AI training data: where labs get it
Labs draw on six sources of training data. Company records are the newest.
AI training data is the text, images, audio, video, code and records that a model learns from. Labs also use it to tune and test the model.
Labs get it from six places: scraped web data, open datasets, platform and user data, made-up data, direct licensing deals and licensed data providers.
Company records are the newest source. They add something public text lacks, a record of real work from start to finish.
In plain words
AI labs have read most of what is public.
Think of it like this
Think of a library. The shelves are nearly read. Now the labs want to borrow how real work gets done.
This guide covers the sources, the limits on public text and what a company's own archive adds. The licensing process lives in the pillar guide on licensing company data to AI companies.
micro1 may pay us a referral fee if a company we introduce goes ahead with it. How we are paid
What counts as AI training data?
Three jobs use it. Pre-training teaches a model from very large amounts of data. Post-training and fine-tuning shape its behavior with smaller, curated sets. Evaluations test it.
Protege's pages describe its datasets as for "pre-training, post-training, fine-tuning and evals."
Human work sits inside the second and third jobs. Experts write model answers, grade output and flag errors. Platforms such as Mercor, Surge AI, Scale AI and Turing pay people to do it. That is labor sold to a lab. Your existing records are a different product.
Where do labs get their data?
Troveo's guide lists six sources: scraped web data, open datasets, platform and user data, synthetic (made-up) data, direct licensing deals and licensed data providers (Troveo). Troveo sells licensed data, so read its ranking with that in mind.
The same guide says made-up data fills gaps but "inherits the ceiling of the models that generate it." Labs use it as a supplement.
On platforms, labs with their own user content use it. Labs without it pay for someone else's. Reuters reported in February 2024 that Google pays Reddit about $60M a year (The Decoder).
Shutterstock's CEO said its AI licensing made $104M in 2023 (Bloomberg Law).
Is public training text running out?
Epoch AI estimated that the supply of good public human text could be fully used between 2026 and 2032, an 80% interval (Epoch).
It put the supply near 300 trillion tokens. It said heavier training on the same data could pull the date earlier.
The range is wide, so treat the dates as a bracket. The count covers public text only. Licensed libraries and private company records sit outside it. That is why private data gets attention.
What does company data add?
It adds a record of how a real problem got solved. Who asked, which tools were used, what approvals came and what the result was.
micro1's CFO described the prize in an undated recorded forum talk as the chain from a communication to a deliverable.
David Revland, who works on human data at micro1, gave the example of a Slack message asking for a finance report, then the report in a spreadsheet.
Fortune calls the recorded steps a trajectory (Fortune).
Labs and vendors use such records to build RL environments. These are simulated workplaces where an agent practices a job (Gizmodo).
Google bid $10M for the Spirit Airlines archive. It held about 100 million emails, 500 million Teams chats and about 30 million lines of code (Time). The sale was still before the bankruptcy court.
What does the demand look like?
Epoch interviewed RL environment vendors. A website-clone environment cost near $20,000. A Slack-grade clone cost near $300,000. Tasks ran $200 to $2,000. Exclusive deals ran roughly 4x to 5x the non-exclusive price (Epoch).
Those prices pay for environments and tasks that someone built. That differs from paying for an existing archive. Epoch's review does not discuss real company archives.
That is the gap in the demand case. Marketplace copy rarely names a buyer.
What happens when rights are unclear?
Content has already produced large claims. Anthropic agreed in September 2025 to pay at least $1.5B to settle an authors' lawsuit.
The suit covered about 500,000 books that the plaintiffs said came from piracy sites. That works out to roughly $3,000 per book (NBC News).
That case is about books, and company records are a different kind of material. It shows why buyers and data owners both care where data came from and who agreed to its use.
For company archives, the live dispute is the Spirit sale. The flight attendants' union and EPIC objected on privacy grounds (EPIC).
What is the owner's part?
Rights come first. Your archive holds customers' and employees' words. Counsel flags third-party confidentiality, trade secrets, legal privilege and employee talk that US labor law protects (Frankfurt Kurnit).
Run the rights check and read the risk map.
Count your systems and years of records using the sheet in how much your business data is worth to AI. Then take the free fit check. The AI data licensing guide explains outright transfer versus license. AI data companies sorts the players.
Talk to a privacy or commercial lawyer in your state before you sign.
Apply to micro1's data partnership (paid link)micro1's own data-partnerships pagePaid link: micro1 may pay us a referral fee if a company we introduce goes ahead with it. You pay nothing extra.
Vaultfish helps you run the process. We are not lawyers and this is not legal advice. If you need counsel who knows this area, we can point you to one. Ask for a referral
