U.S. District Judge William Alsup of California issued the much-anticipated late Monday in a proposed class action by authors, commending Anthropic PBC for the creation of its large language model, Claude, while sharply criticizing the AI developer's conduct and setting the stage for a potential big-money trial.
"The technology at issue was among the most transformative many of us will see in our lifetimes," Judge Alsup said.
However, the judge disapproved of Anthropic acquiring more than seven million books from pirated sites in 2021 and 2022, as it had believed that to be the most cost-effective way to build a world-class large language model. Two years later, Anthropic bought millions of books and digitized them to train its AI, but Judge Alsup said that does not excuse the company's previous conduct, necessitating a trial for those actions.
"That Anthropic later bought a copy of a book it earlier stole off the internet will not absolve it of liability for the theft but it may affect the extent of statutory damages," Judge Alsup said.
And damages could be astronomical, with statutory damages potentially reaching $150,000 per pirated work, despite the transformative nature of Anthropic's technology.
"It gives the plaintiffs a path toward a very positive result here," said Jeremy Elman, a partner at Duane Morris LLP. "The pirated works issue does leave a large amount of damages that can compensate the authors and their attorneys."
Still, Anthropic applauded part of Judge Alsup's decision and quoted from it in a statement to Law360 on Tuesday.
"We are pleased that the court recognized that using 'works to train LLMs was transformative — spectacularly so.' Consistent with copyright's purpose in enabling creativity and fostering scientific progress, 'Anthropic's LLMs trained upon works not to race ahead and replicate or supplant them — but to turn a hard corner and create something different,'" Anthropic said in its statement.
Counsel for the plaintiffs did not immediately respond to a request for comment.
'A Mixed Bag'
The plaintiffs — authors Andrea Bartz, Charles Graeber and Kirk Wallace Johnson — each have at least two books that were included in the millions of pirated books, according to the court's ruling. A motion for class certification is pending because Anthropric moved for early summary judgment on fair use.
Judge Alsup's opinion is unlikely to be the last word on the legality of using copyrighted works for AI training, with dozens of other copyright lawsuits pending against tech companies by artists, authors, musicians and other content creators.
U.S. District Judge Vince Chhabria, also in California, is expected to issue his summary judgment ruling on fair use any day now in a lawsuit by a group of bestselling authors who sued Meta Platforms. The authors in that proposed class action, including Sarah Silverman, Ta-Nehisi Coates and Junot Díaz, have also accused Meta of downloading millions of pirated works.
The major decisions in these cases are also expected to go through years of appeals, likely until the U.S. Supreme Court weighs in on the propriety of using copyrighted works for AI training — a hotly debated question that has divided the legal industry.
In another closely watched case, a Delaware federal court found in February that tech startup ROSS Intelligence infringed copyrighted material from Thomson Reuters' Westlaw platform to build a competing product powered by AI. The Third Circuit accepted the tech company's interlocutory appeal last week. The case does not involve generative AI, however.
"For the copyright community, it's a mixed bag," Keith Kupferschmid, CEO of the Copyright Alliance, said about Judge Alsup's decision. "In many respects, it's a good decision because it makes clear that the AI companies cannot just go to these pirate sites and then use [the material] to train their AI. And since most of the companies are doing that today, this really has a very wide reach."
On the other hand, Kupferschmid said, he disagrees with Judge Alsup's conclusion that using copyrighted material to train AI without a license is fair use.
"The fact that the end product doesn't look the same as the works that are being copied, for an infringement analysis that's important, but for a fair use analysis, that's not the relevant question," Kupferschmid said. "The relevant question is, does the end result serve the same purpose as the works that are being copied? And the court kind of got that part wrong."
Judge Alsup said in his decision that the authors were not alleging that any output from Claude infringed their works, and he noted that Anthropic has guardrails in its platform to prevent infringing outputs. He also rejected the authors' argument that large language models could eventually displace them.
"[The] authors' 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," Judge Alsup said.
Possible Settlement?
James Grimmelmann, a professor at Cornell University's law and tech schools, said Judge Alsup's decision provides a framework for other judges who favor AI innovation but do not want to give developers "a blank check to do whatever they want."
"This provides a roadmap that they could follow, and it also provides a roadmap that AI companies could follow to continue training, but with some guardrails on the process," Grimmelmann said.
Judge Alsup's opinion could also encourage content creators and AI developers to enter into more licensing agreements, attorneys said.
The New York Times, which is suing Microsoft and OpenAI for using its content to train their AI systems, announced last month that it with Amazon to use the newspaper's material for AI products, following other news organizations that have done the same with other tech companies.
"If you're a platform developer, and you have licensed materials that you're using for training, if this decision becomes a law of the land, you're going to be able to operate at a much lower risk than if you're using materials that aren't licensed," said Matt Braunel, a partner at Thompson Coburn LLP.
Judge Alsup's decision may also provide an incentive to settle, especially given the potential monetary liability Anthropic could face after a trial. Before his decision, Judge Alsup announced last month that he could retire before year's end, and he encouraged the parties to begin settlement talks.
"I think it very much depends on what the price tag is and whether some of the plaintiffs are willing to settle at all," Grimmelmann said.
Irene Lee, a partner at Russ August & Kabat, said Judge Alsup's opinion gives AI companies a blueprint to follow if they want to stay within the bounds of fair use when developing their products.
"I think Judge Alsup was giving a bit of a nod to all the tech companies in his backyard in the Northern District of California, saying that this is undoubtedly transformative," she said, but adding that there are limits.
"If you are just wholesale downloading [books] from known pirating sites, that's going to be a problem," Lee said.
--Additional reporting by Elliot Weld and Bonnie Eslinger. Editing by Emily Kokoll and Lakshna Mehta.
For a reprint of this article, please contact reprints@law360.com.
