Meta forced to identify staff behind Shadow Library downloads for Llama models

A US district court judge granted plaintiffs' motion to compel in the Kadrey v. Meta AI-training copyright class action, ordering Meta to turn over the unredacted names and internal communications of the engineers, data curators and managers who downloaded and processed pirated book archives from shadow libraries including Anna's Archive and LibGen. Magistrate Judge Thomas Hixon's discovery order, reported by TorrentFreak, also requires Meta to produce command history files for every server it used to torrent, after Meta argued that no one with authority decided to torrent. The plaintiffs, including authors Richard Kadrey and Sarah Silverman and Entrepreneur Media, aim to establish willful infringement through depositions of the named decision-makers, which could unlock statutory damages reaching millions of dollars per registered work.

The training regimes powering modern large language models have come under intense legal scrutiny from authors, publishers, and intellectual property holders. In an ongoing class-action lawsuit filed by prominent authors—including Sarah Silverman and Richard Kadrey—plaintiffs allege that Meta https://me.mashable.com/tech/76433/everything-announced-at-meta-connect-2026 unlawfully ingested vast repositories of copyrighted literary works without licensing or compensation. Moving beyond general corporate liability, a US district court judge has granted the plaintiffs' motion to compel, formally ordering Meta to turn over the unredacted names and internal communications of the specific engineers, data curators, and managers responsible for downloading and processing pirated book archives. Meta has been ordered to name executives who authorized downloading books from shadow libraries such as Anna's Archive and LibGen, TorrentFreak reported. Magistrate Judge Thomas Hixon's discovery order, in AI-training copyright suits by authors including Richard Kadrey and Sarah Silverman and by Entrepreneur Media, also compels Meta to produce command history files for every server it used to torrent. Meta had argued no one with authority decided to torrent. Authors argue Meta also seeded; Meta says it used mitigations to prevent seeding. Meta's defense is that no one with authority chose to torrent, which leaves the servers as the only executives the command history can name. @Meta — V L S @ V L S October 20, 2026 https://x.com/ V L S /status/2112618398605808019?ref src=twsrc%5Etfw Shadow Library allegations: Books3 and LibGen At the center of the dispute are controversial datasets known in AI research circles as "Books3" and "Library Genesis" LibGen . Court records and leaked internal chat logs revealed that Meta researchers discussed downloading torrent files containing tens of thousands of contemporary books, acknowledging the ambiguous legal origins of the data while prioritizing training volume to keep pace with rival frontier models from OpenAI https://me.mashable.com/tech/58107/openai-is-working-on-an-artificial-intelligence-powered-web-browser-launch-expected-in-the-coming-we and Google DeepMind https://me.mashable.com/tech/59282/google-deepmind-launches-genie-3-to-create-smarter-more-human-like-ai . While Meta previously attempted to protect employee privacy by redacting individual identifiers from discovery disclosures, the presiding judge ruled that plaintiffs have a legitimate legal right to depose key decision-makers directly. Corporate shielding falls: Personal liability and depositions The ruling pierces the veil of anonymity that tech corporations traditionally maintain during early-stage intellectual property litigation. By forcing Meta to name names, individual engineers and team leads can now be subpoenaed for sworn depositions. Plaintiffs' counsel aims to establish whether management directed staff to bypass copyright laws intentionally, which could establish willful infringement and unlock statutory damages reaching millions of dollars per registered work—sharpening the stakes around how commercial foundation models like Muse Spark Open https://me.mashable.com/tech/75355/the-new-apple-mac-studios-are-more-powerful-than-ever-heres-how-to-preorder-ahead-of-the-sept-22-lau and Llama architectures are trained. Uploading Pirated Books via BitTorrent Qualifies as Fair Use, Meta Argues | Ernesto Van der Sar, Torrentfreak In the race to build the most capable LLM models, several tech companies sourced copyrighted content for use as training data, without obtaining permission from content owners. Meta, the parent company of Facebook and Instagram, was one of the companies to get sued. In 2023, well-known book authors, including Richard Kadrey, Sarah Silverman, and Christopher Golden, filed a class-action lawsuit against the company. Meta’s Bittersweet Victory Last summer, Meta scored a key victory in this case, as the court concluded that using pirated books to train its Llama LLM qualified as fair use, based on the arguments presented in this case. This was a bittersweet victory, however, as Meta remained on the hook for downloading and sharing the books via BitTorrent. By downloading books from shadow libraries such as Anna’s Archive, Meta relied on BitTorrent transfers. In addition to downloading content, these typically upload data to others as well. According to the authors, this means that Meta was engaged in widespread and direct copyright infringement. In recent months, the lawsuit continued based on this remaining direct copyright infringement claim. While both parties collected additional evidence through the discovery process, it remained unclear what defense Meta would use. Until now. Seeding Pirated Books is Fair Use Last week, Meta served a supplemental interrogatory response at the California federal court, which marks a new direction in its defense. For the first time, the company argued that uploading pirated books to other BitTorrent users during the torrent download process also qualifies as fair use. Meta’s reasoning is straightforward. Anyone