Questions to ask a training data vendor before you shortlist them

What a first call can settle, and what to leave for written diligence

A first call should establish whether the vendor can answer questions on rights, source, consent, samples, delivery and price specifically enough to justify a sample request; it is not the place to collect evidence. Contracts, consent records and test results come later, through a written due diligence questionnaire (DDQ), a sample evaluation and a dataset-level legal review.

First ask which kind of vendor you are talking to, because the follow-up questions depend on it:

• Data holder: the company whose systems created the records. Rights questions go to its own customer and employee terms.
• Broker or reseller: licenses data obtained from others. Ask for the upstream sublicensing language; see what to verify when buying through a broker or reseller.
• Collection or annotation vendor: commissions new recordings or labels. Ask about participant agreements and IP assignment.
• Synthetic data vendor: generates records with models. Ask which models and seed data.

The formal follow-up belongs in a data provider due diligence questionnaire and SourceX's AI training data due diligence checklist.

The first-call question set and the answers that disqualify

Fourteen questions fit a 30-minute call; the last column lists answers that should end or escalate the conversation. Ask them in order, because a failed rights answer makes the rest moot.

Rights basis: the questions to settle before anything else

The rights questions come first because no sample or price can fix a vendor that cannot show a chain from the record holder to you. Alternative-data buyers ask at this depth: a Lowenstein Sandler guide says investment-firm buyers want detailed provenance covering paid subscriptions, surveys, industry conversations and web scraping, plus any past, current or threatened litigation or enforcement [1]. The FISD Alternative Data Council's illustrative DDQ, whose 2024 edition adds generative AI questions, asks whether a vendor buys data from others and, if so, for the contract terms that permit resale [2].

"Publicly available" is not a rights basis. The Data Provenance Initiative's audit of more than 1,800 text datasets reported license omission rates above 70% and error rates above 50% on popular dataset hosting sites [5], so an inherited license label needs tracing to its origin. Web permissions also shift: the Consent in Crisis authors report that between 2023 and 2024 about 5% of all C4 tokens, and over 28% of its most actively maintained critical sources, became fully restricted from use [6].

For model-generated records, ask about the generator's output terms; Anthropic's help center, for example, states that its terms do not allow outputs to be used to train competing models [7]. For open licenses with use limits, see how non-commercial datasets interact with commercial training.

Source systems and the documentation a vendor should already have

A vendor that knows its data can name the source application, export path and date range, and can send a data dictionary after the call. Ask what the records capture at the process level, not the industry level: a vendor guide to off-the-shelf training data itself notes that such data is generally oriented to foundational business areas rather than specific processes [4], so "customer service conversations" may lack the escalation, refund or tool-call steps your agent needs.

Ask what documentation exists. A datasheet in the sense of Gebru et al. records motivation, composition, collection process and recommended uses [8], and Croissant-RAI adds machine-readable responsible-AI fields to the Croissant metadata format [9]. A vendor with neither can still qualify, but you will write that documentation yourself.

Consent and personal data questions counsel will raise anyway

Questions 7 and 8 decide whether privacy review can clear the data at all, so ask them before anyone loads a sample. The core test is whether the people in the records were told about this use at collection. FTC staff warned in 2024 that adopting more permissive data practices, such as AI training or third-party sharing, through a surreptitious, retroactive change to terms or privacy policies may be unfair or deceptive [10]. Remedies can reach models: the FTC's 2021 Everalbum order required deletion of models and algorithms developed from users' photos and videos [11].

Listen for answers tied to the governing law:

• Call and meeting audio: California Penal Code section 632 prohibits recording a confidential communication without the consent of all parties [12]. See the call-recording consent checks for buyers.
• Faces and voices: Illinois's Biometric Information Privacy Act covers voiceprints and face-geometry scans and requires written notice and a written release before collection [13].
• Health records: HHS describes two HIPAA de-identification methods, Safe Harbor (removing 18 listed identifiers) and Expert Determination [14]. The vendor should name one.

The FISD template asks for copies of consent or other terms agreed with the individuals whose data is collected [2]. Request the notice text in writing and check it against the guide to consent and notice records.

