Meta Ordered to Reveal Employees Who Downloaded Pirated Books for Llama AI

In a courtroom that feels more like a tech‑war zone than a traditional bench, a U.S. district judge just told Meta it has to spill the beans on the people who actually downloaded the books. Not the CEOs, not the PR team, but the engineers, data scientists, and project managers who were on the front lines of the torrent‑driven data‑gathering effort.

The case started as a class‑action lawsuit filed by a handful of well‑known authors—Sarah Silverman, Richard Kadrey, among others—who claim Meta’s Llama large language model was trained on tens of thousands of copyrighted titles that were never licensed. The plaintiffs argue this is outright infringement, while Meta has tried to lean on a “fair‑use” defense, saying the data was used for a transformative AI purpose.

What changed last week was the judge’s discovery order. Magistrate Judge Thomas Hixon said the plaintiffs have a legitimate right to know exactly who authorized and executed the torrent downloads from shadow libraries such as Anna’s Archive and Library Genesis (LibGen). Meta was ordered to hand over unredacted employee names, internal emails, and even the command‑history logs from every server that ever ran a BitTorrent client for the so‑called “Books3” and LibGen datasets.

Meta’s lawyers tried to argue that no single person made the decision to torrent— that it was an automated process run by scripts. The judge wasn’t buying it. By demanding the command‑history files, the court is essentially saying: “If a script did it, tell us who wrote and approved that script.” That opens the door for individual depositions, where engineers could be asked whether they knowingly bypassed copyright law.

For the tech industry, this is a big shift. Corporations have long shielded staff behind layers of legal “privilege” in early‑stage IP cases. Now the veil is being pulled back, and personal liability looms. If the court finds that Meta’s staff acted with willful intent, statutory damages could skyrocket into the millions per work—​a potential financial nightmare for anyone involved in AI training pipelines.

Meta’s previous win on the fair‑use question adds a bitter aftertaste. While the court earlier agreed that using the books to train Llama might qualify as transformative, it never cleared the company of the act of downloading—and, crucially, seeding—the files. The company’s latest argument tries to spin the technical nature of BitTorrent: because the protocol automatically uploads to other peers, the upload was “inevitable” and therefore not a separate infringing act. Critics say that’s a thin line to walk.

At the heart of the dispute are the massive, scraped datasets called Books3 and the LibGen archive. Internal chat logs, obtained by the plaintiffs, show Meta researchers openly discussing the legal gray area, weighing the risk of infringement against the need to keep pace with rivals like OpenAI and Google DeepMind.

What happens next? The discovery phase will likely produce a mountain of emails, server logs, and perhaps a few uncomfortable testimonies. Whether the plaintiffs can prove intentional direction from senior managers remains to be seen, but the very fact that individual engineers can now be subpoenaed is already a warning shot for the whole AI community.

In short, the case underscores a growing tension: the hunger for massive, high‑quality training data versus the legal and ethical boundaries that protect creators. As AI models get bigger and more capable, the question of who pays for the data—​and who gets held accountable—​is only getting louder.