Pith. sign in

REVIEW 3 major objections 4 minor 30 references

AI-generated genre fiction has reached commercial scale and is diluting the market for human-authored books, a study of 14,000 Amazon ebooks finds.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-04 04:09 UTC pith:SDIT5OCE

load-bearing objection The best evidence so far that AI-generated books are diluting the market at scale, though the detector and proprietary panel keep the headline numbers provisional rather than settled. the 3 major comments →

arxiv 2607.20349 v3 pith:SDIT5OCE submitted 2026-07-22 cs.CL cs.AIcs.CY

Generative AI floods and dilutes the market for books

Tuhin Chakrabarty , Xinyue Liu , Jane C. Ginsburg , Paramveer Dhillon This is my paper
classification cs.CL cs.AIcs.CY
keywords generative AIbook marketmarket dilutionself-published fictionAmazoncopyright fair useAI text detectionrevenue per book
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to establish that generative AI books, though dismissed as low-quality 'slop,' have reached commercial scale in self-published genre fiction on Amazon and are diluting the market for human-authored books. Using full-text AI detection on 14,419 ebooks matched to daily sales, the authors find that books with substantial AI text now account for about a fifth of observed sales, take a growing share of scarce top sales ranks, and coincide with falling revenue per selling book across most genres, including for books with no detected AI text. The force is volume: the released catalog grew 38-fold while quarterly revenue grew about 9-fold, so more books compete for a revenue pool that grew much more slowly. The authors argue the findings bear directly on the market-effect question at the center of copyright fair use disputes over AI training.

Core claim

The paper's central claim is that a creative market can be reshaped by scale rather than quality: books with substantial AI text (more than 25% of text windows flagged as not human-written) sell less per title on average, but because they are produced in vast numbers they win a growing share of sales and top-ranked positions, and their entry coincides with lower revenue per book even for books with no detected AI text. Across the study period, the cumulative released catalog grew 38-fold and the number of selling titles grew 19-fold, while quarterly revenue grew about 9-fold, so revenue per selling title fell in most genres. By the second quarter of 2026, AI-heavy books' share of observed sa

What carries the argument

The central object is the full-text AI-detection score: for each book, the percentage of text windows that a commercial detector labels as not 'Human Written,' which places it into one of three bands (no AI text, light AI text up to 25 percent, substantial AI text above 25 percent). The dilution analysis compares books over a fixed launch window (preorder days plus 90 days after release) and tracks revenue per selling title across release cohorts, along with the share of weekly Top-25 rank slots constructed from Amazon sales ranks. A secondary mechanism measures rare-expression overlap: the share of a book's tokens covered by five-word-or-longer expressions that appear in at most a handful o

Load-bearing premise

The load-bearing premise is that the commercial full-text detector correctly recognizes which passages are AI-generated in this specific corpus of self-published genre fiction; if it mistakes human-written prose for AI text, or misses humanized AI text, the band assignments and all dilution estimates change.

