The Burning of Rare Books: A Data Strategy Where the Cost Exceeds the Gain
On February 14, 2025, Crypto Briefing reported that Amazon had purchased rare books and allegedly destroyed the originals to feed its AI training pipelines. The report, based on unnamed sources, paints a picture of escalating data desperation. But the technical and legal rationale for destroying physical copies is, upon dissection, a house of cards.
Context: AI training data scarcity is a well-documented problem. Epoch AI estimates high-quality text data will be exhausted by 2026–2032. Companies are scrambling for unique data sources. Amazon, as the world's largest bookseller, has a structural advantage in acquiring rare texts. The step of destroying originals suggests a strategy of data exclusivity—preventing competitors from scanning the same physical copy. However, the logic collapses under scrutiny.
Core: A systematic teardown reveals three layers of inefficiency. First, the technical argument. Destroying a physical book to prevent others from scanning it only works if the book is a single, unique copy. Most rare books exist in multiple copies across libraries, archives, and private collections. Eliminating one copy does not erase the data—it only removes that artifact. The digital scan is what matters for AI training. The marginal gain from blocking a competitor's ability to scan the same copy is negligible. In fact, if the book is still in copyright, the digital copy itself is a derivative work that may infringe. The destruction does not enhance the AI's performance; it only adds a futile cost. Based on my 2017 audit of Geth, I learned that the most elegant solutions are often the simplest. Digitize and preserve. Do not destroy.
Second, the legal risk. In copyright law, destruction of an original can be interpreted as bad faith, undermining the fair use defense. The Amazon case parallels the Google Books lawsuit, but with a critical difference: Google scanned and stored originals, offering snippets. Amazon's alleged destruction removes the original, which could be seen as willful suppression of evidence. During my work on the SEC Grayscale ETF opposition memo, I observed how minor procedural missteps can amplify liability. Destroying the physical book does not strengthen the fair use argument; it weakens it. The court may view it as an attempt to eliminate the possibility of third-party verification, which is a red flag for judges.
Third, the ethical and cultural cost. Rare books are irreplaceable cultural artifacts. They contain marginalia, binding details, and historical context that digital scans cannot capture. Destroying them for AI training is an irreversible loss. In 2022, when I analyzed the Bored Ape YC floor collapse, I saw how artificial scarcity can be manufactured. Here, the scarcity is real, but the destruction is not a data strategy—it is a public relations disaster waiting to happen. The crypto community understands the value of decentralized data provenance. On-chain records of ownership and digitization would provide transparency without destruction. "Ledger integrity precedes market sentiment."
Contrarian: What bulls got right? The need for high-quality, diverse data is real. Public web data is polluted with low-quality content, spam, and AI-generated noise. Rare books offer unique linguistic structures, historical knowledge, and domain-specific depth. Amazon's infrastructure for digitization is unmatched—they could have created a digital library and licensed the data to the AI community. The destruction is the flaw, not the acquisition. A more ethical approach would mirror Google Books: digitize, preserve the original, and offer controlled access. "Audits reveal what code conceals." In this case, the audit of Amazon's data supply chain should have flagged the destruction as a liability, not a benefit.
Takeaway: The pursuit of data exclusivity through physical destruction is not only inefficient but also invites regulatory backlash. The crypto industry, with its emphasis on decentralized data provenance and immutable ledgers, offers a potential alternative: tokenized data rights, on-chain attribution, and transparent data sourcing. "Precision is the only risk mitigation." If Amazon wants to build a defensible data moat, it should look to cryptographic proofs of uniqueness rather than burning books. The real question is not whether they bought the books, but whether they have the discipline to build a sustainable data infrastructure without destroying the very artifacts they seek to learn from. "Hype evaporates; solvency remains." In this case, the hype of data exclusivity has evaporated, leaving only the solvency of a flawed strategy.