A landmark ruling that training generative AI on legitimately-acquired materials can be transformative fair use under copyright law provides a template for judges facing the same issue in dozens of pending cases.
Amazon-backed Anthropic PBC convinced District Judge William Alsup to grant summary judgment that merely training its “Claude” large language model constitutes fair use, approving the practice of buying a book, training AI with it, and destroying it. But Alsup allowed authors’ infringement claims related to copying works from pirated book websites to go to trial in his opinion in the US District Court for the Northern District of California.
The first judgment on whether generative AI training is fair use—in a class action brought in August by authors Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson—lays down a provisional roadmap for AI companies and judges hearing dozens of copyright cases critical to technology and content industries. Training can qualify as fair use, but Alsup’s ruling included qualifications, such as leaving open the possibility that acquiring pirated training materials may infringe and noting guardrails in Anthropic’s model preventing infringing outputs.
The decision “could be a bellwether,” IP law professor James Grimmelmann of Cornell University said. It “lays down some lines that could be appealing to other judges.”
“Everything in here could go a different way, and a thoughtful judge could write a persuasive opinion disagreeing,” Grimmelmann said. “But I found myself nodding along a lot of times.”
Dozens of AI copyright lawsuits across the country are in various stages, and complaints have been filed as recently as this month. Twelve cases against
Two other district judges—Judge Sidney H. Stein in New York, who’s overseeing the multi-district litigation, and Judge Vince Chhabria in California, who heard the recent Meta argument—are independently minded and could differ from Alsup, said IP attorney Aaron Moss at Mitchell Silberberg & Knupp LLP. Despite the mixed results, the June 23 ruling weighs in favor of AI companies after the content industry celebrated the US Copyright Office’s report rejecting inherent fair use, he said.
“It is a clear win for developers because it essentially provides a roadmap for how to do an LLM that is fair,” Moss said.
Anthropic said in a statement it “acquired books for one purpose only–building LLMs–and the court clearly held that use was fair.” The plaintiffs’ lawyers declined to comment.
Ignoring the Market
Fair use hinges on an evaluation of four factors: the nature of the use of the work, the nature of the original work, the amount used, and the effect on the original work’s market value. Alsup granted the AI industry a clean win on the first factor, which incorporates both whether the use was commercial and whether it was transformative of the original’s purpose and character.
“The technology at issue was among the most transformative many of us will see in our lifetimes,” Alsup said.
Content owners will chafe at Alsup’s treatment of training’s effect on the market, Roy Kaufman of Copyright Clearance Center said. Courts generally treat market impact as the most important factor, and the Copyright Office noted an active licensing market for training materials. But Alsup treated the factor dismissively, Kaufman argued, by saying “such a market for that use is not one the Copyright Act entitles authors to exploit.”
“It’s like a throwaway line about the most important thing according to Warhol,” he said, referring to the 2022 Supreme Court ruling in Andy Warhol Foundation v. Goldsmith. “It was a failure of the court to not really look at the licensing market in reaching a determination.”
While the Supreme Court has twice described the market effect factor as the most important, that doesn’t inherently mean any loosely related market activity counts, Moss said.
“If you don’t have substantially similar output, the fact that it may give you works in a similar style, or that broadly compete, is not a harm that is compensable,” he said. “Historically the way the courts have looked at it focuses on substitutes.”
Pirated Works
By separating the copying and retention of millions of pirated books from the purchase, copying, training and destruction of millions of legitimate copies, Alsup’s ruling may have “split the baby” on the copyright issue, Moss said. But he still called it a “clear win” for AI developers.
The court essentially said “so long as you have bought them legitimately and are not getting them from pirate sites,” it’s fair, Moss said.
The potential infringement by making millions of copies of pirated books, however, means Anthropic still could face substantial liability. Statutory damages range from $750 to $30,000 per work if that work was registered with the Copyright Office prior to infringement. OpenAI and Meta also stand accused of pilfering pirate libraries for training material.
Matthew Sag, an AI law professor at Emory Law School, agreed with Alsup’s overall ruling but echoed Moss’ point that it generated a set of rules instead of evaluating intrinsic harm.
“It seems weird to me that if the end result is Anthropic just goes on to eBay and Amazon, buys a bunch of books and then destroys them in the scanning process, that’s intrinsically better than just going to Libgen and downloading them,” Sag said.
Effectively, “you can do what you did—you just couldn’t go to a bad part of town to do it,” he added.
The opinion is relevant to all of the other AI cases and could be persuasive to judges attempting to reconcile “competing pressures” they’re facing, Grimmelmann said.
“It allows them to say ‘training is transformative, but that doesn’t mean that anything goes,’” he said.
The case is: Bartz et al v. Anthropic PBC, N.D. Cal., 3:24-cv-05417, Order issued 6/23/25
Annelise Levy in San Francisco also contributed to this story.
To contact the reporters on this story:
To contact the editor responsible for this story: