Why AI Companies Are Shredding Millions of Books to Feed Their Models

The Paper That Vanished

A coalition of authors, archivists, librarians, and independent book dealers has mounted what may be the most concrete grievance yet leveled against the artificial intelligence industry. Their complaint is not framed in abstract questions about alignment or synthetic voices — it is documented in paper, adhesive, and ink that once existed and no longer does.

Court filings made public this year unsealed details of a program Anthropic internally called Project Panama. The operation involved purchasing millions of physical books through secondary-market channels, contracting specialized firms to strip their bindings, scanning every page at machine speed, and then discarding the originals. The same operational pattern has since been corroborated at multiple other AI laboratories. The books were never digitized and preserved for posterity. They were consumed as raw material.

The Trail of a Single Book

The full scale of the operation came into sharper focus through investigative reporting by 404 Media, a technology journalism outlet that has focused extensively on the infrastructure behind AI development. Reporters asked a rare-book dealer in California to supply nearly a thousand volumes under conditions of anonymity. Inside one copy, they concealed an Apple AirTag.

The device tracked the package from California to a sorting facility in Milwaukee, onward to Colorado Springs, and ultimately to an Amazon warehouse operating under the internal designation VGT3. The building’s exterior bears a logo depicting a T. rex lunging toward an open book — an image that, perhaps unintentionally, captures the dynamic at the center of this controversy. Workers interviewed by reporters described the job in blunt terms: scan books all day, nothing else. There is no cataloging, no condition reporting, no preservation protocol, and no archival intent. The workflow is ingestion only.

It Was Never Just Bestsellers

Initial reports about the surge in physical book purchases documented before 2022 concentrated on commercial titles — popular fiction, nonfiction bestsellers, and widely available reprint editions. That trend alone was enough to alarm the book trade. Dealers noted anomalous buying patterns that did not correspond to any seasonal or cultural cycle.

But the 404 Media investigation, which incorporated the physical tracking data from the AirTag sample, demonstrated that the procurement extended far beyond mainstream inventory. The books purchased included rare editions, out-of-print scholarly works, local histories with tiny print runs, and volumes for which no surviving copy exists outside the physical original. When such books are bound, guillotined, scanned, and shredded, the loss is permanent and non-recoverable. A scanning facility is not a repository. There is no mirror copy of the physical world waiting if the paper is destroyed.

Why Scanning Destroys Rather Than Preserves

The distinction between digitization and destruction matters because the standard defense of large-scale scanning rests on an assumption that the physical artifact is thereby saved, merely relocated to a different medium. That assumption fails when the workflow is engineered for throughput rather than stewardship.

Cutting a book’s spine allows pages to be fed through high-speed planetary scanners far more quickly than binding-by-binding feeding. For an organization processing millions of volumes, that speed differential is not marginal — it is the difference between operating cost and financial collapse. The economic logic is transparent. The trade-off is that the object ceases to function as a book. It becomes refuse.

The irony cuts deeply: the same companies and technologies now training models on centuries of cultural output are actively dismantling the primary sources those models depend upon. The hunger for training data is so total that preserving the source material is treated as unnecessary overhead — a line item that can be eliminated without consequence.

Who Owns What Is Scanned

The legal landscape remains unsettled. Copyright law in the United States and in many other jurisdictions has not definitively resolved whether the large-scale scanning of legally purchased books for the purpose of training language models qualifies as fair use. Anthropic and several other firms are currently defending against lawsuits filed by authors who contend that this practice constitutes a systemic violation of their exclusive rights. Those cases are still working through the courts.

What is already clear, regardless of the legal outcome, is the material consequence. Even if a court ultimately rules that the scanning activity is lawful, the destruction of the originals means there is no path back to the source. A digitized page is not equivalent to a physical book possessing provenance, marginal annotations, historical binding details, watermarks, or paper composition data. Every one of those contextual layers is obliterated in the cut.

Second-Order Effects on the Book Trade

The ripple effects of this demand are already distorting secondary markets. Prices for older printings, out-of-title inventory, and modestly rare volumes have climbed sharply in regions where dealers report AI-affiliated purchasing activity. Libraries, small academic institutions, and individual collectors with limited budgets are being priced out of collections that were once accessible.

Dealers face a new ethical uncertainty: they can no longer guarantee a buyer that a volume will re-enter circulation. A purchase may vanish into a scanning pipeline. This opacity itself inflicts damage on an industry that has always depended on the integrity of provenance and the trust between seller and buyer.

Some dealers have begun refusing sales to certain brokers or middlemen without asking questions — a practice that would have been unusual and potentially rude a few years ago. Others have started asking buyers directly about their intended use, a conversation that did not exist in this form before 2022.

What Happens Next

The immediate result of the investigations to date has been public anger and journalistic exposure, not policy reform. There is no federal requirement compelling AI companies to preserve the physical copies they acquire for model training. There is no public registry documenting which volumes enter the pipeline and which are destroyed. No transparency mandate exists.

For archivists and librarians, the issue is increasingly framed as a supply-chain crisis. The models being constructed today rely on a physical input stream that is being consumed without replenishment. If current procurement rates hold, the available corpus of pre-2022 paper publications will steadily diminish even as institutional reliance on those texts intensifies.

The authors and advocates organizing the criticism are not calling for a halt to technological development. They are asking that the cost be visible — that the chain from purchase to destruction be traceable, that the physical artifacts being consumed be held to the same preservation standards applied to any cultural material entering a digital database, and that the public and the scholarly community retain access to what was taken.

Right now, none of those safeguards exist. The books are still disappearing. And no one is required to account for them.