It’s fair.
That’s what two California judges recently ruled about Meta and Anthropic’s use of certain copyrighted books in training their respective LLMs. The companies each faced a lawsuit from a different set of prominent authors.
These are the first major rulings to weigh in on the question of whether it should be legal for developers to train GenAI models on copyrighted materials.
Tech companies argue that falls under the fair use doctrine, a legal exception that allows for nonpermissioned use of protected works in certain contexts, like news reporting, research, or parody. But copyright holders like authors, news publishers, and other media creators tend to argue that AI models harm the market for their work or aren’t sufficiently transformative in output, both of which would mean fair use isn’t applicable.
These arguments are playing out over a slew of high-profile lawsuits currently in the court system. Lawyers have told us that it will likely take a Supreme Court decision to settle this for good—and that could take years.
But for now, the Anthropic and Meta rulings offer a window into how judges are thinking about these questions. And so far, they’ve sided with AI companies, albeit with major caveats.
While a judge ruled earlier this year in favor of Thomson Reuters against legal AI startup Ross Intelligence in an AI copyright decision, the case notably did not involve generative AI. And lawyers told us there were some other factors that made it less broadly applicable.
Bartz v. Anthropic: Authors Andrea Bartz, Kirk Wallace Johnson, and Charles Graeber sued Anthropic for the use of their books—both pirated and legally purchased—in the company’s “central library,” a hub of millions of works that Anthropic pulled from to train various LLMs.
Senior US District Judge William Alsup ruled that training Claude on these books was “exceedingly transformative,” thus qualifying as fair use.
Alsup also dismissed the idea of economic harm caused by Claude flooding the market. “[The] complaint is no different than it would be if they complained that training schoolchildren to write well would result in an explosion of competing works,” he wrote in the ruling.
Alsup did, however, grant that Anthropic’s downloading of millions of pirated books for its central library was “theft” and allowed for a trial on that matter specifically.
Kadrey v. Meta: Advocates for copyright protections might be heartened to read the first three pages of District Court Judge Vince Chhabria’s ruling on the legal battle between a group of high-profile authors and Meta.
In “most cases,” non-permissioned training on copyrighted works is likely illegal, he wrote. Copyright law is about preserving the incentive for human beings to create works of value, and a flood of AI-generated media will “dramatically undermine the incentive for human beings to create things the old-fashioned way,” Chhabria wrote.
Unfortunately for these specific authors, including Sarah Silverman and Junot Díaz, theirs are not “most cases.” The judge said the two main arguments their lawyers presented—that Meta’s Llama can reproduce snippets of their books, and that training without permission harms their ability to license their work to AI companies—are “clear losers.” A more persuasive argument would’ve centered on market dilution, but the plaintiffs “barely give this issue lip service.”
Chhabria sought to narrow the ruling to the details of this specific case and these specific authors.
“This ruling does not stand for the proposition that Meta’s use of copyrighted materials to train its language models is lawful,” the judge said. “It stands only for the proposition that these plaintiffs made the wrong arguments and failed to develop a record in support of the right one.”