On Wednesday, a federal judge rejected the allegations that 13 authors, including Sarah Silverman and Junot Diaz, had violated copyright by training AI models in their books.
Judge Vincent Chhabria concluded that Meta was engaged in “fair use” when he used a dataset of nearly 200,000 books, including the plaintiff's works, to train the Llama Language model. The decision follows a similar ruling issued Monday in a lawsuit against humanity over its linguistic model, Claude.
“We are grateful for the decision from the court today,” a Meta spokesperson said in a statement. “The open source AI model is to promote transformative innovation, productivity and creativity among individuals and businesses, and fair use of copyright material is an important legal framework for building this transformative technology.”
Chhabria rejected the plaintiff's claim that the company engaged in “unexempt copyright infringement” when building the model. The judge found that Llama was unable to make copies of more than 50 words, and that the AI model was a “transformation.”
He was more open to the argument that AI could use them to destroy the market for original works. He wrote that even if the output is different from the input, it would not be “fair use.”
“What copyright law concerns us most is to maintain the incentives that humans create, above all, to create.
The judge said: “Fair use” added: “It usually doesn't apply to copies that significantly reduce the copyright owner's ability to make money from their work (and therefore significantly reduces the incentives they create in the future).”
However, Chhabria found that the author in this case suffered from a decline in book sales, or that Llama was not likely to have such an effect.
“Meta introduced evidence that the copy was not causing any harm to the market,” he writes. “The plaintiffs did not present any empirical evidence to the contrary… All plaintiffs presented are speculation.”
The copyright holder has filed dozens of lawsuits against AI companies, claiming that training without a license is illegal.
Chhabria revealed that his ruling was limited to his previous facts, and in most cases the outcome could be different.
“In the grand scheme of things, the outcome of this ruling is limited,” he wrote.
Boys Schiller Flexner LLP, representing the plaintiff, said it disagrees with the outcome.
“The court ruled that AI companies will “supply copyrighted work to the model without obtaining permission from copyright holders or without allowing their payment,” a company spokesperson said. “However, despite the uncontroversial record of pirated copies of Meta's historically unprecedented copyrighted works, the court has ruled in favour of Meta, and we oppose that conclusion.”
The company's lawyers declined to comment on whether they would appeal.
The authors who brought the incident were Silverman, Diaz, Richard Cudley, Christopher Golden, Tanehisi Coates, Andrew Sean Grian, David Henry Hunter, Matthew Crumb, Laura Lipman, Rachel Louise Snyder, Jacqueline Woodson, Lysa Terkelest and Christopher Farnswinswinswien.
The lawsuit alleges that Meta used the “shadow library” to obtain millions of copies of the pirated book. According to the submission, Meta engineers used BitTorrent to download a large amount of data. This involves downloading data from multiple sources and re-uploading in some cases.
The lawsuit alleged that META violated the author's copyright in both the AI training process and the download and re-upload of the illegal library. The judge refused to request training for AI, but was not asked to govern the issue of torrents, so that remains unresolved.
