Amazon Begins Destroying Books for AI Training Amid Data Scarcity

As the public internet becomes increasingly difficult to provide high-quality text that is sufficiently scarce, reliable, and free from AI-generated content contamination, old books are transitioning from cultural artifacts to data raw materials for the AI era.

Article author, source: 0x9999in1, ME News



• This is not just a simple “Amazon burns books” headline. The more significant signal is that as the public internet becomes increasingly unable to provide enough scarce, reliable, and uncontaminated-by-AI-generated-content high-quality text, old books are transitioning from cultural artifacts to data raw materials for the AI era.
• A 404 Media-tracked order of approximately 1,000 books ultimately arrived at Amazon’s LAS8 warehouse in North Las Vegas; employees reported that the VGT3 team would cut off the spines and scan the pages, resulting in the destruction of the original physical copies. Amazon confirmed that it purchases books through commercial channels to improve its products and services, but has not disclosed the scale of the project, its start date, specific purposes, or whether rare or scarce editions are screened out before destruction.
• There is another practical context to “buy the physical book first, then perform a destructive scan”: In the 2025 Bartz v. Anthropic case, a U.S. federal district court recognized fair use for one-to-one digital replacement of legally purchased paper books, but clearly distinguished between legal purchase and pirated acquisition. By July 2026, Anthropic’s class-action copyright lawsuit involving pirated sources was finally approved with a $1.5 billion settlement. The source of data is shifting from a technical issue to a balance sheet issue.
• It’s no surprise that Amazon is entering the scene now. The company expects capital expenditures of approximately $200 billion in 2026, with AI as a key focus area; AWS revenue grew 37% year-over-year to $42.2 billion in the second quarter of this year. Computing power continues to expand rapidly, but the marginal value of high-quality data is also increasing.
• What truly warrants caution is not the act of “deconstructing books” itself, but a procurement and destruction system lacking transparency: ordinary used books can be reprinted and repurchased, but certain out-of-print editions, annotated versions, regional publications, and rare texts may be irrecoverable. Once the screening mechanism is gone, AI companies gain data, while society may lose irreplaceable physical records.

The most striking image of this event is, of course, brutal: a book has its spine cut off, its pages fed into a scanner, and the original book ceases to exist. Even more dramatic is that the company doing this is Amazon—the very same Amazon that started in 1995 by selling books online. More than thirty years later, it is once again buying books in bulk, but this time, the books may not end up in readers’ hands; instead, they are being turned into data that machines can consume.

According to an investigation published by 404 Media on August 17, journalist Emanuel Maiberg collaborated with a bookseller who received a large order via Biblio, placing an AirTag inside one of approximately 1,000 books. The tracker made its way to Amazon’s LAS8 warehouse in North Las Vegas. The report cites employees at the location who said a team called VGT3 is responsible for receiving large quantities of physical books, removing their bindings, and rapidly scanning them, with the original books being destroyed during the process. Amazon later confirmed to Inc. that the company does “purchase books through commercial channels to help develop and improve the products and services used by customers,” but did not answer how many were purchased, when the program began, which products the data feeds into, or whether rare or scarce editions are filtered out in advance.

More notably, this lead did not suddenly appear yesterday. Inc. found that as early as September 2024, Amazon seller forums reported numerous book orders being shipped to LAS8: some sellers encountered a single buyer placing 68 orders at once, others received 48 orders concentrated in the early hours, all of which were university press books, and some sellers said the same account was purchasing books nearly every day. At the time, an Amazon forum representative ultimately confirmed these orders were valid, but Amazon has not now clarified whether those orders are related to VGT3.

My assessment is that what truly matters here is not the emotional shock of “tech giants burning books,” but rather the revelation that the AI industry is undergoing a resource revaluation: for years, the scarcest resources were GPUs, data centers, electricity, and capital; now, high-quality, traceable human-generated text,尽可能 free from AI-generated content contamination, is also being re-priced.

The used book market being suddenly drawn into the AI supply chain is not a whimsical anomaly of a nostalgic era, but a natural and coldly inevitable step following the industrialization of large models.

