Commercial licensing for AI training data

Read for the model, not the data

A training-data licence is unusual because the thing you most need permission for is not the data but the artefact you build from it. A licence that grants generous access to files while leaving your model's status ambiguous has given you nothing you can ship.

The clauses that decide the deal

• Permitted use. Whether commercial training is allowed, and whether that extends to deploying the resulting model in a product.
• Model ownership. Whether you own the weights, embeddings and outputs, and whether you may sublicense the model to your own customers.
• Reconstruction. Almost every licence prohibits a model from which the data can be extracted. Look for a safe harbour for incidental memorisation that survives reasonable mitigation, otherwise ordinary training carries contractual risk.
• Redistribution. Whether contractors and cloud providers acting for you count as third parties, and on what terms.
• Fonts. Typefaces embedded in delivered files stay under their own licences. That is not a formality: it usually bars training a font-generation model on them, and it requires the notices to travel with the files.
• Warranty and indemnity. Whether the licensor warrants chain of title, and whether an infringement claim is defended, with what cap.
• Versions. Which delivered version the licence covers, and what happens when the dataset is updated.

Provenance is a deliverable

Ask for the documentation as part of the contract, not as a favour afterwards: a per-asset record of origin, a machine-readable manifest, and a paragraph you can paste into a training-content summary. A vendor who cannot supply these has not documented their own corpus.