AI companies destroying rare physical books for LLM training

The physical process is deliberately destructive: books are run through hydraulic cutting machines that slice off the spine, then fed through high-speed scanners that digitize every page before the paper is recycled, leaving no physical copy behind ^[1]. That speed and cost advantage - a guillotined book feeds through production scanners far faster than a bound one - is part of what makes buying and destroying books at industrial scale economically attractive to AI labs ^[1]. Anthropic's own vendor proposals reportedly discussed converting between 500,000 and 2 million books over a six-month window ^[2].

What makes this legally viable in the US is a narrow ruling: Judge William Alsup found that destructively scanning a legally purchased physical book into a single digital copy - one copy in, one copy out - is transformative fair use under the first-sale doctrine ^[3]. That ruling only covers books AI companies bought outright; it is explicitly separate from Anthropic's earlier use of pirated books, which was not protected and led to a $1.5 billion settlement, the largest copyright settlement in US history ^[4]. The upshot is a strange incentive structure: destroying the original isn't a side effect of scanning, it's the condition that keeps the practice legal, since scanning without destroying would multiply copies rather than merely change format.

Internal Anthropic planning documents, code-named 'Project Panama,' describe the initiative bluntly as an effort 'to destructively scan all the books in the world' ^[1][2]. The project traces back to 2021, when co-founder Ben Mann personally downloaded millions of pirated books from LibGen and Books3 for early training data ^[1]; Project Panama emerged later as the sanctioned, court-defensible successor to that approach - buying books rather than pirating them, then destroying the evidence of duplication.

The same internal documents show Anthropic anticipated backlash and instructed staff to keep the effort secret: 'we don't want it to be known that we are working on this' ^[1][5]. Anthropic hired the former head of partnerships at Google Books to help run the acquisition effort and spent tens of millions of dollars sourcing books before any of this became public ^[2]. The secrecy wasn't limited to Anthropic - ISBNdb's own internal marketing materials, pitching the bulk-sourcing service to AI labs, conceded the identical reputational risk in writing: 'The optics problem is real. "AI company destroys two million books" is not a headline that generates sympathy' ^[2].

Between AI labs and the shelves of rare-book dealers sits a layer of data brokers and bulk buyers whose role only became visible through reporting, not disclosure. ISBNdb, a company that maintains ISBN metadata, publicly advertised a tiered bulk-sourcing service - offering anywhere from 1,000 to 1 million books, pitched specifically at AI/LLM training customers ^[2][6]. After the pitch was reported, ISBNdb deleted the pages within nine days and denied ever operating the business, claiming 'the service was never launched' and that the pages reflected 'an exploratory concept rather than an active business' ^[3][6].

Zoom Books, a Canadian company, has been repeatedly named by booksellers in Germany, Switzerland, and elsewhere in Europe as the source of unusually large bulk orders, and has also been referenced in Australian reporting on the wider trend, though it frames its purchases as ordinary recycling and trading activity ^[6][7]. A separate Dutch antiquarian dealer who received an order for nearly 3,000 titles from a Singapore-based firm initially assumed it was a scam, since his inventory of expensive, exclusive rare books doesn't typically move in bulk ^[2][6]. Ordinary booksellers are left guessing who is actually buying and why, with brokers at both ends of the chain denying any AI connection the moment it draws scrutiny.

The framing of what's being lost is contested. Public alarm has centered on the idea that AI companies are pulping unique cultural artifacts - rare volumes whose unique physical qualities are gone forever once destroyed, even though a digital scan survives ^[7][8]. Booksellers and organizations like ANZAAB describe genuine unease even when the money is good, and Australian dealers report awareness of similar waves of orders already reported across the US and Europe ^[7].

A more skeptical reading argues that the most visible examples of destruction target mass-market or outdated academic overstock that dealers would have pulped or recycled anyway, not truly one-of-a-kind volumes - making some of the alarm closer to overstatement than an accurate accounting of what's actually disappearing. Both readings can be true simultaneously: some of what's destroyed is disposable stock, and some genuinely is scarce, and the opacity of the supply chain - ISBNdb and Zoom Books both denying AI ties the moment they're named - means neither booksellers nor the public can easily tell which is which before a given book is already cut.

Even setting aside the fate of the physical object, there's a second-order question about where the digitized text goes. A recurring critique in public discussion, echoed in community debate over the reporting, holds that the scans produced by efforts like Project Panama disappear into private, proprietary AI training corpora with no obligation to ever surface publicly - unlike a physical copy sold, donated, or otherwise dispersed elsewhere, which could in theory still resurface later. On this view, a physical copy destroyed to feed a closed commercial model can leave society with strictly less access than before: no shelf copy, no public record, only a training set locked inside a company's infrastructure.

Framed against the 2025 fair-use ruling that made destructive-but-legal scanning possible, and the $1.5 billion settlement that punished the pirated-book shortcut Anthropic took to get into books in the first place ^[3][4], the throughline is that legality and public benefit have been decoupled here: the practice is lawful precisely because it doesn't multiply copies, not because it preserves anything for anyone outside the companies doing the scanning.