What would settle it

Take a sample of self-published genre fiction with known authorship—say, books written before 2023 and books whose authors certify no AI use—run the same full-text detector on them, and check whether false positives concentrate in certain genres or correlate with sales. If most flagged books are actually human-written, the dilution result would not hold.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If correct, market dilution from AI-generated books is observable in a real mainstream book market, not hypothetical.
  • Human-authored books' per-book revenue falls most where AI diffusion is highest, so the harm tracks AI exposure rather than a general trend.
  • AI-heavy books increasingly occupy scarce top sales ranks, so their effect is not confined to the long tail.
  • The results give courts an empirical basis for the market-effect prong of fair use analysis in AI-training litigation.
  • Among successful books, AI-heavy titles draw more heavily on distinctive language from existing books, and that reliance grows with revenue.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the mechanism generalizes, other markets with near-zero entry costs and shared attention pools—music streaming, stock imagery, short-form writing—could show similar crowding-out driven by volume rather than quality.
  • The paper cannot separate nondisclosure from other reasons for AI books' success; an experiment that varies AI-content labels on a platform could measure how much disclosure alone changes demand.
  • Because the results rest on a single detector, re-estimating with an ensemble of detectors or a human annotation sample would reveal how much of the dilution signal depends on detector error.
  • The byline-based author analysis likely understates concentration, as one author publishing under many pen names is split across identities; the true concentration of output and revenue may be higher.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper analyzes 14,419 self-published genre-fiction ebooks sold on Amazon from January 2023 to March 2026, matching full-text AI-detection scores from Pangram v3.3 to a proprietary daily sales panel. Books are classified into no-AI, light-AI, and substantial-AI bands (the latter >25% of windows flagged as not 'Human Written'). The paper reports that substantial-AI books grew from near zero to roughly 20% of observed sales by 2026 Q2, took a growing share of constructed Top-25 rank slots, and accompanied catalog growth (38.3x) that far outpaced revenue growth (8.9x), with revenue per selling book falling for no-AI books in most genres. It also reports that top-selling substantial-AI books have higher rare-expression overlap with existing books than no-AI bestsellers, with overlap rising in revenue. The paper interprets these patterns as evidence of market dilution relevant to copyright fair use.

Significance. If the measurement assumptions hold, this is one of the first market-level, full-text-based documentations of AI output reaching commercial scale in a creative market. Strengths include the use of full-text detection rather than previews, fixed launch-window comparisons across cohorts, explicit robustness of the 25% cutoff, and unusually transparent caveats about the observational nature of the comparisons. The code is released, aggregate data are promised, and the findings bear directly on the market-effect prong of fair use. However, the central estimates are conditional on an externally validated but not corpus-validated detector and on a proprietary panel, so the specific magnitudes and the legal-policy interpretation require additional support.

major comments (3)
  1. [§4.3 and §4.10] All band assignments—no/light/substantial AI text—are derived from Pangram v3.3 window labels. The only validation cited is the vendor's aggregate N-weighted FPR=0.04% and FNR=0.14% on 566,745 texts from 15 domains. This does not establish performance on self-published genre fiction, where formulaic style, non-native English, and humanized AI text are common. A low per-window error rate can still reclassify a meaningful share of books if errors correlate with genre, byline, or sales tier. The robustness table in §4.10 re-thresholds the same scores and cannot detect label error. Please provide corpus-specific validation (e.g., apply Pangram to clearly human-written pre-2022 self-published titles and to known AI-written titles in each genre, reporting confusion matrices by genre and sales tier). Absent that, the paper should consistently speak of 'Pangram-flagged' books and soften claims t
  2. [Title, Abstract, and §3] The central interpretation is causal: the title says AI 'floods and dilutes' the market; the abstract and discussion use 'depressive effect', 'displacement', and 'crowding out'. Yet the comparisons in Figures 3–4 are observational, as the authors themselves note in §3 ('Our comparisons are associational'). The decline in revenue per no-AI book and the rank-slot shifts could be driven by unmeasured confounders such as changes in Amazon ranking algorithms, reader preferences, or human competition. The genre-exposure and Kindle-Unlimited heterogeneity analyses are associational and do not identify a displacement mechanism. Because the legal relevance depends on AI-produced books crowding out human-authored sales, please either add a stronger identification strategy (e.g., an event study or platform-policy discontinuity) or reframe the conclusions as documenting correlational patterns consis
  3. [§4.1 and Fig. 3a] The 'catalog' growth multiple (38.3x) and all band shares are computed on a sample conditioned on sales activity: titles whose Amazon rank never fell below 1.5 million were excluded. Thus 'catalog' denotes the observed, potentially selling universe, not the full catalog. If the never-selling tail grew differently, the growth multiple and market shares could be biased. The paper should clarify whether sampling weights are used to project to the panel universe and should report sensitivity of Figures 1–3 to this exclusion, or at least state explicitly that all figures describe the rank-active sample rather than the full release population.
minor comments (4)
  1. [§2.5, Fig. 5B caption] The interaction between AI status and log10(revenue) has p=0.063, which is not significant at the 0.05 level. The abstract's phrase 'a gradient we do not detect for books with no AI text' is a one-sided description; the statistical evidence for a different slope across groups is marginal. Please report the interaction test prominently and soften the wording accordingly.
  2. [Throughout] The manuscript often writes 'books with substantial AI text' where the operational variable is 'books detected/flagged as having substantial AI text by Pangram'. Adding 'detected' or 'flagged' throughout would prevent reification and align the text with the caveats already present in §3.
  3. [§4.1] The exclusion criterion is phrased as a double negative ('never fell below 1.5 million'). Please clarify: excluded titles are those whose rank stayed above 1.5 million for the entire observation window, i.e., those with no observed sales activity.
  4. [§4.3] Minor typo: 'aiassistance' should be 'AI assistance' in the description of Pangram's window output.

