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

Magistrate Judge Thomas Hixson has refused to let publishers suing Meta pull torrent-client records from OpenAI and Anthropic. He has also ordered Meta to hand over the command logs of every server it used to torrent pirated books. TorrentFreak reported both orders on September 15. The order on Meta’s servers is dated September 11.

The two rulings deal with the same question, and it is now the only live issue against Meta: could Meta have downloaded books over BitTorrent without also uploading them to other people?

Why seeding is what’s left

The class action by authors including Richard Kadrey and Sarah Silverman accused Meta of two things. The first was training its Llama models on pirated books. The second was sharing those books with other BitTorrent users while it downloaded them. On June 25, 2025, Judge Vince Chhabria ruled that the training was fair use. As Loeb notes, that ruling bound only the thirteen named plaintiffs and did not settle the distribution claim, which survives independently.

That second claim is the harder one for Meta. BitTorrent splits a file into small pieces spread across a swarm of users, and a client that has a piece serves it to others. Distribution is a separate right under the Copyright Act, so unauthorized uploading could be a violation of its own even if training is fair use, and it is harder to defend as transformative, Mass Law Blog argues. The blog adds that the plaintiffs say willful infringement could bring enhanced statutory damages of up to $150,000 per infringed work.

Earlier this year Meta added a new line of defense in a supplemental interrogatory response. It said any uploading during its torrent downloads was “part-and-parcel” of a fair use purpose. Meta described BitTorrent as “a more efficient and reliable means of obtaining the datasets” and said that for Anna’s Archive it was the only way to get them in bulk. Torrent clients upload to each other by design, so on Meta’s account any sharing was simply “an inherent characteristic of the BitTorrent protocol.”

Going to the rivals

Both parts of that argument, that torrenting was necessary and that uploading was unavoidable, are being tested in three related lawsuits from Chicken Soup for the Soul, academic publisher Cognella and John Carreyrou’s Cambronne Inc. All three are assigned to Chhabria and target the same shadow-library torrenting. The publishers didn’t wait for Meta to document its client setup. In August they subpoenaed OpenAI and Anthropic for the identity, versions and configurations of every torrent client each company had used since 2019, including any records of efforts to prevent seeding.

The reasoning was simple. Both companies have admitted in similar lawsuits that they used books from shadow libraries, and Anthropic’s case ended in a $1.5 billion settlement that authors and publishers were still dividing up in September. If either company had set a client not to upload, Meta’s necessity argument would be in trouble.

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.

The publishers built that on an earlier finding in the case: a Meta engineer wrote a script to prevent seeding, but apparently not leeching. In torrent terms, seeding is uploading after a download completes, and leeching is the phase while the file is still coming in, when a client normally also passes pieces on.

The publishers also offered a shortcut. If OpenAI or Anthropic simply explained how they acquired the shadow-library data and whether they tried to stop uploading, the document demands would be dropped. Neither company took the offer. Both told the court it would be better to examine the relevant torrent client directly, and that their own practices said nothing about Meta’s.

“Clients are not interchangeable, 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,” Anthropic’s lawyers wrote. “What Anthropic’s client allowed shows nothing about what Meta’s did.” OpenAI said there was no evidence that it used the same torrent clients or “built ‘comparable corpora’ to Meta.”

“Examining the BitTorrent client itself”

Hixson sided with the two companies and did not rule on Meta’s seeding arguments. He found that rivals’ torrent logs are not the best place to look for evidence about what BitTorrent allows.

“To the extent Meta’s fair use defense hinges on the assertion that its use of BitTorrent was the only way BitTorrent can be used, that assertion can be tested by examining the BitTorrent client itself,” he wrote. OpenAI’s or Anthropic’s logs say little about Meta’s setup or about what torrent clients can do. “Any user of a torrent client would be relevant in that sense. Why can’t Plaintiffs’ expert use the torrent clients to show how torrent clients can be used?”

He dealt with the bulk-download claim the same way. Whether shadow-library data could only be obtained in bulk through torrents is something the publishers can check with the libraries directly, without going through other AI companies.

Meta’s own command histories

The other order points the plaintiffs at the more direct evidence. In the Kadrey class action, Hixson granted a motion 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 server keeps of every command an operator types. On a torrenting machine, that presumably includes how the client was installed and any changes 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. To make up for that, Chhabria gave the authors extra discovery, including records of how Meta’s torrent clients were set up and used. Meta argued that the log files it had already handed over were enough. Hixson disagreed.

The AWS servers are central to the plaintiffs’ account. Mass Law Blog says they allege Meta torrented 267 terabytes between April and June 2024 on Amazon Web Services rather than Facebook servers, and that their forensic analysis shows those servers re-seeded files back into the swarm. The authors also hope the command histories will show exactly which copyrighted works Meta torrented. It is not yet known whether the data will show any of that.

Whether Meta could have downloaded the books without seeding is now a question for the plaintiffs’ experts, who will work from Meta’s own server records. Their opening reports in the Meta cases are due later in September.