The five copyright actions against Meta over its use of pirated books to train the Llama models now have a trial date. In a Joint Case Management Statement filed on 27 October 2026 in the Northern District of California (Kadrey v. Meta Platforms, Inc., No. 3:23-cv-03417-VC, Dkt. No. 879), the parties record that, under the court’s 14 October 2026 order, Meta and Entrepreneur Media will prepare for trial on 3 May 2027 before Judge Vince Chhabria. The statement was filed under Civil Local Rule 16-9(a) ahead of a case management conference on 29 October 2026, and covers five captions being managed together: the Kadrey authors’ case, Entrepreneur Media (3:25-cv-09579), and the three publisher matters the parties call the “Three C’s” — Cambronne (3:26-cv-03725), Chicken Soup for the Soul (3:26-cv-02333) and Cognella (3:26-cv-04053).
The parties also propose what happens if the lead case settles: the 3 May slot passes down in order of filing, so Cambronne would go first, then Chicken Soup, then Cognella. One scheduling wrinkle remains in the statement — the pretrial conference set for 20 April 2027 collides with a Delaware Superior Court jury trial for Meta’s lead trial counsel (Conduent State Healthcare, LLC v. AIG Specialty Insurance Company, No. N18C-12-074), and Meta suggests moving the conference to 26 or 27 April while keeping the pretrial submission deadlines keyed to the original date.
The substantive dispute recorded in the statement is about acquisition, not training. Plaintiffs report that on the Friday before filing, former Meta employee Stephen Roller testified at deposition that he torrented and knowingly distributed Books3 — a dataset of some 196,000 pirated books — during the period he worked at Meta (2017 to 2023), and that he deleted his copy about a year ago, “[d]ue to all this litigation.” They note that as recently as Meta’s 24 August 2026 interrogatory responses, Meta had said it did not believe it obtained Books3 via BitTorrent or any other peer-to-peer protocol. Meta’s half of the statement disputes the characterisation squarely: it says Roller did not know when he downloaded the dataset, never testified that the download was for or at Meta rather than for himself or another employer, said he did not know whether more than one copy was downloaded, and did not confirm that the two colleagues plaintiffs name had copies of their own. The deleted copy, Meta adds, sat on Roller’s personal computer and was removed more than two years after he left the company. As TorrentFreak reports, Meta acknowledged in early 2024 that it used portions of Books3 to train Llama 1 and Llama 2 without saying how it obtained the dataset.
What it changes
The useful part of this filing is not the gossip about a book club. It is a set of dates and procedural rulings that tell you, concretely, how an AI-training copyright case is actually won or lost — and almost none of it turns on the merits of fair use.
Your list of asserted works is built out of the defendant’s discovery, not your own records
Entrepreneur Media must file an amended complaint by 6 November 2026, and the statement says it is expected to add more than 400 additional books and more than 140 additional magazine issues. Those works were not found in Entrepreneur’s own archive. They were identified in Meta’s interrogatory responses and search results served between 4 August and 30 September 2026, and in the 2 October 2026 opening report of plaintiffs’ torrenting expert, Dr David Choffnes, which plaintiffs say identifies more than 400 Entrepreneur books inside the shadow-library collections Meta torrented. Entrepreneur’s own position is blunt: only Meta knows which works it acquired and trained on, so the final list depends on Meta’s disclosures.
Read that as a planning rule. If your works may be in a training corpus, the complaint you file first is a placeholder. The real claim emerges from the defendant’s interrogatory responses and its expert’s corpus analysis, twelve to thirty months later, and the amendment window to capture it is measured in weeks. Build the litigation plan around a late conforming amendment rather than around the works you can name on day one.
Registration and chain of title are the gating items — and the price of registration is about to move
Meta’s response to the proposed amendment is not a merits argument. It asks the court to order Entrepreneur and the Three C’s to hand over the proposed amended complaints, the lists of asserted works, all materials establishing their right to assert each work, and copies of copyright registrations for every proposed asserted work. Entrepreneur states it has already produced all author and writer agreements in its possession, with registration certificates possibly still to come.
