Types of AI training data
There are four types of AI training data: private archives (records of real work), produced data (expert demonstrations and labels made to order), synthetic data (model-generated), and public data (the web). Labs combine all four; the price differences between them are explained by scarcity and rights, not by volume.
| Type | Strength | Limit |
|---|---|---|
| Private archives | Real decisions and outcomes; cannot be scraped | Rights, consent and de-identification work |
| Produced expert data | Built exactly to a specification and rubric | Expensive per item; no real-world messiness |
| Synthetic data | Cheap, fast, controllable | Teaches the model's own patterns back to it |
| Public data | Enormous and free to crawl where allowed | Running out (Epoch AI estimates the stock of public human text is exhausted between 2026 and 2032) and increasingly restricted |
Why archives are the growth category
Produced and synthetic data scale, but neither contains what actually happened inside a business. That is why buyers publish programmes paying $20,000 to over $3 million per company for archives (micro1, Mercor — checked November 2026), and why a bankrupt airline's archive drew a $10 million bid in 2026. Archives are also the only type where supply is controlled by thousands of individual owners rather than a handful of vendors.
Company owner reading this? The owner-side process lives on our for company owners page.
Sources — checked November 2026
- Epoch AI — public human text stock estimate, via ddiy.co statistics roundup, checked November 2026.
- Buyer programme pages (micro1, Mercor), checked November 2026.
Cite this entry: Best Data for AI. “Types of AI Training Data — Archives, Produced Data, Synthetic, Public.” Content reviewed 10 November 2026. Figures on this page are published buyer ranges with sources and check dates shown — they are not offers.