The reported operation points to the value AI developers place on human-written material that is difficult or impossible to obtain online. Rare and out-of-print books can contain text unavailable in common digital sources, providing another pool of material for companies seeking data to train large language models.
Books published before 2022 also have a particular characteristic for AI training: their original text predates the current proliferation of content generated by large language models. That distinction matters because repeatedly training models on AI-generated material can contribute to a deterioration in output quality, a problem researchers call model collapse.
Amazon’s approach involves physically modifying the books so they can be processed more efficiently. Removing the spine separates the pages for high-speed scanning, but the procedure means the books themselves do not survive intact. That aspect of the operation has created questions around the treatment of rare material as companies seek additional sources of training data.
Copyright presents a separate issue. Some of the books involved may contain copyrighted material, and the broader technology sector continues to face legal questions over whether using protected works to train AI systems qualifies as fair use. The available reporting does not establish that Amazon’s specific book-scanning operation violates copyright law.
The practice nevertheless illustrates how obtaining AI training material can extend beyond collecting information already available online. Companies developing large language models need substantial quantities of text, while the growing presence of machine-generated material online can make verified human-authored sources more useful for training purposes.
Amazon’s reported book-scanning operation provides one example of that search moving into physical archives. By purchasing hard-to-find books and digitizing their contents, the company can access material that may otherwise remain unavailable to its AI development efforts, even when doing so requires destroying the physical copies used in the process.
This analysis is based on reporting from Whalesbook.
Image courtesy of Unsplash.
This article was generated with AI assistance and reviewed for accuracy and quality.
