Skip to content
DIGITAL RIGHTS & COPYRIGHT LAW

Meta Faces Rising Legal Pressure Over Torrented Books and AI Training Data

As legal battles continue to unfold in a California federal court, Meta is facing mounting scrutiny over its acquisition and use of pirated literary works to train its expansive Llama artificial intelligence models. The multi-front litigation brings together a prominent group of plaintiffs, including bestselling authors like Sarah Silverman and Ta-Nehisi Coates, alongside major media companies and publishers such as Entrepreneur Media and Chicken Soup for the Soul. Overseen collectively by Judge Vince Chhabria, the lawsuits probe the deep mechanics of how major technology enterprises curate massive datasets for machine learning, with a sharp focus on whether copyright infringement occurred during the foundational stages of AI development.

The controversy centers heavily around Books3, a notorious collection of nearly 200,000 books extracted directly from the library of Bibliotik, a private torrent tracker. Back in August, Meta maintained in discussions with Entrepreneur Media that it did not believe it had obtained the Books3 dataset through BitTorrent protocols. However, recent developments in the discovery phase have injected new friction into the proceedings, as rightsholders point to deposition testimony from a former Meta researcher that suggests a vastly different reality regarding how the data was brought into the tech giant’s ecosystem.

Books3 Torrenting Testimony Paints a New Picture

According to a joint status report filed in the U.S. District Court for the Northern District of California, Stephen Roller—a researcher who worked at Meta from 2017 until 2023—admitted during a deposition that he torrented Books3 and knowingly distributed it during his tenure at the company. While Roller could not recall the exact date he acquired the dataset, he acknowledged that Books3 was created in 2020 and confirmed that he was not employed by any other organization while working at Meta.

Ex-Meta Researcher Torrented Pirated Books Dataset, Deleted It ‘Due to All This Litigation’

This testimony strikes at the heart of the ongoing legal arguments. Last year, the court ruled that training Llama models on books constituted fair use under copyright law, effectively narrowing the scope of the litigation. Consequently, Meta’s alleged BitTorrent downloads have emerged as the primary open issue remaining for the authors and publishers. Earlier this month, a judge ordered Meta to explicitly name the employees who authorized the downloading of materials from shadow libraries like Anna’s Archive and Library Genesis.

Meta has previously acknowledged its use of Books3. In early 2024, the company admitted that it utilized portions of the dataset to train its early-generation Llama 1 and Llama 2 models, though that initial admission omitted specific details on how the company acquired the files. Plaintiffs now note that Roller’s testimony indicates Meta possessed multiple copies of Books3 by the time he departed from Meta’s Fundamental AI Research (FAIR) group. Furthermore, Roller reportedly identified two other research scientists who likely acquired their own individual copies.

Interestingly, Roller no longer retains his copy. Plaintiffs state that he deleted the files approximately a year ago, likely after receiving a second subpoena. According to the court filings, the deletion was made "due to all this litigation" because it felt "contentious to have copies of The Pile," a massive text corpus that includes Books3. The revelation that Meta may have maintained multiple copies of the torrented dataset came as a surprise to the plaintiffs, who now intend to press the company for a precise accounting of how many copies were acquired and by what means, threatening to seek judicial intervention if necessary.

Meta Counters With Personal Use Defense

Meta has pushed back firmly against the plaintiffs’ interpretation of the deposition, arguing in court documents that the status report mischaracterizes Roller’s statements. The company emphasizes that this marks the first time the rightsholders have raised these specific concerns and contends that Roller’s testimony does not establish corporate liability for the downloads.

Ex-Meta Researcher Torrented Pirated Books Dataset, Deleted It ‘Due to All This Litigation’

According to Meta, Roller testified that he could not pinpoint when he downloaded a copy of The Pile via torrent, and he explicitly did not state that the download was performed on behalf of Meta or under corporate direction. Instead, Meta asserts that the acquisition could have been executed for personal purposes or on behalf of another employer. Furthermore, the company notes that Roller testified he had no knowledge of whether multiple copies were downloaded or whether the two other researchers he referenced ever actually acquired Books3.

Addressing the deletion of the files, Meta clarifies that the copy in question was stored on Roller’s personal computer. He removed it more than two years after leaving the company simply because he "didn’t need it" and preferred "not to have it" in his possession given the surrounding circumstances.

A Pirated Book Club and Internal Folders

The evidentiary disputes extend beyond AI training datasets into the internal office culture at Meta. The joint status report highlights potential instances where pirated books may have been utilized within the company for purposes entirely unrelated to machine learning research.

Plaintiffs are currently seeking access to the contents of an internal Google Drive folder titled "ML Books," alongside records from a Meta book club that allegedly relied on texts sourced from shadow libraries. Discovery documents produced by Meta show an exchange where one employee shared a "pirated book" with a colleague, accompanying the file with a self-aware remark about the irony of the situation given their line of work. When asked directly if the book was pirated, the employee reportedly responded by sharing a link to the "ML Books" folder. The recipient of the link expressed concern over the potential professional fallout, remarking to another coworker that she was "tryna get me fired" and suggesting that the incriminating messages be deleted.

Ex-Meta Researcher Torrented Pirated Books Dataset, Deleted It ‘Due to All This Litigation’

For the plaintiffs, these internal communications carry significant legal weight. If Meta employees routinely utilized shadow library books for internal activities such as a corporate book club, the rightsholders argue it could undermine the company’s broader fair use defenses by demonstrating a systemic disregard for copyright protocols.

Meta, however, contends that these new document requests are procedurally untimely. The company notes that the plaintiffs first formally requested the linked files on October 9, well after the formal discovery period had closed. Regarding the "ML Books" folder itself, Meta states that it is currently investigating the matter and intends to meet and confer with the plaintiffs once more information is gathered.

Missing Server Logs and Expanding Litigation

Another point of contention centers on missing digital records. Plaintiffs assert that Meta has still failed to produce all command history files associated with its torrenting activities, despite prior court orders requiring their handover. According to the rightsholders, these specific logs are critical because they reveal how Meta configured its BitTorrent software, which torrent files were targeted, and consequently, which specific copyrighted works were downloaded.

Plaintiffs maintain that Meta has neglected to hand over command history files from any of the servers utilized during its active torrenting periods. They are now asking the court to compel Meta to clarify whether any of these command history files were systematically deleted and, if so, the precise timeline of their removal.

Ex-Meta Researcher Torrented Pirated Books Dataset, Deleted It ‘Due to All This Litigation’

Meta rejects the premise that evidence has been improperly withheld or destroyed. The tech firm counters that it has already produced substantial quantities of logs originating from those servers, adding that it continues to comb through data backups and provide new files on a rolling basis. From Meta’s perspective, there is no factual basis to suggest that command history files were intentionally deleted to evade discovery obligations.

The legal horizon for Meta continues to expand as more rightsholders queue up to challenge the company’s data acquisition practices. Entrepreneur Media’s lawsuit is currently scheduled to proceed to trial in the spring of 2027, with additional publisher lawsuits positioned directly behind it in the event of a pretrial settlement. Meanwhile, the roster of plaintiffs continues to grow; earlier this month, major traditional publishers including Elsevier, Hachette, and Macmillan—along with a fresh wave of individual authors—officially joined the legal queue, signaling that scrutiny over AI training data will remain a central battleground for the publishing and technology industries for years to come.