Per a 404 Media report published this morning, Amazon, that evil empire, “is buying massive quantities of books, scanning them for AI training data, and destroying them in the process.”
This news follows slow-creeping revelations that other AI companies have been practicing “destructive scanning” to train their LLMs for at least the past two years.
Dominos stated to fall this January, when The Washington Post broke the story that Anthropic was at work on a secret project to “destructively scan all the books in the world.”
Details about this “Project Panama”—because what is an evil plan without a nom de guerre befitting a Bond villain—emerged during a much-publicized copyright lawsuit authors brought against Anthropic earlier this year.
Though this scoop made the company behind Claude seem uniquely horrible, ongoing lawsuits against Meta, Google, Microsoft, and OpenAI now suggest similar practices.
Everywhere, big corporations are stealing author’s work to train their LLMs. With a special recent focus on dead authors, who cannot protest.
Amazon has previously denied that they practice “destructive scanning.” (Or at least, they have kept very, very quiet about the subject.) But as Emanuel Maiberg reported, new signs point to complicity.
How do we figure? In a neat bit of practical research, 404Media followed a rare book they “suspected would be acquired by an AI company for training data,” and traced its journey around the country:
That final destination was an Amazon warehouse in Las Vegas, Nevada. Amazon employees who work at this location say all they do is receive massive shipments of printed books which they then cut the bindings off in order to scan the books more quickly. The printed book is destroyed in the process.
But wait, there’s even more Bond villain where this came from! Apparently, the offending team’s logo “is a dinosaur, brandishing its teeth and with a book in its hands.”
So the Rex is out of the park, Mr. Bezos.
Perhaps you too are wondering: beyond the silent plaintiff part, what makes rare books so valuable for AI companies?
According to documents WaPo uncovered re: Project Panama, “one Anthropic co-founder theorized that training AI models on books could teach them ‘how to write well’ instead of mimicking ‘low quality internet speak.’”
Maiberg echoed this theory. “Printed books are valuable as training data because a lot of the text they contain is not readily available on the internet, which AI companies have already scraped,” he wrote in this morning’s brief.
The data is also conveniently organized and, if the book was printed before 2022, is guaranteed to be free of AI-generated text, which can make any AI model that is trained on it worse via a recursive process called ‘model collapse.’
Here’s hoping.
There is another silver lining here. The American public is, generally speaking, still pretty against book destruction. On the optics level, if nowhere else.
Certain companies have already seemed to respond to public pressure about this un-beautiful, all-too-on-the-nose practice.
So assuming they don’t collapse on themselves, let us pray the shame spigot will short these machines.
