acceptodds
Under review as a conference paper at ICLR 2027

Generative AI floods and dilutes the market for books

Abstract

Generative AI can produce book-length works of fiction at near-zero cost. These books are often dismissed as low-quality “slop” that buyers will ignore, and are assumed to carry little commercial weight. We test that assumption with full-text AI detection across 14,419 self-published genre-fiction books sold on Amazon from 2023 to 2026, matched to daily sales records through June 2026. None of these books disclose whether or not they contain AI-produced content. We find that books for which we detected substantial AI text ( 25%) make up a large share of the catalog but a smaller share of sales. Even so, they reach commercial scale, winning a growing share of sales over time and taking more of the scarce top-rank positions once held by books with no detected AI text. Over this period, the number of books with observed sales in a quarter grew 19.2-fold, while quarterly revenue grew only 8.9-fold. The market therefore added selling books faster than it added revenue, and revenue per selling book fell across most genres. Books with no detected AI text lost the most ground in genres with high AI diffusion, especially in genres with high Kindle Unlimited availability. Among top-selling books, those with substantial AI text contain more rare expressions found in existing books than do books with no AI text. Generative AI can thus reshape a creative market through scale rather than quality. Our results bear directly on the market-effect question at the center of the fair-use defense to copyright infringement.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Societal impacts of generative AI, Economics of AI, Creative markets, AI-generated text detection, Copyright, Fair Use

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_9AgTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper examines how the entry of AI-generated books reshapes the market for self-published genre fiction, shifting attention from individual books’ quality or sales to market dilution and outcomes for human authors. Combining full-text AI detection for 14,419 titles with daily sales and ranking data from a publisher panel, it finds that books with substantial detected AI content sell less individually but increasingly occupy top-ranking positions. Within the observed sample, the number of selling titles grows faster than revenue, while revenue per title also declines among books with no detected AI text, particularly in genres with greater AI penetration. These observational findings provide evidence relevant to the market-dilution question raised in Kadrey v. Meta, although they do not establish a causal effect of AI entry on human authors’ earnings.

Critical strengths

S1: The paper examines a fresh question on market-level consequences of AI-generated books, with clear relevance to copyright disputes and platform governance.

S2: Broad empirical coverage. The analyses connect AI diffusion, ranking dynamics, per-title revenue, author output, and textual overlap, providing complementary perspectives on market changes.

S3: The authors explicitly characterize the findings as associational and recognize that increased variety and lower prices may benefit consumers, leaving the net welfare effect uncertain.

Critical weaknesses

W1: The zero-price imputation is not rational. Section 2.1 replaces missing or non-positive prices with the title’s median valid launch-window price. However, zero prices may represent genuine free promotions rather than missing data. If free downloads are included in recorded unit sales, this procedure would assign positive revenue to transactions that generated none, potentially biasing revenue rankings and comparisons across AI-content groups. The authors should report the prevalence of zero-price observations, and test robustness to retaining valid zero prices.

W2: The growth comparison might not reflect the market dilution. Because the sample includes only titles released from 2023 onward, later quarters accumulate additional publication cohorts. Older titles may continue to count as selling books even as their revenue declines. Thus, the 19.2-fold increase in selling titles versus 8.9-fold revenue growth may partly reflect cohort accumulation and book-age effects rather than AI-driven competition.

W3: Figure 2D appears to contradict the displacement narrative. The caption and Section 3.2 state that fewer no-AI bestsellers remain in the Top-25 each quarter. However, the retention curve in Figure 2D appears to dip early and then rise steadily, ending above its starting level. Read at face value, incumbent no-AI books became more entrenched in the top ranks toward the end of the period, not less. The authors should reconcile the text with the figure, or explain why rising retention is still consistent with displacement.

W4: Limited reproducibility. The code has not been released, making it difficult to verify the data-processing pipeline, metric construction, and robustness analyses. Also, given the proprietary sales data, releasing the analysis code and a suitably anonymized or synthetic example dataset would substantially improve transparency and reproducibility.

