What is an AI dataset?

A model can only learn patterns that appear in the examples it is shown. An AI dataset is that set of examples, prepared on purpose. A folder of photos is not yet one. The same photos become one when each is labelled, duplicates are removed, the files share a format, and someone has written down where they came from and what they may be used for.

The examples can be almost anything a computer can read: support tickets, X-ray images, call recordings, sensor logs, product listings, source code. What makes the collection useful is not its size alone but whether it matches the job the model will do once it leaves the lab.

• Examples: the individual items, each in a consistent format.
• Labels or targets, where the task needs them: the right answer the model should learn to give, such as a category, a transcript or a fault code.
• Splits: a training portion the model learns from, a validation portion used while tuning, and a held-back test portion used only to measure the result.
• Documentation: who made it, why, how the items were collected and cleaned, known gaps and recommended uses. The research paper Datasheets for Datasets proposed a standard set of questions for exactly this.
• Rights: a record of who owns the material and what licence the buyer receives.

This scenario is illustrative, not drawn from a real deal. Suppose a regional heating and cooling contractor has fifteen years of technician notes: the symptom a customer reported, what the engineer found, the part replaced and whether the fix held. A company building a diagnostic assistant for field engineers wants exactly that pairing of problem and resolution.

1. 1

   Scope it

   The broker and seller agree which years, regions and job types go in, and leave out jobs with missing outcomes.

2. 2

   Strip the people

   Customer names, addresses and phone numbers are removed, because the buyer needs the fault, not the household.

3. 3

   Structure it

   Free-text notes are kept, but each job gets fields for equipment model, fault category and outcome, so the buyer can filter and label.

4. 4

   Document it

   A short datasheet records the collection period, the share of jobs with notes, known blind spots such as one brand being over-represented, and the permitted uses.

5. 5

   License it

   The contract says the buyer may train and evaluate models, may not resell the raw notes, and must delete them if the deal ends early.

The contractor never sold data before. What changed the archive into an AI dataset was the preparation and the paperwork around it, not new collection.

• Ownership: did the seller create the material, or is some of it licensed in from someone else with terms that forbid AI use?
• Personal data: can people be identified, and does the seller have a lawful basis to share what remains?
• Fit: does the material cover the situations the buyer's model will face, or only a narrow slice?
• Contamination: has the same material already been published, so the buyer may already have it?
• Regulation: in the EU, the AI Act sets data governance expectations for the datasets behind high-risk systems, including records of origin and checks for bias. Buyers in scope will ask sellers for that information.

Bigger is not automatically better. A modest, well-labelled collection from a niche that is hard to observe can be worth more to the right buyer than a vast scrape of public pages. The opposite mistake is overselling quality: if a fifth of the labels were guessed by a junior clerk, say so, because the buyer's evaluation will find it anyway. Finally, a dataset is not a model. Selling the examples does not give the seller any claim on what the buyer builds unless the contract says otherwise.