Circularity Check

1 steps flagged

Empirical measurement, not a circular derivation; low-severity definitional tautology in labeling the observed declines 'dilution.'

specific steps
  1. self definitional [Introduction (p.2) and Discussion (p.10)]
    "When a massive number of books that are cheap to produce enter the market at a rate faster than the sales, revenue and top ranks available to absorb them, the return and attention left for the average book falls. We call this dilution. ... By dilution we mean two declines we observe as the catalog fills with AI books, in per-book revenue and in the market share of books with no detected AI text."

    The paper defines 'dilution' as the two declines it then reports as evidence of dilution: falling per-book revenue and falling market share of books with no detected AI text. The later statement that the observed patterns are 'consistent with dilution' is therefore a restatement of the operational definition rather than an independent, falsifiable prediction. The independent content of the paper comes from the specific measurements themselves—cohort comparisons, exposure gradients, rank churn, author output changes—which are not derived from the label 'dilution.'

full rationale

The central derivation chain is not circular in the sense of fitting a model that assumes dilution. Book-level AI bands come from Pangram v3.3 full-text window labels; sales, revenue, and ranks come from an external proprietary panel; rare-expression overlap is computed against Google Books and infini-gram. No parameter is fitted to a target 'dilution' outcome, and the headline title/sales/revenue shares are direct aggregates of observed totals. The main circularity-adjacent point is definitional: the paper names 'dilution' as the observed declines, so calling the results 'consistent with dilution' is tautological. This is a minor labeling issue rather than a load-bearing derivation failure. Self-citations are present but not load-bearing: the rare-expression coverage metric is fully specified in §4.8 and cited prior work (ref. 29) is methodological, not a uniqueness theorem or an unvalidated ansatz imported to force conclusions. The Pangram detector validation concern—corpus-specific false positives or negatives—is a genuine measurement-validity risk that could change band composition, but that is a correctness/robustness issue, not circularity: re-labeling books would alter estimates, not make the conclusions equivalent to the inputs by construction. Overall, the paper is an empirical measurement with largely self-contained derivations; the definitional tautology around 'dilution' justifies a low score rather than zero.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 0 invented entities

The paper introduces no new entities. Its central claims rest on external measurement tools (Pangram, Google Books/infini-gram), a proprietary sales panel, and hand-set analysis thresholds (25% AI cutoff, rare-expression criteria, launch window, Top-25). These are the main inputs a reader must accept to trust the results.

