AI data licensing agreements: the five clauses that actually decide the deal
A data license for model training is not a software license with a new name. The hardest questions concern what happens to a trained model after the data relationship ends. These five clauses carry most of that weight.
The clauses that carry the deal
- Scope Field of use and exclusivity.
- What persists Model survival, derived and synthetic data.
- Your exposure Deletion, audit and warranties.
1. Field of use
Spell out whether the licensee may use the data for pre-training, fine-tuning, evaluation only, internal tools only, or commercially distributed models. A license that just says 'AI purposes' invites disputes later.
2. Model survival
Decide explicitly whether models trained during the term may keep being used, sold or have their weights shared after the license expires. Sellers often assume termination 'unwinds' training; technically and contractually, it usually does not unless you say so.
3. Derived and synthetic data
Can the licensee use your data, or a model trained on it, to generate synthetic records that substitute for your dataset? If uniqueness is your main asset, this clause protects it.
4. Deletion and audit
- Deadline and form of deletion certificate.
- Whether deletion covers intermediate artifacts such as embeddings, filtered subsets and tokenized shards.
- Audit or attestation rights, and who pays for them.
5. Warranties you give
Expect to be asked to warrant that you hold the rights and that collection was lawful. Warrant only what your documentation proves. Where you are unsure, for example whether free-text fields contain personal information, disclose it as a known limitation rather than promise completeness. Have US counsel review indemnity caps.
Exclusivity is a separate decision
| Structure | Trade-off for the owner |
|---|---|
| Non-exclusive | Can license to several parties; each deal may be smaller |
| Exclusive by field or term | Single counterparty; check existing contracts first |
| Hosted access / clean room | Data never leaves you; you carry compute and security costs |
How this affects value
A narrower field of use or shorter model survival is a concession with a price—trade it deliberately, not by default.
See what shapes a dataset's value