Before buying a dataset, an AI lab checks six things: where the data came from, whether the seller has the right to license it, how personal data was handled, whether the data is already public or in other models, whether quality matches the sample, and whether the transfer is secure. Each question needs documentary evidence.
What buyer diligence is for
An AI lab that licenses data takes on two kinds of risk. The first is legal: a claim that the data was never the seller's to license, or that it contains personal information handled unlawfully. The second is technical: paying for data that is thin, duplicated, already public or different from the sample.
Due diligence is the process of reducing both before money moves. It is run by a small group on the buyer's side, usually a data lead, counsel and a security reviewer. They work through a list of questions and ask the seller for evidence on each.
Sellers benefit from knowing the list in advance. A company that can answer each question with a document closes faster and gives the buyer less reason to discount the price. This guide follows the order most buyers use.
Provenance
Provenance is the record of where the data came from and what happened to it on the way to delivery. The buyer wants to know which systems produced the records, over what period, and through which processing steps they passed.
First-party operational data, created by a company in the course of its own business, is the easiest to document. The buyer will ask whether any part was scraped, purchased or received from a partner, because each of those adds a second chain of rights. It will also ask whether any records are synthetic or machine-generated, since model output mixed into a dataset sold as human work changes its value.
The usual evidence is a data card and a processing log: source systems, export dates, filters applied, de-identification steps and the tools used. A buyer may also need this material for its own training-data documentation.
Rights
The rights question is whether the seller can lawfully grant the license it is offering. The buyer checks ownership, contract restrictions and earlier grants.
On ownership, it asks how the data was created and whether employees and contractors assigned their rights. On contracts, it asks whether customer agreements, data processing agreements or non-disclosure agreements restrict use of any records in the set, and how the restricted records were removed. On earlier grants, it asks whether the data has been licensed before, to whom and on what terms.
A data title report answers most of this in one document. Without one, the buyer's counsel reviews the underlying contracts or relies more heavily on the seller's representations and indemnity. Outrich Data supplies a title report with each dataset it sells and can prepare one for deals between other parties.
Privacy
Privacy diligence asks whether the dataset contains personal data and, if it did at source, how that was dealt with. The buyer wants the method, not only the claim.
It will ask which identifier classes were removed, which tools were used, whether free text was reviewed by people as well as by software, and how the result was tested. It will ask which jurisdictions the data subjects are in, because the standards differ. It will ask whether any special categories are present, such as health or financial records, and whether sector rules such as HIPAA applied to the source.
Most buyers then run their own scan on a sample. A seller whose data card describes the method precisely, and whose sample passes the buyer's scan, clears this stage quickly.
Contamination and held-out status
Contamination means the data, or a close copy of it, is already in a model's training set. It matters most for evaluation. A test only measures a model's ability if the model has never seen the test material. Data that was ever public may already have been collected by a web crawler.
The buyer asks whether any part of the dataset has been published, posted to a public repository, indexed by a search engine or shared with another AI company. It asks whether the systems that hold the data were ever open to the internet without a login. For code, it asks whether any repository was public at any time.
Evidence here is partly negative, which makes it harder. The seller describes the access controls on the source systems and gives a written statement of publication and licensing history. Some buyers test for contamination directly by checking whether an existing model can complete passages from the sample.
Quality and security of transfer
Quality diligence compares the full dataset with the sample and the description. The buyer profiles record counts, field fill rates, duplicate rates, language mix and length distributions. It looks for signs of padding, such as templated or automated records counted as conversations. It checks that the time span is evenly covered.
Security diligence covers how the data moves. The buyer asks how files will be transferred, whether they are encrypted in transit and at rest, who on each side has access and how integrity is verified on arrival. A checksum manifest answers the last point. The buyer's security team may also ask whether the files were scanned for credentials and keys, because secrets inside logs and code are a risk to both parties.
The checklist as a table
The table condenses the questions above. Each row gives the question, the reason a buyer asks it and the evidence that settles it.
| Question | Why it matters | What evidence answers it |
|---|---|---|
| Which systems produced the data? | Establishes first-party origin | Data card with source systems and export dates |
| Was any of it scraped or bought? | Adds a second chain of rights | Seller statement and source breakdown |
| Who owns it? | The license is only as good as the grantor's rights | Title report; assignment agreements |
| Do customer contracts restrict it? | Confidentiality and use limits bind the seller | Contract review summary; list of excluded accounts |
| Has it been licensed before? | Affects exclusivity and value | Schedule of earlier licenses |
| How was personal data removed? | Privacy law and reputational exposure | De-identification method statement; scan results |
| Where are the data subjects? | Different laws set different standards | Jurisdiction breakdown in the data card |
| Was it ever public? | Public data may already be in training sets | Publication history; access control description |
| Does the full set match the sample? | Samples can be selected to look good | Dataset profile; buyer's own sample checks |
| How many duplicates and empty fields? | Padding inflates the volume paid for | Duplicate and fill-rate report |
| How is it transferred and verified? | Loss or tampering in transit | Transfer method; SHA-256 manifest |
| Are there secrets in the files? | Credentials in logs and code create liability | Secrets scan report |
How sellers can prepare
A seller does not need to wait for the questions. Assembling the evidence before the first conversation shortens the review and shows the buyer that the data has been handled with care.
- Write a data card that covers source, schema, volume, time span, processing and known gaps
- Keep a processing log of each export, filter and de-identification step
- Collect the contracts and notices a title search needs into an indexed data room
- Prepare a written statement on publication history and earlier licenses
- Run a profile, a duplicate check and a secrets scan, and keep the reports
- Build a checksum manifest for the sample and for the full delivery