Used bookstores were not supposed to be a growth story. For years, the sector struggled as digital reading reshaped habits and foot traffic declined. Bookstore sales peaked in 2007 at $17.18 billion and have been falling largely ever since. Now, some of those same shops are processing bulk orders of thousands of volumes in a single day. The buyers are not readers. They are AI companies, and the books may not survive the encounter.
From Northumberland to Las Vegas: The Scale of the Demand
Barter Books, a used bookstore in Northumberland, England’s least densely populated county, typically moves between 2,000 and 3,000 books per week. Recently, the owner received a single bulk order from a Canadian company matching that entire weekly volume. Several neighboring bookshops report similar patterns.
The orders arrive through platforms like Biblio, which anonymizes buyers, so sellers rarely know with certainty who is purchasing or why. But the scale of the transactions makes the “avid personal collector” explanation implausible. One shipment containing at least one rare book was tracked to an Amazon AI training facility in Las Vegas, a facility that destructively scans books for AI training purposes.
Destructive scanning is exactly what it sounds like. Bound books are physically disassembled so their individual pages can be fed through high-speed scanners. The text is then ingested by a large language model. Once the process is complete, the pages are discarded, recycled, or pulped. The book ceases to exist.
The Legal Landscape That Made This Necessary
The surge in used book purchases is not random. It follows a specific legal development. Anthropic recently reached what has been described as one of the largest copyright infringement settlements in history, agreeing to pay $1.5 billion to more than 300,000 writers who had filed a complaint against the company two years prior. OpenAI continues to face multiple copyright infringement lawsuits from sources including The New York Times, Encyclopedia Britannica, comedian Sarah Silverman, and several nonfiction authors, who allege the company copied their books to train its large language model.
The Anthropic settlement, however, produced a court ruling with significant implications. The court determined that training AI on copyrighted works is not illegal, provided the company paid for the books it used. A separate judge reached a similar conclusion regarding Meta. This ruling effectively created a legal pathway: acquire the physical books, pay for them, and the training is permissible.
Used books offer a cheaper acquisition route than purchasing directly from publishers. Publishers may also decline to supply copies to AI companies if the intended use is known. Buying through anonymized marketplaces sidesteps both problems.
The Anthropic case also surfaced an internal initiative called “Project Panama,” which the company described as its effort to “destructively scan all the books in the world.” Internal documentation revealed that the company used a codename for the training method specifically because, in its own words, it did not want it to be known that it was working on this. Anthropic has since stated to the BBC that none of its data acquisition programs buy and destroy rare or antiquarian books.
What This Reveals About How AI Systems Are Actually Built
Here is what most coverage of AI development underemphasizes: the quality and breadth of a language model’s training data is not a technical footnote. It is the foundation. Books, particularly older ones, contain forms of knowledge, prose structure, and domain expertise that are not easily replicated by scraping the open web. Rare and antiquarian books may hold information that exists nowhere else in digitized form, which is precisely what makes them attractive as training material and what makes their potential destruction a legitimate concern for historians, librarians, and archivists.
Anthropic’s assurances about rare books apply only to Anthropic. It is unlikely to be the only company using destructive scanning, and other companies may not apply the same restrictions on what they process.
For used booksellers, the ethical dimension is real but complicated. Many store owners may have reservations about where the books are going. The cash infusion, however, is difficult to refuse for an industry that has spent years contracting. The sector has seen a modest recovery over the past three years, but many bookstores have still had to restructure or close permanently.
The situation places independent booksellers in an uncomfortable position: they are becoming, without full knowledge or consent, suppliers in an industrial process they cannot fully see or verify.
In Short
AI companies are purchasing used books in bulk, at scale, as a legally defensible and cost-effective way to acquire training data following court rulings that permit training on copyrighted works when the physical copies are paid for. The process, known as destructive scanning, physically destroys the books after digitization. The practice raises legitimate concerns about rare and irreplaceable volumes, the transparency of the supply chain, and what it means when an entire category of human knowledge is being industrially processed and then discarded. For used bookstores, the demand is a financial lifeline. For the broader cultural record, the implications are worth watching carefully.
Based on reporting from Fast Company - Tech.