Judge orders Meta to name everyone who approved downloading pirated books for Llama

What the judge ordered

A federal magistrate judge in San Francisco has ordered Meta to name every person who made, authorised or approved its decisions to download pirated book collections that it later used to train its Llama models. In a ten-page discovery order dated 19 October 2026, Magistrate Judge Thomas S. Hixson granted most of a motion by Entrepreneur Media, the publisher of Entrepreneur magazine, which is suing Meta in one of several related copyright cases before District Judge Vince Chhabria.

The order covers "shadow libraries", the unlicensed online collections such as Library Genesis (LibGen) and Anna's Archive that hold millions of copyrighted books and papers. Meta must identify each decision-maker, give their approximate dates of involvement and say which decision they made, authorised or approved for which dataset. It must also name the people who approved the LibGen acquisition inside FAIR, Meta's AI research group, and those who approved resuming a paused Anna's Archive download, or state under oath that it has found none.

The ruling was first reported by TorrentFreak and then by Ars Technica, which published the order and the parties' earlier filings. It is a discovery ruling, not a finding that anyone at Meta did anything unlawful.

The question Meta tried to narrow

Entrepreneur Media's interrogatory 22 asked Meta to identify "each natural person who made, authorized, or approved any decision to obtain or download any Shadow Library Dataset" containing the publisher's works. According to the judge, Meta's answers kept drifting to a different question: who decided to use the data for training, rather than who decided to obtain it.

For Llama 1 and 2, Meta said it had not identified particular individuals who "formally" made or approved decisions to acquire or use books. Hixson pointed out that the question was never limited to formal decisions. For Llama 3 and Llama 4, he said, the responses were again "mostly about use, not obtaining or downloading, which is what the rog asks about."

Meta also argued that it did not necessarily have a single decision-maker, because choices were often collaborative or the product of team consensus. The judge did not accept that as a reason to name no one. "The rog does not presume that Meta had a single decisionmaker; it asks about all the decisonmakers," he wrote. "If a decision is fairly attributable to a group or a team, everyone in the group or team who participated in making that decision is responsive to this rog."

Who chose BitTorrent, and who decided about seeding

A second interrogatory, number 23, asked who approved obtaining, downloading, uploading or distributing shadow-library data over BitTorrent or other peer-to-peer networks. In a joint letter filed on 11 September, Entrepreneur Media said Meta had answered only that a "decision" was made "only in the sense" that two former researchers carried out LibGen downloads using torrents, and that on Meta's reading "no one at Meta decided to torrent anything, yet Meta torrented shadow libraries repeatedly from 2022-2024."

The publisher's side of that letter quotes deposition testimony it says Meta's answers left out. It cites one witness saying datasets of this kind "required some director or VP level approval to use for training", another describing the Anna's Archive effort being halted as a "huge 'don't touch it'" and later resuming after "some decision changed", and an internal logbook entry rejecting torrents for an earlier download because seeding "would be legally not okay." These are one party's characterisations of the testimony, filed before the judge ruled.

Hixson granted two of the publisher's three requests on this question. Meta must identify everyone who took part in the discussion before an engineer began torrenting, everyone who approved proceeding by torrent, and everyone who directed or approved an April 2024 download. It must also name everyone who decided whether or how to seed, or how to limit seeding, when it obtained LibGen's scientific-paper collection and Anna's Archive content by BitTorrent. The judge refused only a request that Meta say whether it had searched two former employees' files, because the interrogatory had not asked that.

Meta's position, set out in the same letter, is that it has fully responded, that it is not aware of anyone who decided to upload or distribute shadow-library data, and that it is withholding no names on privilege grounds. It has said it used measures to avoid seeding.

Why the names matter after Meta's fair use win

The fight over names matters because of what is left of the case. In 2025, Judge Chhabria ruled that Meta's use of the authors' books to train Llama was fair use on the record before him, a decision the September discovery order quotes. That left the BitTorrent claims: whether Meta, while downloading, also uploaded pieces of the books to other users, which would be distribution rather than training. Courts in other countries are now hearing their own training cases, such as the Canadian newspapers' lawsuit against Anthropic.

