Court Papers Reveal Anthropic Bought, Scanned and Destroyed Millions of Books to Train Its AI — And Tried to Keep It Quiet

Internal filings made public in a recent court order depict a deliberate, large‑scale effort by Anthropic to feed its Claude models with high‑quality written material by purchasing used books, converting them into digital files and recycling the paper originals. Company planners favored physical books as a dependable source of polished prose that could elevate model outputs above noisier web‑scraped text, and the operation relied on industrial scanners and mechanized cutting equipment to process volumes efficiently and at low cost. Anthropic’s legal team anchored its defense in the resale owner’s prerogative to repurpose purchased goods, a justification that featured prominently in court exchanges as judges weighed transformative‑use arguments. The records also disclose earlier use of automated downloads from shadow libraries, a separate procurement channel that executives later characterized as distinct from the book‑scanning program. The book revelations culminated in a significant settlement with authors — reported in filings as roughly $1.5 billion — underscoring both commercial exposure and reputational harm. Parallel litigation made public during discovery has widened the controversy beyond books: music publishers have filed a complaint alleging tens of thousands of songs, lyrics and sheet music were incorporated into training corpora and are seeking more than $3 billion in damages. At the same time, several major publishers have begun blocking automated access to repositories such as the Internet Archive, arguing that organized, machine‑readable archives are being used as indirect routes for mass ingestion by AI firms. Those publisher measures trade reduced archival openness for a tactic aimed at cutting off convenient access to copyrighted reporting and other materials, but they also raise concerns about fragmenting public preservation resources. Taken together, the disclosures and the new complaints illustrate a pattern of aggressive, sometimes opaque data acquisition and prompt questions about when legal defensibility diverges from ethical or reputational risk. For creators, publishers and model builders alike, the episode intensifies pressure for clearer provenance, licensing frameworks, and technical or contractual controls that can prevent unconsented reuse. The affair is likely to push the industry toward auditable supply chains, negotiated licensing deals, and possibly narrower judicial rules on permissible training practices.