Questions for authors

Please address the points that directly correspond to the concerns raised in the Weaknesses section. No new experiments are required. I would consider a higher rating if the authors could clarify these.

2 comments
AuthorsOpenReview ↗

Response to your 9AgT comments

Thanks for you review.

Criticism W1: We did not have any data points in our panel where the quantity was nonzero but it had a recorded price of 0 dollars. However there were very few (0.39%) missing post-release days which didn’t have any quantity or price. We didn’t change their quantity and kept it at 0 but used the title’s median valid price as a fallback when filling the price field. Since revenue is quantity X price, the days with imputed price contributed 0 dollars and the results are unchanged.

Criticism W2 : The 19.2X and 8.9X numbers are quarterly (Fig.3A). All comparisons in Fig 3B-3D rely on release cohorts over the same launch window. This includes any preorders + sales from day 1 to day 90. From 2023 to 2025 cohort the revenue / selling title with no detected AI text dropped from 23877 USD to 19739 USD which is 17.3%. Because based on our design each title can only contribute to its own release cohort the concern about older titles carrying over is not relevant here.

Criticism W3: The way we calculate Top25 depends on how many of its previous titles are retained and how many new titles enter. As mentioned in Section 3.2 (lines 212–215), the % of last quarter’s no-AI books that stayed in the Top 25 moved from 44.1% to 28.2% to 62.5%. But if we look carefully we see that by 2026 Q1 books having substantial AI text made up for 30.7% of entrants into the quarterly Top25 sets. This is a very large increase from the first comparable quarter which was only 3.3%. This also means that no-AI books held fewer spots overall. Infact, Fig 2B shows that share of no-AI book declined from 90.7% to 59.6% while books with substantial AI text increased from 1.2% to 23.3% of the spots. We will make this clearer in camera ready.

We present a simple example

MeasureEarlier quarterLater quarter
No-AI books in Top 252010
No-AI books retained next quarter88
No-AI retention rate40%80%
No-AI share of Top 2580%40%

Criticism W4 : ICLR does not strictly require authors to release code at the time of submission. Our data is huge in GBs and sharing it without due diligence and care might compromise anonymity. We will release all code, scripts and anonymized data for reproducibility upon paper acceptance.

We hope our responses have addressed your concerns. We would really appreciate an increase in overall rating and soundness

I thank the authors for their clarification. The paper’s large-scale data analysis quantifies the impact of AI adoption on the book market and draws attention to its implications. However, I still find the paper’s contribution limited and believe it may be a better fit for a conference such as SIGMETRICS or IMC. I therefore lean toward keeping my original score.

Reviewer_xbtXOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates the impact of generative AI on the self-published fiction market using 14,419 Amazon ebooks published between 2023 and 2026. The authors employ an AI-text detector to classify books and analyze their sales, revenue, market rankings, and textual overlap with existing publications. The results suggest that AI-generated content has expanded rapidly, gained commercial traction, and coincided with declining revenue per book, potentially contributing to market dilution.

Critical strengths
  1. The paper addresses an important question concerning the economic impact of generative AI on creative markets, with potential implications for copyright policy.

  2. The study combines full-text AI detection with longitudinal sales data, providing an interesting perspective on the commercial adoption of AI-generated content.

  3. The analysis reveals substantial growth in AI-associated publications, increasing market penetration, and correlations between AI exposure and declining revenue per book.

Critical weaknesses
  1. The work primarily presents an observational economic study using existing AI detection tools and conventional statistical analysis. Although the research question is relevant, the paper offers limited methodological novelty or advancement in machine learning.

  2. The central findings depend heavily on Pangram's detection accuracy. The reported false-positive and false-negative rates are derived from external evaluations, while the authors' own validation uses only 100 pre-2022 books. This is insufficient to establish classification reliability for contemporary fiction, particularly AI-assisted or heavily edited content.

