Two Ledgers of Lost Books

SunnyJ Shores
4 days ago
8 min read

On the first of August this year, two jet-powered drones hit the same building in Kharkiv twice. The building was a logistics centre belonging to RNK-Ranok, one of Ukraine's larger educational publishers. The fire covered roughly ten thousand square metres. Nobody was killed, which was luck rather than design; one employee was unreachable for several hours before turning up safe.

What burned was eight million books. Twenty-eight thousand separate titles. Around six hundred thousand of them were school textbooks, printed and waiting for the September term.

Ukraine prints something like twenty-five million books in an average year. One night took close to a third of that.

"Two jet-powered drones struck the same location," the company's chief executive, Victor Kruglov, said afterwards. "This means it was a deliberately planned attack by those bastards to destroy textbooks and Ukrainian books."

The first ledger

It would be easier to file the Ranok strike under the general awfulness of war, a warehouse is a warehouse, and warehouses get hit. But the Ukrainian publishing industry has been keeping a list, and the list is hard to read as accident.

March 2022: the Unisoft printing house in Kharkiv, a facility that serves more than two hundred publishers, its facade blown in. April 2022: Ranok's logistics centre, damaged.

March 2024: two more Kharkiv printers, Hurov & K and the Avrora typography works, both burned to the ground, five civilians were killed in that strike as well.

Then 23 May 2024, and the one most people outside Ukraine have heard of, Factor-Druk, a printing plant in Kharkiv. Seven workers were killed and more than twenty injured. Fifty thousand books were destroyed. They were people at their machines in the middle of a shift, printing for Vivat, KSD and Ranok.

March 2025: the Triada-Pak printing house, plus at least three bookshops. June 2025: the Ukrainskyi Prioritet publishing house, warehouse and office gone.

Then this summer, in the space of eighteen days: BookChef's warehouse and the Mega-Polihraf press, eight hundred thousand books. Nash Format's warehouse. Ranok's Kharkiv warehouse. Mamyne Sonechko, a children's publisher, its office and its stock of picture books. The knyholav warehouse and the Konvi press, half a million textbooks and blanks, about seven and a half million dollars of damage across fifteen or more publishers.

One and a half million books in July alone. Across the war, at least thirty-five Ukrainian publishing houses have taken losses.

Five years. The pattern is the argument.

The second ledger

Now set that down and pick up a different one.

Some time in early 2024, an American company began an internal programme it called

Project Panama. The company was Anthropic, an artificial-intelligence firm. The point of the programme was to acquire physical books in bulk, cut the spines off with hydraulic machinery, run the loose pages through high-speed scanners, and send the paper to be recycled. The text went into the training corpus for a language model. The object went into the pulp.

We know the details because the documents were unsealed in litigation and reported at the start of this year. One internal planning document is unusually direct about the scope:

"Project Panama is our effort to destructively scan all the books in the world."

And, a little further on, about the mood:

"We don't want it to be known that we are working on this."

The company hired Tom Turvey, formerly head of partnerships at Google Books, in February 2024. Vendor proposals discussed converting between five hundred thousand and two million books over a single six-month contract. Tens of millions of dollars were committed.

Anthropic was not alone. Booksellers in the United States, the Netherlands, Germany and Switzerland spent much of the past year puzzling over sudden anonymous bulk orders, a thousand copies here, a hundred thousand there, selected apparently at random and specified by ISBN. Pieter de Vries, an antiquarian dealer in Haarlem, was asked for more than three thousand academic titles from 2020 and 2021, Elsevier, Wiley, Routledge,

Oxford University Press, to be shipped to China by a firm called 2077AI. "I was shocked," he said. He assumed it was a phishing attempt. Several of his Dutch colleagues had received the same request and binned it for the same reason.

A book-data company called ISBNdb advertised itself as a broker for exactly these purchases, in lots of a thousand to a million, with the buyer's identity protected. Its own explanation of why anonymity was part of the service is admirably plain:

"The optics problem is real. 'AI company destroys two million books' is not a headline that generates sympathy."

Those pages have since been taken down.

In August, reporters at 404 Media settled the question the simple way. With a rare-book dealer's cooperation, they hid an AirTag in a shipment of nearly a thousand books and watched where it went. It went to a warehouse belonging to Amazon, known to staff there as VGT3, where bindings are cut off and pages are scanned. Amazon's response was that it "purchases books through commercial channels to help develop and improve the products and services our customers use." It declined to say whether the purpose was AI training.

All of this is lawful, and it is worth being precise about why. American copyright gives an author no right against the destruction of a copy somebody else has bought, that protection exists for visual art and not for books. And in June 2025, Judge William Alsup found that training a model on lawfully purchased books was "quintessentially transformative" fair use, and separately that converting those purchased books from paper to digital was permissible as a format change: one copy in, one copy out, nothing redistributed, the original destroyed.

The part he would not protect was the piracy that came first, more than seven million books pulled from shadow libraries before anyone thought to start paying. That is what produced the $1.5 billion settlement, roughly $3,000 a work across the 482,460 titles on the certified list. Alsup gave it preliminary approval; Judge Araceli Martínez-Olguín granted final approval on 20 July this year.

