AI training data: buyers, deals and rights

Early large models were trained mostly on scraped web text. Two things changed the economics. Lawsuits made unlicensed copying a measurable liability, and the open web stopped being enough: labs now want material that isn't freely available, such as archives, specialist corpora, conversation data and real-world code.

That has turned training data from something collected into something bought, and created room for brokers who can find owners and settle the rights.

These are outliers involving famous archives. Most datasets brokered between companies are far smaller and the price depends on how scarce the material is, how clean it is, and what rights come with it. We don't publish a per-token or per-record price because we haven't found a public benchmark that holds across data types.

• Material they can't scrape: private archives, licensed books, paywalled journalism, internal documents released by their owner.
• Domain depth: legal, medical, financial or engineering corpora written by specialists.
• Real-world code with history: repositories, issue trackers and reviews showing how software evolved.
• Multimodal pairs: images, audio or video matched with accurate descriptions or transcripts.
• Clear provenance: a documented chain showing who created the material and who may license it.

Copyright, EU transparency, privacy and contract

• Copyright: the US Copyright Office's May 2025 pre-publication report on generative AI training said several stages of development implicate copyright owners' rights and that fair use depends on the facts. In Thomson Reuters v. Ross, a Delaware court rejected fair use for training a non-generative legal research tool on Westlaw headnotes.
• EU transparency: since 2 August 2025, providers of general-purpose AI models must publish a summary of their training content using the European Commission's template under Article 53 of the AI Act. Models already on the market have until 2 August 2027.
• Personal data: if a dataset includes people's information, privacy law applies to the training use as much as to any other processing.
• Contract: the licence should state permitted uses (training only, or also output and retrieval), exclusivity, term, deletion on expiry and whether derived models survive termination.

1. 1

   Confirm ownership

   Who created the material, under what employment or contributor terms, and does the seller hold the rights or only a licence?

2. 2

   Find third-party content

   Open-source licences in code, quoted works in text, stock media in image sets.

3. 3

   Screen for personal data

   Names, emails, health or financial details. Remove, aggregate or document a lawful basis.

4. 4

   Check earlier promises

   Terms of service, customer contracts or privacy notices that limit reuse of the data.

5. 5

   Write the scope down

   Training, evaluation, retrieval and output uses are different rights and should be priced separately.

DataBroker is a training course with software tools for independent brokers. We don't sell datasets to AI labs and keep no inventory.

What the course covers is the broker's side: finding companies with archives, code or records that labs want, running the rights checklist above, estimating a range, and taking the deal from first call to signed licence. If you are a lab looking for data today, you will need a dataset seller or a labelling company, not us.