  3. The main conclusion relies on correlations between increasing AI-associated publications and declining revenue per book. Other factors, including changes in consumer demand, pricing, platform algorithms, and publishing trends, are not adequately controlled. Consequently, the analysis does not convincingly establish that AI-generated books displaced sales of human-written books.

  4. The dataset excludes books whose Amazon sales rank never falls below 1.5 million and focuses on a single platform. This potentially introduces selection bias, particularly when investigating market flooding by low-selling AI-generated books. The extent to which the findings generalize to the broader publishing market remains unclear.

Questions for authors
  1. How robust are the findings to systematic AI-detection errors, particularly for partially AI-assisted and human-edited books?

  2. How does excluding low-selling books affect the estimated prevalence and market impact of AI-generated content?

1 comment
AuthorsOpenReview ↗

Response to comments by Reviewer xbtX

Critical Weakness 1: Our contribution is new empirical evidence about the economic effects of generative AI. This is very much under the scope of ICLR as the conference allows studying societal consequences of AI without proposing a new model architecture or training algorithm. There are also prior accepted ICLR papers in this vein

Rethinking Artistic Copyright Infringements In the Era Of Text-to-Image Generative Models ; Mazda Moayeri, Sriram Balasubramanian, Samyadeep Basu, Priyatham Kattakinda, Atoosa Chegini, Robert Brauneis, Soheil Feizi; ICLR 2024

Market effect in Fair Use is one of the most important questions in Generative AI training making ICLR a leading ML conference an ideal venue for this work.

Critical Weakness 2: Pre submission we looked into 100 self published books. Since then we have audited 6000 self published and 3000 traditionally published fiction books between 2020 January to 2022 December. Each year 2000 self published books+ 1000 traditional published books. All are Pre ChatGPT so can’t be AI. This corroborates the effectiveness and FPR of Pangram. We also test substantial-AI cutoffs from 10% to 50% (See Appendix B.2). At every cutoff, these books account for a smaller share of sales and revenue than of titles. Our classification includes all windows not labeled “Human Written,” rather than only those labeled fully AI-generated. We explicitly make this clear (line 130)

Critical Weakness 3: We are fully transparent in our paper that our findings are observational and not causal. The results in Fig 3B–D are all over identical launch windows and Figure 4 looks at differences across genres with varying AI exposure.Revenue for books with no AI text in them fell in 7 out of 8 genre except Fantasy/Supernatural/Horror where AI flooding was lowest.Additionally Fig 2B shows declining share of Top 25 slots that once were fully occupied by fully human written books. These findings are consistent with market dilution so while observational our findings are still valuable.

Critical Weakness 4: We exclude books whose sales rank never fell below 1.5M during the entire period because these titles have zero sales and contribute nothing to measured revenue. From the rest of the books, we draw a random sample across genres and release dates, without knowing their AI content ahead of time (Appendix A.1). This does not introduce selection bias within our target population of books since we use the same eligibility criterion to all books irrespective of whether they are AI or not. Last but not least we do not claim our findings will hold through each segment of the publishing market. We study a segment where it's easy for AI to percolate. Acquiring traditionally published titles would also be very expensive to do at a large scale and given the amount of gatekeeping and editorial discretion in those sectors the amount of AI books would be less prevalent.

How robust are the findings to systematic AI-detection errors, particularly for partially AI-assisted and human-edited books?

See response to Critical Weakness 2. We also ran the 14,419 books with Pangram 4.0 released more recently and there was no movement of books between the three bands. All books classified as Humans, Light AI and Substantial AI remained the same. This corroborates the robustness against AI detection errors.

How does excluding low-selling books affect the estimated prevalence and market impact of AI-generated content?

We want to clarify that we do not exclude low selling books but do not include books in our sample that have no recorded sales irrespective of its provenance whether its human or AI. See more details in Critical Weakness 4. Adding books with 0 sales cannot impact the market as they have no sales or revenue

We hope our responses have addressed your concerns. We would really appreciate an increase in overall rating and soundness.

