Copyright is a legal system that gives creators exclusive control over reproducing, distributing, and adapting original works. For AI training, it becomes a practical question: can collected material be used lawfully, especially when images or other online works belong to someone else? section 11.6 section 12.5
When selecting training data, distinguish sources by their copyright and licensing position:
- Public-domain books can be used because their copyright has expired; they also provide edited language and long-form narrative examples. section 11.6
- Web-crawled text and online images offer scale, but collected material can be noisy, copyrighted, or subject to licensing concerns. section 11.6 section 12.5
- Synthetic data and 3D renders can reduce many copyright and privacy issues, although creating them requires time and design expertise. section 12.5
Choosing an image-data source
Suppose a team wants to train an image model. Using an artist’s online work for a commercial model without a license creates a major copyright concern. Synthetic images made in a 3D environment provide a more controlled alternative, with precise labels for lighting, angle, and composition. section 12.5
