AI Companies Bought, Scanned and Destroyed Millions of Physical Books

The artificial-intelligence industry wanted better writing.
Its solution was to buy enormous quantities of physical books, cut off their bindings, scan every page and dispose of what remained.
The practice became widely debated after court documents revealed details of Anthropic’s internal Project Panama, an effort to create a large private digital library for training Claude.
One internal planning document described the project as an attempt to “destructively scan” books at extraordinary scale. A scanning supplier proposed processing between 500,000 and two million books within six months. �
The Washington Post +1
What Is Destructive Scanning?
Scanning an ordinary bound book page by page can be slow.
Industrial scanning operations increase speed by removing the spine so that individual pages can pass automatically through scanning equipment.
The process may involve:
Purchasing physical copies.
Cutting away their bindings.
Feeding loose pages through high-speed scanners.
Converting the text into searchable digital files.
Discarding or recycling the remaining paper.
The digital version becomes available to the organisation.
The original physical copy no longer exists.
Why Anthropic Wanted Books
Books offer something much of the open internet does not consistently provide: edited, sustained and carefully structured writing.
Court documents indicated that Anthropic’s developers viewed books as a source of higher-quality language than large amounts of informal internet text.
The material could help models learn:
Long-form reasoning
Narrative structure
Specialist vocabulary
Coherent argument
Literary style
Historical and cultural information
The controversy is therefore filled with irony.
AI companies valued books because human beings had invested so much labour in producing them—then converted the books into machine-training material through a process that physically destroyed the purchased copies.
Was It Legal?
A US judge previously distinguished between legally purchased books and pirated digital copies.
The court found that using purchased books to build an internal training library could qualify as transformative fair use under the circumstances examined. However, Anthropic separately faced liability questions concerning millions of books obtained from unauthorised shadow libraries before Project Panama. �
The Washington Post +1
Legal permission and cultural wisdom are not identical questions.
A company may have the legal right to destroy an item it purchased.
The public may still object when the item is rare, out of print or part of a shrinking physical record.
Were Rare Books Destroyed?
Online retellings have sometimes implied that AI companies systematically destroyed unique manuscripts or the last surviving copies of priceless books.
The available evidence does not establish that sweeping claim.
The revealed programmes involved mass acquisition of published physical books, including used and out-of-print material. The precise composition of the full collection remains unclear from the public documents.
Responsible coverage should therefore avoid claiming that unique historical artefacts were definitely destroyed unless a specific example is verified.
The established story is already extraordinary enough: physical books were acquired at industrial scale, disassembled, scanned and discarded to create private AI-training libraries.
Why the Story Went Viral Again
The revelations produced images that feel symbolically disturbing.
Society has historically associated book destruction with censorship and the removal of knowledge.
Project Panama involved the opposite stated objective: destroying physical copies to preserve and absorb their contents digitally.
Knowledge was not being erased.
It was being transferred from publicly tradable objects into privately controlled computational infrastructure.
That shift explains much of the unease.
The Ownership Question
When an AI company purchases one copy of a novel, it owns that physical copy.
But does owning the paper provide an ethical right to use its language to build a commercial system capable of generating competing text?
Courts, authors and companies remain divided.
Authors argue that their creative labour becomes part of profitable systems without meaningful consent or compensation.
AI developers argue that models learn statistical patterns rather than storing ordinary searchable copies and that such learning can be transformative.
The debate will shape the future economics of publishing.
Mary Chuks’ Perspective
Books are not merely containers of words.
They are products of human attention.
A model may scan a book in minutes, but the author may have spent years creating it.
The correct business response should not be secrecy or the assumption that purchasing paper settles every moral question.
A better system would involve:
Transparent training-data records
Licensing options
Collective creator compensation
Protection for rare and culturally important works
Auditable rules for data retention
Clear distinctions between public-domain and copyrighted material
AI can learn from human civilisation without pretending civilisation produced itself for free.
Conclusion
Project Panama sounds like a dystopian novel because the physical process is so striking:
Buy the books. Cut them apart. Teach the machine. Recycle the remains.
The deeper question is not whether AI should learn from books.
It is whether the humans who created and preserved those books should remain visible in the value chain after the machine has finished reading.
Original Sources and Further Reading
Detailed investigation: The Washington Post reported on Project Panama using newly revealed legal documents and internal planning records. �
The Washington Post +1
Legal background: Ars Technica examined the court’s distinction between purchased physical books and unauthorised digital copies. �