Reviewer_MB6COpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper studies how books containing AI-generated text affect the market for self-published genre-fiction ebooks on Amazon. The authors run Pangram v3.3 over the full text of 14,419 titles released between January 2023 and March 2026, link each title to daily sales from a publisher-provided panel, and assign titles to three bands by the share of flagged windows (0%, ≤25%, and >25%). They report that books with substantial AI text make up 20.0% of titles but 12.1% of sales, that this share of sales grows over time, and that the catalog grew 38.3x while quarterly revenue grew 8.9x, lowering revenue per selling title even for books with no detected AI text. The authors interpret these patterns, together with higher author output after AI adoption and greater overlap with rare expressions from existing books, as evidence of market dilution relevant to the fair-use analysis in Kadrey v. Meta.

Critical strengths

S1. Important, well-motivated question. Whether low-cost AI production reshapes a creative market is timely, and the paper approaches it through sales rather than through ratings or quality judgments alone.

S2. Popularity. The study links AI detection to daily sales and rank data for the same titles. This combination is rarely available and is the paper's main asset.

S3. Full-text detection. Detection runs on entire books rather than short previews (cf. the 1,000-word samples in Reimers & Waldfogel, NBER w34777), which removes a known source of error. The threshold sensitivity analysis (Tab. 2) and the explicitly associational framing are also appreciated.

Critical weaknesses

W1. No baselines; closest prior work not engaged. The numbers are never compared against another market or a pre-AI period, so it is unclear whether this market behaves unusually. Crucially, Reimers & Waldfogel (NBER w34777), cited only for the finding that AI books sell less, also apply Pangram to Amazon ebooks and conclude that AI has not displaced pre-LLM authors and raised consumer surplus by ~7% in 2025. The paper should explain why its conclusion differs. Baselines also exist for the paper's own metric (AI share of supply vs. share of consumption), e.g., >50% of daily uploads vs. 1–3% of streams on Deezer, and ~50% of new web articles vs. 14% of Google Search results (Graphite). Against these, books (20.0% -> 12.1%) convert AI supply into sales unusually well, which would be a notable finding if it holds. Finally, the panel lacks pre-2023 cohorts and a control group (e.g., traditionally published ebooks in the same panel).

W2. Detector validity is assumed, not tested. The cited error rates (0.04% overall; 0.01% on long-form creative writing) are document-level, while the no-AI band requires zero flagged windows across an entire novel (~150–250 windows of 512 tokens). At a per-window false positive rate of 0.04%, ~7.7% of fully human 200-window books would leave the no-AI band; at 0.1%, ~18%, close to the 17.1% of titles in the light band. Further, (a) S counts every window not labeled "Human Written", so AI-assisted labels (i.e., AI-edited human text) count as AI text; (b) S is a share of variable-length windows, not of text (cf. App. A.2); (c) the pre-2022 audit does not define "flagged", has a 95% upper bound of 3.6%, and likely overlaps with the detector's training data; and (d) the Pangram version is not specified (3.3 and 3.3.1 segment long documents differently). I would like to see window-level false positive rates on post-2022 human genre fiction, agreement with at least one independent detector, and the main results re-run using AI-generated windows only.

W3. No qualitative evidence of what an "AI book" is. The reader never sees what it means for a book to be majority AI-written. I recommend adding (a) the distribution of S (is it bimodal, as Reimers & Waldfogel find on previews?); (b) chapter-level maps of where flagged text sits (e.g., recaps and bonus chapters vs. throughout); (c) 5–10 anonymized case studies across the score range, including the top-revenue title in Fig. 4F; (d) price, length, KU enrollment, and ratings by band; and (e) a random sample of rare expressions to complement the four selected in Sec. 3.5.

