Carreyrou v. Anthropic PBC
OPPOSITION/RESPONSE (re 114 MOTION to Sever, 94 MOTION… — Attachment 126
Issue: Whether plaintiffs asserting copyright infringement claims against multiple competing AI developers—Anthropic, Google, OpenAI, Meta, xAI, Perplexity, Apple, and Nvidia—arising from the alleged use of the same shadow-library piracy ecosystem to train large language models may be joined as defendants in a single action under Federal Rule of Civil Procedure 20(a)(2), or whether those claims must be severed into separate proceedings under Rule 21.
Plaintiffs—professional authors and rights holders—filed an amended complaint alleging that all defendants systematically sourced copyrighted books from the same interconnected network of piracy repositories (LibGen, Z-Library, Bibliotik, Books3, The Pile) and used those works in materially similar LLM training pipelines without authorization. Multiple defendants, including Google, Anthropic, Perplexity, Meta, and xAI, filed motions to sever, arguing that their respective conduct constituted independent acts by separate competitors rather than a common transaction or occurrence. Plaintiffs filed a consolidated opposition contending that severance is impermissible under Rule 20(a)(2) because the claims share a unifying factual nucleus—a single industry-wide scheme with common sourcing, uniform technical copying steps (ingestion, preprocessing, tokenization, iterative training), and parallel commercialization—and that common legal questions predominate, including the scope of fair use for AI training, the market-harm analysis under *Campbell v. Acuff-Rose*, and the willfulness standard. Plaintiffs additionally argued that severance would prejudice them by forcing duplicative discovery and parallel expert proceedings, and requested in the alternative that any severed actions be immediately consolidated under Rule 42(a).
This procedural dispute is an early but consequential test of whether mass AI copyright litigation against industry-wide defendants can proceed in a single forum, with the court's joinder ruling likely to determine whether fair use defenses—particularly the fourth-factor market-harm inquiry, which requires examining the aggregate effect of all defendants' conduct on the licensing market for AI training data—are adjudicated consistently or fragmented across parallel actions. The outcome may signal how courts will structure the wave of generative-AI copyright cases and whether the "industry-wide scheme" theory is sufficient to sustain multi-defendant joinder in AI training-data litigation.
Amended Complaint — Attachment 119
Issue: Whether Anthropic, Google, Meta, xAI, Perplexity, Apple, NVIDIA, and OpenAI are liable under the Copyright Act for willful infringement by downloading plaintiffs' copyrighted books from shadow libraries (including LibGen, Z-Library, Anna's Archive, and The Pile/Books3) and reproducing those works during LLM training, preprocessing, and fine-tuning without license or permission.
Plaintiffs — individual authors and a corporate plaintiff — filed this amended complaint in the Northern District of California against thirteen AI and technology defendants, alleging that each defendant obtained pirated copies of plaintiffs' registered, copyrighted books from shadow-library websites and made additional unauthorized reproductions during LLM ingestion, preprocessing, and iterative training. Plaintiffs allege the infringement was willful, citing internal industry warnings that these sources were illegal and defendants' deliberate decision to proceed for competitive advantage. Plaintiffs expressly declined class-action treatment, invoking their right under the Copyright Act to pursue individualized statutory damages — up to $150,000 per willfully infringed work — and a jury determination of willfulness, contrasting their approach with a pending class settlement they characterize as yielding approximately $3,000 per work.
This complaint advances the unsettled question of whether the use of pirated training datasets constitutes willful copyright infringement by LLM developers at each stage of the AI development pipeline, potentially establishing that liability attaches not only at initial download but also at preprocessing, deduplication, and iterative fine-tuning; the plaintiffs' deliberate individual-action strategy, if successful, could foreclose industry efforts to resolve mass AI copyright claims through low-value class settlements.