Inside the court case that could change AI's future

Anthropic's decision to settle a class action lawsuit poses significant questions about how AI firms have dealt, and will deal, with copyrighted material, writes Chris Stokel-Walker

- Chris Stokel-Walker

The biggest copyright case of the AI era won't go to trial, we learnt last week, as Anthropic, the company behind the Claude chatbot, chose to settle what the firm said was “possibly” the largest class action lawsuit around copyright ever.

Three authors - Andrea Bartz, Charles Graeber and Kirk Wallace Johnson - filed a lawsuit against the firm in August 2024, alleging that Anthropic had used pirated versions of their books to train its AI model.

The authors also claimed that Anthropic made its business work by “largescale theft of copyrighted works” and alleged the firm “has taken multiple steps to hide the full extent of its copyright theft”. They presented evidence in their claim that the firm had downloaded 7 million books from the websites LibGen and PiLiMi, which host pirated versions of published works.

AI models work by ingesting data that they are given by their makers, which becomes their ground base of “knowledge” on which they can generate their responses. Much of this data is drawn from the internet, and some of it is alleged to come from pirated websites. Over the last year, Anthropic has repeatedly attempted to avoid the case coming to trial, but was on the ropes after the judge overseeing the case said in a ruling earlier this month that the firm had “refused to come clean” about which pirated works were used in training its systems. Judge William Alsup added in his early August ruling: “If Anthropic loses big, it will be because what it did wrong was also big.”

The case - and the decision to settle - poses significant questions about how AI firms have dealt with copyrighted material and how the future of large language models may develop. There are multiple other lawsuits currently underway, and all will have been watching this case very closely.