free parameters (4)
  • Substantial AI text threshold (>25%) = 25%
    Chosen cutoff for the substantial-AI band; the paper shows robustness across 10–50%, so it does not drive the main results, but it is a hand-set discretization of the continuous detector score.
  • Rare-expression length and frequency thresholds = >=5 words; <=5 Google Books volumes; absent from infini-gram
    Hand-chosen definition of 'rare existing-book language' used in §4.8; changing these thresholds would change coverage values, and the paper does not report sensitivity to them.
  • Launch window length = 90 days plus preorder days
    Fixed observation window for comparing cohorts; chosen to balance follow-up completeness, and only cohorts with full windows are used.
  • Top-K rank slots = 25 (Top-25)
    Used to define scarce rank positions; choice affects churn and exposure analyses but is a standard reporting choice.
axioms (6)
  • domain assumption The proprietary Big Five sales panel accurately records daily unit sales, revenue, and rank for ~95% of Amazon ebook volume.
    Invoked in Methods §4.1 and throughout; the central market statistics are computed from this panel, which is not publicly verifiable.
  • domain assumption Pangram v3.3's window-level labels correctly identify AI-generated text, with negligible false positives in this corpus.
    Invoked in Methods §4.3; all AI-band assignments and downstream comparisons depend on this detector's accuracy.
  • domain assumption Google Books and infini-gram are valid reference corpora for 'rare existing-book language'; absence from them indicates distinctiveness.
    Invoked in Methods §4.8; the rare-expression overlap measure and its interpretation rely on this.
  • domain assumption The 90-day launch window plus preorders is a comparable early-performance window across cohorts.
    Invoked in §2.3 and Methods §4.5; cohort comparisons assume no systematic seasonality or algorithm changes within the window.
  • domain assumption Byline identity approximates author identity for output and revenue concentration.
    Invoked in §4.4; the paper acknowledges pen names can split one author across bylines, so concentration is a lower bound.
  • standard math Ordinary least squares and Welch's t-test provide valid inference for the reported comparisons.
    Used in §4.9 for KU and rare-expression analyses; no clustering or multiple-testing corrections are described.

pith-pipeline@v1.3.0-alltime-deepseek · 16394 in / 12060 out tokens · 408800 ms · 2026-08-04T04:09:29.383384+00:00 · methodology

0 comments
read the original 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 AI text lose the most ground in genres with high AI diffusion, and most of all where Kindle Unlimited availability is high. Among top-selling books, those with substantial AI text draw on more distinctive language from existing books than do books with no AI text; for these books overlap rises with revenue, a gradient we do not detect for 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.

Figures

Figures reproduced from arXiv: 2607.20349 by Jane C. Ginsburg, Paramveer Dhillon, Tuhin Chakrabarty, Xinyue Liu.

Figure 1
Figure 1. Figure 1: Books with substantial AI text spread across every genre yet earn a smaller share of sales. (a) [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 1
Figure 1. Figure 1: Books with substantial AI text spread across every genre yet earn a smaller share of sales. (a) [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Books with substantial AI text reach commercial scale and the top sales ranks. (a) [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Releases outrun sales and revenue, and the decline reaches books with no AI text. (a) [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Losses for books with no AI text concentrate where AI exposure is greatest, and a broad base of authors scales [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 4
Figure 4. Figure 4: Losses for books with no AI text concentrate where AI exposure is greatest, and a broad base of authors scales [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: (A) Books with substantial AI text (>25%) show consistently higher coverage of rare 5+ word expressions than no-AI￾text books, and the gap widens among the highest earners (+7.2, +5.5, and +4.3 pts for the Top 50/100/200). Books are ranked by total revenue within each category. Error bars = 95% CI for 𝑛 = 50/100/200 per group; all differences significant (Welch’s 𝑡-test, 𝑝 < 10−8 ). (B) Among best-selling … view at source ↗

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

30 extracted references · 2 linked inside Pith

  1. [1]

    M., Chang, Y

    Pham, C. M., Chang, Y. & Iyyer, M. Autofiction (2026). URLhttps://www.autofiction.ai/. Web platform

    2026

  2. [2]

    & Walsh, M

    Gupta, N., Antoniak, M. & Walsh, M. Ai fiction in the wild.arXiv preprint arXiv:2606.22748(2026)

