Resources
Datasheets for datasets: a buyer’s template
Turn the academic datasheet idea into a practical diligence form. Blank fields become findings.
Published 2026-07-22 · 7 min read
Key takeaways
- A datasheet turns diligence into a form the supplier fills in.
- The academic version documents; the buyer version adds rights fields.
- Licence, consent, and chain of title are the fields you cannot skip.
- Blank fields are findings, not formatting problems.
- Ask for one datasheet per dataset, not one per vendor.
In 2018, a group of researchers proposed a simple idea. Every dataset should ship with a datasheet, the way an electronic component ships with a spec sheet. The datasheet would record why the data was collected, what is in it, how it was gathered, and what it should and should not be used for.
The original paper, Datasheets for Datasets by Gebru and colleagues, was about transparency between the people who make datasets and the people who use them. For a buyer licensing AI-training data, the same document doubles as a diligence tool. If you make the datasheet a condition of purchase, the blank fields become your findings.
Where the idea comes from
The paper proposed that a datasheet cover a dataset’s motivation, composition, collection process, recommended uses, distribution, and maintenance. The goal was to prompt creators to think through provenance and limitations, and to give consumers enough to decide whether a dataset fit their use.
That framing maps almost directly onto training-data diligence. You want the same things: what is in the set, where it came from, and what you are allowed to do with it.
What a buyer adds: rights fields
The academic datasheet was written mostly for documentation and fairness. A commercial buyer needs to bolt on a rights layer. Three fields do most of the work: the AI-training licence and its scope, the consent status for identifiable people, and the chain of title.
Without those, a datasheet can describe a beautifully documented set that you still cannot lawfully train on. Documentation and permission are different things. Ask for both.
The fields that matter most
Composition tells you what you are buying: modality, volume, languages, and who appears in the data. Collection process tells you how it was sourced and when. Consent tells you whether identifiable people agreed to training use. Rights and chain of title tell you whether the seller can grant what they are granting.
Recommended and excluded uses matter more than they look. If a datasheet says a set was collected for speech recognition, that is a signal about its shape, and a hint about gaps if you plan to use it for something else.
How to use it in a deal
Send the template before you buy. Ask the supplier to complete it per dataset, not per company. A blank or hand-waved field is not a paperwork gap. It is a finding, and often the most useful thing the process surfaces.
On fiund, the underlying records exist already: a signed AI-training licence on each asset, consent where people are identifiable, and provenance available in diligence. The datasheet is where you collect them into one page you can keep on file.
Buyer datasheet fields
- Motivation: why the dataset was created, and for whom
- Composition: modality, volume, languages, demographics, what each record contains
- Collection process: how and when it was sourced, and from whom
- Consent: how voice and likeness consent was obtained for identifiable people
- Rights: the AI-training licence, its scope, and exclusivity
- Chain of title: the line from creator to seller
- Recommended and excluded uses
- Distribution and retention: who else holds it, and deletion terms
- Maintenance: updates, versioning, and takedown handling
Sources
Frequently asked questions
Is a datasheet the same as a data card?
They are close cousins. Both document a dataset in a structured form. Data cards are often lighter and product-facing; datasheets are more thorough. For diligence, use whichever captures rights, consent, and provenance, and add those fields if the template lacks them.
Who fills out the datasheet, the buyer or the supplier?
The supplier completes it, because they hold the facts. The buyer supplies the template and treats missing answers as diligence findings. If a supplier will not complete one, that itself is information.
Does a datasheet replace a licence?
No. A datasheet documents the dataset. The licence grants the rights. The datasheet should point to the licence and summarise its scope, but it does not substitute for the signed grant.
Related resources
Want data that clears this in diligence?
Whether you're building a model or sitting on an archive, the first conversation is short and specific.
Send a brief