Torrent clients normally share pieces of a file with other users while they download it. According to TorrentFreak, Meta argued this year that any uploading during its downloads was "part-and-parcel" of a fair use purpose and "an inherent characteristic of the BitTorrent protocol", and that torrents were the only practical way to get Anna's Archive in bulk. Entrepreneur Media told the court that the identities of the people who chose torrents, and who set or did not set the seeding parameters, bear on willfulness and on its contributory infringement claims.

Willfulness matters for money. US copyright law lets courts raise statutory damages when infringement is wilful, so evidence about who knew what, and who signed off, can change the stakes of a case even after the training question has been decided. Ars Technica reports that the plaintiffs' earlier filings cite 80.6 terabytes downloaded from LibGen and 35.7 terabytes through Anna's Archive; those figures come from the plaintiffs, not from the order.

What else the order decided

The 19 October order also dealt with other disputes in the coordinated cases. Meta must explain what, if anything, it did with plaintiffs' books that were downloaded but never used for model training, which the judge called "relevant and discoverable information." It must answer an interrogatory about the removal of copyright management information in those terms, rather than recasting it as removal of "repetitive text", and it must say how roughly nine terabits of shadow-library data deleted on or about 13 July 2023 were acquired.

Meta also has seven days to hand redacted documents from its internal AI Data Catalog to the court for private review of its privilege claims. Other requests went Meta's way. The judge refused to make Meta answer questions for every internal, unreleased model rather than the public Llama 1 to 4, calling that disproportionate, and denied extra deposition time for a corporate witness.

One ruling involved a licence. Fox Corporation licenses multimedia content to Meta for AI training, not books. Meta accidentally produced an unredacted copy of that agreement and the negotiation documents to the plaintiffs. The judge granted Fox's request that the plaintiffs destroy them, finding that the price and terms of a deal with a different rights holder for a different format were too distant from the case to outweigh the harm to Fox's other negotiations.

OpenAI and Anthropic stay out of it

A month earlier the same judge closed another route. In a discovery order dated 10 September, in the cases brought by Chicken Soup for the Soul, the academic publisher Cognella and the journalist John Carreyrou's company, Hixson refused to compel OpenAI and Anthropic, which are not parties, to disclose which torrent clients they used since 2019 and how they tried to prevent uploading.

The publishers had argued that if OpenAI or Anthropic torrented shadow libraries without seeding, Meta's claim that seeding was unavoidable would fall apart. The judge said that claim could be tested by examining the torrent clients themselves, asking: "Why can't Plaintiffs' expert use the torrent clients to show how torrent clients can be used?" He also rejected requests for the two companies' licensing records and tests of their models' ability to write literary content, citing Judge Chhabria's ruling that a market for licensing books as training data is not one rights holders are entitled to.

The September order records each company's own account. OpenAI said its use of shadow libraries was very limited, took place several years ago and has not been connected to its products for years. Anthropic said no publicly available Claude model was trained on text from the shadow-library repositories at issue in its own case, which it says were used only for early research models.

What the order does not settle

A discovery order decides what Meta must disclose, not who is liable. Meta may still argue that some people it names made no decision in any meaningful sense, and the order does not say whether any seeding occurred. It also does not set a public deadline for Meta's new answers; interrogatory responses are exchanged between the parties and may never be filed in full, or may be filed under seal.

Reports have speculated about whether senior executives, including Mark Zuckerberg, could be named. Ars Technica points to messages unsealed in 2025 that referred to "a prior escalation to MZ" before a LibGen decision. The order itself names no executives, and whether anyone senior is identified depends on Meta's sworn answers. Meta and Entrepreneur Media's lawyers did not respond to Ars Technica's requests for comment.

Why it matters

The ruling turns a broad dispute about AI training data into a narrow, personal question: who said yes. That is harder to answer with arguments about fair use, and it is the kind of evidence that can shape settlement talks. The pressure on how labs source books is visible elsewhere too, from a Korean bookseller refusing a 100,000-title order it suspected was for AI training to Meta's own new commercial terms on crawling business websites.

For other AI developers the lesson is about records. Courts are now treating informal, team-level choices about data as decisions that someone owns, and are willing to make companies trace them back to named people. Labs that can show who approved each dataset, and why, will be in a stronger position than those that say no one decided.