Ex-Meta Researcher Torrented Pirated Books, Deleted It 'Due to Litigation'

A former Meta AI researcher, Stephen Roller, testified that he downloaded the Books3 collection through BitTorrent while employed at the company and knowingly shared it. Books3 contains nearly 200,000 books taken from a private tracker and was used, in part, to train Meta’s Llama 1 and Llama 2 models. Authors and publishers suing Meta say Roller’s account suggests the company had multiple copies, including one obtained by torrent, making the acquisition method a key issue: a court has already ruled that training on books was fair use, but the torrenting question remains unresolved. Roller said Meta had Books3 copies when he left its research group, and deleted his own copy about a year ago because of the litigation.

Meta disputes that Roller downloaded the dataset for work, arguing he could have obtained it for personal use or for another employer; it says the copy was on his personal computer and was deleted because he no longer needed it. The plaintiffs are also seeking records about an internal “ML Books” folder and a possible employee book club, citing messages about sharing a pirated book. Meta says some requests came after discovery closed and is investigating the folder. The plaintiffs further allege Meta has not produced all server command histories that could identify torrents and works downloaded; Meta says it has supplied extensive logs and continues searching backups. The disputes add to lawsuits scheduled for trial beginning in 2027.