Anthropic trained Claude on illegally obtained Stanford professors' books
- ID
- 40301
- Status
- summarized
- Published
- 27 Oct 2026, 11:28 PM
- Fetched
- 28 Oct 2026, 2:26 PM
- Provider
- Hacker News
- Category
- dev-community
- Original URL
- https://stanforddaily.com/2026/10/27/professors-books-pirated-to-train-claude/
- Source URL
- https://hnrss.org/best
Summary
- Score
- 7.5
- Created
- 28 Oct 2026, 2:26 PM
- Tags
- Audience
- developersvibe_codersai_ml_learnersai_agent_userssaas_founders
What happened
A May 2026 federal ruling in Bartz v. Anthropic found that Anthropic illegally scraped and pirated books by over 10 Stanford professors to train Claude, while Judge William Alsup treated legally acquired books used for training as fair use. The $1.5 billion settlement covers nearly 500,000 registered works, about $3,000 per work, but Tobias Wolff and Chang-rae Lee opted out and filed Cruz v. Anthropic, seeking up to $150,000 per infringed work. Wolff had 10 publications used without permission, and the new suit alleges Anthropic downloaded and stored pirated books from shadow libraries including LibGen and PiLiMi.
Why it matters
For builders shipping on hosted LLMs, this is a concrete legal-risk datapoint about training-data provenance: pirated-library use was ruled illegal, and the opt-out suit seeks up to $150,000 per work versus roughly $3,000 per work in the settlement. If you depend on a model vendor, ask what provenance and indemnity terms you actually have instead of assuming the training data is clean; the text shows no Malaysia-specific impact.
Discussion angle
Should model-provider training-data provenance become a standard procurement question, given the $1.5B settlement at roughly $3,000 per work versus the opt-out suit seeking up to $150,000 per infringed work?