The internet may seem endless, but data suitable for training cutting-edge models is not infinite.

In its 2024 research, Epoch AI estimated that, after accounting for quality and deduplication, the effective available stock of public human-generated text for training was approximately 300 trillion tokens; if the data growth trends at that time continued, this stock could be fully utilized between 2026 and 2032. This projection is not a countdown to “data running out tomorrow,” as synthetic data, data efficiency, post-training methods, and multimodal approaches continue to evolve—but it reveals a trend: relying solely on further web scraping will increasingly yield diminishing returns in terms of new information quality.

However, old books have several characteristics that AI companies cannot ignore.

First, they are often subjected to multiple rounds of screening by authors, editors, and publishers, resulting in greater stability in information density and structure compared to random web pages. Second, a large volume of local histories, specialized handbooks, niche academic works, older textbooks, industry materials, and non-English publications have never been fully digitized and are simply inaccessible via public web pages. More subtly, content published before the widespread adoption of generative AI is far more likely to have originated from a purely human writing environment.

This value has been directly stated by data providers. In July of this year, in a report on the book procurement market, the book database company ISBNdb once tested demand from AI companies by claiming that “the world’s best AI training data is sitting on bookshelves,” emphasizing that books contain organized, edited, and professional human knowledge. ISBNdb later stated that the page was merely a market demand test and has since been taken down.

Why does this sentence sting? Because it lays bare the entire industry's anxiety.

The more powerful generative AI becomes, the more text on the internet will be written, rewritten, or aggregated by AI. If models continue to ingest this content indiscriminately, the sources of data will grow increasingly complex. As a result, the things that were already printed on paper before the wave of generative AI have suddenly gained a new layer of value they didn’t have before: at least in terms of time, they are closer to an information environment untouched by large-scale generative AI.

So, buying used books on Amazon doesn’t mean it’s “returning to the print era”—on the contrary, it reflects a deeply data-driven mindset. For model trainers, a book’s cover, binding, and collection history may have no direct training value; what’s truly valuable are the tokens inside that can be tokenized, indexed, deduplicated, annotated, and trained.

Books have shifted from commodities to corpora, and their pricing systems have changed accordingly. A decades-old professional book that no one cared about before may suddenly generate new machine demand due to the absence of a complete digital version online.

This is the most noteworthy point.

The reason is first and foremost simple: efficiency.

Digitizing tens of thousands of books industrially by photographing each page individually is extremely slow. After removing the spine, the pages can be flattened and separated more effectively for high-speed scanning and subsequent optical character recognition. Destructive scanning is not a technology that emerged in the AI era, but AI may be the first to push it toward such direct data industrial applications.

There is also a second, more sensitive layer: legal.

In June 2025, the U.S. District Court for the Northern District of California, in the case of Bartz v. Anthropic, disclosed that Anthropic had purchased millions of physical books, disassembled their bindings, cut the pages, scanned them into digital copies, and discarded the original physical copies after generating the digital versions. Judge William Alsup determined, under the specific facts of the case, that this format conversion—where one physical copy is replaced by an internal digital copy and the digital copy is not distributed externally—constitutes fair use; the court also found that the specific large model training purpose in this case qualifies as fair use.

But the same ruling also drew a very clear line.

Anthropic previously downloaded over 7 million books from pirated databases. The court ruled that establishing and maintaining a central digital library of pirated sources cannot automatically qualify for fair use protection merely because some of the content was later used for AI training. The court clearly distinguished between legally purchased materials and those obtained through piracy.

Why is this line important? Because it taught the entire AI industry an expensive lesson: it’s not “any data can be used as long as it’s for training”—where the data comes from, whether companies paid for copies, and how they store and replicate it can all alter legal risk.

And this lesson later became very expensive.

On July 20, 2026, a U.S. federal court finally approved a $1.5 billion class-action copyright settlement between Anthropic and authors and publishers. The settlement covers over 482,000 books, and approximately 91% of the relevant works have already had claims filed for compensation. It must be clearly stated that this settlement addresses claims related to pirated sources and does not mean the U.S. judicial system has issued a universal permit for all AI training; other cases, different data sources, and alternative uses may still yield different outcomes.

