What should an AI data license cover? A checklist

Short answer. A license should say exactly which records are included, what the lab may do with them and for how long, whether anyone else may receive the same data, how delivery is accepted and paid for, who removes personal information, and what happens to the data and to models when the license ends. Use the list below to read a draft with your own counsel; we do not give legal advice.

General information, not legal advice.

Contact usWhat's my data worth?


Scope: which records, and for what use

The license should point to the manifest, row by row, so both sides agree on what is being delivered. It should then name the permitted use. Training a model is one use. Evaluating a model, building a test environment, or reselling the set are others, and a broad clause can cover all of them without saying so.

Ask whether the lab may share the data with affiliates, contractors or customers. Labs often build training environments from licensed records and sell access to those environments, so the question of who else ends up holding a version of your data belongs in this section.


Exclusivity and term

An exclusive license promises that the company will not license the same records to anyone else for a period. It usually commands a better price, and it limits every later deal. Check the length, whether it covers only these records or the whole category, and whether it starts at signing or at delivery.

The term of the license is a separate question from exclusivity. A license may be exclusive for a short time and run for many years, or never end. Check whether the right to use the data, and anything built from it, survives the end of the term.


Delivery, acceptance and payment

Most licenses include an acceptance window: after delivery, the lab checks the files against the manifest and agreed quality tests, then accepts or rejects. Read how long the window is, what counts as a failure, whether you get a chance to fix it, and whether payment is tied to acceptance, to milestones or to signing.

Look at how payment is calculated. A fixed sum, a rate per unit delivered, and a share of later revenue each behave differently if the counts change. If any part depends on revenue, ask for a floor and for the reports that show how the figure was worked out.


Personal information, deletion and liability

The agreement should name who removes personal information, to what standard, and who checks a sample before anything leaves the company. It should also say what happens if a person is later found in the data. That clause decides where the risk sits, and some drafts quietly move it to the seller.

Check the end of the deal too: whether the lab must delete the raw files, whether it may keep models already trained, and how deletion is confirmed. Finally, read the warranties you are asked to give about ownership of the records and the indemnities that follow from them. These are the clauses to spend counsel time on.

The checks on the contract itself are between the company and the lab, with the company relying on its own lawyer. The US Copyright Office has said it expects a licensing market for training data to keep developing without new government rules, which means the contract is where most of these questions get settled.


Reading a draft in the right order

Read the scope and use clauses first, since every other term depends on them. Then the money: price, how it is calculated, and what triggers each payment. Then the risk clauses on personal information, warranties and indemnities. Leave governing law and notice addresses for last; they rarely decide whether a deal is good, while the first three groups almost always do.


Follow-up questions

Is a short license safer than a long one?

Not by itself. A short draft often leaves use, sharing and deletion open, and an open question tends to be read in favour of the party holding the data. Length matters less than whether each item above is answered.

Can the license limit what a trained model may do?

Some do, for example by barring a model from reproducing records word for word. Whether such a clause can be checked in practice is worth asking the lab directly.

What should be attached to the license?

At least the manifest, the de-identification standard, the acceptance tests and the payment schedule, so that each can be pointed to if there is a dispute.

Didn't find your question? Ask us directly.


Sources


Keep reading

The slides and script for the video on this topic