Meta tries to scapegoat staff who downloaded torrents to spare execs

Magistrate Judge Thomas Hixon has ordered Meta to name the executives who authorized downloading books from shadow libraries for AI training, according to TorrentFreak's reporting. The discovery order lands in two overlapping cases: a class action brought by authors including Richard Kadrey and Sarah Silverman, and a separate suit from publisher Entrepreneur Media. Meta has spent weeks arguing that the torrenting was the work of individual staffers rather than a leadership decision, and Hixon largely rejected that framing.
The order requires Meta to identify each person who made, authorized or approved any decision to obtain or download shadow library datasets containing the plaintiffs' works, and to specify for each person the dates of involvement and which dataset the decision covered — or to state under oath that no such person exists.
Meta's central defense has been procedural as much as factual. In a court filing, the company described its early AI dataset building as an "informal approach" in which nobody with authority made decisions, and said a decision was "made only in the sense" that staffers "technically" downloaded torrents. It also said leadership did not commonly make formal decisions when acquiring data for early models like Llama 1 and 2, and that this informal approach later gave way to structured review and approval processes.
Hixon was not persuaded on several fronts. He noted that the authors' request was not limited to formal decisions. When Meta argued that approvals came out of "collaborative" discussions and therefore had no single decision-maker, the judge replied that the motion does not presume one decision maker but asks about all of them. Meta had proposed supplying only names; Hixon said that would not do.
The internal record cuts against the no-decision story. Staff testified that downloads required director- or VP-level approval for training use, and one employee wrote that "torrenting from a corporate laptop doesn't feel right." Another staffer testified that leadership halted and then resumed downloading torrents from Anna's Archive. A string of unredacted messages unsealed early last year allegedly showed that the decision to download from LibGen came after an escalation to Mark Zuckerberg, which is why his name may yet enter the case.
The scale numbers matter. Authors argue Meta could not have pulled 80.6 terabytes from LibGen or 35.7 terabytes through Anna's Archive quickly without also seeding — uploading pieces of files to other swarm members. Meta says it relied on "mitigations" to prevent seeding and that there is no basis to suspect someone approved seeding that those mitigations were meant to avoid. Under a contributory infringement theory, the authors need only show Meta facilitated transfers of their works, so evidence of seeding is central rather than incidental.
That is where the command history files come in. Hixon's order also compels Meta to produce command history for every server used to torrent, including virtual machines and AWS instances. In practice, shell history and instance logs can show what commands ran, from which host, and in what order — which is why they are often the most concrete evidence available about how a download pipeline actually behaved, and potentially about whether seeding occurred.
Meta also came under pressure over pirated works it obtained but did not use for training; Hixon said the company must answer questions about those too. The plaintiffs put the torrenting window at 2022 to 2024, repeatedly, across multiple shadow libraries.
A lawyer for Meta and Entrepreneur Media did not immediately respond to Ars' request for comment. The discovery fight is not the merits ruling, and naming people is not the same as proving willfulness, but the order narrows the gap between what authors suspect and what they can put in front of a jury. This suggests the practical risk for Meta is less about the individual names than about the chain-of-command documents and server logs that naming will pull into the record — material that could support a finding that infringement was deliberate rather than accidental.
This is a familiar pattern in AI copyright litigation: plaintiffs chase the training-data pipeline because models themselves reveal little about provenance, and defendants argue that data acquisition was ad hoc engineering rather than corporate policy. Courts have generally allowed broad discovery into how datasets were assembled, and this order fits that trend by refusing to let Meta define "decision" narrowly enough to exclude everyone with authority.
Why it matters: Developers and product teams building on scraped or downloaded datasets should read this as a warning that informal data practices become the central evidence years later. If a court can compel names, dates and server logs for a training run, then approval trails and data provenance records are a legal asset, not bureaucracy. The immediate effect is on Meta's discovery burden and exposure; the broader inference is that companies treating dataset acquisition as undocumented engineering will face harder questions than those that logged who approved what and when.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Meta argues nobody “decided” to pirate books to train AI.