The Bartz Architecture Goes East
Five publishers and Scott Turow sued Meta Platforms and Mark Zuckerberg on 5 May 2026 in the Southern District of New York for willful copyright infringement in the development of the Llama models1. The complaint pleads that Zuckerberg and other Meta executives authorised and directed the torrenting of more than 267 TB of pirated material, including from LibGen, Anna’s Archive, Sci-Hub and other pirate sites. Three points are doing the analytical work in the complaint. Forum is the most important.
The Forum Is the Strategy
Three cases now frame the book-training branch of the AI copyright docket. Bartz v Anthropic2 in the Northern District of California, where Judge Alsup held that Anthropic’s training use was fair use but refused to excuse the separate acquisition and retention of pirated works from LibGen and PiLiMi. The $1.5 billion settlement received preliminary approval on 25 September 2025 and remains subject to final approval. Kadrey v Meta3, also NDCal, where Judge Chhabria granted Meta summary judgment on the record before him on 25 June 2025 with the express warning that the ruling did not establish that Meta’s use of copyrighted works to train Llama was lawful as a general proposition. And now Elsevier v Meta in the Southern District of New York.
The publishers’ choice of venue is not coincidental. SDNY now has its own active AI copyright docket, including The New York Times v OpenAI before Judge Stein4. SDNY judges have not opined on the acquisition/use split that Judge Alsup applied in Bartz. Filing east of the Sierra Nevada gives the publishers a clean bench.
The Pleading Choices
Two pleading choices structure the case.
First, the complaint centres acquisition. Judge Alsup’s reasoning in Bartz turned on a clean conceptual division. Training on legally acquired copies was treated as transformative under Campbell v Acuff-Rose. The antecedent unlawful download of seven million books from LibGen and PiLiMi was independently actionable. The Elsevier complaint mirrors that division. It alleges over 267 TB of pirated material was torrented after Zuckerberg and other Meta executives authorised and directed the process, an executive-level decision in April 2023 to abandon publisher licensing negotiations and a knowing infringement theory under 17 U.S.C. § 504(c). Statutory damages on willful infringement can run to $150,000 per eligible work. If the allegations are proved, even at a fraction of that ceiling the exposure dwarfs the Anthropic settlement of $3,000 per work across roughly 500,000 books.
Second, the complaint names Zuckerberg personally. Naming Zuckerberg is not cosmetic. The complaint pleads direct infringement against him and, in the alternative, contributory infringement. The Rule 12 fight will therefore be whether the pleaded facts show knowledge, authorisation and material contribution, rather than merely apex status at Meta. The doctrinal frame remains Gershwin Publishing v Columbia Artists Management (2d Cir. 1971), Sony, Grokster and In re Aimster.
The Comparative Lens
UK practitioners are watching Getty Images v Stability AI for a different reason. The High Court’s 2025 judgment did not decide the central training-copying question under UK copyright law. It rejected important parts of Getty’s case and found only limited trade mark infringement. The comparison with Elsevier v Meta is therefore structural, not doctrinal: both cases turn on source material, provenance and what can be proved about training data. The US dispute is filtered through fair use. The UK dispute is not, but that does not mean UK claimants can ignore proof of copying, territoriality or subsistence.
EU regulatory attention has centred on the AI Act’s training-data transparency obligation under Article 53 of Regulation (EU) 2024/16895, in force for general-purpose AI providers since 2 August 2025 with transitional rules for pre-existing models. Article 53 may produce public-facing and regulatory material relevant to provenance disputes, but it is not a substitute for US discovery. It is an evidential pressure point, not an evidence pipeline.
What to Watch in the Next Ninety Days
Three near-term decisions matter. First, the motion-to-dismiss schedule on the personal claims against Zuckerberg. Second, any consolidation or transfer motion under 28 U.S.C. § 1404 that could pull the case back to NDCal. Third, whether any party seeks coordination or consolidation of related SDNY AI copyright actions under Rule 42(a).
Anthropic showed why the acquisition record can drive settlement economics. The Bartz architecture has now travelled east. The economic question is whether SDNY will produce a Bartz-style holding without trial, or whether Meta will litigate the alleged willful-infringement record and the corporate-officer point to judgment.
Notes
1 Elsevier Inc., Cengage Learning, Inc., Hachette Book Group, Inc., Macmillan Publishing Group, LLC, McGraw Hill LLC, Scott Turow and S.C.R.I.B.E., Inc. v Meta Platforms, Inc. and Mark Zuckerberg, Civil Action No. 26-cv-3689 (S.D.N.Y., complaint filed 5 May 2026).
2 Bartz et al. v Anthropic PBC, No. 3:24-cv-05417 (N.D. Cal.) (Alsup, J.). Summary judgment opinion on fair use 23 June 2025. Settlement preliminary approval 25 September 2025; final approval pending (fairness hearing listed 14 May 2026).
3 Kadrey et al. v Meta Platforms, Inc., No. 3:23-cv-03417 (N.D. Cal.) (Chhabria, J.). Summary judgment for Meta on training fair use 25 June 2025.
4 The New York Times Co. v OpenAI, Inc. and Microsoft Corp., No. 1:23-cv-11195 (S.D.N.Y.) (Stein, J.).
5 Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 (the AI Act). Article 53 obligations on providers of general-purpose AI models entered into application on 2 August 2025, with transitional rules for models placed on the market before that date.