A federal judge has ruled that Anthropic illegally scraped and pirated books from more than 10 Stanford professors to train Claude, a finding that now sits underneath a $1.5 billion settlement and a fresh round of litigation that enterprise teams building on Claude cannot treat as background noise. The ruling, issued by Judge William Alsup in Bartz v. Anthropic, splits the legal question in two: training on legally acquired books is fair use, but building a permanent library from pirated copies is not. That distinction is now the line enterprise legal teams need to understand before they sign a Claude deployment contract.
The mechanics of the ruling matter for anyone doing data-provenance diligence. Alsup's opinion, as reported by the Stanford Daily, draws an analogy between how Claude's large language models trained and how a human reader internalizes a book: "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," he wrote. That framing protects the training method itself. What it does not protect is the acquisition method — downloading and storing copyrighted texts from sources the Stanford Daily names as "shadow libraries" like LibGen and PiLiMi, which is the conduct both Bartz and the newer Cruz v. Anthropic complaint target directly.
The settlement figure is concrete: $1.5 billion distributed to authors and publishers of nearly 500,000 registered works, which the Authors Alliance calls the largest copyright settlement involving an AI company, according to the Stanford Daily. Stanford professors named in the piece as having had their work used without permission include Tobias Wolff, Abraham Verghese, Chang-Rae Lee, Daniel Mason, Hayagreeva Rao and Myra Strober. Wolff says 10 of his own publications were used for AI training without permission. He and Chang-rae Lee opted out of the Bartz settlement and filed a separate non-class suit, Cruz v. Anthropic, while Wolff and 27 other authors have said the settlement isn't enough.
The economics of opting out are the operational number architects should sit with. The Bartz settlement works out to roughly $3,000 of damages per infringed work, per the Stanford Daily's reporting. Wolff and Lee's separate suit, Cruz v. Anthropic, seeks a jury trial for full statutory damages of up to $150,000 per infringed work — a fifty-fold difference that signals plaintiffs believe the class settlement undervalued the claims, and that a jury verdict could reset the price of copyright exposure for every AI company with a scraped training corpus.
What breaks here isn't the model — it's the assumption that a settlement closes the liability question. Cruz v. Anthropic seeks a permanent injunction against using the plaintiffs' works for Claude model training if successful, which would mean an enterprise relying on an existing Claude model has no guarantee the underlying training corpus survives future litigation intact. Paul Goldstein, a Stanford Law intellectual property scholar and one of the plaintiffs represented in the Bartz class action, frames the structural problem in an email quoted by the Stanford Daily: "Millions of individual copyrighted works may be copied in the course of any training incident, and it's simply impracticable for the platform to obtain licenses from the copyright owner of every — or, indeed, any — single work."
This is not an isolated dispute. The Stanford Daily reports that over 70 infringement lawsuits have been filed against AI companies by authors, visual artists and news outlets including the New York Times and Getty Images, and that the Authors Guild is separately pursuing litigation against OpenAI and Microsoft while lobbying Congress for licensing legislation. Authors Guild CEO Mary Rasenberger, quoted in the piece, says licensing "allows authors and publishers to be compensated and to require the use of filters that ensure their works do not appear in outputs unless separately licensed" — a regime that, if enacted, would require AI vendors to expose training-data manifests that most currently treat as proprietary.
For architects, the takeaway is not that Claude is uniquely exposed — Bartz's fair-use carve-out for legally trained models actually gives Anthropic more cover than companies facing suits with no settlement at all — but that provenance litigation is live and unresolved industry-wide: before locking a vendor into a long-term enterprise contract, get contractual indemnification language for training-data claims, because the per-work damages Cruz is seeking are fifty times what the current settlement priced in, and no court has yet decided which number holds.