How AI Is Squeezing the Book World

Books Are Being Squeezed by AI on Both Ends

Across the book world, a new pattern is coming into focus: artificial intelligence is no longer just borrowing from books, but increasingly reshaping how books are bought, made and sold.

Secondhand booksellers in Britain, Ireland and elsewhere say they have been fielding unusual bulk orders for eclectic, often low-demand used titles from hard-to-identify buyers, raising suspicions that physical books are being acquired not to be read, but to be scanned as training material for AI systems. At the same time, a new study of Amazon’s self-publishing marketplace suggests that AI-generated books are proliferating fast enough to erode sales opportunities for human authors, even when those AI-written titles earn less on average.

Taken together, the developments point to a publishing ecosystem under pressure on both the input and output sides. Printed books, especially inexpensive used copies, are increasingly viewed as raw material for machine learning. And digital storefronts are becoming crowded with rapidly produced AI-assisted or AI-generated works competing for rankings, visibility and readers’ attention.

A Strange Run on Used Books

Booksellers in the United Kingdom and Ireland have described a recent wave of large, odd orders for secondhand books — a mix of obscure, backlist and seemingly random titles that do not fit the patterns of ordinary collectors, dealers or institutional buyers. Similar reports have surfaced in Australia, continental Europe and the United States.

No public evidence has yet tied those purchases to a specific AI developer. But the speculation has gained force because the practice itself is no longer hypothetical.

In court disclosures this year, Anthropic was found to have spent millions of dollars buying millions of physical books, scanning them into digital files and keeping those copies for AI-related “data acquisition.” A federal court concluded that training on lawfully purchased and scanned books qualified as fair use, while drawing a line against the retention of pirated books. The ruling gave AI companies a clearer legal path to using purchased print books as training data, even as it deepened anger among authors and publishers.

That legal backdrop matters. Once a court signaled that bought-and-scanned books could be treated differently from pirated libraries, physical books acquired in bulk became more valuable as compliant training inputs. Reporting in recent months has also shown that intermediaries have marketed printed-book sourcing services to AI companies, further professionalizing the trade.

For secondhand sellers, that can mean sudden demand for inventory that had long sat untouched. For authors and publishers, it raises a more unsettling possibility: that books are being stripped of their traditional cultural and commercial role and repurposed as feedstock for systems that may eventually compete with them.

Amazon’s Shelves Fill Up

If the used-book trade reflects the appetite for AI training material, Amazon’s marketplace shows what happens at the other end of the pipeline.

A July 2026 study examining 14,419 self-published genre-fiction titles sold on Amazon found that books with substantial detected AI-generated text accounted for 20 percent of sampled titles but only 12.1 percent of sales and 11.3 percent of revenue. On average, the AI-heavy books underperformed human-written ones.

But the study’s warning was not that AI books were individually outselling everyone else. It was that they were multiplying quickly enough to dilute the market.

By the second quarter of 2026, the sales share of those AI-heavy books had risen to about 20 percent, the study found. And when combined with books containing lighter AI use, their share of high-ranking positions approached 40 percent. In effect, AI-written and AI-assisted works were taking up more of the digital shelf, more of the recommendation slots and more of the discovery space through which readers encounter books.

That matters because the economics of self-publishing depend heavily on visibility. On Amazon, rank is often destiny: the higher a title appears in search, charts and recommendation carousels, the more likely it is to sell. A flood of low-cost AI-generated titles can therefore hurt human authors not only by winning sales directly, but by making it harder for their books to be found at all.

More Books, Less Money

The study found another troubling signal. From January 2023 to March 2026, the number of books with observed quarterly sales rose 19.2-fold, while quarterly revenue increased only 8.9-fold. In plain terms, many more books were selling, but the market was not growing nearly fast enough to support them all. Revenue per selling title fell across seven of the eight genres studied.

That finding could prove important far beyond Amazon’s storefront. In copyright cases over AI training, plaintiffs have struggled to show concrete market harm. The study’s authors explicitly connected their results to that unresolved legal question, arguing that AI may damage authors not simply by copying their work, but by helping produce a glut of competing titles that weakens earnings across a category.

The evidence is still bounded. The paper looked at self-published genre fiction ebooks sold through Amazon, not the broader trade publishing business, nonfiction or print. And because it relied on AI-detection methods, its estimates carry uncertainty. Still, it offers one of the clearest quantitative snapshots yet of how generative AI may be changing the economics of online bookselling.

A Visibility Problem for Readers, Too

Amazon already requires publishers using Kindle Direct Publishing to disclose AI-generated content to the company. But those disclosures are not shown to shoppers.

That means readers often cannot tell whether a book was substantially written by a person, assembled with AI assistance or largely generated by software. For authors who argue that they are being forced into direct competition with machine-made books, the lack of visible labeling adds another layer of frustration. For readers, it obscures information that some may consider relevant to what they are buying.

The tension reflects a broader problem in the AI era: platforms can gather information about how content is made without necessarily sharing that information with the public. As a result, the marketplace can be transformed by automation before consumers fully understand what has changed.

Why This Moment Matters

Publishing has faced technological disruption before, from ebooks to print-on-demand to the rise of self-publishing. But the current shift is unusually comprehensive.

On one side, AI developers need vast amounts of text, and books remain among the richest and most structured sources of long-form language. On the other, the same technology can be used to generate fresh titles at a speed and scale no human author can match. The result is a loop: books are mined to build models, and models are then used to produce more books.

Whether that loop becomes a permanent feature of publishing may depend on decisions that are still unresolved. Courts must decide how far “market dilution” counts as copyright harm. Platforms may face pressure to tighten labeling, ranking and spam controls. Publishers may pursue new licensing deals, or more aggressive litigation, to preserve control over their catalogs.

For now, the signs of strain are appearing in places far apart but connected by the same technology: in the dusty stacks of secondhand shops, where old books are suddenly in demand for mysterious reasons, and on the bright digital shelves of Amazon, where human authors are finding it harder to stand out.

Sources

Further reading and reporting used to add context:

AI News