Anthropic has settled a copyright lawsuit that risked exposing the company to billions of dollars’ worth of damages.
The case, which concerns Anthropic’s use of millions of pirated books to train its large language model, Claude, was due to go to trial on Dec. 1. However, in a Tuesday court filing, Anthropic and the authors jointly informed a San Francisco federal court that they had reached a class-wide settlement. They requested that discovery be paused and existing deadlines vacated while the agreement is finalized.
The amount of the settlement was not immediately disclosed, but legal experts not involved in the case said the figure could easily reach into the hundreds of millions. It’s also still unclear how the settlement will be distributed among various copyright holders, which could include large publishing houses as well as individual authors.
Justin Nelson, partner at Susman Godfrey and lawyer for the authors, told Fortune the “settlement will benefit all class members” and they looked “forward to announcing details of the settlement in the coming weeks.”
The case was the first certified class action against an AI company over the use of copyrighted materials, and the quick settlement, which came just one month after the judge ruled the case could proceed to trial as a class action, is a win for the authors, according to legal experts.
“The authors’ side will be reasonably happy with this result, because they’ve essentially forced Anthropic into making a settlement,” Luke McDonagh, an associate professor of law at LSE, told Fortune. “This might be the first domino to fall.”
While the settlement doesn’t set a legal precedent, it could serve to legitimize the authors’ claims and may set a business precedent for similar cases, he said.
Crucially, the case didn’t rest on the use of copyrighted material for AI training generally—and, in fact, the trial judge ruled that training AI models on legally acquired works would likely qualify as “fair use”—but on how AI companies acquired the works they used for training.
The plaintiffs, who include authors Andrea Bartz, Charles Graeber and Kirk Wallace Johnson, alleged that millions of these works were obtained from piracy websites and shadow libraries in direct and knowing violation of copyright law.
The judge presiding over the case, William Alsup, had suggested in a ruling made in July that even if the AI training use might be considered fair use, this initial acquisition of works was illegitimate and would need compensation. If the court found that Anthropic willfully violated copyright law, statutory damages could have reached $750 to $150,000 per work, a scenario that Santa Clara law professor Ed Lee said could have “at least the potential for business-ending liability.”
“Even though they will be paying out a considerable amount of money, from Anthropic’s point of view, they’ll be seeing it as simply dealing with a legal problem, that they weren’t careful enough about the way they used copyright works,” McDonagh said.
He noted that Anthropic had, in addition to allegedly using libraries of pirated books, also purchased some copies of books that it later digitized and fed to its AI models. Alsup had ruled that the use of these books for AI training was “fair use,” which didn’t require additional licensing from the copyright holders. “If they bought the 7 million books, then they could have actually used them to train their programs, and that would have been a fair use. The judge is objecting to the fact that they didn’t pay for the books in the first place,” McDonagh said.
But Anthropic wasn’t the only company to acquire large numbers of books by using pirated book libraries. Meta has also been accused of using the same shadow library, Library Genesis (LibGen), to download books to train its AI models.
If the settlement is approved and the risk to Anthropic of further legal action is sufficiently mitigated, it could create a road map for how other companies deal with similar cases. AI companies have not previously acknowledged they have any liability when it comes to the use of copyrighted materials, so the settlement could set a precedent for how much copyright owners are compensated for works used for AI training.
“Some companies are already engaging in licensing content from publishers. So this may push us forward towards a situation in which the AI companies reach settlements that effectively retrospectively pay license fees for what they’ve already used,” McDonagh said. “Perhaps [it’s] the beginning of a licensing future for the AI industry, where authors can be paid for the use of their copyrighted works in the training of AI.”
Representatives for Anthropic did not immediately respond to a request for comment from Fortune.
