Meta secretly trained its AI on notorious hacking databases, new unredacted court documents reveal

“Meta treated the so-called ‘public availability’ of shadow datasets as a get-out-of-jail-free card, even though internal Meta memos show all relevant decision-makers, including Meta CEO Mark, were involved.” Zuckerberg knew that LibGen had “a data set that we knew to be pirated,” the plaintiffs allege in the motion. (The motion originally filed in late 2024 is a request to file a third amended complaint.)

In addition to the plaintiffs’ information, another filing was not redacted in response to Chhabria’s order — Meta’s opposition to a motion to file an amended complaint. It claims the authors’ attempts to add additional claims to the case are an “eleventh-hour gambit based on a false and inflammatory premise,” and denies that Meta expects important information to emerge in discovery. Instead, Meta claims it first disclosed to the plaintiffs that it was using the LibGen dataset in July 2024. (Since much of the discovery material remains classified, it’s difficult for WIRED to confirm this claim.)

Meta’s argument is based on the claim that the plaintiffs should not have been granted additional time to file the third amended claim when they already knew about LibGen’s use and had ample time to do so before discovery ends in December 2024. “The plaintiffs knew about the loading of Methane. and using LibGen and other alleged “shadow libraries” since at least mid-July 2024,” attorneys for the tech giant to argue.

In November 2023, Chhabria granted Meta’s motion to dismiss some of the lawsuit’s claims, including a claim that Meta used the authors’ work to train artificial intelligence. Digital Millennium Copyright ActIn 1998, a US law was passed to stop people from selling or reproducing copyrighted works on the Internet. Then the judge agreed According to Meta’s position, the plaintiffs did not provide sufficient evidence to prove that the company deleted information known as “copyright management information,” such as the name of the author and the title of the work.

The unredacted filings argue that the plaintiffs should be allowed to amend their complaint, arguing that the information Meta discovered is evidence that the DMCA claim has merit. They also say that the discovery process has uncovered grounds for adding new claims. “Meta, through its corporate representative who testified on November 20, 2024, has now admitted under oath to uploading (aka ‘seeding’) pirated files containing Plaintiffs’ works to ‘torrent’ sites,” it said. (Seeding is when torrent files are shared with other peers after they have finished downloading.)

“This torrenting activity made Meta itself a distributor of the same pirated copyrighted material that it downloaded for use in commercially available AI models,” one of the new unredacted documents claims, which Meta, in other words, did not do. not only used copyrighted material without permission, but also distributed it.

Founded in Russia in 2008, LibGen, an archive of online books, is one of the world’s largest and most controversial “shadow libraries.” In 2015, a New York judge ordered the initial ban against the site, a measure that was theoretically intended to temporarily shut down the archive, but its anonymous administrators simply changed its domain. A different New York judge in September 2024 ordered LibGen will pay rights holders $30 million for copyright infringement, despite not knowing who runs the piracy hub.

Meta’s discovery challenges for this work are far from over. In the same vein, Chhabria warned the tech giant against any excessive redaction requests in the future: “If Meta files an unreasonably broad unsealing request again, all material will simply be sealed,” he said.