Skip to content
TORRENT & P2P MEDIA NEWS

Meta Faces New Scrutiny Over Torrented ‘Books3’ Dataset and Internal Shadow Libraries in AI Copyright Lawsuits

Meta is facing mounting pressure in a consolidated series of copyright lawsuits in California federal court as rightsholders zero in on the tech giant’s alleged acquisition and use of pirated books to train its Llama artificial intelligence models. The legal proceedings, which are being managed collectively by U.S. District Judge Vince Chhabria in the U.S. District Court for the Northern District of California, feature an expanding roster of high-profile plaintiffs. Among them are well-known authors such as Sarah Silverman and Ta-Nehisi Coates, alongside major publishing entities including Entrepreneur Media and Chicken Soup for the Soul.

The central controversy revolves around the origin of the "Books3" dataset, a massive collection containing nearly 200,000 digitized books originally harvested from the library of Bibliotik, a private BitTorrent tracker. Back in August, representatives for Meta informed Entrepreneur that the company did not believe it had acquired the controversial dataset through BitTorrent protocols. However, recent developments in the discovery phase of the litigation have brought forth deposition testimony from a former Meta researcher that rightsholders argue tells a radically different story.

Books3 Torrenting Testimony

According to a joint status report recently filed by the plaintiffs and Meta, Stephen Roller—a researcher who worked at Meta from 2017 to 2023—admitted during a deposition that he personally torrented the Books3 dataset and knowingly distributed it during his tenure at the company. While Roller could not recall the exact date he downloaded the repository, the plaintiffs point out that he acknowledged Books3 was created in 2020 and that he maintained no other corporate employment while working at Meta.

This testimony strikes at the heart of the ongoing legal battle. Last year, the presiding court ruled that training the Llama models on copyrighted books constituted fair use under copyright law. This crucial legal determination left Meta’s alleged BitTorrent downloads as the primary remaining battleground in the authors’ case. Compounding these issues earlier this month, a federal judge ordered Meta to identify the specific employees who authorized the downloading of material from shadow libraries like Anna’s Archive and Library Genesis (LibGen).

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

Meta has previously acknowledged utilizing portions of Books3. In early 2024, the company conceded that it used segments of the dataset to train its early Llama 1 and Llama 2 iterations, though that initial admission omitted details regarding how the company originally secured the data.

The recent deposition adds further complexity. The plaintiffs state that Roller testified that Meta retained copies of Books3 by the time he departed from the company’s Fundamental AI Research (FAIR) group. Furthermore, Roller identified two other research scientists who he suggested likely acquired their own independent copies. Roller himself no longer retains the data, explaining that he deleted his personal copy approximately a year ago—likely after receiving a second subpoena—because of the contentious nature of the litigation surrounding repositories like The Pile, which incorporates Books3.

For the plaintiffs, the revelation that Meta may have housed multiple internal copies of the dataset, at least one of which was sourced via torrenting, came as entirely new information. They intend to press Meta for a complete accounting of how many copies of Books3 were acquired and through what channels, signaling that they are prepared to ask the court for intervention if necessary.

Meta: Could Have Been for Himself

Meta strongly disputes the plaintiffs’ interpretation of the deposition, maintaining that the joint status report represents the first time these specific concerns have been formally raised and arguing that Roller’s testimony fails to substantiate the claims.

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

In its written response, Meta emphasized that Roller could not pinpoint when he downloaded The Pile via torrent. Crucially, the company contends that Roller did not testify that the download was performed on behalf of Meta, under Meta’s directives, or for corporate utilization, noting that the acquisition could have been executed for personal purposes or on behalf of another employer.

Furthermore, Meta pointed out that Roller testified he was uncertain whether multiple copies were ever downloaded or whether the two other researchers he referenced actually acquired Books3. Roller departed from Meta in January 2023 and has since held positions at three other companies operating in the artificial intelligence sector.

Regarding the deletion of the file, Meta clarified that the copy was stored entirely on Roller’s personal computer. He reportedly removed the data more than two years after leaving the company simply because he no longer required it and preferred not to retain it.

A (Pirated) Book Club

The dispute over Books3 is not the only copyright-related issue brought to light in the recent status report. Plaintiffs have also raised concerns that pirated books may have been utilized within Meta for purposes entirely separate from core artificial intelligence training regimens.

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

The legal team representing the rightsholders is currently pushing for Meta to produce the contents of an internal Google Drive folder labeled "ML Books," alongside records concerning an internal Meta book club that allegedly relied on texts sourced from various shadow libraries.

Discovery documents previously produced by Meta revealed an internal exchange where an employee shared what was described as a "pirated book" with a colleague. Recognizing the irony given the company’s ongoing legal challenges, the employee remarked, "i know for vellum sake this is prob very ironic." When questioned about the provenance of the book, the sender provided a direct link to the "ML Books" repository. The recipient subsequently expressed alarm to another colleague, writing, "bro she tryna get me fired" and adding that he "may need her to delete those messages."

The plaintiffs argue that this internal chatter carries significant legal weight. If Meta employees routinely utilized shadow library materials for casual activities like a corporate book club, the plaintiffs believe it could undermine Meta’s broader fair use defenses.

Meta counters that this request is untimely, noting that the plaintiffs formally requested the linked files on October 9, well after the formal discovery period had already closed. Regarding the contents of the "ML Books" folder itself, Meta stated that it is currently investigating the matter and intends to confer with the plaintiffs once more information is uncovered.

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

Missing Server Logs

Beyond the disputed files and internal communications, the plaintiffs claim that critical digital evidence remains unaccounted for. Specifically, Meta has allegedly failed to surrender complete command history files associated with its historical torrenting activities, despite prior judicial instructions to do so. According to the rightsholders, these technical logs are vital because they detail how Meta’s torrenting software was configured, which specific torrent files were targeted, and consequently, which copyrighted works were accessed.

The plaintiffs assert that Meta has withheld command history files from every server utilized during its active torrenting periods. They are now asking the court to compel Meta to officially disclose whether any of these server logs were systematically deleted and, if so, the exact timeline of those deletions.

Meta maintains that it has already handed over substantial volumes of server logs and continues to comb through archival backups, providing new files on a rolling basis. The tech firm insists there is no evidence to suggest that any command history files were deliberately destroyed.

The overarching legal battle continues to expand on multiple fronts. Entrepreneur Media’s lawsuit is currently scheduled to proceed to trial in the spring of 2027, with a growing line of publishers prepared to follow should a settlement be reached beforehand. Meanwhile, the docket of related litigation continues to swell. This month, major publishing houses including Elsevier, Hachette, and Macmillan, alongside an increased number of independent authors, officially joined the queue of plaintiffs challenging the tech industry’s data acquisition practices.