Destructive scanning of rare books for AI training

Court filings unsealed in early 2026 revealed that Anthropic had been running a codenamed effort, 'Project Panama,' since early 2024 to acquire books at massive scale for AI training. The company hired Tom Turvey, the former head of partnerships for Google Books, to lead outreach to distributors and retailers, and internal vendor proposals discussed converting between 500,000 and two million books over a six-month period ^[1]. What makes the documents notable isn't just the scale, it's the candor: an internal planning document stated plainly, 'Project Panama is our effort to destructively scan all the books in the world,' followed by an explicit instruction to keep the effort quiet - 'We don't want it to be known that we are working on this' ^[2].

That internal admission sits awkwardly next to Anthropic's public posture. The company has told reporters it purchases books through regular commercial markets and that none of its data acquisition programs acquire or destroy rare or antiquarian books ^[3]. The gap between a leaked memo built around secrecy and a public statement built around reassurance is itself the story - not because the underlying legal purchase-and-scan model is illegal, but because the company that built it clearly anticipated the backlash it's now facing.

The legal foundation for all of this rests on a June 2025 ruling by Judge William Alsup in Bartz v. Anthropic, which split the company's book-acquisition practices into two very different buckets. Digitizing legally purchased print books and training on the resulting text was ruled transformative and protected under fair use. Retaining a library built from pirated ebooks, by contrast, was not, and that distinction cost Anthropic $1.5 billion in a settlement covering more than 400,000 registered works at roughly $3,000 apiece, the largest copyright settlement on record, finalized in July 2026 ^[4].

That legal architecture is what makes the physical destruction mechanically central rather than incidental. Hydraulic cutters slice the spine off a purchased copy so the loose pages can run through scanners at 80 to 120 pages per minute ^[5]. The book doesn't survive the process by design: once the original copy is gone, the company holds one lawfully purchased instance in digital form rather than an unauthorized duplicate sitting alongside a printed original. The unsettling part isn't merely that the books are destroyed, it's that destruction is closer to a compliance requirement than a byproduct - the mechanism by which digitization avoids looking like piracy in the eyes of a court.

Behind the scale that Project Panama required sits a supply chain nobody involved wants to fully own. ISBNdb, a book-metadata company, openly pitched bulk orders ranging from 1,000 to 1 million books as a 'compliant path' to LLM training data - explicitly free of AI-generated contamination since the books predate the AI boom - but later told Fortune the service was never actually launched and that it never purchased or scanned books for AI companies ^[3]. Zoom Books, a Canadian firm that calls itself a book recycler, has independently been flagged by secondhand booksellers across the Netherlands, Germany, Australia, and New Zealand as the source of unusually large, oddly timed, high-priced orders for obscure and local titles. Zoom Books denies digitizing books for AI purposes and refuses to name its customers ^[6].

Reporting has described this structure as AI companies outsourcing to middlemen specifically to buy up books for training material without directly identifying themselves to sellers ^[7]. Every party in the chain - the AI lab, the metadata broker, the recycler - has a story that ends with someone else responsible, which is precisely the dynamic the Project Panama secrecy memo anticipated: a program designed from the outset so its footprints wouldn't be traceable back to a single company's front door.

Cultural-heritage advocates argue the loss goes beyond text. First editions, handwritten marginalia, ownership inscriptions, and original bindings carry historical evidence that a scan can't capture, so when the physical copy is pulped, that evidence is gone regardless of whether the words survive digitally ^[8]. Compounding that, the digital copies that do survive typically enter private databases used exclusively for AI training rather than public archives, so a scanned book can be simultaneously 'preserved' and effectively unavailable to anyone outside the company that scanned it ^[9].

But the loudest online pushback challenges the premise itself: a vocal contingent of commenters on Reddit and elsewhere insists most of the destroyed volumes are mass-market overprints already headed for pulping and recycling anyway, not museum-grade rarities - arguing genuinely unique or valuable first printings would already be preserved in libraries or existing digital archives rather than routed through a bulk book recycler. Separately, some commentary has pushed back on the framing from the opposite direction, arguing the destruction narrative undersells a more mundane economic driver: existing digital scans of these books, where they exist at all, are often too low-quality for training, making a fresh, clean scan of a physical copy the more practical option regardless of how 'rare' any individual book is. Neither read fully resolves the tension, since nobody outside the AI labs and their brokers can verify exactly what's being cut apart. Even Musk's public response split the difference rather than rejecting destructive scanning outright: he directed his own team to preserve any genuinely rare books it encounters by scanning them the harder, non-destructive way instead of cutting off the spine ^[10]- implicitly conceding that most books in these bulk lots probably don't warrant that extra care, while insisting the ones that do shouldn't be sacrificed for scanning speed.