Who licenses AI training data, and what each type of buyer asks for
A model developer, an aggregator and a research lab may ask for the same files but want very different rights. Match the counterparty's questions to its proposed use before discussing access or a fee.
Same files, different counterparties
- Proposed use Training, fine-tuning, evaluation or resale.
- Onward rights Sublicensing and redistribution.
- Evidence asked Provenance, labels, privacy method.
Four archetypes
| Archetype | Typical interest | Likely first questions |
|---|---|---|
| Frontier and foundation model developers | Large, clean, diverse corpora; rare domains | Provenance at scale, commercial training rights, documentation for public disclosures |
| Vertical model builders | Narrow expert data: legal, clinical, industrial | Label accuracy, expert annotation, privacy method |
| Data vendors and aggregators | Inventory to annotate or resell | Sublicensing rights — the clause that most often hurts owners |
| Academic and nonprofit labs | Research and benchmarking | Non-commercial terms, IRB considerations |
Why disclosure laws matter to buyers
California's AB 2013 requires developers of generative AI systems offered to Californians to post high-level documentation about their training data. Buyers subject to it will want metadata from you that they can summarize publicly. A seller who can provide it reduces friction.
Red flags in a counterparty
- Cannot describe the model or use case.
- Wants the full dataset before signing an NDA or license.
- Refuses any field-of-use or deletion terms.
- Asks you to remove provenance metadata.
How this affects value
Sublicensing rights requested by intermediaries can be worth more than the delivery itself—price them separately.
See what shapes a dataset's value