The Books Cut Apart for AI: What Destructive Scanning Means
16 October 2026 · 3 min read
A book is more than its sequence of words to the person holding it. It is paper, an edition, a binding, perhaps a note left by a previous reader. That is one reason accounts of books being cut apart for AI have travelled so widely. The physical process is real, and the details matter.
What the court record describes
The 23 June 2025 order in Bartz v. Anthropic describes the company purchasing millions of print books for an internal library. Anthropic or its vendors removed bindings, cut pages for scanning, created digital copies and discarded the paper originals. This is destructive scanning: the physical copy is dismantled during digitisation.
The order describes an internal research library from which material could be selected for model training. It does not establish that every book in the library was used to train a model. Nor does this record establish that the destroyed volumes were the last surviving copies of their titles.
Three activities with different questions attached
There are three activities to keep distinct: buying a printed book and scanning it, acquiring a digital copy from a pirate source, and using a copy for model training.
In that June ruling, the US district court treated the training use at issue as fair use and separately accepted the replacement of purchased print copies with internal digital copies. It rejected the same justification for building a permanent library from pirated sources. That was a ruling on particular facts, rather than a worldwide permission covering every use of books by AI companies.
The Association of American Publishers’ September 2025 settlement FAQ explains that the claims remaining after the ruling concerned copying from illegal shadow libraries. This helps explain why the settlement story and the physical destruction story should be read separately.
A digital copy is not automatically public preservation
A scan can preserve readable text while remaining unavailable to the public. A digital research collection held inside a company serves a different purpose from a library where readers can locate, consult or borrow material.
That distinction gives the story a cultural question beyond the legal dispute. Who can use the resulting collection? Under what conditions? What happens to details of the physical object that a text extraction does not capture? These questions are relevant even when another print copy exists elsewhere.
Destruction of a copy is a loss of that object. Establishing the loss of an entire work requires further evidence about surviving copies and access. Keeping those claims separate makes the criticism more accurate and makes preservation needs easier to identify.
What readers can ask for
Ask for clear accounts of where training material comes from, how authors’ interests are addressed and whether digitised collections will remain accessible beyond a company’s immediate needs. A purchased volume, permission to reproduce a work and permission to train a model raise distinct questions.
At a smaller scale, supporting libraries and archives helps sustain places where books can be found and read. Keeping an unusual edition in circulation, donating an appropriate book to a collection that wants it, or consulting a librarian about preservation can give a physical copy a future.
The disturbing image is a severed spine. The lasting question is how the work, the object and the people who made it are treated once the words become data.
Ready to start tracking?
Join readers who track, discover, and share their reading life with Nocturn. Free forever, no credit card needed.
Create free account