Legality of Training Large Language Models on Copyrighted Materials (US/EU)

Clarify whether the use of copyrighted materials to train large language models without explicit permission constitutes copyright infringement under European Union and United States law.

Background

The authors discuss the widespread practice of training LLMs on internet-scale corpora that include copyrighted works. They highlight uncertainty in the legality of such training in the EU and US, contrasting it with a more permissive stance in Japanese law.

This unresolved legal question has significant implications for both AI developers and the scientific community, as it affects the permissibility of using scholarly literature and other copyrighted content in training datasets.