Destructive scanning: digitizing at the expense of the book
Cut spine, removed binding, and pages acquired one by one: a quick and efficient method, but entirely irreversible
A few weeks ago, Anthropic, the company behind Claude, faced a lawsuit. It was accused of training its AI models on thousands of pirated books. The case ended with a $1.5 billion settlement. But it raised an even more interesting question. Where can AI companies get enough text to train their models?
The case made one thing clear. AI companies can't illegally obtain pirated copies of books. But they can legally buy physical books and scan them. They can then use those texts to train their models. The difference matters. They bought the original book legally. And the digital copy replaces it, rather than adding another copy.
That's where destructive scanning comes in. The book is taken apart. Its spine is cut off, and the binding removed. Then each page is scanned. Automatic scanners can process the loose pages quickly. The scans become digital files for text recognition. It's efficient, but there's no going back. The original book is no longer usable.
That makes books especially valuable to AI developers. They contain huge amounts of text. More importantly, much of it predates widespread generative AI. Books published before 2022 may be especially valuable. Back then, tools like ChatGPT weren't part of everyday life.
Newer content brings a different problem: "AI slop." That's the growing flood of text written by AI. Some of it is only partly AI-generated. Train new models on that material, and the risk grows. AI could end up learning from AI-generated text, not human writing. So an old library can be surprisingly useful today. It holds writing from before generative AI took off.
