Why Is Anthropic Destroying Books?

The idea of destroying books to obtain training data for artificial intelligence has emerged from court documents in the Bartz v Anthropic PBC case. The case, decided in late July, highlighted “Project Panama”, an effort by Anthropic to destructively scan books for use in improving its large language model, Claude.

An internal memo described Project Panama as an effort to “destructively scan all the books in the world”. It also advised discretion about the project. The episode raises questions about how AI companies obtain high-quality training material and how copyright law treats that process.

Anthropic’s Search For Training Data

Anthropic needed a large and high-quality language dataset to train Claude. The company reportedly preferred material created before 2022, before generative AI began significantly influencing contemporary text.

Books offered an especially valuable source of long-form language. They contain carefully developed facts, organised analysis and fictional narratives. These qualities could help Claude produce writing that is accurate and compelling.

Anthropic could have obtained permission to use existing e-books. However, that approach would have required extensive work involving copyright owners, licensing and business arrangements.

Instead, the company initially used pirated sources. That decision later contributed to a $1.5 billion out-of-court settlement with authors.

When Anthropic became less willing to use pirated books for legal reasons, it turned towards destructive scanning. The company effectively sought to convert physical books into digital copies while disposing of the originals.

How Destructive Scanning Worked

The process involved more than simply scanning books. Anthropic hired a logistics manager, sourced books through vendors, employed staff, housed the material in a warehouse and engaged a digitisation company.

The vendors removed books from their bindings, cut their pages to size and scanned them into digital form. They then discarded the paper originals.

The court’s decision described the process as a transformation of copyrighted works into a new electronic format. Anthropic subsequently disposed of the physical copies.

The operation turned book destruction into a logistical exercise. Warehouses contained neatly organised shelves of books awaiting scanning, while workers handled the material before it was converted into digital files.

The process also involved preparing books for efficient scanning. Vendors sliced spines and edges so that pages could be processed more easily.

The court found no evidence that the new digital copies were shown, shared or sold outside Anthropic. That distinction became significant in considering whether the company’s use of the material qualified as fair use.

Copyright And Fair Use

Under US copyright law, fair use can permit certain transformative uses of copyrighted works without permission from the copyright owner. Anthropic’s process relied on the argument that printed books could be transformed into a digital format for training purposes.

The court also ruled that using proprietary material to train a large language model did not, by itself, constitute copyright infringement. The decision therefore treated the training process differently from simply distributing or selling copies of the works.

However, the case raises broader questions about the relationship between AI development and authors. Anthropic’s decision to destroy physical books rather than negotiate with their creators has become a central concern surrounding the practice.

The issue also extends beyond copyright. Books have cultural value as physical objects, not merely as containers for text.

Books have long carried meanings connected to their physical form, history and role in culture. Yet the United States has relatively few legal provisions governing the treatment or export of American cultural heritage, including books.

What Happens To Human-Created Text?

The court’s decision makes destructive scanning an effective way to remove direct authorial involvement from printed works while retaining their content for AI development.

If other generative AI companies adopt similar practices, the scale of the activity could become much larger. That possibility raises questions about how many books could be scanned and destroyed before society considers the practice unacceptable.

There is no endangered list for printed works. There is also little regulation defining when rare or unique books should be preserved. Furthermore, there is no formal recognition of human-generated textual works as a distinct category of cultural heritage requiring protection.

The question becomes more pressing if 2022 is treated as an important point when AI-generated text began entering the wider textual record. In that context, wholly human-authored works could eventually become an important category for collections, preservation and stewardship.

The case also highlights the continuing value of books and long-form writing. For centuries, people have read and reread books, absorbing their themes, ideas and approaches to writing.

That history makes Anthropic’s approach particularly striking. The company apparently found it easier to scan and destroy physical books than to undertake the legal, practical and commercial work required to deal with their creators.

The Value Of Books And Readers

The case ultimately raises questions about why complex, long-form and largely uncorrupted text remains so valuable to AI companies.

One response is to resist the devaluation of reading and writing. Another is to recognise the cultural spaces in which people create, exchange and preserve ideas and words.

The broader concern is that generative AI could shift control over the means of producing and distributing written culture. Readers and writers could increasingly become consumers of systems trained on their work rather than active participants in creating and preserving that culture.

The Anthropic case therefore extends beyond the question of whether books can legally be scanned. It asks what society values in books, who should control the knowledge they contain and how human creativity should be treated as AI systems increasingly depend on it.

With inputs from Reuters