Anthropic $1.5 B Settlement Over Unauthorized Use of Stanford Professors' Books for Claude Training

Anthropic’s $1.5 B Settlement Confirms Unauthorized Use of Stanford Professors’ Books

A U.S. district court found that Anthropic AI illegally downloaded and stored copyrighted books by ten Stanford professors to train its Claude language models, and ordered a $1.5 billion settlement that covers authors and publishers of roughly 500,000 works.

Court Ruling Distinguishes Piracy from Fair‑Use Training

Judge William Alsup ruled that while the act of pirating the books – copying them from shadow libraries such as LibGen and PiLiMi into a central repository – was unlawful, the subsequent use of those copies for model training qualified as fair use. He likened AI training to a human reader internalizing a book’s themes and style, stating:

"Like any reader aspiring to be a writer, Anthropic’s LLMs trained upon works not to race ahead and replicate or supplant them — but to turn a hard corner and create something different."

The judgment therefore blocks future piracy for training but does not prohibit training on legally obtained copies.

Key Plaintiffs and Their Claims

  • Tobias Wolff, National Medal of Arts laureate, and 27 other authors objected that the settlement amount was insufficient. Wolff described the unauthorized use as "a form of slave labor" and warned that unchecked AI training could undermine writers’ livelihoods.
  • Wolff and fellow author Chang‑Rae Lee rejected the class‑action settlement and filed a separate suit, Cruz v. Anthropic, seeking statutory damages of up to $150,000 per infringed work—far higher than the roughly $3,000 per work offered in the Bartz settlement.
  • Stanford IP scholar Paul Goldstein highlighted the systemic "transaction‑cost" problem: copying millions of works for training makes it impractical for AI firms to obtain individual licenses, but impracticality is not a legal defense.

Settlement Details and Open Issues

  • The $1.5 billion pool is earmarked for authors and publishers of the 500,000 registered works covered by the class action (Bartz v. Anthropic).
  • The settlement does not grant Anthropic a blanket license; the court barred the company from creating a permanent central library of pirated books for future training.
  • If Wolff’s Cruz case succeeds, Anthropic could face a permanent injunction requiring the removal of the disputed books from any training corpus and possibly the retraining of existing Claude models.

Community Reaction on Hacker News

  • @piltdownman warned that Anthropic’s CEO could face felony charges under the Computer Fraud and Abuse Act, likening the case to the prosecution of Aaron Swartz.
  • @legitster summarized the practical reality: Anthropic likely scraped a mix of wholesale‑purchased books and pirated copies, and the legal dispute now hinges on licensing rather than proving direct influence of any single text.
  • @CPUHall clarified that the ruling’s problem was the source of the books (pirated from LibGen), not the act of training itself, noting that a $3,000‑per‑work settlement would be negligible compared to the scale of the infringement.
  • @ceroxylon asked whether a permanent injunction would force Anthropic to re‑train Claude without the disputed works, a question that remains unresolved pending the outcome of Cruz v. Anthropic.

Broader Legal Landscape

  • Over 70 infringement lawsuits have been filed against AI firms, including actions by The New York Times, Getty Images, and the Authors Guild against OpenAI and Microsoft.
  • The Authors Guild, led by Mary Rasenberger, is pursuing legislative reforms that would require AI companies to obtain licenses before using copyrighted works, arguing that licensing would also enable authors to demand content‑filtering safeguards.
  • The U.S. remains an outlier among industrialized nations for lacking a statutory moral right of attribution, a point emphasized by Goldstein as a systemic disadvantage for authors.

What This Means for AI Development

  1. Technical Impact – If a permanent injunction is issued, Anthropic may need to purge the disputed books from its training pipelines and possibly re‑train Claude models, incurring significant compute costs.
  2. Business Model Shift – The settlement underscores the financial risk of relying on unlicensed data; future AI projects will likely prioritize pre‑licensed corpora or develop robust data‑audit frameworks.
  3. Policy Momentum – The case adds pressure on Congress to enact national licensing mandates, which could standardize compensation for authors and reduce litigation uncertainty.
  4. Industry Precedent – While the court recognized fair‑use training, it also affirmed that pirated acquisition is illegal, setting a clear boundary for AI firms: obtain data legally, or face massive liability.

Takeaway

The Anthropic ruling confirms that pirating copyrighted books for AI training is illegal, even though the subsequent use of those copies may be deemed fair use. The $1.5 billion settlement and ongoing lawsuits signal a turning point: AI developers must now confront the licensing and attribution challenges that have long plagued the industry, or risk costly legal setbacks and forced model retraining.

Sources

Related