK, Getty lost most of its case against Stability AI but won on trademark where the output carried Getty’s watermark.Getty v. Stability (UK) Second, the scariest risk is a real person’s face or voice, because that is barely a copyright question at all.

Right of publicity is state law, and it is a patchwork. Using someone’s face or voice to sell something without consent is actionable in a majority of states, by statute, common law, or both, with California and New York among the strictest.Right of publicity You do not even need the exact face; a look-alike that clearly evokes a real person can be enough. And the usual escape hatches do not hold: when a face-swap app argued its AI output was “transformative” and that copyright law overrode the claim, a federal appeals court rejected both and let the publicity case proceed.Young v. NeoCortext Tennessee’s 2024 ELVIS Act went further and added voice outright, aimed straight at AI clones.ELVIS Act

And dead does not mean free. The George Carlin estate sued over an AI-generated comedy special in his voice; the case settled with a permanent injunction and the special was pulled.Carlin estate Postmortem likeness rights run long: 27 states protect a person after death, from 20 years in Virginia to 100 in Indiana, and potentially forever in Tennessee if the likeness is kept in use.Copyright Office

The floor is still rising. The federal NO FAKES Act, which would create a nationwide right against unauthorized AI replicas of anyone’s voice or face, cleared a key Senate committee in 2026. It is not law yet, but the direction only points one way.NO FAKES Act

Gate 3: will the platform even run it?

Even a clean clip has to clear the new disclosure rules. And this is where the tools split hardest on who covers you when something goes wrong.

Say a clip cleared the first two gates. It still has to get published, and the rules on labeling AI content went from nice-to-have to enforced in about two years.

The rule that bites fastest is fake AI testimonials. The FTC’s 2024 rule bans fake and false reviews and testimonials, including AI-generated ones attributed to people who do not exist or never used the product; its enforcement sweep put it plainly, that there is no AI exemption from the laws already on the books.FTC It does not ban AI avatars outright. What it bans is a synthetic spokesperson vouching for a product nobody really tested, which is now a federal problem, not a growth hack.

FTC · Operation AI Comply

“There is no AI exemption from the laws on the books.”

Here is the trap that catches careful people. Ad platforms lean on that provenance data, a standard called C2PA, to tell what is AI.C2PA Export or heavily compress a video and that hidden data often gets stripped, so a compliant clip can suddenly read as undisclosed and get rejected. TikTok says plainly that undisclosed AI content will be rejected or restricted.

The punishment usually is not a lawsuit. It is faster and dumber than that: the ad gets rejected, the budget burns on nothing, and the launch loses its early momentum. And the enforcement is mostly automated, so there is no person to argue with.

Which is the whole reason the tool choice matters. When something does go wrong, a few tools stand behind the customer and most just leave them holding it. That promise is called indemnity: the vendor agrees to defend you if the output triggers a third-party claim.

Read worst to best. “Commercial use allowed” is not indemnity. Every check in the last column is a paid or enterprise plan with caps, exclusions, or conditions attached, and terms change often, so confirm your exact tool and tier. Sources, each vendor’s own August 2026 terms: Adobe, Midjourney, HeyGen, Runway, Canva, OpenAI, Google.