Anthropic and a group of US authors have settled a class action suit in a way that “will benefit all class members,” said Justin Nelson, the authors’ attorney, in a statement acquired by Reuters. The details of the settlement will be shown in “the coming weeks,” Nelson said.
This establishes the first settlement in several copyright cases brought against generative AI companies that train their models using text written by humans. Authors suing Anthropic would damage both parties, some testified.
Timeline of this AI copyright case against Anthropic
On July 17, US District Judge William Alsup certified a class action allowing three lead authors, Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson, to represent all US writers whose works were allegedly downloaded by Anthropic from pirate libraries such as LibGen and PiLiMi without licenses. This ruling meant millions of authors could pursue their copyright claims against the company in a single, unified case.
When the class action suit was certified, groups representing the AI and publishing sectors said allowing up to seven million authors to sue Anthropic would work against both their industries. They warned it could bankrupt AI firms by triggering massive litigation costs and destabilise publishing by undermining existing licensing agreements.
Anthropic swiftly petitioned to appeal the class certification, arguing that Alsup brushed aside the company’s argument that its use of the books was legal “fair use.” The company also said that figuring out who owns each work and whether it was actually copied would require complex case-by-case decisions, and that the looming threat of “hundreds of billions of dollars in potential damages” could force a settlement before it could mount a defence.
“Several other pending putative class actions likewise challenge companies’ use of allegedly copyrighted books to train large language models,” Anthropic wrote in its petition, referring to various ongoing disputes between copyright holders and OpenAI and Meta. “These cases raise important fair-use questions that will establish the ground rules for the industry moving forward.”
Nevertheless, judges ruled that Anthropic’s and Meta’s use of copyrighted books to train their AI models qualified as “highly transformative” and thus protected by fair use. At the time, they decided Anthropic must stand trial on December 1, 2025.
Groups representing tech companies and artists were in agreement, for once
The Consumer Technology Association and the Computer and Communications Industry Association, which represent the US tech and communications industries, were both publicly in support of Anthropic. They agreed that the broad-brush certification prevented AI companies from fully investigating and contesting each claim, and that the prospect of billions in statutory damages without proof of actual harm exerted overwhelming pressure on Anthropic to settle.
“Class certification for alleged copyright infringement — involving millions of putative class members — will chill investment in AI companies as investors will avoid the threat of such catastrophic liability,” the trade bodies wrote in an August 7 court filing. “The United States currently may be the global leader in AI development, but that could change if litigation stymies investment by imposing excessive damages on AI companies.”
While industry groups representing authors, including Authors Alliance, the Electronic Frontier Foundation, American Library Association, Association of Research Libraries, and Public Knowledge, ultimately agreed that the class certification was misguided, they opposed it for different reasons.
In their amicus brief, they said the judge assumed the authors and publishers would be able to “work out the best way to recover” damages should they win the case. However, because publishers and authors can disagree on how works should be used, doing so would be extremely complicated and could result in “hundreds of mini-trials” to determine ownership.
They also objected to the assumption that the three named plaintiffs spoke for all rightsholders, as views on how AI models should be trained vary widely. They argued the case should be resolved in a way that ensured a win for the rightsholders does not create further conflict.
This article was updated by Megan Crouse to reflect the settlement decision.
US President Donald Trump dismissed calls to compensate writers whose works are used to train AI models, calling the idea “impractical.”