Where AI video asset rights actually end — five questions to ask before you buy

Supply of AI-generated video has exploded over the past two years, and prices have fallen hard with it. But in actual deals, we find that almost no dispute is about price. They are about rights boundaries nobody nailed down. What the buyer typically gets is the phrase "cleared for commercial use" — which, as a legal matter, promises close to nothing.

Here are five questions worth settling before payment, and worth writing into the agreement.

1. Who carries the risk on training data

The provenance of a model's training data is where the disputes are concentrated right now. Where a model was trained on unlicensed works, whether the output infringes is answered differently across jurisdictions — and the answers are still moving.

The question to ask is not "is your training data properly licensed?" The vendor will always say yes. Ask instead: if a third party asserts rights over the training data, who indemnifies whom, and up to what cap? Unless that sits in the contract, you have not actually asked.

2. Are you buying copyright, or a license to use

These are very different things. An assignment of copyright lets you sub-license and lets you sue infringers. A use license only lets you use the material within an agreed scope, and usually cannot be passed on.

Many asset platforms grant the second while marketing it in the language of the first. Settling this before signing saves a great deal of trouble later.

3. What does the license scope actually cover

The usual grey areas:

• Whether paid media placement is permitted — many base licenses exclude it
• Whether you may modify, re-cut, or composite the material with other assets
• Whether you may embed it in the product itself, such as default assets inside an app
• Which territories are covered
• Whether your client's clients may use it — agencies should read this one closely

4. Is there exclusivity, and how far does it go

Most AI assets are licensed non-exclusively, which means a competitor can buy the identical frames. If the material is going into brand-level creative, price that risk in before you commit.

Genuine exclusivity is expensive in this category. Usually the more realistic structure is custom generation plus a time-limited exclusivity window, rather than an outright buyout.

5. Is the output itself protectable

This one gets skipped. In several jurisdictions, material generated purely by AI, without substantive human authorship, may not attract copyright protection at all.

Which means the assets you paid for may give you no standing to stop anyone else from using identical or near-identical frames. If your business model depends on content exclusivity, this goes straight to valuation.

How we handle it

When acquiring AI assets for clients, we require a training-data compliance representation and an indemnity clause from the supplier, and we itemize the license scope in a schedule rather than letting "commercial use" stand in for it.

These are not comfortable clauses to negotiate. They are far cheaper than answering an infringement claim after the fact.