AI’s Need for Content Has Put Rare-Book Dealers in a Bind

James Payne was stumped. The Brooklyn, N.Y.-based rare-book seller saw orders surge earlier this year for titles that even he considered to be obscure.

By unloading volumes like “Home Long-Term Oxygen Treatment in Italy” and “Marriage Litigation in Medieval England” on third-party marketplaces, he has roughly doubled his revenue from this point in 2025.

Payne had no idea who was buying these books through Alibris, Biblio and Amazon.com—and he wasn’t alone. Sellers of used books worldwide have been swapping stories of curious sales patterns for months. The buyers went by unusual names including Red Sparrow Project, Blue Finch Project and Green Parrot Project. They ordered dozens, even hundreds of books at a time, on completely unrelated subjects, and routed them through logistics warehouses. And they didn’t haggle over prices.

At first, dealers had practical concerns: How does a small shop that has never packaged an order bigger than 20 books even ship a batch of 200? And would these brand new buyers pay their bills?

Beneath those questions lurked a growing suspicion that the orders had something to do with artificial-intelligence companies’ insatiable need for new content to train their large language models. As evidence mounted that many of the mystery orders were landing in facilities where books are sliced up, scanned and destroyed, booksellers were forced to consider their participation in an unusually on-the-nose example of digital technology devouring literary culture.

“Let’s quote Dickens,” said Allan Stypeck, whose Washington, D.C.-area Second Story Books started receiving bulk online orders this spring. “ ‘It was the best of times, it was the worst of times.’ ”

Second Story’s sales volume has skyrocketed. The shop, with 75,000 books online and more in store, is in talks to sell directly to one of the bulk buyers. But Stypeck is uneasy about the bigger implications, including whether he may be pulling from circulation the last copy of a truly scarce title.

“These books probably have a purpose that will be totally contradictory to the longevity of my business,” he said.

Books have long been used to train large language models, the technology underlying systems such as OpenAI’s ChatGPT, Anthropic’s Claude and Google’s Gemini. Anthropic used pirated digital copies of books and also bought physical copies of millions more in bulk, stripped them from their bindings to scan the contents, then destroyed the physical copies, according to court filings in a copyright infringement lawsuit a group of authors filed against the AI company in 2024. The Washington Post reported in detail on the process behind what was known as Project Panama earlier this year. Anthropic agreed to pay $1.5 billion to settle the portion of the case connected to the pirated works.

A federal judge ruled last year that the material in the books Anthropic bought was fair use under copyright law in part because the company’s digital versions had only replaced the destroyed print ones, and didn’t introduce more copies into the market.

As AI models grow more sophisticated, the need for additional authoritative data on which to train them has become more urgent—even as the supply of undigested, commercially available writing dwindles.

The International Standard Book Number, which provides a unique identifier for each edition of books published since the late 1960s, gives AI firms and vendors that specialize in supplying training data a way of tracking which books they have ingested. The mystery buyers’ strong preference for books with an ISBN helped tip off sellers that the bulk buys were tied to the care and feeding of large language models.

A seller based in the Pacific Northwest took a virtual cross-country journey with her books after noticing more bulk orders earlier this year. She told The Wall Street Journal that she fixed credit card-sized location trackers inside the back covers behind the dust jacket flaps of books in two separate shipments in late June and early July.

One went to an Ohio warehouse for the marketplace Alibris before heading to an industrial complex in Addison, Ill., whose address matched a location for a company with a large book-scanning business. While some books are scanned for various purposes without losing their spines, so-called destructive scanning is considered more efficient.

“I think that’s where it died,” the seller said.

The other book went first to a warehouse in Kenosha, Wis., then to a spot in northern Las Vegas at an intersection with at least two Amazon facilities. It eventually moved to a position in Mexicali, Mexico. While the seller couldn’t determine the exact address, a Google map of the location shows a sprawling recycled-paper manufacturing facility.

404 Media, a tech-industry news site, published a story earlier this month saying it had tracked a shipment from a different seller around the country, ultimately to an Amazon facility in Las Vegas known as VGT3. The symbol for that facility, 404 Media reported, was a sharp-toothed green dinosaur holding an open book.

In a Reddit forum for Amazon employees, a post from about a year ago described the work done at VGT3. “All we do is scan books,” the post read. “Some are assigned to cut books, and others go to receive where they get books and scan the bar codes.”

An Amazon spokesman said the company purchases books through commercial channels “to help develop and improve the products and services our customers use.”

As outrage built this summer over the potential destruction of irreplaceable volumes, Elon Musk said in an X post that he had asked the SpaceX AI team “to preserve any rare books in a library and scan them the hard way vs just cutting off the spine and scanning.” SpaceX didn’t respond to a request for additional comment.

An Anthropic spokesperson said sourcing books is a widely used approach for AI training, and “none of our data acquisition programs buy and destroy rare or antiquarian books.”

Mike Floreani, a Texas divorce lawyer who runs online-only Palos Verdes Books and Ephemera, recently completed large orders that included a textbook about the Russian economy’s transition after the fall of the Soviet Union, a Victorian novel, a 2012 Texas civil procedure guide and a book about a Swedish comedy team popular in the 1960s.

Floreani calls these books dead stock—titles that sat on his shelves, collecting dust, for years. He had considered bringing some to the recycling center.

“Cultural artifacts?” he said. “I can categorically tell you that they’re not.”

Floreani hit sales records each of the past three months, thanks to bulk orders. As long as the AI companies are recycling the books after scanning them, Floreani said he doesn’t have much problem with their activities—provided they aren’t violating intellectual property rules.

Scott Brown also fulfilled some orders for what he believed at the time were AI-related customers. But the proprietor of Downtown Brown Books in Portland, Ore., drew the line at an inscribed copy of Ray Bradbury’s “Futuria Fantasia,” a magazine the iconic writer produced as a teenager that was published as a collection nearly 20 years ago. The buyer’s shipping information included the word “SCAN” in the name.

Brown, who reflected on the situation in a recent Substack post, was willing to sell the book. After all, it was listed on a marketplace for $90. But he knew the late owner personally, a customer named Tom, and “I know Tom would not want that book sent to the guillotine,” Brown told The Journal.

He ultimately asked another seller to fill the order with an unsigned copy. “If we’re going to save the signed copies, we have to sacrifice the unsigned,” Brown said. “I didn’t realize that I saw my job as selling to humans until robots started buying books from me.”