The Stanford Daily reported on Oct. 27 that Anthropic pirated the work of at least 10 Stanford professors to train its Claude models, a finding confirmed in the Bartz v. Anthropic copyright case. Tobias Wolff, professor emeritus of English, had 10 of his own publications in the training set. Court filings in the case disclosed that Anthropic downloaded more than 7 million book copies from the shadow libraries LibGen and PiLiMi, including works by Abraham Verghese, Chang-Rae Lee, Daniel Mason, Hayagreeva Rao and Myra Strober.
Judge William Alsup drew a hard line: training on legally purchased books counts as fair use, but downloading and storing pirated copies does not. Anthropic agreed to a $1.5 billion class settlement covering roughly 500,000 works, which works out to about $3,000 per book. Wolff and 28 other authors opted out, arguing the payout is too small and leaves the company free to keep mining their writing.
In May 2026 those authors filed Cruz v. Anthropic, seeking a jury trial, statutory damages of up to $150,000 per infringed work, and a permanent injunction barring their books from future training runs. For founders building on top of AI models, the takeaway is blunt: training data provenance is a line item with seven figures of downside, and the gap between fair use and infringement can come down to how a file arrived on a disk.