A data provenance checklist for training sets

A provenance problem rarely shows up on the day you buy data. It shows up a year later, when a customer’s procurement team asks what your model was trained on, or a rightsholder writes to ask the same question less politely. By then the person who signed the order may have left, and the vendor’s sales contact certainly has. The only defense is paper you collected at the start.

Why provenance became a buying question

For years, training data was treated as a technical input. Two changes have moved it into contracts. First, rightsholders now license data for training openly and at scale, which means unlicensed use is easier to identify and harder to excuse. Second, the data on offer has widened from public text to private material: company records, recordings, and the business files of companies that have closed. Reuters, in August 2026, covered rival bids by Google and Mercor for Spirit Airlines’ internal mail, chat and files, which were to be de-identified ahead of any sale.

The wider the data, the more people have a stake in it. Each of those stakes needs an answer in writing.

The five documents to request

1. A source statement for every component

Not one line for the whole set. A dataset assembled from a licensed archive, crowd tasks and a generator has three provenance stories, and each needs its own: who created it, when, and under what agreement.

2. A license that names AI training

Look for the words. “Internal use”, “research” and “analytics” do not obviously cover training a commercial model. Check four further points in the same document: whether the license is exclusive, how long it lasts, what happens to models trained during the term, and who owns derived outputs such as labels or rewritten text.

Exclusivity varies by seller. Three sellers, three answers: Labelbox Prism defaults to exclusive licenses, CCC’s collective training license is non-exclusive, and publishers who join RSL Collective pay nothing and keep the freedom to license elsewhere.

3. Contributor agreements

Annotators, writers, speakers and participants should have signed terms that cover AI training. Ask for the template, not just an assurance. For voice or face data, ask specifically, because voiceprints and face geometry can be biometric identifiers under Illinois BIPA, Texas law (Business and Commerce Code chapter 503) and Washington’s biometric statute.

4. The method used to remove personal data

“We anonymize everything” is a promise; a method is evidence. Ask what is removed (names, contact details, account numbers, faces), how (rules, models, human review), where it happens (on the seller’s systems or after transfer) and how the vendor measures what it missed.

5. A delivery manifest

A list of files with hashes, or an equivalent record, lets you show exactly which data trained which model version. Without it, a later deletion request or dispute becomes guesswork.

Questions that separate prepared vendors

Question	Prepared answer	Warning sign
Who created this component?	A named source and collection period	“Various public sources”
Can you sublicense it to us for training?	Yes, with the upstream clause quoted	“Our customers use it for that”
How did contributors consent?	A template agreement covering AI training	A link to general website terms
When the license lapses, what about our models?	Models already trained may stay in use; raw files are deleted	No answer about models
What is warranted?	Title and rights, with an indemnity	“Provided as is”

A vendor that answers the left column quickly has been asked before. One that needs a call to explain, or offers answers only to its largest customers, is telling you where the gaps are.

What changes for synthetic and generated data

Generated data does not escape provenance; it moves the questions up one level. Ask what seed data the generator learned from and whether that seed data could be used this way. Ask which model produced the output and whether that model’s terms let its output train another model. And ask how the vendor checked that generated records do not reproduce real people’s details from the seed set, because a generator trained on personal data can echo it.

Records to keep after you buy

• The source statements, license and contributor template, filed with the model’s training run.
• The delivery manifest and the dataset version used for each model version.
• Any de-identification report the vendor provided.
• Dates: when the license started, when it ends, and when you confirmed deletion if required.

Store them where an auditor, not just an engineer, can find them.

Where to go from here

The full provenance checklist breaks these five documents into 21 specific checks you can send as a questionnaire. Every entry in the directory of AI training data companies links to the company’s own site, which is the right place to start asking. Neither replaces your own lawyer reading the agreement before you sign.

Provenance is cheap to collect on the first day and expensive to reconstruct on the last.