The plaintiffs filed the documents in the Kadrey v. Meta lawsuit, which is one of several AI copyright cases that are gradually making its way through the American legal system. Meta, the defendant, argues that it is “fair use” to train models on intellectual property, especially books. Authors Sarah Silverman and Ta-Nehisi Coates are among the plaintiffs who disagree.
In the earlier filings in the lawsuit, Meta CEO Mark Zuckerberg authorized Meta’s AI team to train on copyrighted material and stopped negotiations with book publishers for the licensing of AI training data. However, it is now more evident than ever how Meta may have come to use copyrighted material to train its models, including models in the company’s Llama family, according to the latest papers, the majority of which include excerpts of internal work communications between Meta employees.
Melanie Kambadur, a senior manager for Meta’s Llama model research team, and other Meta staff members talked about training models on works they were aware would be legally sensitive in one discussion.
“[M]y opinion would be (in the line of ‘ask forgiveness, not for permission’): we try to acquire the books and escalate it to execs so they make the call,” wrote Xavier Martinet, a Meta research engineer, in a chat dated February 2023, according to the filings. “[T]his is why they set up this gen ai org for [sic]: so we can be less risk averse.”
Instead of negotiating licensing agreements with individual book publishers, Martinet proposed purchasing e-books at retail prices to create a training set. Martinet doubled down after another employee raised the possibility of a legal battle for utilizing unapproved, copyrighted materials, claiming that “a gazillion” businesses were likely already using pirated books for training.
Martinet reported, according to the filings
“I mean, worst case: we found out it is finally ok, while a gazillion start up [sic] just pirated tons of books on bittorrent,”. “[M]y 2 cents again: trying to have deals with publishers directly takes a long time …”
Kambadur warned that although utilizing “publicly available data” for model training would necessitate approvals, Meta’s lawyers were being “less conservative” than they had previously been with such approvals in the same chat. Kambadur also mentioned that Meta was in discussions with document hosting platform Scribd “and others” for licenses.
“Yeah we definitely need to get licenses or approvals on publicly available data still,” Kambadur said, according to the filings. “[D]ifference now is we have more money, more lawyers, more bizdev help, ability to fast track/escalate for speed, and lawyers are being a bit less conservative on approvals.”
