A US judge has ruled that using copyrighted books to train AI models can qualify as a transformative use under copyright law, but also allowed a trial to proceed regarding the alleged use of pirated copies.

The case involves AI company Anthropic, accused by authors including Andrea Bartz and Charles Graeber of unlawfully using their books to train its Claude AI system. Judge William Alsup found that the AI’s training process, which involves creating new works rather than copying original texts, is protected as transformative use. He wrote, “Like any reader aspiring to be a writer, Anthropic’s models train on works not to replicate but to create something new.”

However, Alsup denied Anthropic’s request to dismiss the lawsuit, stating the company must face trial over its storage and use of more than seven million unauthorized copies of books in a “central library.” Anthropic, supported by investors like Amazon and Alphabet, could face significant damages if found liable.

The ruling is one of the first in the ongoing global debate over how AI systems may legally learn from copyrighted content. It distinguishes between fair use in training AI and outright piracy, noting no evidence was shown that the AI produced infringing copies.

Similar lawsuits are emerging against AI firms across various creative sectors, including a recent case against image generator Midjourney. Anthropic said it welcomes the recognition of transformative use but disagrees with the trial decision and is considering its next steps.

‍