Therefore, Amazon's current practice of purchasing physical books through commercial channels is very easy to understand from a risk control perspective.

Is it directly modeling its process after the legal path established by the Anthropic case? There is currently no public evidence to draw such a conclusion. But by 2026, with the Anthropic ruling and the $1.5 billion settlement already on the table, any major AI company will be paying closer attention to one thing: can you prove where your data came from?

The price of buying a physical book might be very low; explaining the origin of millions of training materials in court years later might not be cheap at all.

The AI industry once treated data as an endless supply of air on the internet; now, it is learning to pay for "sources."

Here, you must separate emotion from fact.

The entry of mass-market bestsellers, widely printed popular books, and old books with many remaining copies into recycling systems does not in itself constitute a cultural disaster. The used and antiquarian book industry has historically included practices such as dismantling books to sell illustrations and disposing of unsold inventory. The emergence of AI should not lead us to suddenly pretend that every old book is a unique copy.

The issue is that Amazon has not yet answered whether it will identify rare, out-of-print, or special editions before performing the scan.

The book tracked by 404 Media was reported to be a rare book, and Inc. specifically asked Amazon whether it screened for rare or scarce editions before destruction, but received no response.

This leaves a gap that cannot be easily dismissed with “but I already paid for it”—legal ownership is fundamentally not the same as cultural preservation value.

The main text of a book can be scanned. However, signatures, reader annotations, bookplates, special bindings, correction notes, local printing marks, and even the physical traces left by a book’s decades-long journey may not be captured by OCR, let alone included in the final training dataset.

The text remains, but physical evidence may disappear.

This concern is not limited to the United States. This year, several secondhand booksellers in the UK and Ireland told The Guardian that over the past few months, they have received large-volume purchases with no thematic connection and very low price sensitivity: one buyer purchased Estonian-language novels, specific editions of Brontë works, and old magazines from 1983; one bookseller said they have received orders totaling approximately 6,000 books from similar buyers since the beginning of the year. Similar unusual bulk orders have also been reported by booksellers in Australia.

But here too, we must stay within the bounds of fact: there is currently no public evidence that all these overseas orders are from AI companies. The Guardian’s report itself explicitly presents them as speculation and concern from booksellers, not confirmed AI purchases.

What was actually tracked all the way to the specific AI scanning facility by the tracking device was 404 Media’s investigation into Amazon.

This is exactly why transparency is important.

If a company believes this procurement is legal, reasonable, and genuinely improves its product, then these few basic questions are not excessive: What types of books will be subject to destructive scanning? Are rare, unique, signed, or physically valuable editions excluded? How will the physical copies be handled after scanning? How long will the digital copies be retained? And for what products or model categories will they generally be used?

Companies certainly do not need to disclose all training formulas.

But "we paid for it, so we can silently cut everything" is not a sufficiently mature answer.

Because it is betting on AI at an unprecedented pace.

In February, Amazon stated that it expects full-year capital expenditures of approximately $200 billion in 2026, with key focus areas including AI, chips, robotics, and low-earth-orbit satellites. By the second quarter, the company’s revenue reached $200.6 billion, a 20% year-over-year increase; AWS revenue reached $42.2 billion, up 37% year-over-year, marking the fastest growth in 18 quarters. Over the past 12 months, free cash flow turned into a net outflow of $7.6 billion, with Amazon explicitly attributing the significant increase in property and equipment purchases primarily to AI investments. AWS’s AI and chip businesses have each surpassed annualized revenue run rates of $25 billion.

More notably, the relationship between Amazon and Anthropic.

In April this year, Amazon announced an additional $5 billion investment in Anthropic, on top of its previous $8 billion investment, with the potential to add up to an additional $20 billion upon achieving specific business milestones. Meanwhile, Anthropic has committed to spending over $100 billion on AWS technologies over the next decade and will receive up to 5 GW of existing and future Trainium compute capacity.

In other words, Amazon is now both a cloud infrastructure provider and a capital participant in the frontier model industry, while also operating Nova, Alexa, Rufus, and a wide range of enterprise AI services.

