How to sell data to AI companies

Models have already been trained on much of the public internet, so the data labs pay for tends to be what they can't collect themselves: private archives, specialist material, and records of real work. Public reporting gives a sense of the range, from large publishing and community deals to purchases of codebases and workplace data from companies that have closed.

• Provenance: who created the data, when, and under what terms. Labs want a chain they can show to their own lawyers.
• Rights: whether you can grant a licence for training, and whether that covers derived models and outputs.
• Uniqueness: whether the content is already on the open web or in common datasets. Public code and text add little; private history adds more.
• Personal and third-party material: what was removed, by what method, and what remains.
• Quality and structure: consistent formats, metadata, deduplication, and for code, intact commit history.
• Volume and coverage: enough to matter for the lab's purpose, with a clear description of what is and isn't included.

1. 1

   Confirm you can license it for training

   Read the terms under which the data was created: employee and contractor IP clauses, customer contracts, user terms, contributor agreements. Training use is a specific right; a general licence to "use" data may not cover it.

2. 2

   Check how much is already public

   If parts are on public websites, in public repositories or in widely shared datasets, say so and price accordingly. Buyers will check, and finding public material you called private damages trust.

3. 3

   Remove personal and third-party material

   Strip or de-identify names, contact details, health and financial information and anything about customers. For code, scan for secrets and keys, and list open-source components and their licences.

4. 4

   Package it in formats labs use

   Plain text or JSON lines for text, standard audio and video formats with timestamps, Git repositories with history for code. Include a data card: what it is, how it was made, what was removed and known gaps.

5. 5

   Prepare a sample and statistics

   A small representative sample plus counts (records, tokens, hours, repositories, date range, languages). Labs often decide from statistics before looking at content.

6. 6

   Reach the right team

   Large labs have data partnership or acquisition teams; data companies supply labs too. A broker who already knows what specific buyers are asking for can shorten the search.

7. 7

   License with clear limits

   Agree whether use is for training, evaluation or both, whether the licence is exclusive, how long data may be kept, and what warranties you give about rights. Have a lawyer review it.

Workplace archives are where the most money and the most risk sit. MIT Technology Review's reporting on sales of data from shut-down startups noted objections from the Spirit Airlines flight attendants' union and quoted a lawyer calling such data "legally radioactive"; emails and calendars can hold sensitive details about employees who never agreed to a sale. If your data involves people, decide what you will remove before you price it, not after a buyer asks.