    Pith/arXiv arXiv 2026

  3. [3]

    & Wallace, B

    Shaib, C., Chakrabarty, T., Garcia-Olano, D. & Wallace, B. C. Measuring ai” slop” in text.arXiv preprint arXiv:2509.19163(2025)

    arXiv 2025

  4. [4]

    InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 14554–14580 (2026)

    Russell, J.et al.Ai use in american newspapers is widespread, uneven, and rarely disclosed. InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 14554–14580 (2026)

    2026

  5. [5]

    Horror novel ‘shy girl’ canceled over suspected a.i

    Alter, A. Horror novel ‘shy girl’ canceled over suspected a.i. use.The New York Times(2026). URLhttps: //www.nytimes.com/2026/03/19/books/shy-girl-book-ai.html

    2026

  6. [6]

    Alter, A. A.I. is writing fiction. Publishers are unprepared.The New York Times(2026). URLhttps: //www.nytimes.com/2026/03/19/books/ai-fiction-shy-girl.html

    2026

  7. [7]

    The new fabio is claude: The romance industry, always at the vanguard of technological change, is rapidly adapting to A.I.The New York Times(2026)

    Alter, A. The new fabio is claude: The romance industry, always at the vanguard of technological change, is rapidly adapting to A.I.The New York Times(2026). URLhttps://www.nytimes.com/2026/02/08/busin ess/ai-claude-romance-books.html

    2026

  8. [8]

    Book output topped four million in 2025.Publishers Weekly(2026)

    Milliot, J. Book output topped four million in 2025.Publishers Weekly(2026). URLhttps://www.publis hersweekly.com/pw/by-topic/industry-news/publisher-news/article/99943-boo k-output-topped-4-million-in-2025.html

    2025

  9. [9]

    Amazon publishing statistics.https://wordsrated.com/amazon-publishing-stati stics/(2023)

    WordsRated. Amazon publishing statistics.https://wordsrated.com/amazon-publishing-stati stics/(2023). WordsRated industry statistics

    2023

  10. [10]

    M., Iyyer, M

    Russell, J., Rajendhran, R., Pham, C. M., Iyyer, M. & Wieting, J. Storyscope: Investigating idiosyncrasies in ai fiction.arXiv preprint arXiv:2604.03136(2026). 17

    Pith/arXiv arXiv 2026

  11. [11]

    & Wu, C.-S

    Chakrabarty, T., Laban, P., Agarwal, D., Muresan, S. & Wu, C.-S. Art or artifice? large language models and the false promise of creativity. InProceedings of the 2024 CHI Conference on Human Factors in Computing Systems, 1–34 (2024)

    2024

  12. [12]

    & Waldfogel, J

    Reimers, I. & Waldfogel, J. Ai and the quantity and quality of creative products: Have llms boosted creation of valuable books? Tech. Rep., National Bureau of Economic Research (2026)

    2026

  13. [13]

    Kadrey et al. v. meta platforms, inc. (2025). URLhttps://law.justia.com/cases/federal/dis trict-courts/california/candce/3:2023cv03417/415175/598/. Order denying plaintiffs’ motion for partial summary judgment and granting Meta’s cross-motion on fair use grounds

    2025

  14. [14]

    InInternational Conference on Learning Representations, vol

    Lu, X.et al.Ai as humanity’s salieri: Quantifying linguistic creativity of language models via systematic attribution of machine text against web text. InInternational Conference on Learning Representations, vol. 2025, 91005– 91056 (2025)

    2025

  15. [15]

    178–188 (2025)

    Liu, J.et al.Olmotrace: Tracing language model outputs back to trillions of training tokens. 178–188 (2025)

    2025

  16. [16]

    Chakrabarty, T., Ginsburg, J. C. & Dhillon, P. Readers prefer outputs of ai trained on copyrighted books over expert human writers.arXiv preprint arXiv:2510.13939(2025)

    arXiv 2025

  17. [17]

