AI companies are destroying rare books for training data

Books into the shredder

According to the account given by the author of the thread, who goes by the handle Hedgie and cites a report by the outlet 404 Media, AI companies are buying up rare books in bulk, running them through high-speed scanners that slice off the spines, and destroying the originals themselves. The middleman is a service called ISBNdb: it lets you place orders of up to a million copies and keeps buyers anonymous.

Editions published before 2022 are especially prized — their text is guaranteed not to contain fragments generated by neural networks, and so it is considered “clean” material for training.

A legal indulgence

The practice has already been ruled lawful. According to the author, a federal judge deemed it fair use on the grounds that, once the original is destroyed, only a single copy of the work exists in the world at any one time. It is separately mentioned that Anthropic hired the former head of Google Books' partnership programs, setting the goal of acquiring “all the books in the world.” In court documents, Hedgie claims, the project even has a name — Project Panama — tens of millions of dollars in budget, and a plan that directly calls for every book on earth to be scanned and destroyed.

What troubles the author

Hedgie admits that this particular story got to him more than any other tale about AI. A bookseller told 404 Media that rare editions of which only a handful of copies survive are ending up in the pipeline — texts that outlasted wars, fires and centuries of readers' hands are now being fed into the shredder so that a model can learn to write more persuasive marketing emails.

A separate irony lies in ISBNdb's own rhetoric. According to the quote cited, the service's website says outright that the headline “AI company destroyed two million books” won't stir any sympathy — and an entire business, operating quietly, is built around this. Clients are offered non-disclosure agreements as a separate service and advised to call the process “digital preservation.”

The key difference from earlier grievances against the industry is irreversibility, the author stresses. A website can be re-uploaded, a bestseller reprinted, but the last three copies of an 18^th-century botanical atlas can no longer be recovered. When no physical original remains, there is nothing left to check the digital copy against.

The discussion under the post naturally drifted toward comparisons with the Library of Alexandria and the book bonfires of the past. There were practical remarks too: why not, having scanned a book, hand it over to a library — after all, storage costs mere pennies, while destruction is free. The answer, judging by the picture described, has already been given by economics.

Source: Hacker News