Judge keeps OpenAI and Anthropic logs out of Meta's case

Meta's answer to the last live claim against it is that uploading pirated books while downloading them was never a choice, because BitTorrent works that way. Publishers in three related cases tried to break that by subpoenaing OpenAI and Anthropic, and last week a magistrate judge told them to go test a torrent client instead, TorrentFreak reports.

At a glance

• In August the three publishers subpoenaed OpenAI and Anthropic for the identity, versions and configurations of every torrent client the two have used since 2019, including any records of attempts to stop uploading.
• The logic ran like this: if a rival configured its client not to seed, then the redistribution Meta calls inherent to BitTorrent was a setting Meta simply chose not to change.
• Magistrate Judge Thomas Hixson quashed the subpoenas without ruling on the seeding defense itself, while a separate September 11 order sends Meta's own server command histories to the plaintiffs.

If you haven't followed it: authors including Richard Kadrey and Sarah Silverman sued Meta over Llama training on pirated books, and over sharing those books with other BitTorrent users in the process. Last summer Judge Vince Chhabria ruled the training itself was fair use, leaving the BitTorrent distribution claims as the last live part of the case. In the parallel Anthropic case, Judge William Alsup ruled in June 2025 that downloading from pirate libraries was not fair use.

Meta calls the uploading "an inherent characteristic of the BitTorrent protocol"

Earlier this year Meta added a line of defense to the distribution claims in a supplemental interrogatory response. Any uploading of pirated books during its downloads, the company argued, was part-and-parcel of a fair use purpose: BitTorrent was "a more efficient and reliable means of obtaining the datasets," and in the case of Anna's Archive the only way to get them in bulk.

The protocol part is plain enough. BitTorrent moves files in pieces among everyone who holds them, so a client that is downloading normally hands out the pieces it already has, like a potluck where taking a plate means putting one on the table. Taking is leeching, giving is seeding, and most clients do both by default.

One earlier finding in the litigation sits next to that argument. A Meta engineer wrote a script to prevent seeding, but apparently not leeching.

Three publishers asked the rivals for every client configuration since 2019

Both of Meta's torrenting claims are being tested in three related lawsuits, filed by Chicken Soup for the Soul, academic publisher Cognella, and John Carreyrou's Cambronne Inc. All three are assigned to Judge Chhabria and all three target the same shadow-library torrenting.

Rather than wait for Meta to document its own client setup, the publishers went to the two AI companies that could potentially disprove the necessity claim. In August they subpoenaed OpenAI and Anthropic for the identity, versions and configurations of every torrent client each has used since 2019, specifically including any records of efforts to prevent seeding. Both companies have admitted in similar lawsuits that they used books from shadow libraries.

If OpenAI torrented but configured its clients to suppress uploading, then the redistribution Meta calls an "inherent characteristic" of the protocol was a setting Meta declined to change.

Hixson: the assertion "can be tested by examining the BitTorrent client itself"

The publishers also offered a trade. Explain how you acquired the shadow-library data and whether you tried to prevent uploading, they said, and the torrent document demands go away. Neither company took it.

Anthropic told the court that clients are not interchangeable, that they differ in their default upload settings, in whether those defaults can be reconfigured, and in their capacity to suppress uploading during and after a download. "What Anthropic's client allowed shows nothing about what Meta's did," its lawyers wrote. OpenAI made the same point, noting there is no evidence it used the same clients or built "comparable corpora" to Meta.

Hixson sided with them last week. If Meta's defense hinges on the assertion that its use of BitTorrent was the only way BitTorrent can be used, he wrote, the client itself can be examined; any user of a torrent client would be relevant in that sense. The bulk-download claim, he added, can be checked with the shadow libraries directly.

Command histories for every server Meta used to torrent

On September 11, Hixson granted a motion in the Kadrey class action covering the command history files for every server Meta used to torrent, including its virtual machines and AWS instances. A command history is the log a machine keeps of every command an operator typed, in order. On a torrenting box that presumably includes how the client was installed and any changes made to its upload settings.

The order goes back to early 2025, when Meta admitted it had held back relevant documents until after the discovery deadline had passed. Chhabria gave the authors extra discovery to make up for it, including records showing how Meta's torrent clients were set up and used. Meta argued the log files it had already handed over were enough; Hixson disagreed. The authors also hope the histories will reveal exactly which copyrighted works Meta torrented.

Hixson didn't touch Meta's seeding defense; he ruled only that the rivals' logs are the wrong place to test it. Whether the command histories show upload settings, or which books were pulled, has yet to be seen; the order covers the files, not what is in them. In our view the sharper fact is already in the record, since a Meta engineer wrote an anti-seeding script, which makes "inherent characteristic" read as a description of defaults.

When the expert reports land. Opening expert reports in the Meta cases are due later this month, and they will rest on Meta's own server records rather than anyone else's logs. Hixson also pointed the publishers at two places nobody has blocked: the torrent clients themselves, and the shadow libraries, which can say whether bulk downloads were available any other way. The seeding question itself remains undecided.