W4. The headline growth contrast is produced by the panel design. The panel contains only titles released after January 1, 2023, indexed to 2023 Q1. Hence, the catalog grows by construction, and selling titles outgrow sales because older titles keep selling a few copies as their volume decays. I simulated this design with no AI and no dilution (each title's sales depend only on its own age and appeal): with releases tripling, the catalog reaches ~24x, selling titles 12–21x, and sales 6–11x by 2026 Q1, while revenue per selling title falls by half or more. The abstract's contrast (19.2x vs. 8.9x) is therefore expected without any AI entry. Relatedly, under the stated uniform sampling across release dates, the 2023 Q1 cohort should be ~7.7% of the sample, whereas 38.3x implies ~2.6%. The rank filter (<1.5M) also admits different parts of the sales distribution as the catalog grows.

W5. Revenue is mismeasured for Kindle Unlimited titles. Revenue is computed as units × consumer price, but KU borrows are paid per page read from the KDP Select Global Fund, and the panel cannot separate borrows from purchases (App. A.2). The KU result (Sec. 3.3, Fig. 4B) rests on this measure. Moreover, because the per-page rate equals the fund divided by total pages read, any new KU title (AI or not) lowers the payout of all others by construction, while the fund itself grew ~40% over the window ($47.9M in January 2023 -> $66.9M in May 2026).

W6. No economic model or identification strategy. Most evidence is in shares (sales, revenue, Top-25 slots), which move with composition even when no sales are diverted. I recommend (a) levels of no-AI sales against a control group, decomposed into market expansion and substitution, as in Goldberg & Lam (2025); (b) a null for Top-25 slots, where each band's expected share equals its share of titles on sale (notably, Fig. 1B suggests AI books are under-represented at the top, and Fig. 2D shows no-AI retention rising to ~62%); (c) within-author comparisons, since successful authors who adopt AI leave the no-AI pool and lower its average without any dilution; and (d) a demand model with an outside option (e.g., the nested logit in Reimers & Waldfogel) together with plausibly exogenous variation in AI entry (e.g., frontier model releases, or the KDP cap of three new titles per day introduced in September 2023).

W7. Inconsistencies in the author panel. If the 15,000 titles are a sample of a much larger population, per-byline output is observed only through sampled titles, yet Fig. 4D shows bylines with 33 sampled titles within a year of adoption. In addition, "newly observed" authors have zero pre-adoption output by construction, and conditioning on authors who released further AI books (287 of 385) conditions on the outcome. Please report the population size and sampling fraction, and reconcile the 824, 385, and 466 byline counts.

W8. Rare-expression analysis. The Google Books index plausibly contains other AI-generated books from 2023–2026, so overlap may reflect shared model phrasing rather than borrowing from human-written books; restricting the index to pre-2022 books would resolve this. The award-winner comparison is confounded, since passages from prize-winning novels are widely quoted online and fail the "absent from the web" criterion. The slope difference between groups is also not significant (p = 0.063).

W9. Framing and provenance. (a) The legal discussion cites only Kadrey v. Meta and omits Bartz v. Anthropic, in which Judge Alsup rejected the market-dilution theory; many harmed titles also postdate the training data of the models used to write competing titles. (b) KDP has required disclosure of AI-generated content to Amazon since September 2023, so "none of these books disclose" should read "none disclose to readers." (c) The Falade anecdote rests on an unproven allegation that the author denies; I recommend removing it. (d) Please disclose the data provider's role, whether unit sales are reported or imputed from rank, and how full texts were extracted from DRM-protected library and Kindle copies.

Questions for authors

Q1 (W1). How do you reconcile your findings with the conclusion of Reimers & Waldfogel that AI has not displaced incumbent authors?

Q2 (W1, W4). Can you add 2019-2022 cohorts built with the same rules, so that the growth indices and revenue per title have a pre-AI benchmark?

Q3 (W2). What is the window-level false positive rate on post-2022 human-written genre fiction, and what is the full distribution of S on the 100-book audit set?

Q4 (W2). What share of flagged windows in each band is labeled AI-generated rather than AI-assisted, and do the results hold using AI-generated windows only?

