U.S. Magistrate Judge Thomas S. Hixson ordered Meta Platforms to identify every manager and executive who authorized downloading over 160 terabytes of pirated books from shadow libraries like Library Genesis via BitTorrent to train its Llama AI models. While court rulings established that AI training itself constitutes fair use, Meta faces massive legal exposure for torrent seeding and unauthorized distribution. Internal records indicate CEO Mark Zuckerberg and AI chief Ahmad Al-Dahle approved the acquisition despite explicit employee warnings, leaving Meta exposed to statutory damages of up to $150,000 per work.
Judge Hixson's discovery disclosure orders
- ▪Judge Thomas S. Hixson ordered Meta Platforms to reveal what happened to pirated works that were obtained but not ultimately used for AI model training.
- ▪U.S. Magistrate Judge Thomas S. Hixson of the Northern District of California ordered Meta Platforms to identify every manager and executive who approved BitTorrent downloading, seeding, and seeding mitigation for Anna's Archive and Library Genesis Scimag datasets.
Rejection of Meta's team-consensus defense
- ▪Meta Platforms defended against discovery by asserting that torrenting decisions resulted from informal team consensus rather than explicit, top-down executive directives.
- ▪Judge Thomas S. Hixson rejected Meta Platforms' team-consensus defense, ruling that every team member participating in group choices is responsive to the court's discovery requests.
Internal Meta conduct and approvals
- ▪Meta engineer Nikolay Bashlykov authored a script to remove copyright notices from downloaded Library Genesis ebooks to obscure their origin.
- ▪Internal filings unsealed in January 2025 allege that Meta Platforms CEO Mark Zuckerberg personally approved using Library Genesis to train Llama models despite internal executive objections.
- ▪Meta Platforms Head of Generative AI Ahmad Al-Dahle approved BitTorrent downloads after engineer Nikolay Bashlykov warned in writing that using the protocol could violate copyright law.
Volume of Meta's shadow library downloads
- ▪Between 2022 and 2024, Meta Platforms acquired over 160 terabytes of copyrighted books and academic papers to train Llama AI models, downloading 81.7 terabytes via Anna's Archive and 80.6 terabytes from Library Genesis.
- ▪Internal Meta Platforms communications described a single operation updating the download status of 10 terabytes from Library Genesis, 54 terabytes from Z-Library, and 126 terabytes from the Internet Archive.
Legal liability from BitTorrent seeding
- ▪Plaintiffs in the Kadrey v. Meta litigation argue that BitTorrent seeding constitutes contributory copyright infringement by distributing material to unauthorized third parties.
- ▪BitTorrent seeding automatically uploads file fragments to external network users during downloads, creating a distinct act of unauthorized distribution beyond basic downloading.
- ▪District Judge Vince Chhabria ruled that training AI models on books qualifies as fair use, leaving acquisition and BitTorrent seeding as surviving grounds for copyright liability.
Northern California AI copyright lawsuits
- ▪Judge Thomas S. Hixson's discovery order applies to two consolidated lawsuits in Northern California federal court, including the Kadrey v. Meta class action by authors and a complaint by publisher Entrepreneur Media.
- ▪The Kadrey v. Meta class action and Entrepreneur Media complaint are among five jointly managed lawsuits before the Northern District of California accusing Meta Platforms of unlawfully training AI models on protected writing.
Story comments
Loading comments…