Case Studies

Japan's Used Bookstores Are Being Drained by AI's Hunger for Training Data

A single 50-ton shipment of books recently left Japan for the US, with bulk buyers suspected to be AI companies scrambling for high-quality text corpora.

Font size

Source: Tom's Hardware, 2026-08-24

50 Tons of Books, Shipped Overnight to the US

Tom's Hardware ran a story yesterday with the headline "5x sales surge," documenting how Japanese used-bookstores have seen their sales quintuple almost overnight.

The trend runs against conventional wisdom. Used books should be a dying business: young people don't read much, Kindles are everywhere, and libraries lend for free. Who still buys secondhand?

But starting in the first half of this year, major Japanese used-book chains (the Bookoff type) began receiving unusual bulk orders, not by the copy, but by the metric ton.

The report flags one specific shipment: 50 tons, shipped directly to the United States in a single transaction.

To put that in scale: at roughly 300 grams per paperback, that's about 160,000 books, enough to fill an 11-meter cargo truck. Orders like these wipe out half a store's inventory in one go, forcing staff into overnight restocking marathons.

Who Is Buying? A Buyer Profile

The report doesn't name a specific company, but the buyer profile fits the template of AI training-data procurement almost perfectly:

  • No quality discrimination: everything from out-of-print novels to outdated almanacs is acceptable.
  • No questions about provenance: bulk volume alone justifies the price.
  • Fast payment: no haggling, no delays.
  • Opaque delivery addresses: many orders route through third-party freight forwarders, with final destinations undisclosed.

This is textbook corpus-acquisition behavior, the same logic behind the public data-procurement RFPs from Anthropic and OpenAI: define the category, specify the language, then compare bids.

The report uses the phrase "multitude of suspicious bulk buys," and multiple used-book chains have reported similar experiences.

The middleman layer is critical here. The actual buyer rarely contacts used-bookstores directly. Orders flow through one or two trading companies that sign NDAs, collect deposits, and arrange ocean freight. Once the chain breaks at the trading company, the trail goes cold.

The Data Drought Is Real: High-Quality Corpora Are Depleted

Why Japan? Three reasons.

First, the quality of Japanese-language text. Japanese fiction (novels, light novels) and non-fiction (academic, technical) rank among the best non-English corpora available. Japanese grammar is complex and expensive to annotate, but the payoff in multilingual model capability is significant.

Second, the copyright gray zone around out-of-print works. Many books published in Japan before the 1990s have no digital footprint anywhere. Common Crawl doesn't index them. Google Books never digitized them. Physical copies sit on bookstore shelves. These "orphan works" are exactly the kind of incremental data LLM training needs most.

Third, physical format convenience. AI companies don't want the book; they want the words inside it. Scan, OCR, extract the text for training, then destroy the physical object. The whole flow sidesteps direct "reproduction right" infringement (Japan's copyright law has an ongoing discussion around OCR-specific exemptions, covered below).

This isn't speculation. In 2025 court filings, Anthropic acknowledged that "high-quality data is the limiting factor in AI training." OpenAI's Ilya Sutskever said something similar back in 2023. The entire industry has been warning "we are running out of data" since 2024.

The quality ceiling of publicly scraped sources like Common Crawl is already locked in. The next growth area has to be content that exists physically but not online: books, newspapers, internal archives.

Japanese used-bookstores became the most visible mine.

The Copyright Gray Zone: Are Out-of-Print Books Lawless Territory?

Japanese copyright law has an awkward gap.

When a book is still under copyright (life of the author plus 70 years in Japan) but the publisher has gone out of print and stopped selling it, the law doesn't clearly designate who has the authority to license reproduction.

This is the "orphan works" problem in copyright law. The rights technically exist, but there's no practical path to securing a license.

Buying a used book grants ownership of the physical object, not the reproduction right. You can read it, resell it, or burn it, but you can't legally scan it, OCR it, and convert it into training data.

But consider:

  1. Who enforces this? Scanning happens in the US. The data owner isn't in Japan.
  2. Once the physical book changes hands, there's no way to trace it from the Japanese side.
  3. Even if traced, cross-border enforcement costs far outweigh any recovery.

This is the foundation that the gray supply chain rests on.

The buyer places an order through a trading company. The books arrive and are legally classified as "personal property." Scanning happens in a third country. The rights holder never learns they've been infringed upon. There is no single legal loophole in this chain; it's a chain of loopholes. Three individually legal actions, combined, bypass copyright protection entirely.

The Other End of the Supply Chain: Scan and Shred

The report uses the phrase "AI scan and shred facilities."

After books arrive in the US (or a third country), they enter specialized facilities: conveyor belts, high-speed scanners, OCR processing, text extraction. Each scanned book goes straight into a shredder.

These facilities aren't new. Similar operations have been reported in Texas and New Mexico. A 2024 report described one company doing dedicated "book-to-training-corpus" work, processing hundreds of tons annually.

The physical destruction step is essential: what remains afterward is "a destroyed book," not "duplicated data."

This distinction matters legally.

Digital scraping leaves server logs, which can be traced. Physical destruction leaves shredded paper, which cannot. From an evidence standpoint, this path is far harder to track than pure digital scraping.

These scanning facilities often masquerade as "document destruction services," indistinguishable on paper from outfits like Shred-it. Inside, though, they run industrial-grade scanners, machines that can process over ten thousand pages per day.

The shredded paper goes to recycling. The text goes into the training set. Physical evidence vanishes.

Editorial Take

How does this differ from older data-scraping methods? Three ways.

First, the shift from "grab what's online" to "haul what's in homes." For the past decade, AI training data came primarily from Common Crawl, Wikipedia, Reddit, and arXiv, all content already on the internet. Now the industry is mining content that exists physically but not online: out-of-print books, defunct journals, internal archives. This is a qualitative shift.

Second, the shift from "digital duplication" to "physical destruction." Scanning a book and then destroying it upgrades the evidentiary problem from "data was duplicated" to "physical evidence has vanished." This is a tactic previous copyright litigation hasn't encountered.

Third, the shift from "individual enforcement" to "cross-border industrial chain." A single book's rights holder might be scattered across dozens of countries: a Japanese author, a US AI company, a third-country intermediary, a Southeast Asian freight forwarder. Legal jurisdiction on this chain doesn't close.

My assessment: this won't end quickly.

The training data shortage is structural. It won't be solved once a particular model finishes training. As long as base-model training scales continue to grow, demand for incremental corpora won't stop. Japan won't be the last country targeted. Europe's antiquarian book market, India's used textbook supply chain, Latin America's library clear-outs, all are next.

The sustainability of this approach, though, is questionable. Legally, a "personal property + cross-border + destruction" chain can hold up for a few more years. But from an industry standpoint, a single landmark lawsuit (similar to the NYT v. OpenAI case in 2023) could collapse the entire chain.

Uncertainties remain: the report doesn't name specific AI companies, the reliability of the source chain hasn't been fully verified, and whether the "5x sales surge" at used-bookstores is an absolute or relative figure still requires cross-checking.

Either way, this development deserves attention. It's not just a Japanese used-bookstore problem. It marks the transition of AI training data acquisition from "digital scraping" to "physical extraction."

Source: https://www.tomshardware.com/tech-industry/artificial-intelligence/japanese-used-bookstores-see-5x-sales-surge-as-books-are-being-bought-by-the-ton-one-50-ton-order-sent-to-the-us-for-ai-scanning-and-destruction-multitude-of-suspicious-bulk-buys-thought-to-end-up-in-foreign-ai-scan-and-shred-facilities