Q5 (W4, W7). How many sampled titles fall in each release quarter, and how large is the population sampled from?

Q6 (W5). Are daily unit sales reported or imputed from sales rank, and how are KU borrows counted in units and in revenue?

Q7 (W6). How many authors moved from the no-AI band to the substantial band during the study, and what were their pre-adoption sales?

3 comments
AuthorsOpenReview ↗

Response to comments by Reviewer MB6C

Dear Reviewer MB6C

Please see our response to all Critical Weaknesses one by one. We took significant effort to address almost all of your points and would really appreciate an increase in overall rating and soundness.

W1. No baselines; closest prior work not engaged.

Reimers and Waldfogel study the number of books released by incumbent authors, which is conceptually distinct from our measure of sales per new book. They find that authors active before LLMs continued to release roughly the same number of books classified as non-AI, while we document changes in sales per selling title at launch. Their findings on incumbent output and consumer surplus thus speak to different research questions, and they don’t use the same notion of “displacement” like us. Even if authors continued to write books, it doesn't mean that they earned less. In terms of other differences, they use AI text detection on 1,000-word previews from Amazon whereas we perform full-text AI detection and can distinguish cases of light AI use from cases of substantial AI use. Short previews can miss AI text elsewhere in a book.

Our analysis focuses on sales of predominantly self-published genre-fiction ebooks. Deezer and Graphite provide comparisons from other contexts. Deezer compares the share of fully AI-generated tracks among new uploads with their share of streams across the entire catalog. Graphite reports the share of primarily AI-generated articles among newly published articles and, in a separate study, among articles appearing in Google search results. So, it is hard to generalize those findings to books as the economics of music and news is different.Although our empirical analysis focuses on comparing launch sales of books released in 2025 to those released in 2023, future work could study changes in sales relative to a pre-LLM benchmark; however, that is beyond the scope of the current paper.

W2. Detector validity is assumed, not tested.

Pre submission we looked into 100 self published books. Since then we have audited 6000 self published and 3000 traditionally published fiction books between 2020 January to 2022 December. Each year had 2000 self published books+ 1000 traditional published books. Not a single window was flagged as AI giving an observed book-level FPR of 0%. We additionally confirmed with the Pangram team that none of these 9000 books were in Pangrams training data. We appreciate reviewer concerns that anything between 7.7% to 18% of the fully human-written novels could be incorrectly placed in our light-AI category. But we find no evidence of such false positives in our 9,000-book audit. We also already add evidence by varying the substantial-AI cutoffs from 10% to 50% (See Appendix B.2). At every substantial AI cut off, these books account for a smaller share of sales and revenue than of titles.

Pangram version used was 3.3.2 Post 2022 we do not have gold labels. But assuming traditionally published genre fiction books from Big 5 do not have any AI in them we looked into 500 books from PRH, MacMillian, Simon and Schuster, Hachette and Harper Collins (100 each). Not a single window was flagged as AI giving an observed book-level FPR of 0% We also calculated window level agreement with GPTZero (second most accurate AI detector) for 200 randomly sampled books (100 human+ 100 substantial AI) and found a very high cross detector agreement of 96.7%.

Results change very little if we just consider AI generated windows and ignore AI-assisted windows

MeasureOriginalAI-generated only
Substantial-AI books2,880 (20.0%)2,692 (18.7%)
No-AI books9,068 (62.9%)9,304 (64.5%)
Substantial-AI share of sales10.0%8.9%
Substantial-AI share of revenue9.6%8.5%
Substantial-AI share of 2026 releases36.8%34.9%

W3. No qualitative evidence of what an "AI book" is

S is indeed bimodal as Reimers and Waldfogal. Most books sit on one end. 63% have no detected AI text and 6.7% are above 90%. Among books detected as substantial AI, two thirds of them are in the range of 25% - 90%. Previews would miss this. Looking at just chapter 1, 28% of the books would not be counted as substantial AI as the AI text comes later. We also looked at 10 substantial AI generated books with the most earned revenue. We find the AI text appears throughout the body and not just recaps or bonus.

