For those of us who live in the world of editing, writing, and publishing, the ability of the newest generations of AI tools to produce rivers of grammatically correct prose is a deep shock. But is there actual evidence on AI and the quality of published work? Imke Reimers and Joel Waldfogel offer a starting point in their research paper, “AI and the Quantity and Quality of Creative Products: Have LLMs Boosted Creation of Valuable Books?” (NBER Working paper 34777, May 2026, also available here).
The obvious questions here are how to measure quantity and especially quality of new books. Quantity is easier. The authors offer some highly suggestive evidence: “Using data on new books offered for sale at Amazon, we document that the number of new titles appearing each month nearly tripled between 2022 and late 2025 and rose by a factor of nearly ten in some categories; and the increase in new titles 2022-2025 coincides with both the diffusion of LLMs and the incidence of detected AI in books.”
The Amazon data also provides information on the number of ratings for any given book, the sales rank, and the number of “stars” for any given book. These kinds of measures offer a data-driven way of getting at the question of quality. In addition, Reimers and Waldfogel describe:
[W]e take a direct approach employing AI detection on over 50,000 randomly selected titles. We document the growth of AI usage and compare the quality of “AI books” and “non-AI books” (books with and without detected AI). We have four findings. First, the timing of AI growth tracks the growth in releases: Detected AI use is roughly zero through 2022, rises to 30 percent in 2023, to 45 percent in 2024, and surpasses 60 percent during 2025. Second, AI books garner substantially less usage per title. Most have very little usage, and a modest share is somewhat used, both when measured by the number of ratings and eventual sales ranks. AI books are also worse in the sense of having lower star ratings. Third, the human-AI usage gap narrows substantially between 2023 and 2025. Finally, the number and quality of human-authored books has remained stable.
Perhaps unsurprisingly, authors with relatively low sales and rankings in the past much more likely to take up the AI tools.
At some level, the results of the study are unsurprising. AI seems to have led to many more books, but the average quality of these AI-assisted or -generated books is lower than the previous average. The authors write: “The effects of this influx [of new books] on consumer welfare depend on the quality of the additional books. The average quality of new books has fallen with the LLM-induced influx, and books with detected AI are substantially worse than human-authored books, so that much of the new work is of little value to consumers. Still, the LLM influx has delivered some books in the middle range of the usage/quality distribution …”
Consumers vary in their tastes, and more books means more to choose from. At least so far, AI-generated fiction is not hitting the top of the measures of quality. But it’s worth remembering that a large proportion of book says are in the middle range of the quality distribution, including categories like romance, SF, and mystery. Personally, I read a lot of mysteries, and while I try to pick high-quality ones, I read a fair number that could be fairly characterized as in the middle range of quality; indeed, some of the mysteries I read could probably be improved with an assist from AI.
This broad pattern probably characterizes a lot of AI-assisted work in creative areas. It can produce vast quantities very quickly, which will mostly be low-quality, but some of it will reach middle-quality. If the low-quality output can be ignored at low cost, a greater choice among middle-quality output is a modest social gain.
