AI training data due diligence: a checklist for buyers.

Problems in training data travel into the model. They cannot easily be removed later, because a trained model cannot simply forget what it learned. Four developments have made checks before purchase a standard step.

• EU AI Act documentation. Providers of general-purpose AI models must keep technical documentation covering the type and provenance (origin) of training data, how it was obtained and selected and how it was curated (Annex XI), have a policy to comply with EU copyright law (Article 53(1)(c)) and publish a training data summary (Article 53(1)(d)).
• US copyright. The US Copyright Office concluded in May 2025 that whether training is fair use depends on what works were used, from what source, for what purpose and with what controls on outputs, and that using works obtained through illegal access weighs against it.
• Data protection. The European Data Protection Board's Opinion 28/2024 says a model trained on personal data cannot in all cases be considered anonymous, and that a company deploying a model should assess whether it was developed by unlawfully processing personal data.
• Location. In Getty Images v Stability AI (High Court, November 2025), the claim about training was abandoned because there was no evidence the training took place in the UK. Where data is processed affects which law applies.

Most company data starts as personal data. The question is whether it has been lawfully and effectively turned into data that no longer identifies anyone.

• Lawful basis for de-identification. In the UK, anonymising personal data counts as processing, so the supplier needed a lawful basis and transparent privacy information.
• Effectiveness. Ask how the supplier tested the result. The ICO's motivated intruder test asks whether a reasonably competent person could identify someone using public information. Ask for samples from every source, including free text.
• Pseudonymised is not anonymous. If the supplier still holds a key linking codes to people, the data is personal data in its hands, and may be in yours too. See anonymisation vs pseudonymisation.
• US requirements. For California, information is deidentified only if recipients are bound by contract to the deidentification conditions. For health data, ask which HIPAA method was used, as set out in our HIPAA de-identification guide.
• Special category data. Under UK and EU GDPR, information about health, ethnicity, religion, political opinions, trade union membership, sex life, sexual orientation, and genetic and biometric data used to identify people carries extra protection. Confirm it was excluded.

1. Send the questionnaire early. Share the checklist with the supplier before price is agreed, so gaps surface while terms can still change.
2. Test a sample. Check a representative sample against the data sheet, and read free text from every source for names and other identifying details.
3. Rate each area. Mark each of the ten areas as cleared, cleared with conditions, or not cleared, with the evidence for each decision.
4. Fix or price the gaps. Ask the supplier to remove material, provide records or give specific warranties. Where a gap cannot be closed, narrow the use or decline the dataset.
5. Keep the file. Store the evidence with the dataset, because it feeds technical documentation, the EU training data summary and any later enquiry.

Due diligence reduces risk; the licence allocates what remains. The key protections are warranties on rights, lawful collection and de-identification; an indemnity against third-party claims; audit rights; a clear end-of-term rule that lets trained models continue in use; and the right to disclose licensed data in the EU AI Act summary. Our guide to the AI data licensing agreement covers each term.

Buying from a supplier that has already done this work saves time. Every dataset supplied through our AI training data service comes with the licence, a rights record, an exclusion record, a de-identification report and a data sheet.