who uses BitTorrent to transfer files automatically uploads content to other people, as it is inherent to the protocol. In other words, the uploading wasn’t a choice, it was simply how the technology works. Meta also argued that the BitTorrent sharing was a necessity to get the valuable but pirated data. In the case of Anna’s Archive, Meta said, the datasets were only available in bulk through torrent downloads, making BitTorrent the only practical option. “Meta used BitTorrent because it was a more efficient and reliable means of obtaining the datasets, and in the case of Anna’s Archive, those datasets were only available in bulk through torrent downloads,” Meta’s attorney writes. “Accordingly, to the extent Plaintiffs can come forth with evidence that their works or portions thereof were theoretically ‘made available’ to others on the BitTorrent network during the torrent download process, this was part-and-parcel of the download of Plaintiffs’ works in furtherance of Meta’s transformative fair use purpose.” In other words, obtaining the millions of books that were needed to engage in the fair use training of its LLM, required the direct downloading, which ultimately serves the same fair use purpose. Authors and Meta Disagree over Fair Use Timing The authors were not happy with last week’s late Friday submission and the new defense. On Monday morning, their lawyers filed a letter with Judge Vince Chhabria flagging the late-night filing as an improper end-run around the discovery deadline. They point out that Meta had been aware of the uploading claims since November 2024, but that it never brought up this fair use defense in the past, not even when the court asked about it. The letter specifically mentions that while Meta has a “continuing duty” to supplement discovery under Rule 26 e , this rule does not create a “loophole” allowing a party to add new defenses to its advantage after a court deadline has passed. “Meta for understandable reasons never once suggested it would assert a fair use defense to the uploading-based claims, including after this Court raised the issue with Meta last November,” the lawyers write. Meta’s legal team fired back the following day, filing their own letter with Judge Chhabria. This letter explains that the fair use argument for the direct copyright infringement claim is not new at all. Meta pointed to the parties’ joint December 2025 case management statement, in which it had explicitly flagged the defense, and noted that the author’s own attorney had addressed it at a court hearing days later. “In short, Plaintiffs’ assertion that Meta ‘never once suggested it would assert a fair use defense to the uploading-based claims, including after’ the November 2025 hearing, is false” Meta’s attorney writes in the letter. Authors Admit No Harm, No Infringing Output Meanwhile, it’s worth noting that Meta’s interrogatory response also cites deposition testimony from the authors themselves, using their own words to bolster its fair use defense. The company notes that every named author has admitted they are unaware of any Meta model output that replicates content from their books. Sarah Silverman, when asked whether it mattered if Meta’s models never output language from her book, testified that “It doesn’t matter at all.” Meta argues these admissions undercut any theory of market harm. If the authors themselves cannot point to infringing output or lost sales, the lawsuit is less about protecting their books and more about challenging the training process itself, which the court already ruled was fair use. These admissions were central to Meta’s fair use defense on the training claims, which Meta won last summer. Whether they carry the same weight in the remaining BitTorrent distribution dispute has yet to be seen. ‘U.S. AI Leadership at Stake’ In its interrogatory response, Meta added further weight by stressing that its investment in AI has helped the U.S. to establish U.S. global leadership, putting the country ahead of geopolitical competitors. That’s a valuable asset worth treasuring, it indirectly suggested. As the case moves forward, Judge Chhabria will have to decide whether to allow this “fair use by technical necessity” defense. Needless to say, this will be of vital importance to this and many other AI lawsuits, where the use of shadow libraries is at stake. For now, the BitTorrent distribution claims remain the last live piece of a lawsuit filed in 2023. Whether Judge Chhabria will allow Meta’s new defense to proceed has yet to be seen. Read more: https://t.co/y1A9oRrGKy — Owen Gregorian @OwenGregorian March 8, 2026 https://x.com/OwenGregorian/status/2030630802967462099?ref src=twsrc%5Etfw Industry-wide ramifications for AI training data Legal analysts describe the decision as a watershed moment for the generative AI sector https://me.mashable.com/creative-software/68432/compare-chatgpt-gemini-and-more-side-by-side-with-this-ai-tool . For years, foundational model developers treated web scraping and shadow torrents as industry-standard research methods protected under the fair use doctrine. By opening individual employees to deposition and cross-examination, the court’s order is likely to prompt AI laboratories to overhaul data acquisition policies and demand strict provenance audits for all ingested training corpuses, echoing broader corporate crackdowns on cybersecurity and algorithmic trust https://me.mashable.com/tech/77087/bill-gates-rejects-voluntary-ai-self-regulation-we-are-not-doing-enough . The federal order compelling Meta to identify employees involved in torrenting copyrighted books represents a turning point in AI copyright enforcement. Forcing individual researchers into the witness chair establishes that the race for model capability cannot outrun legal accountability for the data that powers it. Read More: Mark Zuckerberg named in lawsuit over Meta’s use of pirated books for AI training https://me.mashable.com/tech/51231/mark-zuckerberg-named-in-lawsuit-over-metas-use-of-pirated-books-for-ai-training