Court revelations have shed light on how AI companies acquire books to train artificial intelligence models, including large-scale purchases from second-hand booksellers.
Anthropic, the company behind the Claude AI model, reportedly launched an internal initiative known as “Project Panama” to obtain and digitise vast quantities of books. Court documents in the Bartz v. Anthropic case described a plan to scan books on an enormous scale.
The company faced a fundamental challenge shared across the generative AI industry: advanced language models need large amounts of high-quality human writing. Books provide detailed narratives, analysis, knowledge and diverse language that online content often lacks.
Older books have become particularly valuable as AI-generated material increasingly fills the internet. Training models on content produced by other AI systems can cause them to reproduce errors and repetitive language.
From piracy to physical books
Anthropic initially turned to pirated digital libraries, a strategy that led to legal action and eventually a $1.5 billion settlement with authors.
The company then adopted another approach: purchasing physical books, sending them for digitisation and destroying the copies afterward.
According to the court documents, workers removed the bindings, cut the spines and prepared the pages for scanning. Once digitised, the physical books could be discarded or recycled, leaving the company with a digital copy for internal use.
A US court subsequently ruled that, under certain circumstances, this process could qualify as transformative useunder copyright law.
Second-hand booksellers report unusual orders
The practice has raised broader concerns about the large-scale acquisition of books by AI companies.
Second-hand booksellers in the UK and Ireland have reported an increase in unusual bulk orders from unidentified buyers. The Guardian reported that similar activity has emerged among booksellers in the US, Australia and Europe.
Some requests involve highly specific editions and seemingly unrelated subjects.
One bookseller in Northumberland, for example, received requests for an Estonian translation of John le Carré’s The Mission Song, a particular edition of Anne Brontë’s Agnes Grey and an October 1983 issue of Warship magazine.
An Irish bookseller also reported sporadic, apparently automated orders ranging from books about 18th-century agricultural tools in Africa to biographies of 1950s racing drivers.
Some orders reportedly came from different buyers but went to the same delivery address, including a warehouse complex near Heathrow Airport.
The unusual purchasing patterns have fuelled speculation that AI companies or contractors are acquiring physical books primarily for their training data, although the identities and intentions of many of the buyers remain unclear.
A growing ethical debate
The controversy goes beyond copyright. Critics question whether companies should be able to acquire books on a massive scale, extract their intellectual content and commercially exploit that knowledge without directly compensating the authors.
The debate also raises questions about the physical destruction of books that may contain unique editions, annotations, dedications or historical value.
For AI companies, the books represent valuable training material. For authors and booksellers, however, the practice raises a fundamental question: who should benefit when human knowledge becomes the raw material for increasingly powerful commercial AI systems?