paper web signal

Study: AI books gain Amazon share as catalog outpaces revenue

TL;DR

  • A new arXiv study of 14,419 self-published genre-fiction titles found Amazon's catalog with observed quarterly sales grew 19.2-fold from 2023 to June 2026 while quarterly revenue grew only 8.9-fold.
  • Books with more than 25% detected AI text make up 20% of the sample but earn only 12.1% of sales and 11.3% of revenue, yet capture more top-rank positions over time.
  • In genres where more titles are on Kindle Unlimited, the sales-share lead held by non-AI books is 8.5 percentage points smaller than in less-KU-heavy genres.

The interesting thing about the new empirical study of AI-authored fiction on Amazon is that it does not need to argue AI books are winning to make its point. It just needs to show the pie is being cut thinner. A group from Stony Brook, Columbia Law School and the University of Michigan and MIT set out to demonstrate that in a paper posted to arXiv, running full-text AI detection over 14,419 self-published genre-fiction titles and matching them to daily Amazon sales through June 2026.

The headline is the divergence. Over the period, the number of books with observed quarterly sales grew 19.2-fold while quarterly revenue grew only 8.9-fold. Titles with more than 25% detected AI text account for 20% of the sample but earn only 12.1% of sales and 11.3% of revenue, so per-title, AI books still under-earn their non-AI peers. What is changing is share and rank. According to the paper, substantial-AI books are taking a growing slice of sales and showing up in more of the scarce top-rank positions once held by fully human-written books.

Subscription economics look like an accelerant. In genres where more titles are available on Kindle Unlimited, the lead non-AI books hold in sales share is 8.5 percentage points smaller. The authors also flag a stylistic tell: the top-selling AI books are, in their words, saturated with distinctive rare language from existing books, at 45.0% versus 37.7% for the non-AI comparators, an overlap that correlates positively with revenue for the AI titles but not for the human ones.

The honest caveat is that this is a preprint, revised late July 2026, covering one platform and one segment, self-published genre fiction on Amazon. It does not size AI dilution across trade publishing, non-fiction, or non-English markets, and it does not tell you how concentrated the AI supply is among a handful of prolific accounts versus a long tail. Detection at scale is a claim, not a settled instrument, so take the specific percentages as reported, not settled.

What it does give operators is a real per-title revenue trend to plan against. If you run a self-publishing platform, an author-services business, or a subscription reading product, the working assumption for the rest of 2026 should be that catalog growth and reader spend have decoupled, and that the discovery and payout systems tuned for a smaller pool are the surface most exposed.