Turning quality talk into numbers you can test

Accept only quality claims that a sample can confirm or refute. A Hugging Face forum discussion on selling datasets draws the line: "the JSONL is valid" or "the exact-duplicate rate is below 1%" can be checked, while "good for fine-tuning" cannot [3]. The thread also separates three claims: that a specific revision was tested, that the producer runs a repeatable correction process, and that the data was tested for a named buyer's intended use [3]. Ask which of the three the vendor can document.

Ask about three measures by name:

• Duplicate rates. Lee et al. found many near-duplicates in common language-modeling datasets; models trained on deduplicated data emitted memorized text about ten times less often [15].
• Audited label error rate. Northcutt et al. estimated an average label error rate of at least 3.3% across the test sets of ten widely used datasets [16], so ask how any vendor figure was measured.
• Prior exposure. For evaluation data, ask whether any subset has been published or shared. In a February 2026 post, OpenAI said it no longer reports SWE-bench Verified because the benchmark is increasingly contaminated [17].

Then request a sample sized for those tests under evaluation license or NDA terms that permit them.

Delivery mechanics and the pricing unit

Format, transfer route and billable unit can change a cost comparison more than the headline number, so pin them down before any quote. JSON Lines files must be UTF-8 with one valid JSON value per line [18]; a vendor that says "JSON" may mean one multi-gigabyte array your loader cannot stream. For multi-terabyte deliveries, check that the plan uses a service still on offer: as of October 2026, AWS no longer offers Snowball Edge to new customers and points them to DataSync, Data Transfer Terminal or partner solutions [19].

Per-token prices depend on which tokenizer counts; per-record prices depend on whether duplicates and filtered records are billable. Get both answers, then normalize offers to cost per usable record and check them against common data license pricing structures.

Disclosure duties that make vendor answers part of your records

If you build or modify models, several laws can require you to describe training data publicly or to deployers, depending on the type of model and where it is offered, and your description will rest on what vendors tell you. As of October 2026:

• California AB 2013 required developers of generative AI systems made available to Californians to post training-data documentation by January 1, 2026, covering items such as dataset sources or owners, copyrighted or licensed material, personal information, collection periods and synthetic data [20]. See the records to collect for AB 2013.
• EU AI Act Article 53(1)(d) requires general-purpose AI model providers to publish a sufficiently detailed training-content summary using the AI Office template, an obligation in effect since August 2, 2025 [21].
• Colorado SB26-189 requires developers of automated decision-making technology that materially influences consequential decisions to give deployers documentation including training data categories, with its main duties starting January 1, 2027 [22].

Question 14 tests whether a vendor's confidentiality terms let you meet these duties.

When an intermediary sources the data for you

With an intermediary, ask the same questions about the underlying record holder, plus two more: did the holder approve this release, and under what agreement does the intermediary act? SourceX is one such intermediary: it sources operational datasets from US companies on request, buyers describe the data rather than the businesses, and every release is approved by the supplying company. Each dataset goes through rights review, which checks that the business owns or may share the records and that required consents are in place, and diligence materials on source, rights, preparation and allowed use are prepared per dataset for the buyer's review. A request does not guarantee a matching dataset; you can describe the data you need on SourceX's buyer page.

After the call: what to put in writing

Send a written recap within a day and ask the vendor to correct it; a spoken "training is fine" is not a license term. Move vendors that pass to the written DDQ, collect the evidence to request from a data vendor before you sign, and rank them on the data vendor evaluation scorecard. SourceX's guide to evaluating data supplier quality and the AI training data procurement hub cover the remaining steps.

This page is general information, not legal advice. Confirm requirements with counsel for your jurisdiction and use case.

Qualifying a data supplier for your next purchase?

At SourceX's buyer page you can submit a data request, talk to SourceX or browse the dataset catalog. Describe the records you need; SourceX looks for US businesses that hold them, checks the data and the supplier's licensing permissions, and manages the license and delivery, and nothing is contracted until a supplier agrees. See how SourceX works with data buyers.