Judge sympathetic to famous authors fighting Meta over AI

SAN FRANCISCO (CN) — Meta said at a summary judgment hearing Thursday morning that it did not violate copyright law when it used books from 13 acclaimed authors, including comedian Sarah Silverman and award-winning author Ta-Nehisi Coates to train its flagship AI model, Llama. But a federal judge didn’t seem convinced.

The authors sued in 2023. Silverman is also a party to a similar lawsuit against ChatGPT developer OpenAI, which was moved from California federal court to New York federal court last week.

Kannon Shanmugam, arguing for Meta, defended the company’s conduct as falling under fair use. He claimed there is no precedent in a case like this, and that Meta does not heavily train Llama on one particular book to the point that the technology can reproduce expressions in the book.

“It would be impractical, and I don’t think there’s really any dispute about this, for a company like Meta to go to authors one by one for purposes of this training, in order to have an effective large language model,” Shanmugam said. “When it comes to transformational technological tools like generative AI, the court should be cautious about imposing liability. And it really is a matter, I think, ultimately, for Congress.”

Shanmugam argued that any economic harm caused by transformative use of the works does not violate copyright law because such uses, by definition, do not serve as substitutes for the original work.

U.S. District Judge Vince Chhabria, a Barack Obama appointee, conceded that that was a fair point but also said all copyright cases are context-sensitive and fact-dependent.

“You’re giving the product the ability to produce a million articles. You are dramatically changing — you might even say obliterating — the market for that person’s work, and you’re saying that you don’t even have to pay a license to that person for using their work to create the product that’s destroying the market for their work,” Chhabria said. “It’s not inconceivable to imagine that the market for the copyrighted works that were fed into the model could be eliminated entirely. How could that be fair use?”

To hammer home his point, Chhabria brought up a hypothetical: If an AI company were to copy news articles without permission en masse, it would seem very likely to lead to a creation of a product that could produce an infinite amount of content that would diminish demand for the copyrighted work, in this case the newspaper.

Shanmugam replied that the scenarios Chhabria was describing were speculative, and the plaintiffs have not pleaded any record for that theory.

“I don’t know if you need to wait to have the impact," Chhabria said. “Potential effects can count.”

Shanmugam said that expressions, not ideas, are protected by copyright law.

“How can you say that the expression is not being reproduced when the copying of the expression gives the machine the ability to engage in similar expression?” Chhabria asked.

Shanmugam responded that Llama does not retain the expressions it is trained on, and cannot reproduce them.

David Boies, counsel for the plaintiff authors, pushed back against Shanmugam’s assertion that Llama does not retain expressions it is trained on, telling Chhabria that Llama can complete a partial quotation from the authors’ work.

“The fact that it can regurgitate that snippet indicates that what they have done in that model is they have copied into the model the creativity and expression of the author.”

“I think I agree with you about that,” Chhabria said.

Boies also said that it’s reasonable to infer that Llama will do a better job of producing works in particular genres by being trained on copyrighted works, affecting the market for the copyrighted works.

“The record may not be as robust as either of us like, but I think there is a lot in the record about the importance to Meta of having long-length books” to train Llama, Boies said, adding “the burden is on the alleged infringer.”

Before heading back to his chambers at the conclusion of the three-hour hearing, Chhabria joked that he would issue a ruling today, before saying he would need “a lot longer” time to study the case.