Anthropic’s Destructive Book Digitization for Model Training- A Modern "Book Burning"?

A source is testimony, not proof. Citing one is a provenance fact — it records who said what, where — and never a verification that the claim is true.

Who is named

The characters and things this document talks about — 4 in alls. Each links to everything else the record knows about them.

Claims made

Every claim in this document. Firsthand = asserted on its own basis; secondhand = recounted from elsewhere, carrying no independent weight.

• Anthropic acquired millions of physical books, removed their bindings with industrial cutters, scanned the pages into digital files, and discarded or recycled the paper.

  Anthropic acquired millions of physical books, removed their bindings with industrial cutters, scanned the pages into digital files, and discarded or recycled the paper.

  Anthropic

  Untested — nothing else bears on it yet

• The resulting corpus was stored in an internal research library and used to train Claude.

  The resulting corpus was stored in an internal research library and used to train Claude.

  Claude

  Untested — nothing else bears on it yet

• The process is ordinary high-volume digitization performed at industrial scale.

  The process is ordinary high-volume digitization performed at industrial scale.

  Anthropic

  Untested — nothing else bears on it yet

• Court filings confirm the purchases and the destruction.

  Court filings confirm the purchases and the destruction.

  Untested — nothing else bears on it yet

• Judge William Alsup treated the lawfully bought physical copies as fair use because the books were transformed into training data and the digital files were not distributed.

  Judge William Alsup treated the lawfully bought physical copies as fair use because the books were transformed into training data and the digital files were not distributed.

  William Alsup

  Untested — nothing else bears on it yet

• A parallel collection of millions of pirated digital books was ruled outside fair use and later resolved by a $1.5 billion class-action settlement.

  A parallel collection of millions of pirated digital books was ruled outside fair use and later resolved by a $1.5 billion class-action settlement.

  Untested — nothing else bears on it yet

• No public licensing of the physical volumes is documented.

  No public licensing of the physical volumes is documented.

  Untested — nothing else bears on it yet

• Google Books digitized library holdings under non‑destructive methods and returned the originals.

  Comparable earlier projects, most notably Google Books, digitized library holdings under non‑destructive methods and returned the originals.

  Google Books

  Untested — nothing else bears on it yet

• Anthropic’s approach permanently removed the physical objects from circulation while concentrating the digital versions under private control.

  Anthropic’s approach permanently removed the physical objects from circulation while concentrating the digital versions under private control.

  Anthropic

  Untested — nothing else bears on it yet

• Historical digitization has expanded from research access to commercial model training; the present case adds irreversible physical loss at volume.

  Historical digitization has expanded from research access to commercial model training; the present case adds irreversible physical loss at volume.

  Untested — nothing else bears on it yet

• The concrete stakes are ownership of the resulting knowledge base and the absence of any public archival obligation.

  The concrete stakes are ownership of the resulting knowledge base and the absence of any public archival obligation.

  Untested — nothing else bears on it yet

• Oversight presently rests with copyright courts and settlement distribution; independent non‑destructive digitization projects and collective licensing remain the practical counters available to authors and libraries.

  Oversight presently rests with copyright courts and settlement distribution; independent non‑destructive digitization projects and collective licensing remain the practical counters available to authors and libraries.

  Untested — nothing else bears on it yet

No claims match this filter.

What it cost

Reading this document into the graph cost <$0.01.

12 claims in all. Content fingerprint 99397caa98df… — the document's text itself is not republished here; follow the link above to read it at its source.