Here are also 5 anonymized case studies

CaseS (%)Ch. 1 AIGenrePriceRatingRevenue
10.00%Fantasy romance$8.704.6$6.6M
217.356%Hockey romance$4.974.3$0.94M
3 (Fig. 4F)45.7100%Women's fiction$7.994.5$0.66M
460.29%Dystopian$11.994.6$0.68M
593.265%Mafia romance$7.994.4$0.27M
AuthorsOpenReview ↗

Response to comments by Reviewer MB6C (Part 2)

W4. The headline growth contrast is produced by the panel design.

The 19.2X and 8.9X numbers are quarterly (Fig.3A). All comparisons in Fig 3B-3D rely on release cohorts over the same launch window. This includes any preorders + sales from ONLY day 1 to day 90. From 2023 to 2025 cohort the revenue / selling title with no detected AI text dropped from $23877 to $19739 which is 17.3%. Because based on our design each title can only contribute to its own release cohort the concern about older titles carrying over does not apply here.

We are really not sure how the reviewer ran the simulation to obtain those numbers. We also want to add that Uniform random sampling does not imply equal numbers of books per quarter. So reviewers 2023 Q1 cohort should be ~7.7% of the sample is not correct

In regards to the rank filter point, we exclude books whose sales rank never fell below 1.5M during the entire period because these titles have zero sales and contribute nothing to measured revenue. From the rest of the books, we draw a random sample across genres and release dates, without knowing their AI content ahead of time (Appendix A.1). This does not introduce selection bias within our target population of books since we use the same eligibility criterion to all books irrespective of whether they are AI or not.

W5. Revenue is mismeasured for Kindle Unlimited titles.

The revenue figures are estimates and not direct author royalties and Appendix A.2 mentions already that by design our panel cannot separate KU activity from purchases. But KU-enrolled books also remain available for purchase and simply being enrolled does not mean every reader borrows the book. Even NYTimes bestsellers are on Kindle Unlimited. For instance Margaret Atwood's, The Handsmaid Tale is on KU but that doesn't mean a reader won't buy it. Our data provider also calibrated its KU estimates with another publisher and checked them against realized June KU sales for several hundred titles, reporting an error below 3% on that sample. This supports the accuracy of the sales estimates. We also want to add that Figure 4B shows the same pattern in sales shares, independently of the revenue calculation. Last but not least, adding a new KU book does not automatically reduce other authors’ earnings.The payments typically rely on how much readers read and the total money available. So even if the fund grows any individual book can still earn less.

Please refer to https://kdp.amazon.com/en_US/help/topic/G201541130 for better understanding.

W7. Inconsistencies in the author panel.

We thank the reviewer for these comments. Upon checking, we confirmed that two bylines each published as many as 33 distinct titles within twelve months of their first observed substantial-AI release in our sample. The different numbers of bylines (i.e., 824, 385, and 466) refer to distinct groups. Specifically, 824 is the number of bylines with a first observed substantial-AI release and at least three months of follow-up, 385 is the number that published at least one further substantial-AI title, whereas 466 is the number with twelve months of sales follow-up and positive revenue from substantial-AI titles. We agree with the reviewer that conditioning on subsequent AI releases does not allow us to conclude that producers increase their productivity after AI adoption. In fact, among 178 previously observed bylines for which we can match twelve-month windows before and after their first substantial-AI release, we find that only 69 increase their output. In addition, the average number of sampled titles these bylines published per month decreases slightly from 0.350 to 0.320. We will therefore change our claim regarding productivity and instead emphasize that a subset of producers remain highly active after their first observed substantial-AI release. Finally, we will report the eligible title population and sampling fractions using our original sampling records.

