In a response to my recent post on compiling AI books, Thad McIlroy sent a link to a paper on the topic, “Generative AI floods and dilutes the market for books” by Tuhin Chakrabarty, Xinyue Liu, Jane C. Ginsburg, and Paramveer Dhillon. He has himself written about it, less than flatteringly, at The Future of Publishing.
Chakrabarty et al assure us “When a large supply of cheap-to-produce titles enters a market faster than the sales, revenue, and top ranks available to absorb them, the return and attention left for the average title both fall. We call this dilution.” Mr McIlroy’s most potent point is that it’s hard to see some new thing flooding a market which is already flooded. “In 2025 there were roughly 475,000 new adult fiction titles published, from a total of over 4 million new titles and editions. That’s 71 times more new fiction titles and editions as were published in 1995.”
We do need to acknowledge that the use of the computer in the identification of AI content is already a disputedly imprecise art. Even the best program, Pangram, manages to register false positives. However let us just accept the authors’ estimates: they say that 20% of their total 14,419 universe of books show substantial AI involvement* and 17.1% light AI content.

In this paper there’s a whole lot of hairy looking mathematics being applied to Amazon’s sales data, derived Mr McIlroy reveals from Kindle Unlimited data — which data is page-rental, not sales data. But hey, it’s hard to get these sorts of numbers! I’m sorry, but all this analysis just seems to me to be busy work. It’s what academics do. But the beam in their eye which they are of course overlooking is the fact that books are not like gravel. One load is NOT just the same as any another load. A gain in sales of this book does not mean a loss of sales for some other book. Book purchases are individual purchases, and I suspect that market dilution just isn’t a factor in the book business — that requires some sort of equivalency between the objects being sold. There is no equivalency between The Elements of Hydrodynamics and Shy Girl. I am perhaps willing to accept that some such equivalency may be available in some genre markets — though even there I’d need to see more real evidence. The authors are, to be fair, focussing on genre books plus what they call General/Contemporary Fiction. See their diagram for the categories they use in their analysis of Amazon “sales” data.
They claim to have detected a drop in sales of real books after the addition to Amazon of lots of AI books. “The study rules out the conclusion that the average revenue per book fell merely because the catalog filled up with a plethora of low-earning AI books, cumulatively depressing the average without deleteriously affecting the sales of human-written books. In other words, bloating the catalog with poorly selling AI books is not akin to the effect on individual students in a college class when weaker students enroll: there, the overall grades fall, but the weaker students’ results do not drag down the stronger students’ grades. Rather, this study shows that books with ‘no AI text’ (the good students) also, as a category, earned less than human-written books had earned before AI-content books began to compete with them. AI books’ entry into the market on a vast scale does not just overpopulate the bottom tier of sales; it allows at least some of the entrants to vie for the higher echelons of revenue generation, to the displacement of human-authored content. We note that this study does not attribute a reason for competition through scale; it merely documents the phenomenon.” . . . Our authors slide effortlessly from sales numbers (which are actually really numbers of pages read) to dollars earned. Just what “earned less than” really means is not clear. — Do bear in mind this is just a pre-print, an academic paper which has not been looked at (and by implication approved) by other scholars in the field. It may all end up having to be revised or abandoned.
Do not panic about market dilution — if your book’s any good, chances are it’ll find its market. Twas ever thus. A fixation on AI is once again obscuring the important issues.
_________________
* Substantial is defined by Chakrabarty et al. as greater than 25%. One might argue about the substantiality of 26% if one really knew what counted as evidence of AI use.

