Skip to content

Korra team · · 3 min read

What should an AI training data licence cover? A buyer’s clause list

An AI training data licence should settle the uses, exclusivity, the term, the seller’s right to license, personal data, ownership of what you build, and what happens at the end.

An AI training data license agreement should answer eight questions before you sign it. The one buyers skip most often is the last: what happens to the data, and to a model trained on it, when the licence ends.

Why is a training data licence different from other licences?

Because you cannot easily give the data back.

HHQ, a law firm, wrote that parties often overlook how to unwind a licensing deal. It said complete removal of a dataset from a trained AI model may not be possible with current technology, and that where it is possible it may require retraining the model, which is expensive and takes time. Its conclusion for sellers was blunt: the licensor may have to be prepared to leave its data with the model.

For a buyer, this means the end of the licence has to be agreed at the start.

A licence has three parts in time
  1. The grantWho may use what, for which uses, and whether anyone else may too.
  2. The useWhere the data sits, who has access, and what records you keep.
  3. The endWhat is deleted, what stays inside a trained model, and what survives.

What should the grant say?

The uses. State them: training, fine-tuning, evaluation, building test environments for agents. A use that is not listed is a use you do not have. HHQ noted that sellers want to control what type of AI is trained on their data, so expect limits here.

Exclusivity. HHQ’s advice is that exclusivity should be clarified and agreed up front. Decide whether you need the data to yourself, to yourself for one use, or not at all. Exclusive terms cost more and need a clear end date.

The term. Public deals give a sense of length. TechCrunch reported from Reddit’s IPO filing that its data licensing arrangements had an aggregate contract value of $203.0 million and terms ranging from two to three years. Those are deals for a live platform. A licence for a fixed copy of past records can be simpler, but it still needs a start, an end and a renewal rule.

Which warranties matter most?

The right to license. HHQ called the licensor’s ability to grant the licence crucial, and said it is often reflected as a key warranty in the agreement. It warned that the party offering the data may not hold rights to everything in it, and that individual permissions may be needed.

Lawful collection, as a continuing promise. Best Lawyers advises requiring ongoing compliance representations, with the example of a vendor confirming that training data was lawfully acquired and used. It also advises scoping indemnity to cover data privacy violations and regulatory penalties, not only intellectual property claims.

Bricker Graydon Wyatt made the related point for acquisitions: problems with data rights may not emerge until months after closing, so the documents should be written with that in mind. The same is true of a licence.

What should the licence say about personal data and passing data on?

Say where the data is stored, who has access, and what security applies. Best Lawyers suggests clear retention and deletion rules that reach backups, caches and shadow copies, and audit logs and reporting for data usage and access.

Then deal with onward transfer. The Spirit Airlines bankruptcy sale gives a model term. TNW reported that the buyer there must publicly commit not to try to re-identify the data, and may not pass it on unless the recipient makes the same commitment. A promise that stops at the first buyer protects nobody.

Who owns what you build, and what happens at the end?

HHQ advises the parties to define who retains ownership of any new intellectual property created by using the licensed dataset. Settle three things:

  • The model. You will want to own the model and its weights. Say so.
  • Derived data. Say whether cleaned, labelled or synthetic versions of the data are yours, the seller’s, or shared. Best Lawyers suggests defining “data” to include metadata, embeddings, synthetic data and derivative datasets.
  • The end. Say which copies are deleted and when. Say plainly that models already trained may keep running. If the seller does not accept that, you need to know before you train.
Eight clauses to settle before you sign
  • The uses you may make of the data
  • Exclusive, non-exclusive, or exclusive for one use
  • The term, and what renews
  • The seller’s right to license, as a warranty
  • Personal data, storage and security
  • Who owns what you build
  • Whether you may pass the data on
  • The end: deletion, and trained models

Where does a marketplace fit?

On Korra, the licence is between you and the seller. We hand both sides a template licence and a delivery checklist, and you can use your own paper instead. We are not a party to the licence and we do not deliver the data.

Your offer already carries the main terms: the amount, the use, the term and the delivery you want. The clauses above turn that offer into a contract. Have your own adviser read it. The buyer agreement and how buying works show where each step sits.

Questions

Can licensed data be removed from a trained model?
Often not in practice. HHQ wrote that complete removal may not be possible with current technology, or may require retraining the model. Agree what happens to trained models before you sign.
Should a buyer ask for exclusivity?
Only if it is worth the price. Exclusive terms stop the seller from licensing the same data to others, so sellers charge more for them. Exclusivity for one use or for a set time is a common middle course.
How long does a data licence run?
It varies. TechCrunch reported that Reddit’s data licensing arrangements had terms ranging from two to three years. A licence for a fixed copy of past records can be set differently.
Is a template licence enough?
A template is a starting point. The uses, the warranties and the end-of-term rules depend on the data and on your plans, so have your own adviser read the final text.

Sources

  1. Licensing of Data for AI Model Training: Things to Take Note of — HHQ
  2. Key IP Licensing Considerations in AI — Best Lawyers
  3. Reddit says it’s made $203M so far licensing its data — TechCrunch
  4. Spirit Airlines data risks ‘eliminated or mitigated’, ombudsman says — TNW
  5. AI in M&A: Emerging Due Diligence and Liability Considerations — Bricker Graydon Wyatt

Figures are as reported by the publishers above. Discussion threads show what people said; they are not facts or Korra results.

See the shelf working.

Book a call. We open a card, an offer and the paper with you.

How buying works

Keep reading.