Japan's Used Bookstores See 5x Sales Surge: The AI Training Data Buying Frenzy Explained

An Unusual Market Signal

A recent report on Hacker News has sparked widespread attention: used bookstores in Japan are experiencing a fivefold surge in sales, with books being bought out in bulk — measured in tons. The fact that purchases are being priced by weight rather than by volume is itself a red flag — ordinary readers don't buy books by the ton. Only institutions with a voracious appetite for large-scale text content would operate this way.

Although the original report was light on details (the Hacker News post received 42 upvotes and 20 comments), community discussion has largely converged on one obvious hypothesis: these books are most likely being acquired for AI large language model training data. As copyright compliance and data regulation tighten globally, physical books — as a source of high-quality, well-structured text — are being rediscovered as a valuable resource.

Why Physical Books Have Become Hot Commodities

The Scarcity of High-Quality Training Corpora

Large language models are deeply dependent on high-quality text. Publicly available internet data, having been scraped through multiple cycles, is exhibiting declining marginal quality — riddled with duplicates, low-value content, and machine-generated text. By contrast, formally published books have undergone editorial review, proofreading, and typesetting. They are linguistically consistent, information-dense, and represent the premium tier of training corpora.

Japanese, as a non-English language, has relatively scarce high-quality web corpora to begin with. For models seeking to improve Japanese-language capabilities, bulk-scanning physical books may be the most direct path to authentic Japanese text.

Researchers refer to this phenomenon as "web data exhaustion." Estimates suggest that mainstream web crawl datasets like Common Crawl had largely covered accessible high-quality English web pages by around 2020, and the proportion of AI-generated content in newly added web material has been rising rapidly ever since — creating a feedback loop where models consume outputs generated by other models. Research shows that models trained on low-quality or machine-generated text exhibit measurable degradation in linguistic diversity and reasoning accuracy, a phenomenon known as "model collapse." For this reason, the vast corpus of printed books published before the internet era represent relatively "pure" samples of human writing — spanning long historical periods, covering diverse subjects, and offering unique value for deepening a model's linguistic comprehension.

Navigating Copyright Risk

Copyright litigation surrounding AI training data has proliferated in recent years. Purchasing physical books at least establishes legitimate ownership at the acquisition stage. Compared to directly scraping copyright-protected online content, the physical book pipeline appears legally more defensible — even though whether scanning and using books for training constitutes fair use remains hotly contested.

Notably, Japan's copyright law takes a relatively permissive stance on AI training data use. Prior amendments explicitly permitted the use of copyrighted works for information analysis purposes, which may help explain why acquisition activity appears concentrated in the Japanese market.

Article 30-4 of Japan's Copyright Act (amended in 2018) is the pivotal provision here: it explicitly permits the use of copyrighted works for "information analysis" (情報解析), including machine learning training, in principle without requiring authorization from or payment to rights holders. This provision is among the most permissive AI training data legislation among major economies globally. Unlike the United States, which relies on the ambiguity of "fair use," Japan provides relatively clear legal authorization for the use of AI training data. However, whether this exemption applies to large-scale, general-purpose models trained for commercial purposes remains a subject of debate among Japanese academics and legal scholars. The provision also does not extend beyond Japan's borders — meaning models acquired and scanned in Japan but deployed internationally still carry complex legal risks.

"Buying Books by the Ton" and the Intensifying Data Race

If the speculation holds, this phenomenon reflects that the AI industry's data race has entered an all-out phase. With the low-hanging fruit of web data already divided up, institutions are being forced to turn offline — toward traditional media — to mine incremental training material.

This trend raises several questions worth considering:

• The physical turn in data sourcing: Physical books, once considered obsolete, have become strategic assets in the AI era.
• The value dilemma for knowledge creators: Authors' and publishers' works are being bulk-purchased for training, yet whether original creators receive fair compensation remains an unresolved ethical question.
• The unexpected windfall for second-hand markets: Behind the surge in bookstore sales lies the reality of content being "commoditized" — books are no longer objects to be read, but raw materials to be processed.

A Note of Caution

It must be emphasized that current reporting does not directly confirm that these books are in fact flowing into AI training pipelines. A fivefold sales increase is an observable fact; the "AI training" explanation is more a reasoned community inference based on circumstantial logic. Other explanations cannot be ruled out — such as inventory liquidation by publishers, niche collector demand, or digitization archiving projects.

In the absence of clear information about who is doing the buying, this phenomenon is better treated as an industry signal worth tracking rather than a definitive conclusion. But whatever the truth turns out to be, it serves as a reminder: AI's appetite for data is reshaping the operating logic of traditional content industries in ways we have yet to fully understand.

Conclusion

"Buying books by the ton" is a strikingly vivid image — it makes the abstract concept of "data hunger" concrete, conjuring mountains of physical books stacked high. As algorithms need to be fed, the printed knowledge humanity has accumulated over centuries is being repriced. The scramble for training corpora may have only just begun.