Model Behavior
The successful motion to dismiss did not challenge the key claim that Meta’s language model was improperly trained on copyrighted material
A federal judge has tossed several claims in a copyright infringement case brought by authors — including Sarah Silverman — against Meta over the way it trains its artificial intelligence language models.
Silverman and her co-plaintiffs had accused the tech giant of improperly using their copyrighted books to train its LLaMA language models. The Meta suit was filed alongside a similar copyright claim against OpenAI over the way it trained ChatGPT; that case also remains ongoing.
Judge Vince Chhabria’s ruling wasn’t exactly a surprise, as he said he would grant Meta’s motion to dismiss several of the lawsuit’s claims at a hearing earlier this month (per Reuters). And while Meta’s motion to get these claims tossed was successful, the company notably did not challenge the core argument brought by the authors: that Meta violated copyright law by using their books to train LLaMA.
“I understand your core theory,” Chhabria told the plaintiffs’ attorneys earlier this month. “Your remaining theories of liability I don’t understand even a little bit.”
The claims Meta successfully challenged were rather broad arguments about the scope of Meta’s alleged liability. To start, the plaintiffs alleged that the LLaMA models were inherently infringing because they “cannot function without the expressive information extracted” from the authors’ books. Chhabria, in his ruling, called this argument “nonsensical,” adding, “There is no way to understand the LLaMA models themselves as a recasting or adaptation of any of the plaintiffs’ books.”
Additionally, Chhabria threw out the plaintiffs’ argument that every LLaMA output was “an infringing derivative” work and “constitutes an act of vicarious copyright infringement”; that LLaMA was in violation of the Digital Millennium Copyright Act; and that LLaMA “unjustly enriched Meta” and “breached a duty of care ‘to act in a reasonable manner towards others’ by copying the plaintiffs’ books to train LLaMA.”
Despite the dismissals, the authors’ case is not completely dead, especially with the core argument still unchallenged. They will have 21 days to file an amended complaint.
Reps for the plaintiffs, as well as Meta, did not immediately return Rolling Stone’s requests for comment.