PRESIDIO

Guide · Licensing data for AI training

What an AI lab asks for before it licenses your dataset.

If you want to sell or license a dataset to AI companies, the price is rarely the hard part. The lab's legal and data teams typically ask six sets of questions before they sign. This guide lists them, says what answers them, and where the answer is a legal one for your counsel.

By Vladimir Stantchev, PRESIDIO · 2026-11-08

1 of 6

Rights: can you license it at all?

The first question is not about privacy. It is whether you may license the data for training. The lab will ask where each source came from, under which contract or terms, and whether those terms allow onward licensing for AI training. Data bought from a company that no longer exists, or collected under terms that never mentioned training, is where most deals stall.

What helps: a schedule of sources with the contract or terms that cover each one, the date it was acquired, and any restriction (territory, field of use, term). Your counsel drafts the warranties from that schedule; nobody can draft them without it.

2 of 6

Personal data: what is left after de-identification?

Direct identifiers such as names, email addresses and phone numbers are the easy part: you drop them or replace them with pseudonyms. The harder parts are quasi-identifiers (dates, postcodes, rare job titles) that single people out in combination, and free text that mentions people by name.

Since the Court of Justice's ruling in EDPS v SRB (C-413/23 P, September 2025), whether pseudonymised data is personal data depends on what that recipient can realistically do with it; for you, holding the key, it remains personal data. So the question is asked per lab: what else it holds, what the licence forbids, how access is controlled. Pseudonyms keyed per recipient, so that two labs cannot join their copies, are one practical response.

What helps: a column-by-column policy (drop, pseudonymise, generalise, scrub free text, keep) with a reason for each, and counts of how many rows share each combination of quasi-identifiers. Whether the result is personal data for that lab is a legal question; the policy and the counts are the facts your counsel needs to answer it.

3 of 6

Opt-outs and deletion requests: which list, and when?

People who opted out or asked for deletion should not be in the release. The lab will want to know which suppression lists you applied and as of which date, and what happens with the next deletion cycle after it already holds a copy.

In California, data brokers registered under the Delete Act have had to process the state's DROP deletion lists at least every 45 days since 1 August 2026, and at least every 30 days from 1 January 2027 (AB 883). DROP distributes hashed identifiers, so matching needs its exact standardisation and hash format. Whether you are a data broker under the Delete Act is for your counsel.

What helps: each list identified by its hash and date, the number of rows it removed, and a re-release that links to the earlier one, so the lab can see which release is current.

4 of 6

Provenance file: one document the lab can file

Some labs ask for a provenance file: where the data came from, what was done to it, by whom and when. Two public formats are common reference points: the Data Provenance Standards of the Data & Trusted AI Alliance (formerly Data & Trust Alliance), and MLCommons Croissant, with its responsible-AI fields, for dataset metadata.

What helps: a file generated from the actual processing run rather than written afterwards, with hashes that tie it to the released files.

5 of 6

Training-content summary: the lab's own obligation

Providers of general-purpose AI models in the EU must publish a summary of the content used for training, on the Commission's template, including licensed and private data sources. The AI Office can enforce this since 2 August 2026. That obligation sits with the lab, but the lab can only fill in its summary from what the seller documents.

What helps: the fields the template asks about (modality, nature of the content such as personal or copyright-protected data, languages, the dataset's name where it is publicly known, and optionally period and size) answered in the provenance file, so the lab does not have to ask again.

6 of 6

Evidence: a record, not a spreadsheet

Warranties are only as good as what stands behind them. A spreadsheet written after the fact cannot show that the released files are the ones it describes. A signed record of the processing run can: it names the policy, the suppression lists and the counts, and it carries the hash of every released file.

What helps: a record the lab can check itself, offline, without an account with the seller or a third party, and that never contains the values it describes.

Checklist

What to have ready before the first call with a lab.

  1. A sources-and-rights schedule, one line per source, with the terms that allow training use.
  2. A column policy with a reason for every column, and the quasi-identifier counts for the output.
  3. The opt-out and deletion lists applied, each with its hash, date and number of rows removed.
  4. A provenance file generated from the processing run, mapped to the fields the lab's training-content summary needs.
  5. A signed release record that binds all of it to the released files, verifiable by the lab offline.

Data marketplaces and enterprise-data programmes that connect sellers with labs tend to ask the same questions. Having the answers ready shortens those conversations.

How PRESIDIO helps

Dataset Release.

We help you prepare the answers: a workshop on your schema, contracts and opt-out lists, a column policy and suppression procedure, and a signed release record the lab can verify offline with an open-source verifier (in development). The tools run in your environment; the data does not leave it.

See the Dataset Release offer Book a 30-minute call

This guide is general information, not legal advice. Whether data is personal for a given recipient, and whether you fall under a given law, are legal questions for your counsel.