Literary Hub, citing a report from 404 Media, says Amazon has been buying books, scanning them for AI training data, and destroying the copies afterward. The sequence is brutally efficient: acquire the object, extract what the system can use, discard the remainder.
It also clarifies a distinction that digital culture often blurs. A dataset is not an archive. The two may contain some of the same words, but they preserve those words for different purposes, under different conditions, and with different ideas about who should be able to consult them.
Nicholson Baker’s Double Fold: Libraries and the Assault on Paper belongs in this argument. Published in 2001, the nonfiction book is cataloged under subjects including libraries, newspapers, paper, conservation, restoration, preservation, and special collections. Even without importing the details of Baker’s case into the Amazon story, his polemical subtitle identifies the relevant fault line: institutions can believe they are preserving information while eliminating the physical things that carried it.
That problem looks newly sharp when the institution doing the extracting is building an AI model.
Preservation leaves something to return to
Scanning is often described as preservation because it produces another version of a work. Sometimes that version is more searchable, portable, and widely accessible than the original. Those are real advantages. But they do not settle what has been preserved.
An archive keeps material available for future consultation. That last word matters. Consultation means that someone can return to the source, examine it under a new hypothesis, notice what an earlier reader ignored, or test whether a transcription and interpretation were accurate. The object remains capable of contradicting the account made from it.
A physical book contains more than a sequence of words. Its paper, typography, binding, illustrations, layout, signs of ownership, and annotations can all become evidence, depending on the question being asked. Most copies will not be singular treasures, and no library can save everything. Still, destroying a volume after scanning closes off questions that the scanning process was not designed to answer. The loss is not mystical. It is a loss of possible inspection.
A dataset has another purpose. It converts works into material that can be processed at scale. Its value comes partly from reducing differences among objects: this edition and that edition, this clean copy and that marked-up one, all become inputs of a kind the system can absorb. Features irrelevant to training are noise. The book has been valued, but only according to what can be extracted from it.
This is why the reported Amazon practice is not simply an accelerated version of library digitization. A library archive is, in principle, organized around future access to records. A corporate training corpus is organized around producing a model. The model may retain statistical relationships learned from books, but it is not a reading room. A user cannot ordinarily ask it to produce the acquired copy, show how that copy differed from another, or permit an independent inspection of the source material. Once the book is destroyed, the surviving product is not necessarily an accessible scan. It may be only the capability built from the scan.
The copy is also an audit trail
The distinction matters because preservation is not just storage. It is an arrangement of accountability.
Keeping source material allows later people to challenge the choices made during collection and conversion. Was a page missed? Was an illustration excluded? Did optical character recognition distort a passage? Did the scan come from a particular edition? A retained volume does not answer those questions automatically, but it makes answers possible. It is an audit trail in paper form.
Destroying the source transfers authority to whoever performed the extraction. The company decides what counts as the book’s useful content, how that content is transformed, and what remains available afterward. Future readers inherit the result rather than the evidence behind it.
That does not mean every scanned copy must be preserved forever, or that paper is always superior to digital storage. Preservation requires selection, money, space, and maintenance. Physical objects decay; digital files can multiply and circulate. The harder and more useful question is whether a conversion expands future access or merely captures present value.
Double Fold is worth resurfacing because its subject vocabulary refuses to treat paper as a sentimental accessory to information. Libraries, conservation, restoration, and special collections all begin from the premise that the carrier can matter. The book’s title also suggests that destruction may arrive under the administrative language of care and modernization, not only through obvious neglect.
AI training gives that tension a new institutional setting. A volume can survive as influence while disappearing as an object. Its language can help shape a commercial system even when no future reader can handle the copy, inspect the scan, or determine exactly what was taken from it.
Calling that process preservation would be misleading. It is extraction. The difference is what remains available after the extractor has finished.