AI training data due diligence checklist

This checklist is for AI teams licensing data from companies, marketplaces or vendors. It follows the order in which problems usually surface, starting with who can sign and ending with the records you need to keep after delivery.

It draws on Nixon Peabody's October 2026 alert on buying AI training data, guidance from Frankfurt Kurnit Klein and Selz on company data sales and the Commission's template for EU AI Act training data summaries. It is not legal advice, and you should take independent advice on any deal.

What is AI training data due diligence?

AI training data due diligence is the review a buyer runs to confirm that a dataset can lawfully be used for training and matches its description. It covers the seller's authority and title, restrictions inherited from contracts and privacy law, the handling of personal data and third-party material, and the quality of what will actually be delivered.

The review matters more for training data than for most purchases because the use is hard to undo. Nixon Peabody notes that deleting source files may not remove their influence from a trained model, and that the FTC has required companies to delete models and algorithms built on improperly obtained data.

Which documents should you ask a seller for?

The documents to ask a seller for are the ones that turn each rights claim into evidence. Statements in an email or a license tag on a listing are a starting point, not proof.

How do you confirm the seller can sign?

You confirm the seller can sign by checking that the named signatory can bind the legal entity that owns the records. That sounds basic, but data often sits in group companies, former subsidiaries or businesses that have closed.

• Registered entity name and number for the company that holds the records
• Evidence that the signatory has authority, such as a board resolution for a larger deal
• For a closed or insolvent company, the appointment of the liquidator, administrator or trustee
• For a US bankruptcy sale, whether the sale is consistent with the privacy policy or has court approval after a consumer privacy ombudsman review under Section 332

Nixon Peabody points out that Section 363(b)(1) of the US Bankruptcy Code restricts sales of personally identifiable information that conflict with the debtor's privacy policy, and that court approval does not cure every restriction inside the records. Our guide to selling data from a closed company covers the seller's side.

How do you check ownership and chain of title?

Ownership and chain of title are checked by tracing each part of the dataset to the person or system that created it and the agreement that passed the rights on. A company owns its own operational records in most cases, but staff, contractors and clients may hold rights in parts of them.

• Employment and contractor agreements with IP assignment clauses, especially for code, designs and written content
• Client contracts stating who owns deliverables and working files
• Inbound licenses for any data the company bought or received from third parties
• Open-source license terms for any code in a repository

The Data Provenance Initiative found license omission rates above 70% on popular dataset hosting sites in its 2024 audit, which is why a license label is never enough on its own. Data provenance explained shows how to build the trail.

Which contracts can restrict a data sale?

The contracts that can restrict a data sale are client agreements, NDAs, vendor terms, data processing agreements and employment arrangements. Frankfurt Kurnit Klein and Selz notes that email and chat archives often contain other companies' confidential information covered by NDAs.

• Client confidentiality clauses that cover working documents and correspondence
• Data processing agreements where the seller acted as a processor, which usually bar use for its own purposes
• SaaS terms that limit how exported data from a platform may be used
• Union collective bargaining agreements, which may restrict a sale of workplace communications, according to the same firm

In the US, the firm also notes that employee communications about pay and working conditions may be protected concerted activity under Section 7 of the National Labor Relations Act. Ask whether staff representatives have been consulted where communications data is involved.

How do you check privacy notices and lawful basis?

Privacy notices are checked by comparing the notice in force at the time of collection with the proposed AI use. Frankfurt Kurnit Klein and Selz notes that companies must honor the statements made when the data was collected, and that the FTC treats quietly changing them as potentially deceptive.

• Dated copies of each privacy notice that covered the records
• The lawful basis relied on for disclosure where UK GDPR or EU GDPR applies
• Whether the data counts as a sale under the CCPA, and whether it meets the CCPA de-identification standard
• Whether any sector rules apply, such as HIPAA for health records or GLBA for financial records
• Whether employee and applicant data is included, which Nixon Peabody notes is generally subject to the same CCPA sale restrictions as consumer data

Our guide to UK GDPR, EU GDPR and CCPA when selling data for AI training explains each regime from the seller's side.

De-identification should be checked as a documented method with test results, not as a statement that names were removed. Nixon Peabody warns that job titles, dates and linked records can still identify people after names are gone.

• A written method, ideally applied by an independent provider and informed by NIST guidance
• Record-level exclusions for categories that should never transfer
• Re-identification testing against external data
• Certification before each transfer, not only at the start
• Contractual bans on re-identification backed by technical controls

Frankfurt Kurnit Klein and Selz goes further and recommends excluding high-risk data entirely rather than relying on de-identification. For US health data, the Department of Health and Human Services recognizes two methods: expert determination, or safe harbor removal of 18 types of identifier.

How do you check quality and the manifest?

Quality and the manifest are checked together, because the manifest is the seller's promise about what you will receive. Every count, date range and language it lists should be testable in the sample and again at delivery.

• Record counts, file counts and total size by source system
• Date ranges and languages, with the share of each
• File hashes tied to a specific version of the dataset
• Duplicate detection method with disclosed thresholds
• Label guidelines and reviewer agreement for any annotated data

How to evaluate a dataset sample sets out the tests, and how to build a data manifest shows what a good manifest contains.

What security and delivery terms should you agree?

Security and delivery terms should limit access to what each stage needs. Nixon Peabody recommends starting with a limited sample in a controlled environment and expanding access only as agreed conditions are met, and for sensitive archives keeping source data in a secure environment rather than handing over unrestricted copies.

Frankfurt Kurnit Klein and Selz advises sellers to avoid giving buyer tools direct access to company systems and to use a secure transfer method. As a buyer, a fixed export delivered that way is also easier to hash, version and accept.

Which contract protections should the license include?

The license should include protections that match who controlled collection, who can test the data and who decides how it is used. Nixon Peabody lists the main tools.

The AI data licensing agreement checklist covers the same clauses from the seller's side.

What records do you need for the EU AI Act?

For the EU AI Act you need enough information about each dataset to complete the Commission's public summary of training content if you provide a general-purpose AI model in the EU. Collect it during diligence, because it is much harder to reconstruct later.

• Modality and approximate size of the data
• Languages and any regional or demographic characteristics
• The latest date the data was collected
• Whether you licensed it directly from the rights holder or obtained it through an intermediary
• For any scraped portion, how text and data mining opt-outs were respected
• Measures taken to remove illegal content

Our guide to the EU AI Act and training data transparency explains each section of the template.

How does Sell My Data To AI handle seller diligence?

Sell My Data To AI runs seller diligence before you see anything. Every seller passes our eight-point vetting standard before any buyer sees their data, and the eight checks match the core table above. Buyers pay nothing to submit a requirement or review a shortlist.

We are a broker, not a law firm, so your own lawyers should still review the license. Book a buyer consultation to discuss a requirement.