Bartz v. Anthropic PBC

Why This Case Matters

Bartz v. Anthropic PBC is the most consequential AI copyright ruling to date. For the first time, a federal court drew a clear, constitutionally grounded line between lawful and unlawful AI training data — and the $1.5 billion settlement that followed is the largest copyright recovery in United States history. This case answers the question that NYT v. OpenAI, Authors Guild v. Google, and dozens of other suits have been circling: when does AI training on copyrighted works cross the line?

The Piracy Timeline

Anthropic’s liability began long before any lawsuit was filed. To train its Claude family of AI models, the company used datasets containing millions of pirated books sourced from three shadow libraries:

• Books3: A dataset of 196,640 books scraped from Bibliotik, a private torrent tracker, and widely shared in the AI research community.
• Library Genesis (LibGen): An underground repository hosting approximately 5 million pirated books, academic papers, and other copyrighted works.
• PiLiMi: A lesser-known shadow library containing roughly 2 million pirated texts.

In total, the training corpus included over 7 million pirated works. Plaintiffs Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson — all published authors — filed their class action on August 19, 2024, alleging that Anthropic knowingly incorporated stolen copies of their books into its training pipeline.

The Fair Use Ruling

On June 23, 2025, Senior U.S. District Judge William Alsup issued the ruling that split the case — and the AI industry — in two. His opinion turned on a single, decisive distinction:

Purchased copies = fair use. Judge Alsup found that when an AI company lawfully purchases copyrighted books and uses them to train a model that produces entirely new, non-substitutive output, the use is “quintessentially transformative.” The model does not reproduce the books; it learns patterns from them and generates something fundamentally different. This analysis followed the Supreme Court’s framework in Campbell v. Acuff-Rose Music (1994) and the Second Circuit’s reasoning in Authors Guild v. Google (2015), while carefully distinguishing Andy Warhol Foundation v. Goldsmith (2023) on the grounds that AI output does not substitute for the original works.

Pirated copies = per se infringement. For books obtained through Books3, LibGen, and PiLiMi, the court held that the analysis was far simpler. Training on pirated copies is “inherently, irredeemably infringing” because the foundational act — unauthorized reproduction — taints every downstream use. No amount of transformation can launder a pirated source. Judge Alsup rooted this holding in the Constitution’s Progress Clause (Art. I, Sec. 8, Cl. 8), arguing that copyright’s role as “the engine of free expression” requires that unauthorized copying be treated as categorically different from licensed use.

The $1.5 Billion Settlement

Following the fair use ruling, the parties entered settlement negotiations. On September 5, 2025, Anthropic agreed to pay $1.5 billion to resolve claims on behalf of a class covering 482,460 works — approximately $3,000 per book. The settlement was preliminarily approved by the court on September 25, 2025, with final approval scheduled for April 2026.

The settlement structure reflected the court’s dual ruling: Anthropic accepted liability only for works whose copies were traced to pirated sources, while maintaining that its use of lawfully purchased copies was protected fair use.

Constitutional Foundations

What makes this ruling particularly durable is its constitutional grounding. Rather than relying solely on statutory fair use factors, Judge Alsup anchored his analysis in the Copyright Clause itself. The Constitution grants Congress the power to secure exclusive rights to authors in order to “promote the Progress of Science and useful Arts.” Judge Alsup reasoned that this purpose is served by allowing transformative AI training (which creates new knowledge tools) but defeated by piracy (which deprives authors of their incentive to create).

This constitutional framing gives the ruling persuasive weight well beyond the Northern District of California and may influence how other circuits — and ultimately the Supreme Court — approach AI training questions.

The Broader Impact

Bartz v. Anthropic establishes three principles that will shape AI copyright law for years to come:

1. The provenance principle: Where your training data comes from matters as much as what you do with it. Clean data pipelines are not just good ethics — they are legal necessities.

2. The valuation benchmark: At approximately $3,000 per work, the settlement gives publishers, authors, and AI companies a starting point for licensing negotiations. Future deals will be priced against this number.

3. The compliance roadmap: AI companies now have a clear framework — purchase or license your training data, document provenance, and avoid shadow libraries. Companies that follow this path can rely on the transformative fair use defense; those that don’t face existential liability.

The case also sends a signal to the global debate over AI and copyright. While the EU AI Act and other regulatory frameworks take different approaches, the Bartz ruling demonstrates that existing U.S. copyright law, properly applied, can accommodate AI innovation without sacrificing creators’ rights.