AI giants reportedly outsourcing to middlemen to secretly buy and destroy millions of books for training data

To sidestep mounting copyright lawsuits and a severe shortage of clean digital data, major AI companies are secretly purchasing millions of physical secondhand books through anonymous third-party middlemen.

Using industrial hydraulic cutters and high-speed scanners, firms undergo “destructive scanning”—slicing book spines and digitizing the pages before shredding the physical remains to train next-generation large language models (LLMs).

The Scramble for “Clean Data” and the Internet Shortage

As AI models get bigger, tech giants are hitting a major wall: the open internet is running out of high-quality, human-written text.

The problem is model collapse. When companies keep training large language models on web-scraped data, more and more of it is AI-generated “slop.” Feeding synthetic text back into new models causes them to degrade, hurting both quality and logic over time.

That’s why pre-2022 books are becoming the new gold standard. Physical books published before the generative AI boom are uncorrupted, expert-edited, and long-form — exactly the kind of clean human data needed to train more sophisticated AI systems.

Rather than licensing digital catalogs directly from major publishers or buying directly under their own corporate names, tech firms rely on covert supply networks and third-party logistics firms to buy physical literature en masse.

Key mechanics behind the operation include:

Anonymous Middlemen: Tech companies hire third-party vendors and specialized book buyers (such as Canadian buyer Zoom Books) to set up bulk purchasing accounts across used bookstores, library sales, and online platforms like Alibris and Biblio.

“Project Panama”: Unsealed court documents from legal filings against AI developer Anthropic revealed covert initiatives like “Project Panama,” where tens of millions of dollars were allocated to acquire over two million physical books.

Hydraulic Slicing and Shredding: Once shipped to temporary warehouses, industrial hydraulic cutting machines slice off the book bindings. The loose pages are fed through production-level, high-speed scanners to extract digital tokens before the paper remains are sent for pulping and recycling.

The “First-Sale” Legal Loophole and Copyright Battles

The push toward physical book destruction stems directly from recent legal rulings and evolving copyright dynamics in the United States.

AI companies are leaning on an old copyright rule to get around legal risk: the first-sale doctrine. Under U.S. copyright law, this lets you resell, give away, or physically alter a book you legally bought.

Tech firms argue that if they legally buy a physical book and then digitize it for training, it counts as “transformative use” and is protected under Fair Use.

It’s also a way to avoid piracy lawsuits. Scraping digital “shadow libraries” like LibGen or Books3 has already led to multi-billion-dollar class-action suits from authors and publishers. By contrast, destructively scanning physical copies they own gives AI developers a more legally defensible path to building training datasets.

Cultural Concerns and Long-Term Feasibility

Despite providing a temporary influx of pristine human text, industry analysts point out that destructive scanning is an unsustainable, short-term fix.

Irreversible Cultural Vandalism: Antiquarian booksellers and archivists worry that rare, out-of-print, or foreign-language titles—some among the last surviving physical copies—are permanently lost when pulped for data extraction.

Finite Supply: With roughly 130 million unique books ever published historically yielding around 30 to 40 trillion text tokens, the physical literary canon alone cannot indefinitely sustain the exponentially growing data demands of future frontier AI models.