For AI data buyers: questions to send a dataset owner, and the evidence to request
A useful diligence request pairs every question with evidence the owner can supply. Start with source rights and collection promises, then test whether the offered version fits the intended model use.
Question, evidence, decision
- Risk question Title, repurposing, biometrics, health.
- Evidence Assignments, archived notices, consents.
- Fit decision Version, use and disclosure fields.
Diligence by risk
| Risk | Question | Evidence to request |
|---|---|---|
| Title | How did you obtain each source? | Assignments, vendor contracts |
| Repurposing | What were data subjects told at collection? | Archived privacy notices with dates |
| Biometrics | Are faces or voiceprints present? | Consent records or exclusion method |
| Health | Is any data PHI, and how was it de-identified? | Safe Harbor checklist or expert determination |
| Contamination | Has the set ever been public? | Storage and access history |
| Disclosure | Can you supply fields for our public training-data summary? | Provenance dossier |
Signals of a well-prepared seller
- Frozen version with hash.
- Volunteers known limitations.
- Accepts field-of-use and deletion terms.
How this affects value
Sellers who can answer with evidence reduce your clearance cost—a legitimate factor in what a license is worth to you.
See what shapes a dataset's value