The most valuable resource for AI is not compute—it's human-generated text. And the latest method to obtain it involves buying, scanning, and then destroying physical books. This is not a dystopian novel; it is happening right now, driven by court rulings that treat rare volumes as disposable data mines.
Hook
In 2025, a California court confirmed that legally purchased physical books could be destructively scanned—as long as the originals are thrown away. The logic: one digital copy replaces one physical copy, so no net loss of expression occurs. Since then, a quiet industry has emerged. ISBNdb, a data broker, now offers to buy, scan, and shred any book in its catalogue—for AI training. Anthropic, one of the leading AI labs, reportedly spent millions to acquire millions of physical books this way. Social media erupted with images of pallets of books being dumped into industrial grinders, but the full cultural cost remains hidden behind nondisclosure agreements.
Context
Let me step back. I have spent years building crypto education platforms, and I have seen how quickly the line between progress and destruction blurs. In 2017, I watched ICO teams burn tokens to create artificial scarcity. Today, we are burning books to create artificial data purity. The court's reasoning was narrow: a library that scans a book for preservation and then discards the physical copy does not harm the copyright holder because the work is still accessible in digital form. But AI companies are not libraries. They are building profit-driven models that will serve billions of queries. The digital copies they create are never meant to be accessed by readers—they are fed to neural networks and then stored in black boxes.
Core
Let me be clear: this is not a technological breakthrough. It is a legal and ethical arbitrage. The data engineering here involves physical supply chains—warehouses, industrial scanners, shredders—and a sophisticated understanding of copyright loopholes. The core insight is that pre-2022 books are considered “clean” data: they contain no AI-generated text, no modern data poisoning. They are, in the words of ISBNdb's marketing, “human-generated text at its purest.” But this purity comes at a price. The books being destroyed are often out-of-print, rare, or culturally significant. The court only cared about “protected expression”—the words on the page. It ignored the physical artifact: the binding, the marginalia, the edition's provenance. As someone who has worked with blockchain-based provenance for digital art, I can tell you: destroying the original erases a layer of meaning that no digital copy can replace.
Code is law, but ethics is conscience. The AI industry now operates in a legal framework that permits the annihilation of physical culture for the sake of model performance. I have audited data pipelines for DeFi protocols, and I know that once a dataset is destroyed, you cannot audit it. There is no way to verify whether a particular book was truly “clean” or whether it contained inaccuracies that now pollute the model. The one-to-one replacement logic also ignores the fact that digital copies are infinitely reproducible. Once scanned, there is no technical mechanism to prevent unauthorized copying—only legal threats. The court’s reasoning is a house of cards.
Culture on-chain, heart on-screen. I have curated digital art collectives that sold NFTs to fund blockchain literacy in Cape Town townships. I have seen how tokenization can empower creators while preserving cultural heritage. The destructive scanning model does the opposite: it centralizes control and erases physical history. Could we not instead use blockchain to create a transparent registry of licensed digital archives—where AI companies pay a fair price to digitize books without destroying them, and where the public can verify that no unauthorized copies exist? This is not a naive dream. It is a market opportunity that aligns incentives: publishers get revenue, AI companies get clean data, and the physical books survive.
Contrarian
Now, let me test this pragmatically. Some argue that the scale of AI training makes destruction inevitable. “There are billions of books in warehouses,” they say. “Destroying a few million is a drop in the ocean.” But we don't know which books are being destroyed. ISBNdb and Anthropic have not disclosed titles. The danger is not the number—it is the selection. If rare first editions, indigenous language texts, or hand-annotated scientific volumes are being shredded, the loss is irreversible. The 'drop in the ocean' argument also ignores power dynamics: AI companies can outbid libraries, museums, and researchers for physical books. In a market where destruction is legal, the highest bidder wins the right to erase. That is not progress. It is predation.
Takeaway
We stand at a crossroads. The court's ruling was a legal convenience, not a moral compass. The AI industry can choose to embrace a new social contract—one that values both innovation and heritage. Blockchain-based provenance, transparent data licensing, and mandatory public registries of destroyed books could be part of the solution. Or we can continue down this path, where the most advanced intelligence of the 21st century is trained on the ashes of our libraries.
Solidarity over speculation. The same ethos that drives decentralized governance should also guide our relationship with data. We must demand that AI companies disclose what they destroy, and we must build systems that reward preservation over extraction. The choice is ours.