- AI training data is the examples a model learns from: text, images, audio, video, code, actions and human judgments.
- Pre-training data teaches broad knowledge; fine-tuning data teaches specific tasks; evaluation data tests the result.
- Most public web data has already been collected, so labs now license private and specialist data.
- Licensing gives AI developers permission and provenance; it gives data holders a new source of revenue.
The main stages
| Stage | Purpose | Typical data |
|---|---|---|
| Pre-training | Broad knowledge and language | Very large text, code, image and video corpora |
| Fine-tuning | Specific tasks and behavior | Curated examples, demonstrations, domain data |
| Preference training | Helpful, safe responses | Human rankings and critiques (RLHF) |
| Evaluation | Measure ability | Held-out tasks with checked answers |
| Retrieval | Ground answers in sources | Licensed documents searched at answer time |
Main types
See our data catalog for 24 types, from conversational data to agent trajectories.
Why licensing has grown
AI developers have collected most freely available web data and now face legal challenges and disclosure duties around how they sourced it. Licensing private and specialist data from its owners solves both problems, which is why deals such as Reddit's reported $60 million a year with Google (The Decoder) and publisher deals like Wiley's (The Bookseller) have followed.
Questions
Is my personal data used to train AI?
This site is about business data licensing. For your rights over personal data, see the privacy policy of each service you use and your state's privacy law.