US courts have made rulings on AI training data cases involving Meta and Anthropic, with differing conclusions on using pirated materials. A US district judge, Vince Chhabria, ruled that Meta's use of copyrighted books to train its LLaMA models was protected under fair use. However, the court refused to dismiss claims related to Meta's alleged unlawful distribution of copyrighted books, which will proceed to trial.
In a separate case, Judge William Alsup held that Anthropic's training of its Claude model on copyrighted works was fair use, but building a research library from pirated material was not. The court divided the company's conduct into three phases: training on copyrighted works, scanning and storing lawfully purchased books, and mass downloading and permanent retention of pirated works.
Both courts analyzed the fair use factors, including purpose and character of the use, nature of the copyrighted works, amount of the work taken, and effect on the market. While the courts found that training AI models was highly transformative and favored AI developers, they also emphasized that using pirated books may not be considered fair use.
The rulings suggest that AI training on copyrighted material can be fair use if the outputs are non-infringing and the use is transformative. However, the cases will proceed to trial on specific issues, including piracy-related damages and distribution claims.
