A pallet of textbooks can disappear faster than anyone notices. In a warehouse, a guillotine trimmer slices the spine off each volume, a scanner flips through the loose pages at high speed, and the paper goes into recycling. The words, digitized, end up somewhere else entirely: inside a training corpus for a large language model. This is not a crime scene. It is procurement, and it has become one of the stranger supply chains in the artificial intelligence boom.
The reporting that surfaced this practice through outlets like France 24 is worth pausing over, because it exposes a gap between how we talk about AI and how it actually gets built. The public conversation is dominated by chips, data centers, and power bills. The raw material question, where the text comes from, is quieter and messier. Models need enormous volumes of high-quality language. Web scraping is cheap but increasingly polluted with spam, AI-generated slop, and licensing disputes. Books, by contrast, are edited, structured, and long. That makes them unusually valuable. It also makes them unusually awkward to acquire legally.
So a secondary market has formed. Some of it is legitimate. Publishers license backlists, universities digitize collections under agreements, and companies pay for access. Some of it is gray. Scanned books circulate through datasets whose provenance is hard to trace. And some of it is bluntly destructive, at least to the physical object: the book is bought, cut, scanned, and pulped, because scanning a bound book without damaging it is slower and more expensive than simply removing the binding.
There is a certain irony here. The technology that can summarize a novel in seconds is being fed by the physical destruction of novels. But the deeper issue is not sentimental. It is economic and legal. If a model is trained on a book, who was paid? The author almost certainly was not, unless a licensing deal covered it. The publisher may or may not have been. The person who bought the copy at a thrift store owns that copy, but owning a book has never meant owning the right to reproduce it at scale. Copyright law distinguishes between reading and copying, and scanning is copying.
The training-data debate has mostly played out in courtrooms and comment threads, often reduced to a binary: either AI companies are thieves or rightsholders are Luddites. Reality is duller and more interesting. Some authors and publishers have signed deals, treating their catalogs as licensable assets. Others have sued. Still others are experimenting with collective licensing schemes that would pay creators when their work is used, similar to how music royalties work in some markets. None of these approaches is fully settled, and the law is moving slowly relative to the technology.
What the book-scanning trade reveals is that the industry has a demand problem it cannot solve with engineering alone. You can build a bigger cluster. You cannot conjure a million clean, well-edited, legally unambiguous books out of thin air. That scarcity pushes companies toward whatever supply exists, including the used-book market, where a title can be acquired for a dollar and turned into training tokens worth far more than that. The economics are compelling precisely because the legal exposure is uncertain.
There is also a quality argument that gets lost. Synthetic data, generated by models to train other models, is cheap but risks compounding errors and flattening style. Human-written books carry nuance that scraped web text rarely does. If the best human writing becomes a finite resource that gets consumed and not replenished, the long-term pipeline looks fragile. Pulping a book to train a model is, in a small way, eating the seed corn.
None of this means the practice is universally sinister. Digitization preserves fragile texts, enables search, and can widen access. The problem is consent and compensation, not scanning as such. A library that digitizes a rare manuscript for scholars is doing something different from a company that buys remaindered copies to bulk up a commercial dataset without telling anyone.
The uncomfortable takeaway is that the AI industry’s most celebrated products rest on a supply chain that is partly informal, partly contested, and largely invisible. Chips get fabs and headlines. Books get shredded in warehouses. If the next few years bring clearer licensing norms and better provenance tracking, the pulping may stop being necessary. Until then, the quiet trade will continue, one spine at a time, and the models will keep learning from pages that no longer exist anywhere but in a tensor.