Venture capital firm Andreessen Horowitz (a16z) recently cited the research report “Generative AI floods and dilutes the market for books,” published in July. The report found that books containing detectable AI-generated text accounted for approximately 40% of observed sales in the Amazon self-publishing market.
Research Methods and Core Data from Stony Brook, Columbia, Michigan, and MIT
The study was jointly conducted by Stony Brook University, Columbia University, the University of Michigan, and the MIT Initiative on the Digital Economy. It used 14,419 self-published genre-fiction books sold on Amazon between 2023 and 2026 as its sample. The methodology combined full-text AI detection with daily sales data through June 2026. Works in which AI-generated content accounted for more than 25% of the full text were defined as containing “substantial AI text.”
The core quantitative results are as follows:
Share of sales: Books containing any detectable AI-generated text accounted for approximately 40% of total observed sales
Penetration among the Top 25 bestsellers: Works with substantial AI text made up at least 10% of the Top 25 bestsellers across all genres covered by the study
Top 25 in the science fiction, adventure, and dystopian genres: Works with substantial AI text made up nearly 40%
Number of books with sales: Increased approximately 19.2-fold since 2023
Quarterly revenue growth: Approximately 8.9-fold
Overall catalog size: a16z observed that the catalog expanded nearly 40-fold
The study results show that although books containing substantial AI text still had lower average sales than works with no detected AI text, they have already formed a real market rather than representing a negligible fringe phenomenon.
a16z, Bookslop, and the “Volume Approach”
a16z points out that AI is changing the publishing market not by enabling individual books to achieve higher quality, but by reducing the marginal cost of producing books to nearly zero. Under the traditional publishing model, authors spend months or even years completing a novel, so they must maximize the likelihood that each work will succeed. Generative AI makes the business model of “publish at scale and let the market choose” possible: the hit rate of AI-generated works may be lower, but if enough shots are taken, some may find readers willing to pay. a16z calls this strategy the “volume approach.”
Viewed through the overall figures, the number of books in the catalog grew nearly 40-fold, while market revenue increased only approximately 9-fold. This means that the massive increase in supply did not drive a corresponding expansion in demand; instead, it diluted readers’ limited attention and purchasing power. a16z uses the coined term “Bookslop” to describe these AI-generated publications that replace quality with scale.
Kindle Unlimited’s Market-Dilution Effect and the Fair Use Controversy
The study found that in genres with higher AI-book penetration, human-authored works with no detected AI text also lost a more noticeable share of the market. This phenomenon was especially pronounced in markets with higher Kindle Unlimited penetration. The finding reveals a competitive mechanism of “dilution” rather than “direct replacement”: although human-authored works still achieved higher average sales, the explosive growth in supply reduced the likelihood that readers would discover them. The researchers precisely described this dynamic with “floods and dilutes” in the paper’s title rather than “replaces.” Structurally, it closely resembles the content-farm problem that previously emerged in the SEO field—the dimensions of competition are gradually expanding from “who writes best” to include the new question of “who can produce the most.”
Regarding copyright disputes, the study also found that among best-selling books containing substantial amounts of AI-generated text, the degree of overlap in distinctive language between those works and existing books increased as revenue rose. The study’s authors explicitly noted that this does not prove that AI-generated books plagiarized specific works. However, it is directly related to the market-effect factor in Fair Use determinations under U.S. copyright law: the potential impact of the use on the market for the original work is one of the important bases for determining Fair Use. If AI-generated works can produce large numbers of competing works at extremely low cost and actually take rankings and revenue away from existing works, the legal controversy over “market substitution” will no longer remain merely theoretical.
Frequently Asked Questions
What was the sample size and methodology of this study, and what was the standard for defining “substantial AI text”?
The study was jointly conducted by Stony Brook University, Columbia University, the University of Michigan, and the MIT Initiative on the Digital Economy. The sample consisted of 14,419 self-published genre-fiction books sold on Amazon between 2023 and 2026. The methodology combined full-text AI detection with daily sales data through June 2026. The study defined works in which AI-generated content accounted for more than 25% of the full text as containing “substantial AI text.”
The book catalog grew 40-fold while the market grew only 9-fold. What was the practical impact on human authors?
The study found that in genres with higher AI-book penetration, human-authored works with no detected AI text lost a more noticeable share of the market. This was especially pronounced in markets with higher Kindle Unlimited penetration. The explosive growth in supply diluted readers’ limited attention, so although human-authored works still had better average sales performance than AI-generated books, the likelihood that readers would discover works by human authors also declined.
What impact does the overlap in language between the best-selling AI books identified by the study and existing works have on copyright law?
The study found that among best-selling books containing substantial amounts of AI-generated text, the degree of language overlap with existing books increased as revenue rose. The study’s authors explicitly noted that this is directly related to the “market effect” factor in U.S. Fair Use determinations. If AI-generated works can produce large numbers of competing works at extremely low cost and seize market share, the controversy over “market substitution” will no longer be merely theoretical. The study does not amount to a finding of plagiarism, but it could make existing copyright disputes in the publishing industry more complex.