What it needs is no longer just more chips.

It requires more high-quality knowledge, broader long-tail knowledge coverage, more reliable data sources, and a more easily explainable data ownership chain.

At Amazon’s current scale of capital expenditure, bulk purchasing of used books is clearly not a major cost that determines the company’s financial position. What truly matters is the signal this action sends: a giant willing to invest approximately $200 billion annually in expanding its technological infrastructure still finds it worthwhile to extend its supply chain into the fragmented, inefficient, and individually priced secondhand book market.

Why?

Because some data really doesn't have shortcuts.

This is where the competition in AI is shifting. Previously, it was about who could scrape more web pages; then, it was about who had more GPUs; now, it’s increasingly about who possesses cleaner, more unique, more legitimate, and harder-to-replicate data.

Data is no longer just a pile of "raw materials" before training begins.

It is becoming an asset that must be acquired, cleaned, verified, tracked, and audited.

The real data moat may not be about who owns a copy of the internet, but who possesses high-quality content that others cannot access, cannot hold securely, or dare not use confidently.

I don't think the conclusion should be "ban AI companies from scanning physical books."

Digitization itself can preserve knowledge and bring out-of-print content back into search, research, and machine understanding systems. For the vast number of ordinary used books, destructive scanning may not be worse than ultimately ending up in paper recycling. Reducing this issue to simply “AI is burning books” is emotionally compelling but factually inadequate.

There are two true boundaries.

The first is a boundary of rights: purchasing a physical book does not equate to acquiring all copyrights to the work, nor does it make all training, copying, and output activities inherently legal. The Bartz case is merely an important ruling by a U.S. federal district court under specific circumstances—it is not a permanent license for AI companies worldwide.

The second boundary is preservation: if a company cannot demonstrate that it can identify and avoid books that are rare, uniquely editioned, or hold special historical value, then所谓“高效扫描” may turn a data engineering problem into an irreversible cultural loss.

What Amazon should really answer is not "Why do you need books?"

This answer is becoming increasingly clear: because the vast amount of high-quality knowledge accumulated by humans over the past few centuries is not all available on the open internet.

The real question is: when these books move from shelves into AI training pipelines, who decides what can be sacrificed and what shouldn’t be? Who ensures that companies are acquiring content, not inadvertently erasing the history beyond it?

The irony of this situation is strong.

Amazon became one of the most important internet companies by delivering books directly to people’s hands; today, it buys books back, takes them apart, and feeds the text into machines.

Technology can certainly change the medium of knowledge, but that doesn't mean we should pretend the medium has no value.

AI needs things written by humans, and there’s nothing shameful about that. What truly warrants caution is when an entire industry increasingly relies on “high-quality text left behind by humans,” yet continues to bury the acquisition process deep within its supply chain, treating copyright, provenance, scarcity, and preservation responsibilities as afterthoughts.

A model can compress a book into tokens, and a company can turn tokens into parameters, but some things, once broken down, cannot be restored by any future training.

Civilization can be digitized, but responsibility cannot be "converted" away.

1. Emanuel Maiberg. “We Tracked a Shipment of Rare Books. It Ended at an Amazon AI Training Facility.” 404 Media, 2026-08-17.
2. Georgia Fearn. “Amazon Is Buying Up Older Books and Destroying Them After Scanning. The Orders Go Back to 2024.” Inc., 2026-08-17.
3. Amazon.com, Inc. “Amazon.com Announces Second Quarter Results.” 2026-07-30.
4. Amazon. “Amazon and Anthropic expand strategic collaboration.” 2026-04-20.
5. United States District Court, Northern District of California.Bartz v. Anthropic PBC, Order on Fair Use and Infringement, 2025-06-23.
6. Reuters. “US judge approves Anthropic’s $1.5 billion settlement of copyright lawsuit.” 2026-07-20.
7. Pablo Villalobos et al. “Will we run out of data? Limits of LLM scaling based on human-generated data.”Epoch AI, 2024-06-06.
8. Dan Milmo. “Secondhand booksellers in UK and Ireland suspect AI firms behind ‘strange’ bulk orders.”The Guardian, 2026-08-15.