A quiet project with a telling name
Buried inside a 2025 copyright ruling was a detail that turned into one of the most talked-about AI stories of the year: Anthropic had spent tens of millions of dollars buying millions of used print books, then paid contractors to slice off the bindings, feed the loose pages through industrial scanners, and throw away what was left. Internally, the effort was called Project Panama, and according to unsealed planning documents, employees candidly wrote that the goal was to destructively scan every book they could find, adding that they didn’t want this work to become public.
That last line is the crux of the story. Anthropic knew the optics were bad, went ahead anyway, and got caught only because of an unrelated lawsuit.
What actually happened
According to court filings reviewed by the Washington Post, Anthropic hired Tom Turvey — who had previously run book-scanning partnerships for Google Books — in February 2024 and gave him a simple mandate: obtain “all the books in the world.” His team bought books in bulk, often used, from sellers like Better World Books and the UK’s World of Books, sometimes acquiring tens of thousands of titles at once. A vendor proposal described wanting the capacity to convert between 500,000 and 2 million books over a six-month period.
The physical process itself was straightforward and brutal for the books involved. Contractors used a hydraulic-powered cutting machine to neatly slice the bindings off each book, then ran the loose pages through high-speed, production-level scanners. Once a book had been digitized, the paper original was discarded — a recycling company was brought in to collect what remained.
This wasn’t a small pilot. The Washington Post’s review of more than 4,000 pages of unsealed filings found the company had spent tens of millions of dollars on the broader book-acquisition operation.
Why destroy the books at all?
The destruction wasn’t incidental — it was the point, legally speaking. Anthropic had also trained earlier models on pirated text libraries (Books3, Library Genesis, and the Pirate Library Mirror), which is the part of the case that actually got the company sued and ultimately cost it a $1.5 billion settlement with authors. The purchased-and-shredded books were a separate, later effort designed specifically to be defensible.
The legal logic, as laid out by U.S. District Judge William Alsup, comes down to the first-sale doctrine: if you buy a physical copy of a book, you’re allowed to do largely what you want with that one copy, including destroy it. By scanning a purchased book and then discarding the paper version, Anthropic ended up holding exactly one copy of the work — just in a different format — rather than doubling up on copies it didn’t pay for. Keeping the paper copy around alongside a new digital one would have looked a lot more like unauthorized copying. Judge Alsup ultimately ruled that this kind of purchase-scan-destroy approach counted as a transformative, and therefore legal, use of copyrighted material — the physical books were converted into a searchable, trainable digital format rather than redistributed or reproduced for others to read.
Executives reportedly saw well-edited, professionally published books as valuable training material in their own right, not just a source of legal cover. One co-founder is said to have argued that books would teach Claude to write with real craftsmanship, rather than absorbing the more casual, lower-quality style typical of internet text. Buying used books in bulk and scanning them destructively was also simply cheaper and faster than negotiating individual licensing deals with publishers — Anthropic’s own CEO reportedly opted to acquire pirated text early on specifically to avoid what he called the “legal/practice/business slog” of licensing, according to the judge’s order, before the company later shifted toward the more defensible purchase-and-destroy model.
Why it caused an uproar anyway
Even after the legal question was settled in Anthropic’s favor, the practice drew sharp public criticism once the details became widely known in mid-2026, following further unsealing of court documents. Part of the backlash was aesthetic and emotional: the image of rare or out-of-print books being run through a blade and pulped struck many people as needlessly destructive, especially given that Anthropic could have kept the physical books in storage rather than discarding them. Investor Michael Burry called the practice “evil incarnate” in a social media post, and Elon Musk said he’d asked his own AI team to preserve rare books by scanning them without cutting the spines.
Part of it was about candor. The internal documents showing that employees knew this would look bad if it got out, and tried to keep it quiet, undercut any argument that this was simply a routine digitization project done in the open.
And part of it is a broader worry about what happens next. Anthropic wasn’t alone — Meta faced similar accusations over how it sourced training text — and brokers are reportedly now marketing large-scale physical-book acquisition and destructive-scanning services to other AI companies. Because the “buy one copy, destroy it, keep the scan” approach is the version of AI training that has (so far) survived a fair-use legal challenge, some observers worry the incentives created by this ruling could push more of the industry toward the same practice, including toward rarer or harder-to-replace books, even though no confirmed case of an irreplaceable rare book being destroyed this way has yet surfaced.
Where things stand now
The purchased-book scanning was found lawful; it was the earlier use of pirated text libraries that actually triggered Anthropic’s $1.5 billion settlement with a class of authors, the largest copyright settlement of its kind in the U.S. to date. The two threads — one legal, one merely unpopular — often get collapsed together in public discussion, but they were legally distinct issues that happened to surface in the same case.
This account is based on court filings and reporting from the Washington Post, Ars Technica, Futurism, and other outlets following the unsealing of documents in the Anthropic authors’ lawsuit.
Sources
- Ars Technica, “Anthropic destroyed millions of print books to build its AI models”
- Futurism, “Anthropic Knew the Public Would Be Disgusted by How It Was Destroying Physical Books, Secret Documents Reveal”
- Novara Media, “AI Firms Are Buying up Old Books, Then Scanning and Destroying Them”
- Yahoo News, “Anthropic Destroyed Millions of Books to Train Claude: Was That Legal?”
- NewsNation, “AI companies, including Anthropic, accused of buying, ripping pages from books to train models”
- Yahoo Finance, “Is Anthropic Destroying Rare Books After Training AI Models On Them? Elon Musk, Michael Burry And Others React Amid Online Outrage”
- CryptoSlate, “AI firms are shredding physical books because copyright law is quietly rewarding them”
- MLQ News, “What the Evidence Actually Shows About AI Companies Destroying Books”