    & Dhillon, P

    Chakrabarty, T. & Dhillon, P. S. Can good writing be generative? expert-level ai writing emerges through fine- tuning on high quality books. InProceedings of the 2026 CHI Conference on Human Factors in Computing Systems, 1–27 (2026)

    2026

  18. [18]

    & Krizan, C

    Foster, L., Haltiwanger, J. & Krizan, C. J. Aggregate productivity growth: Lessons from microeconomic evidence. In Hulten, C. R., Dean, E. R. & Harper, M. J. (eds.)New Developments in Productivity Analysis, 303–372 (University of Chicago Press, 2001)

    2001

  19. [19]

    N., Hulten, C

    Baily, M. N., Hulten, C. & Campbell, D. Productivity dynamics in manufacturing plants.Brookings Papers on Economic Activity: Microeconomics1992, 187–267 (1992)

    1992

  20. [20]

    Landes, W. M. & Posner, R. A. An economic analysis of copyright law.The Journal of Legal Studies18, 325–363 (1989)

    1989

  21. [21]

    Landes, W. M. & Posner, R. A.The Economic Structure of Intellectual Property Law(Harvard University Press, Cambridge, MA, 2003)

    2003

  22. [22]

    Copyright Office

    U.S. Copyright Office. Copyright and artificial intelligence part 3: Generative AI training report. Tech. Rep., U.S. Copyright Office (2024). URLhttps://www.copyright.gov/ai/Copyright-and-Artificia l-Intelligence-Part-3-Generative-AI-Training-Report-Pre-Publication-Versi on.pdf. Pre-publication version analyzing copyright implications of AI training

    2024

  23. [23]

    & Lam, H

    Goldberg, S. & Lam, H. T. Generative AI and creative goods: Market expansion, crowd-out, and copyright. Stanford University Graduate School of Business Research Paper, Stanford University Graduate School of Business (2025). URLhttp://dx.doi.org/10.2139/ssrn.5152649. Available at SSRN:https://ssrn.com/a bstract=5152649

    doi:10.2139/ssrn.5152649 2025

  24. [24]

    & Telang, R

    Dai, J., Li, B. & Telang, R. Generative ai content governance and market outcomes: Evidence from food delivery platforms.Available at SSRN 6711039(2026)

    2026

  25. [25]

    & Imas, A

    Jabarian, B. & Imas, A. Artificial writing and automated detection. Tech. Rep., National Bureau of Economic Research (2025)

    2025

  26. [26]

    & Iyyer, M

    Russell, J., Karpinska, M. & Iyyer, M. People who frequently use ChatGPT for writing tasks are accurate and robust detectors of AI-generated text. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 5342–5373 (Association for Computational Linguistics, Vienna, Austria, 2025)

    2025

  27. [27]

    & Levy, J

    Shah, A. & Levy, J. Access to justice in the age of ai: Evidence from us federal courts (2026). 18

    2026

  28. [28]

    & Thai, K

    Masrour, E., Spero, M., Emi, B. & Thai, K. Pangram 3.3 model card.https://www.pangram.com/rese arch/model-card/pangram-3-3(2026). Accessed: 2026-07-05

    2026

  29. [29]

    Liu, X., Mireshghallah, N., Ginsburg, J. C. & Chakrabarty, T. Alignment whack-a-mole: Finetuning activates verbatim recall of copyrighted books in large language models.arXiv preprint arXiv:2603.20957(2026). 19

    arXiv 2026

  30. [2026]

    title” as the ASIN level catalog entry that we are analyzing and refer to the “book

    and then reduced the search space by excluding titles whose Amazon sales rank never fell below 1.5 million 13During publisher verification we identified a small number of traditionally published editions that accounts for roughly 0.3% of the corpus. We retain them in the analysis. 11 at any point during the observation period. This got rid of listings tha...

    2026