Amazon Is Destroying Rare Books to Train AI and Why LLMs Need Them

TL;DR

• Viral claims allege Amazon is purchasing and destructively scanning rare physical books to create exclusive AI training data, but as of August 2026 there is no verified evidence of a systematic destruction program, and Amazon has not confirmed such a practice.
• AI companies are increasingly desperate for rare, out-of-print, and high-quality physical texts because frontier LLMs have largely exhausted high-quality public internet data, leading to a hunt for untouched, human-curated language.
• The controversy highlights a growing collision between AI development and cultural preservation, raising serious questions about copyright, the irreversible loss of physical artifacts, and who controls access to human knowledge.

The Allegation That Set the Internet on Fire

In recent weeks, a disturbing claim has spread across X, Reddit, and tech newsletters: that Amazon is quietly buying up rare and antiquarian books — first editions, out-of-print academic texts, and fragile archival volumes — only to have them guillotined, scanned, and destroyed to feed its large language models.

The posts, which often feature videos of industrial book scanners slicing spines off books, allege that physical copies are being sacrificed for a one-time digitization. The implication is that irreplaceable cultural artifacts are being permanently erased to create a proprietary dataset that no competitor can replicate. The outrage was immediate, with authors, librarians, and archivists accusing the tech giant of cultural vandalism.

So far, however, no major investigation or whistleblower report has substantiated that Amazon is running a large-scale, destructive scanning operation specifically targeting rare books. The videos circulating online appear to be generic examples of destructive scanning — a real, long-standing practice used by mass digitization services — repurposed to illustrate the claim. Amazon has not announced such a program, and its known book digitization efforts have historically focused on publisher partnerships and non-destructive scanning for its retail and Kindle platforms.

Why the Rumor Felt So Believable

Even without confirmation, the story gained traction because it fits into a very real and well-documented crisis in AI development: we are running out of data.

Hitting the Data Wall

For the last three years, leading AI labs including OpenAI, Google DeepMind, Anthropic, and Meta have warned about the "data wall." Models like GPT-4, Claude, and Gemini were trained on essentially the entire high-quality public internet — Wikipedia, news archives, GitHub, Reddit, and trillions of tokens of web crawl data.

Research estimates published in 2024 and 2025 suggested that the stock of high-quality, human-written public text would be fully utilized between 2026 and 2028. Once that happens, scaling laws break down. Training on more low-quality, repetitive, or AI-generated content doesn't make models smarter — it can actually make them worse, leading to a phenomenon researchers call model collapse.

That has forced AI companies to look elsewhere for fresh, untouched, high-signal data.

Why a Dusty Old Book Is Worth More Than a Million Tweets

Rare and physical books represent the perfect solution to the data problem, and that is why the Amazon rumor resonated.

First, they are pristine. A 19th-century history text, a 1970s out-of-print philosophy monograph, or a limited-run collection of poetry has never been posted online. It is not contaminated by SEO spam, bot comments, or AI-generated filler. It is long-form, edited, fact-checked, and linguistically rich.

Second, they are legally attractive. Many rare texts are in the public domain or out of print, making them far less risky than scraping modern copyrighted works, which has triggered massive lawsuits against OpenAI, Anthropic, and others from publishers and The New York Times.

Third, they contain rare knowledge. LLMs are already fluent in general internet knowledge. What they lack is depth — the nuanced arguments, obscure facts, and specialized vocabularies found in academic, technical, and literary works that were never digitized. Training on this "long-tail" knowledge is how companies hope to make the next generation of models genuinely more capable, not just more fluent.

For a company like Amazon, which owns AbeBooks, has the world's largest logistics network for used books, and operates Amazon Web Services and its own Titan and Olympus AI models, the idea of turning physical inventory into an exclusive data moat seems strategically logical, even if unproven.

Destructive Scanning: A Real Practice With Real Consequences

The most alarming part of the claim is the word "destroying," but destructive scanning is not science fiction. It is a common, if controversial, industry practice.

To digitize a book quickly and cheaply at scale, high-speed scanners often require the spine to be cut off so pages can be fed through an automatic feeder like a photocopier. This produces a perfect, flat scan in minutes, compared to hours of careful, non-destructive, page-by-page photography required for preservation. Companies that offer to digitize personal libraries often give customers the choice: pay more for non-destructive scanning, or pay less and get your book back as a stack of loose paper or not at all.

For a mass-market paperback, the loss is trivial. For a rare first edition where only a few hundred copies exist, it is irreversible. Librarians and archivists argue that a physical book is more than just its text — its paper, binding, marginalia, printing errors, and provenance are historical data in themselves. Once destroyed, that artifact is gone forever, even if its words live on as tokens in a neural network.

The Bigger Fight: Preservation vs. Progress and Who Owns Knowledge

Whether or not Amazon is actually doing this, the debate has exposed three unresolved tensions that will define the next phase of AI.

1. Preservation and Access. If tech giants are incentivized to lock up the digitization of rare texts as proprietary training data, that knowledge may never enter the public digital commons. Unlike the Google Books project, which aimed to make scanned books searchable for everyone, a private AI dataset is a black box. The public loses the book and never gains a readable digital copy.

2. Copyright and Consent. Even for out-of-print books, copyright often still applies. Authors and estates are increasingly arguing that training an AI on their work without permission or compensation is infringement, even if the physical book was legally purchased. The legal doctrine of first sale allows you to buy and destroy a book you own, but it does not automatically grant you the right to exploit its contents commercially in a new medium.

3. The Future of Training Data. The hunt for rare books is a symptom of a larger shift. The era of free, infinite internet scraping is over. The next era will be defined by expensive, negotiated, and physical-world data acquisition — licensing deals with publishers, partnerships with libraries and universities, and yes, the buying up of private collections. How that data is acquired, and whether it is preserved or consumed, will determine whether AI development enriches our cultural record or quietly erodes it.

What Happens Next

For now, the Amazon claim remains an allegation without hard proof, but it has served as a warning shot. Archivists are now calling for greater transparency from AI labs about the provenance of their training data, and for legal guardrails that require non-destructive digitization for any work of cultural significance.

If the next leap in artificial intelligence truly depends on the knowledge trapped in our rarest books, the tech industry will have to decide whether it wants to be the steward of that knowledge — or the last one to ever hold it.

AndroGuider Team

Articles written by the AndroGuider team. We try to make them thorough and informational while being easy to read.