That is the whole claim reduced to a paperwork exercise. A US infringement claim cannot be asserted on an unregistered US work, and every added title has to arrive with a registration and a clean chain from author to claimant. For publishers, agencies and in-house teams, the practical consequence is that the back catalogue has to be registration-complete and contract-complete before the corpus list lands, not after — which is where disciplined IP contract and rights-record management stops being administrative hygiene and starts being the difference between a 400-work claim and a 40-work one. There is also a cost clock: the Register of Copyrights submitted a new fee schedule to Congress on 14 July 2026 under 17 U.S.C. §708(b), which can take effect as early as 12 November 2026. If a bulk catch-up registration programme is on your list, the arithmetic changes this month.
Timeliness, not relevance, is killing the best evidence
Plaintiffs want Meta ordered to produce documents hyperlinked inside documents it already produced, including a Google Drive folder titled “ML Books” and records of a Meta book club that may have used shadow-library books. Their legal theory is sound and they cite Andy Warhol Foundation for the Visual Arts v. Goldsmith, 598 U.S. 508, 533 (2023) for it: fair use is assessed separately for each use, so books taken from a shadow library and then read for pleasure are a second, unexcused use that sits outside any training-related defence.
Meta’s answer is a calendar. The two documents containing those links were produced on 17 August 2026; plaintiffs first asked for the linked files on 9 October 2026, which Meta says is more than a month after fact discovery closed and over three weeks after the Local Rule 37-3 deadline to raise discovery disputes. On that reasoning the folder never gets produced, whatever is in it. The same timeliness objection is aimed at the request for Meta’s torrenting expert’s prior reports from other BitTorrent matters, where Meta points out that Judge Hixson already denied an equivalent request directed at third-party OpenAI and that no objection was filed, making the ruling final.
The lesson generalises well beyond AI. Hyperlinked attachments, shared-drive folders and chat threads are now where the decisive admissions live, and they are not captured by ordinary document review — you have to notice the link, ask for the target, and ask while the discovery window is open. Diarise the dispute-raising deadline, not just the discovery cut-off.
Provenance logs and personal devices are the real exposure for anyone training a model
The command-history fight is the most transferable part of the record. Back in March 2025 the court adopted a limited-discovery proposal that expressly defined “Documents” to include command-line history logs, naming the .bash_history file in the user home directories of servers used to download data. Plaintiffs say Meta produced none from the servers used during its torrenting periods until one surfaced in search-term results, that Judge Hixson then ordered full compliance — Meta dates that order to 11 September 2026 — and that they now want the court to order Meta to state whether any such files were deleted and when. Meta says it has produced substantial logs from those servers and is still searching backups on a rolling basis, and that there is no basis to allege deletion.
If you build or fine-tune models on third-party content, assume that shell history, torrent client configuration, and the identity of every engineer who ever pulled a corpus are discoverable and will be read aloud. Three things follow. Keep a dataset provenance register that records, per corpus, the source, the acquisition method, the licence or permission relied on, and the approving manager — because the alternative is reconstructing it from .bash_history under court order. Extend litigation holds to personal machines and former employees: the most damaging testimony in this filing came from someone who had left the company three years earlier and deleted a dataset from his own computer. And govern secondary internal uses of any acquired corpus, because under the per-use rule a reading group can forfeit ground that the training defence had already won.
The dates to hold
From the statement and the discovery orders it describes: amended complaint by 6 November 2026; supplemental interrogatory responses on the schedule Meta agreed (30 October and 6 November 2026); privilege-log disputes to Judge Hixson by 13 November 2026; a further joint discovery letter brief on the remaining interrogatory disputes, or a notice that none remain, by 17 November 2026; pretrial conference 20 April 2027 (possibly 26 or 27 April); trial 3 May 2027. TorrentFreak reports that Elsevier, Hachette and Macmillan, along with further authors, joined the queue of related cases this month — so the sequencing the parties proposed is likely to matter more, not less, as the trial date approaches.
For the fair-use half of the argument, the contrast with the Third Circuit’s decision in the ROSS Intelligence case is worth holding in mind: there the defendant’s lawful access did not save it, and here it is unlawful acquisition that survived an otherwise favourable training ruling. Either way, the door that stays open is the one marked “how did you get it”.