Meta tries to scapegoat staff who downloaded torrents to spare execs

In copyright suits brought by authors including Richard Kadrey and Sarah Silverman, and by publisher Entrepreneur Media, Magistrate Judge Thomas Hixon ordered Meta to identify the people who made, authorized or approved decisions to download shadow libraries such as LibGen and Anna's Archive, including their roles and dates. Meta must also hand over command history files for every server it used to torrent, including virtual machines and AWS instances, and explain what happened to pirated works that were never used for training.

Meta had argued that early dataset building was so informal that no one with authority really decided anything; staff simply downloaded. Plaintiffs counter that Meta torrented shadow libraries repeatedly from 2022 to 2024, pulling 80.6 terabytes from LibGen and 35.7 terabytes via Anna's Archive, and that this scale implies seeding, which means distribution.

This matters because accountability for how AI training data was gathered is moving from abstract fair-use debate to named decisions and server logs.

We think Meta's defense was always going to be hard to sustain, and the order shows why. A company does not move tens of terabytes of pirated books through BitTorrent across multiple years by accident. Arguing that a decision was made only in a technical sense, by whoever pressed the button, asks the court to believe that one of the most process-heavy companies on earth had no process for its flagship AI effort.

The record makes that framing worse. Internal messages surfaced in the litigation show staff pushing back; one worker wrote that torrenting from a corporate laptop did not feel right. Another testified that some downloads needed director or VP approval. Plaintiffs also point to messages they say show the LibGen decision followed an escalation to Mark Zuckerberg. Those are allegations, not findings, but they cut directly against the idea that leadership was absent.

That is the uncomfortable heart of this story. If Meta's position holds, the people who carried out the work absorb the blame while the people who set direction stay unnamed. Employees who flagged discomfort would end up as the face of a decision they questioned. Whatever happens on the copyright merits, that is a bad precedent for how companies treat staff who raise concerns.

The server logs may matter even more than the names. Contributory infringement turns on whether Meta facilitated transfers, and command histories could show whether its claimed mitigations actually prevented seeding. If they did not, the distribution question becomes much harder for Meta to escape.

We are firmly in favor of capable open models, and Llama has done real good for researchers. But useful AI built on knowingly pirated work, with no one willing to own the call, erodes the trust the whole field needs. Clear accountability now is better than a settlement that leaves the questions unanswered.

The case grows out of authors' lawsuits over the books used to train early Llama models. Meta maintains it used mitigations to avoid seeding and says it is not aware of anyone approving uploads or distribution. Meta did not immediately respond to Ars Technica's request for comment.