Amazon Is Buying and Destroying Rare Books to Train Its AI. Here's What That Means for Professionals.

TL;DR. On August 17, 2026, investigative outlet 404 Media published the results of a tracking experiment: place a GPS device in a shipment of rare books sold through Biblio, and follow it. It ended up at VGT3, an Amazon warehouse in Las Vegas, where workers cut off spines, scan the pages, and discard what's left. Amazon is systematically acquiring physical rare books to train its AI models — acquiring knowledge that doesn't exist anywhere online. Here's why AI companies are doing this, what it reveals about how AI tools work, and what professionals should understand.

---

The Investigation

On August 17, 2026, Emanuel Maiberg at 404 Media published an investigation that put a tracking device into a shipment of rare books listed for sale on Biblio, a rare-book marketplace. The device led to VGT3, an Amazon fulfillment facility outside Las Vegas, Nevada. An anonymous Amazon employee confirmed the operation simply: "All we do is scan books."

TechCrunch corroborated the investigation in independent reporting. The process: workers at the facility cut off book spines, feed loose pages through high-speed scanners, and discard the originals after capture. Amazon's official statement offered no specifics: "Amazon purchases books through commercial channels to help develop and improve the products and services our customers use."

The investigation tracked roughly 1,000 books in one traced shipment. The systematic logging of ISBN barcodes at the facility suggests the actual scale is considerably larger — possibly a systematic attempt to digitize rare printed works across many categories.

Why Physical Books, and Why Rare Ones?

AI labs need training data — enormous quantities of text for language models to learn from. But the internet, once a seemingly inexhaustible source of human-written text, has a problem: since large language models went mainstream in 2022, it has been rapidly filling with AI-generated content. Articles, summaries, forum posts, documentation — an increasing share of what's published online was written by AI.

Training the next generation of AI models on that AI-generated content causes what researchers call model collapse: a gradual degradation where the model loses nuance, hallucinates more frequently, and converges on bland, generic-sounding outputs. The model's knowledge narrows toward whatever average patterns the prior AI had already baked in.

Rare physical books solve this problem cleanly. Books published before 2022, especially ones never digitized, are guaranteed to be:

• Human-written. No AI touched them.
• Historically deep. They contain knowledge, reasoning styles, and terminology that reflect decades or centuries of human thought in a domain.
• Not already in a competitor's training set. Common digital text (Wikipedia, web crawls, digitized public domain books) has already been consumed by everyone. Rare out-of-print books represent genuinely novel data.

For AI companies, rare physical books are one of the last reserves of high-quality, unduplicated, human-generated training material. The economics are simple: a rare book that costs $50 at auction may be worth far more as training data for a model trained at $100 million.

What This Means for AI Quality — and for You

Training data composition is a real differentiator in AI output quality. Providers rarely disclose what's in their training sets, but the pattern is consistent: AI tools are more reliable on topics that are heavily documented in digitized text, and weaker on topics that live in physical archives, specialist literature, and out-of-print sources.

Consider what that means for professional domains:

• A lawyer asking about obscure pre-digital case law gets a different quality answer depending on whether historical legal treatises and court reports were in the training data.
• A physician researching rare syndromes first described in pre-digital medical literature may find that AI knowledge becomes thin or approximate past a certain historical depth.
• An engineer or architect working with legacy systems or historical construction methods will encounter limits in AI knowledge when the reference material exists only in printed manuals never transferred to digital form.

This doesn't mean AI tools are unreliable — they're generally excellent for well-documented topics. It means the accuracy gradient matters: AI answers about specialized, historically rich, or niche professional topics deserve more verification than AI answers about mainstream subjects with abundant online coverage.

The Legal and Ethical Picture

The legal status of physical-book scanning for AI training sits in a gap that courts haven't fully addressed. Scanning and discarding a physical book — rather than copying a digitized version — is factually different from the digital piracy cases that have been litigated. A prior lawsuit against Anthropic (for training on pirated digital copies) established some related precedent, but digital piracy and physical-to-digital conversion are legally distinct. Amazon appears to be operating in this ambiguity deliberately.

What's harder to dispute is the practical effect on book communities. One rare-book dealer told investigators: "They just want the content as a bunch of words strung together." Rare books — historical records, specialized treatises, texts that document now-vanished fields of practice — are being acquired as consumable raw material, not preserved as cultural artifacts.

Publishers, libraries, and archivists have pushed back on AI companies' training data practices generally. Some are negotiating licensing agreements. Others are exploring legal challenges. But for now, the physical secondary market appears to remain an open channel.

What Professionals Should Do With This

Nothing urgent. This story doesn't require any immediate action. But three things are worth holding:

1. Calibrate your trust by domain. AI tools are not uniformly knowledgeable. For questions where the authoritative literature is mostly pre-digital, old, or specialized, verify AI answers against primary sources. This is especially true in law, medicine, engineering, and academic research.

2. Ask about training data when it matters. If you're evaluating an enterprise AI subscription for a domain with deep specialist literature, asking about training data sourcing is a fair question. Some providers disclose more than others, and for niche professional domains, it can meaningfully affect output quality.

3. Understand that this is what AI progress costs. Rare physical texts are being consumed to make the AI tools professionals use every day more capable. That's a real tradeoff — and understanding it helps you think more clearly about what AI systems are, what they cost, and who pays.

---

Sources

1. 404 Media — "Amazon Buys and Destroys Rare Books to Train AI" — Emanuel Maiberg, August 17, 2026. Primary investigative source; includes the GPS tracking experiment.

2. TechCrunch — "Amazon, which started off selling books, is destroying rare texts to train AI models" — August 17, 2026. Independent corroboration with process details and scale context.

3. Futurism — "Amazon Is Destroying Rare Books to Train AI" — August 17, 2026. Additional reporting including scale estimates and bookseller reactions.