AuthorsOpenReview ↗

Response to comments by Reviewer MB6C (Part 3)

W6. No economic model or identification strategy.

We performed some additional analyses to answer reviewer concerns.

(a) Sales levels. Our paper already reports sales levels for books with no detected AI. Over the first 91 days after release, excluding preorders, recorded units per selling title decreased by 29.4%, from 5,099 in 2023 to 3,601 in 2025. Similarly, revenue per selling title decreased by 19.9%. These declines cannot arise mechanically from adding AI books to the sample, since both measures concern only books with no detected AI.

(b) Top-25 rankings. We conducted the exercise suggested by the reviewer, comparing the observed share of Top-25 slots held by books containing substantial AI-detected text with the share expected if all on-sale books with valid ranks had an equal chance of reaching the Top 25. We find that substantial-AI books held a smaller share of slots than expected (10.1% versus 13.3% in 2025 Q1, and 16.1% versus 18.8% in 2026 Q1), although their share increased over time. The 62.5% retention statistic highlighted by the reviewer concerns books with no detected AI that were already in the previous quarter's Top 25. It does not contradict an increasing share of AI-containing books among new entrants. Separately, using fixed commercial benchmarks, we find that the fraction of substantial-AI releases achieving an Amazon sales rank of 5,000 or better for at least 46 of their first 91 days increased from 7.8% in 2023 to 18.3% in 2025.

(c) Within-author comparisons. We also conducted the exercise suggested by the reviewer. Restricting attention to 73 continuing bylines with selling titles containing no detected AI in both 2023 and 2025, and holding their relative weights fixed at their 2023 publication frequencies, we find that launch revenue per selling title with no detected AI decreased by 15.2%. Restricting further to the 35 bylines with no observed substantial-AI release through March 2026, the decline was 17.9%. Thus, these declines cannot be explained by successful bylines dropping out of the comparison or changes in their relative weights. (d) Demand model and identification. A nested-logit demand model with an outside option does not by itself identify substitution effects or provide identification by itself. Unfortunately, we don’t have a clean policy shock similar to Goldberg and Lam. Nor do the suggested alternatives provide straightforward identification: frontier-model releases affect production across the market, while Amazon's KDP cap on new titles applies to both AI and non-AI books.

Taken together, our results show that a growing fraction of books containing substantial AI-detected text achieve commercial success, while recorded units and revenue per selling title with no detected AI decline, including among continuing bylines with fixed weights. These patterns cannot be explained simply by changes in market shares or the composition of publishing identities. Estimating the number of sales diverted from existing books by AI entry is a separate causal question. It is not necessary to establish the observed market dilution documented in our paper.

W8. Rare-expression analysis

Our rare-expression analysis isn’t aimed at showing whether expressions are borrowed from human-written books. Rather, it's a proxy to understand differences across different types of books. In fact, footnote 6 already reflects this. Restricting the index to pre-2022 books would have been possible, but this would ignore a significant number of books from the last three years and would also require a pre-2022 Infinigram corpus. Infinigram itself was released in 2024, and as authors, we do not have the computational infrastructure to build and maintain a 4.7-trillion-token index. We understand that certain award-winning passages are widely quoted online, but we perform rare-expression analysis on entire books, not just a few passages. We further inspected all 200 award-winning books and found that very few are bestsellers. Among these 200 books, the highest amount of exposure from Goodreads quotes was 4% of the total book. Therefore, we do not think this would meaningfully impact the results. We are already transparent about the difference in slopes not being statistically significant.

W9. Framing and provenance.

Because this paper is about market dilution (a newly proposed theory in copyright law) it makes more sense to talk about Kadrey vs Meta and not about Bartz v Anthropic here. Most LLM companies have actively continued to pretrain their models so we definitely cannot say with certainty that many harmed titles postdate training data of models. Unlike downloading from shadow libraries many providers have found a way to scan and digitize books now. We will reframe the disclosure statement to reflect disclosure to readers. Unit sales are estimated from sales rank.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.