01
Factual background & dispute
- Authors challenged Anthropic’s acquisition and copying of books for model training and its library.
- Purchased books were scanned one-for-one, the originals destroyed and the digital copies used as replacements without outside distribution. The pirated library was considered separately.
- Partial summary judgment was entered on 23 June 2025, settlement was finally approved on 20 July 2026, and implementation arrangements were addressed on 15 September.
02
Core issues & judicial focus
- Whether training large language models on copyrighted books constitutes fair use
- How the conversion of purchased print books into internal digital copies should be evaluated
- Whether downloading and retaining books from unauthorized repositories creates independent liability
03
Judicial finding & holding
- The 2025 partial judgment found fair use for training and replacement digitisation on those limited facts, without extending that defence to the permanent pirated-source library.
- On 20 July 2026, the court finally approved USD 1.5 billion for 482,460 listed works. Implementation continued; approval does not establish that distributions are complete.
- Settlement approval was not a trial finding of every alleged infringement. The release excluded past AI outputs and future conduct from 25 August 2025.
Dates & procedural record
- ECF231 distinguished training, replacement digitisation and the pirated library.
- ECF680 granted final settlement approval.
- ECF692 addressed website updates and further distribution arrangements.
04
Practical risk implications
01Fair use assessments for AI training must examine data acquisition methods alongside subsequent data retention practices.
02Enterprises should maintain comprehensive records of purchases, licenses, downloads, deduplication, deletions, and training runs.
03Settlement agreements resolve only defined past input activities; downstream output disputes and future conduct require independent assessment.