For years, Meta employees have internally discussed using copyrighted works obtained through legally questionable means to train the company’s AI models, according to court documents unsealed on Thursday.
The documents were submitted by plaintiffs in the case Kadrey v. Meta, one of many AI copyright disputes slowly winding through the U.S. court system. The defendant, Meta, claims that training models on IP-protected works, particularly books, is “fair use.” The plaintiffs, who include authors Sarah Silverman and Ta-Nehisi Coates, disagree.
Previous materials submitted in the suit alleged that Meta CEO Mark Zuckerberg gave Meta’s AI team the OK to train on copyrighted content and that Meta halted AI training data licensing talks with book publishers. But the new filings, most of which show portions of internal work chats between Meta staffers, paint the clearest picture yet of how Meta may have come to use copyrighted data to train its models, including models in the company’s Llama family.
In one chat, Meta employees, including Melanie Kambadur, a senior manager for Meta’s Llama model research team, discussed training models on works they knew may be legally fraught.
“[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.”
Martinet floated the idea of buying e-books at retail prices to build a training set rather than cutting licensing deals with individual book publishers. After another staffer pointed out that using unauthorized, copyrighted materials might be grounds for a legal challenge, Martinet doubled down, arguing that “a gazillion” startups were probably already using pirated books for training.
“I mean, worst case: we found out it is finally ok, while a gazillion start up [sic] just pirated tons of books on bittorrent,” Martinet wrote, according to the filings. “[M]y 2 cents again: trying to have deals with publishers directly takes a long time …”
In the same chat, Kambadur, who noted Meta was in talks with document hosting platform S …