Read that sequence carefully, because it is the whole story in miniature. The theft was punished. The shredding was blessed.

What the two ledgers do and do not share

These are not the same thing, and I am not going to pretend they are.

One of them killed seven people in a printing house in Kharkiv on an ordinary Thursday. The other is a procurement process. Nobody at a scanning warehouse is trying to erase a nation's language, and no one dies. A blog post that flattens that distinction isn't worth writing.

But there is a shape common to both, and it is worth naming. In each case the book has stopped being a thing with standing. In Kharkiv it is a target, a legitimate one, in the attacker's reckoning, precisely because it carries a culture. In the warehouse it is feedstock an inefficient container that happens to be wrapped around the content someone actually wants. Both positions agree on something the last five centuries mostly didn't: that the physical object is in the way.

The consequences differ in kind. They rhyme in direction.

And here is the part that unsettles me most about the second ledger. What makes a book worth buying for this purpose is that its text is not already online, so the buying skews towards the old, the out-of-print, the less common. Snopes, which looked at these reports last month, was careful on this point, and I will be too: the destructive scanning at one company is confirmed, but how far the practice extends, and how genuinely scarce the books being pulped are, is not established. Some firms named as intermediaries have denied any involvement.

What is not in doubt is the direction of travel. The scarcer a title is, the more useful it is as training data, and the less recoverable it is once it has gone through the cutter. A scan is not a book. It is a photograph of one. Ask any archivist what happens to a digital format after forty years of drift, then ask what happens to a paper codex left alone in a dry room for four hundred.

Don't blame the machine

It would be easy, and it would be wrong to end, this by pointing at the artificial intelligence.

The model has no view on any of this. It cannot want a book, or resent one, or decide that a first edition is an inconvenient wrapper. A hydraulic cutter is not wicked. A scanner is not wicked. Neither is the thing they are feeding.

What is worth judging is the sequence of ordinary human decisions: to buy in bulk through an intermediary so nobody could trace it, to write down we don't want it to be known that we are working on this, to conclude that a rare book's highest use is as a few megabytes of clean pre-2022 prose. Each of those was made by a person in a meeting who could have said something else. The secrecy is the tell. You do not hide a thing you are comfortable with.

Dark ages do not announce themselves. Nobody at Alexandria wrote today we begin the dark age. What actually happens is more boring: the people who hold the books stop believing the books are owed anything, and the loss looks, from the inside, like logistics. A purchase order. A recycling contract. A quarterly target met.

The drones are one way. The scanners are another. Both ledgers are still open, and both are still being added to.

Sources

The video this started from

Forbes, "AI Companies Are Buying—And Destroying—Antique Books. Here's Why." - video, 17 August 2026.
Mary Roeloffs, "AI Companies Are Buying—And Destroying—Antique Books. Here's Why," Forbes, 17 August 2026.

Project Panama, Amazon and the book trade

"Anthropic 'destructively' scanned millions of books to build Claude," The Washington Post, 27 January 2026 (subscription).
Emanuel Maiberg, "We Tracked a Shipment of Rare Books. It Ended at an Amazon AI Training Facility," 404 Media, 17 August 2026.
Fortune, on Pieter de Vries of Haarlem and the 2077AI order, 31 July 2026.
Futurism, on anonymous bulk orders and ISBNdb.
Snopes, "Are AI companies scanning and destroying millions of books, including rare titles?" - the fact-check that confirms the practice at one company while hedging on its scope.

The law

Eileen McDermott, "Judge Calls Anthropic's Training of LLMs with Authors' Works 'Quintessentially Transformative' But Gives No Pass on Piracy," IPWatchdog, 24 June 2025.
Kluwer Copyright Blog, "The Bartz v. Anthropic Settlement: Understanding America's Largest Copyright Settlement."
Kirsten Korosec, "Anthropic's landmark $1.5B copyright settlement is approved," TechCrunch, 20 July 2026.

Ukraine

"Russian strikes destroy 1.5 million Ukrainian books and key printing hubs," NV - the strike-by-strike chronology used for the timeline above.
"A single Russian attack destroyed one-third of Ukraine's annual book print volume," Chytomo - the RNK-Ranok logistics centre, 1 August 2026, and the Kruglov quote.
"Around 800,000 of Ukrainian publisher's books destroyed in overnight Russian attack," Kyiv Independent, 2 July 2026 - BookChef and Mega-Polihraf.
"Russian attack destroys 50,000 books at Kharkiv printing house, kills 7 workers," Euromaidan Press, 24 May 2024 - Factor-Druk.
Joanna Kakissis, Polina Lytvynova and Laurel Chor, "Ukraine's bookstores are thriving, but a top printing plant took a deadly hit," NPR, 18 July 2024.

A note on sourcing: every figure above was checked against at least one primary report. Two items remain open. The date of the strike on Nash Format's warehouse is given as 4 July 2026 by NV, while other outlets describe a comparable strike on 3–4 July 2025; it is possible the warehouse was hit in consecutive years. And Snopes, reviewing the book-scanning reports, confirms the destructive scanning at one company but notes that the scale of the practice across the industry is not established, and that some firms named as intermediaries have denied involvement.