Judge Rejects Authors’ Claim That Meta AI Training Violated Copyrights

A federal judge on Wednesday rejected a claim brought by 13 authors, including Sarah Silverman and Junot Díaz, that Meta violated their copyrights by training its AI model on their books.

Judge Vincent Chhabria concluded that Meta had engaged in “fair use” when it used a dataset of nearly 200,000 books — including the plaintiffs’ works — to train its Llama language model. The decision follows a similar ruling issued on Monday in a case against Anthropic over its language model, Claude.

“We appreciate today’s decision from the Court,” a Meta spokesperson said in a statement. “Open-source AI models are powering transformative innovations, productivity and creativity for individuals and companies, and fair use of copyright material is a vital legal framework for building this transformative technology.”

Chhabria rejected the plaintiffs’ claim that the company engaged in “unmitigated piracy” when it built the model. The judge found that Llama cannot create copies of more than 50 words, and that the AI model is thus “transformative.”

He was more open to the argument that AI could destroy the market for original works by using them to create millions of cheap knockoffs. That likely would not be “fair use,” even if the outputs were different from the inputs, he wrote.

“What copyright law cares about, above all else, is preserving the incentive for human beings to create
artistic and scientific works,” the judge wrote. Fair use, he added, “typically doesn’t apply to copying that will significantly diminish the ability of copyright holders to make money from their works (thus significantly diminishing the incentive to create in the future).”

But Chhabria found that the authors in this case had not shown that they suffered any decline in book sales, or that Llama is likely to have such an effect.

“Meta introduced evidence that its copying hasn’t caused market harm,” he wrote. “The plaintiffs presented no empirical evidence to the contrary… All the plaintiffs presented is speculation.”

Copyright holders have brought dozens of lawsuits against AI companies, alleging that training on copyrighted work without a license is illegal.

Chhabria made clear that his ruling is confined to the facts in front of him and that the outcome could be different in most cases.

“In the grand scheme of things, the consequences of this ruling are limited,” he wrote.

Boies Schiller Flexner LLP, which represented the plaintiffs, said it disagrees with the outcome.

“The court ruled that AI companies that ‘feed copyright-protected works into their models without getting permission from the copyright holders or paying for them’ are generally violating the law,” said a spokesperson for the firm. “Yet, despite the undisputed record of Meta’s historically unprecedented pirating of copyrighted works, the court ruled in Meta’s favor. We respectfully disagree with that conclusion.”

The firm’s attorneys declined to comment on whether they will appeal.

The authors who brought the case are Silverman, Díaz, Richard Kadrey, Christopher Golden, Ta-Nehisi Coates, Andrew Sean Greer, David Henry Hwang, Matthew Klam, Laura Lippman, Rachel Louise Snyder, Jacqueline Woodson, Lysa TerKeurst and Christopher Farnsworth.

The lawsuit alleges that Meta used “shadow libraries” to obtain millions of copies of pirated books. According to the filings, Meta’s engineers used BitTorrent to download the large volume of data, which involves downloading data from multiple sources and, in some cases, reuploading it.

The suit alleged that Meta violated the authors’ copyrights both in the AI training process and in the process of downloading and reuploading the illicit libraries. While the judge rejected the AI training claim, he was not asked to rule on the torrenting issue, so that remains unresolved.