Amazon is buying rare books to scan them, then throwing them away

Someone placed an AirTag inside a rare book listed for sale on AbeBooks. The book was purchased. The tracker moved to a fulfilment-adjacent warehouse. Then the signal went quiet in a place that handles waste. The book, presumably, no longer exists.

That is not a theory. It is a documented movement trail from a deliberate experiment, repeated across multiple titles. The conclusion it points to: Amazon is systematically buying rare and out-of-print books, scanning them for AI training data, and then discarding the physical copies.

This is not a leak. No whistleblower came forward. The evidence was spatial and timestamped, gathered by attaching consumer tracking hardware to items moving through a supply chain that was not designed to be watched. It is the kind of investigation that produces primary evidence rather than inference, which is why it matters more than another think-piece about AI and copyright.

What the pipeline looks like

The mechanics, as far as can be traced, are not complicated. Amazon operates and owns AbeBooks, the dominant global marketplace for used, rare, and out-of-print books. That gives it both a purchasing interface and a pricing signal for every title in the catalogue.

Buyers connected to Amazon's acquisition operations place orders through AbeBooks. Independent booksellers, many of them small shops or individual dealers, ship the books in good faith. The books arrive at receiving locations, get scanned, and then move on to disposal rather than resale or archiving.

The sellers do not know this is happening. They believe they are selling to a reader, a collector, or a library. They are, in effect, feeding a training pipeline. The transaction looks like any other used-book sale. The outcome is not.

For the AI industry, this kind of acquisition is genuinely useful. Out-of-print books represent text that does not exist in any crawlable digital form. Academic monographs, regional histories, limited-edition literary fiction, specialist technical manuals from the 1960s and 1970s: this is material that never made it to the open web and is therefore absent from the datasets that models have been trained on so far. If you want to move model quality forward on rare or domain-specific knowledge, physical books are one of the few remaining reservoirs.

Why this is a harder problem than the copyright debate

Most public discussion about AI and intellectual property focuses on digital text: Common Crawl, scraped news archives, GitHub repositories. Courts in the US and Europe are working through what training on that material means under fair use and similar doctrines. The legal landscape is unsettled but at least the terrain is visible.

Physical book acquisition is different in two ways that the copyright conversation has not yet caught up with.

First, the act of scanning a physical book and ingesting the result into a training corpus is, under most copyright frameworks, a reproduction. The doctrine of first sale in US law allows someone who buys a book to resell it or lend it. It does not grant the right to reproduce it. The same distinction holds in most European jurisdictions. Buying a copy gives you that copy. It does not give you the text.

Second, the destruction of the physical object after scanning removes any possibility of the book returning to the cultural commons. A book that is sold, shelved, and resold circulates. A book that is scanned and binned is gone. For rare titles with small surviving print runs, that loss is permanent. There is no digital surrogate that the original seller or the book's author controls or benefits from. The cultural object disappears and the extracted text becomes a corporate asset.

This is not a small technical detail. It is a fundamentally different kind of extraction than scraping the open web.

What it means if you work in design or creative services

I want to be precise here, because the easy response is to reach for outrage and the harder response is to think clearly about the implications.

If you run an agency, a studio, or a creative practice, your work generates outputs that are increasingly being used to train models. That is already true for anything published digitally. What this book investigation reveals is that the appetite extends further, into physical artifacts, into materials that predate the internet, into knowledge that was never meant to be machine-readable.

The practical consequence is that the competitive advantage of having produced original, published, specialist work is being arbitraged away faster than most practitioners realise. A monograph you wrote in 2004. A catalogue essay for a design retrospective. A technical manual for a niche manufacturing process. These are the kinds of texts that now have material value to model builders, even if they have no active digital footprint.

That is not an argument against AI. It is an argument for understanding what is actually being taken, from whom, and on what terms.

There is also a secondary effect that deserves attention. If rare books are being removed from the secondhand market at scale, the availability and pricing of that material changes for everyone else: researchers, archivists, independent scholars, and the specialist booksellers who sustain the trade. A buying programme that operates invisibly inside the largest used-book marketplace in the world has the power to reshape that market without disclosing that it is doing so.

The strategic question this raises

We spend a lot of time at Studio Hyra thinking about where data comes from and what it costs. Not in an abstract ethics-seminar way, but in the practical sense of: what are the actual inputs to the models our clients use, and what are the conditions under which those inputs were gathered?

Most organisations cannot answer that question in detail. They use foundation models from major providers, and the data provenance of those models is largely undisclosed. The Amazon book investigation is a useful data point not because it is unique, but because it is unusually concrete. For once, there is a tracker, a route, and a disposal location. Most of what feeds large language models is far more opaque than that.

For agencies advising clients on AI adoption, the honest position is this: the legal and ethical questions around training data are not resolved, they are accumulating. Courts will clarify some of it over the next several years. Regulatory frameworks in the EU, through the AI Act, will impose disclosure requirements that do not yet have teeth. In the meantime, the companies building the most capable models are moving fast and sourcing aggressively.

That is not a reason to avoid AI tools. It is a reason to hold those questions clearly rather than outsourcing your conscience to a terms-of-service document.

What comes next

The AirTag investigation is a proof of method as much as it is a specific finding. It demonstrates that consumer-grade hardware can surface what corporate opacity tries to obscure. Expect more of this kind of forensic accountability work as AI data sourcing becomes a more contested public issue.

For Amazon, the specific exposure here is reputational and legal. AbeBooks depends on trust from independent booksellers. If the platform is perceived as a covert acquisition tool for AI training, that relationship degrades. There is also the copyright reproduction question, which is not hypothetical once you have a documented scan-and-destroy pipeline.

For the broader AI industry, the book case points to a pattern that will keep surfacing in different domains: physical archives, museum collections, private correspondence, specialist databases. Anywhere that valuable, structured human knowledge exists in a form that has not yet been digitised is a target. The methods will vary. The logic will not.

The question worth asking, for anyone building with or advising on AI, is not whether this is happening. It clearly is. The question is what obligations follow from knowing it.