Data provenance for AI training

Data provenance is the documented history of where a dataset came from, who created it and what rights attach to it. For AI training data, provenance decides whether a buyer can use the data safely, and it affects the price. This guide lists what to document and how buyers check it.

What should a provenance record contain?

• Source. Where each part of the data was collected, and when.
• Creators and contributors. Who produced it and under what agreement.
• Consent and terms. What contributors or users agreed to, including any restrictions.
• Rights. The license or ownership basis for each source.
• Transformations. Cleaning, filtering, annotation and de-identification steps.
• Exclusions. Content removed because of opt-outs, takedown requests or privacy rules.

Package this with a data card that adds contents, limits and intended use.

How do buyers check provenance?

Buyers sample records and trace them to sources, review contributor agreements, run duplicate checks against known public sets and ask for warranties. For sensitive data they check PII handling and, where relevant, clean-room access.

Why does provenance affect price?

Clean provenance reduces legal risk, so buyers pay more and close faster. Datasets with disputed or unknown rights sell at a discount or not at all. The deal tracker includes a reported settlement that shows how costly a weak rights chain can be.

How do you keep records over time?

Log each source and change when it happens rather than reconstructing later. Version datasets, keep contributor agreements and record every delivery. These records also support the clauses in the licensing checklist.

Where does provenance fit in a sale?

Sellers should prepare it before outreach, as step one of the seller process. Buyers should request it in the RFP and check it during the pilot.