Pith. sign in

REVIEW 4 major objections 5 minor 26 references

GPT Editors, Not Authors: The Stylistic Footprint of LLMs in Academic Preprints

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read LLM use in arXiv preprints is stylistically uniform, the paper argues: authors edit or translate across the whole manuscript rather than generating isolated sections, making partial LLM generation rare.

desk verdict A real null result and a promising segmentation method, but the title's 'editors not authors' claim outruns the evidence. read the letter →

arxiv 2505.17327 v1 pith:LLKCJQ5D submitted 2025-05-22 cs.CL cs.ITcs.LGmath.IT

Soren DeHaan , Yuanze Liu , Johan Bollen , Sa'ul A. Blanco This is my paper
classification cs.CLcs.ITcs.LGmath.IT
keywords LLMdetectionacademicpreprintsarXivstylisticsegmentationchangepointPELTBayesianclassifierhallucinationrisk
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper asks whether large language models are being used to write isolated chunks of academic preprints or only to smooth over human-authored text. It argues that on arXiv, LLM use is stylistically uniform: the words a classifier associates with LLMs are spread evenly through a manuscript rather than clumped into sections. Across 2,408 preprints from 2023-2025, once text length is statistically removed, a paper's total LLM signal and the strength of its internal stylistic segmentation are unrelated (r = -0.0233, p > 0.2). From this the paper concludes that partial LLM generation—human text with only some sections machine-written—is highly uncommon, and that nearly all LLM use in the sample is editing or translation. The stakes are hallucination risk: editing preserves a human author's content claims, while partially generated sections are where fabricated statements would most easily enter the literature.

What carries the argument

The engine of the argument is a pairing of a naive Bayesian word classifier with PELT (Pruned Exact Linear Time), a changepoint detection algorithm that finds where a cumulative sequence shifts. The classifier assigns each word $W$ a log-odds score $\text{LogOdds}(W) = \log\left(\frac{P_{\text{LLM}}(W)+10^{-4}}{P_{\text{Human}}(W)+10^{-4}}\right)$ using frequency rates from GPT-3.5 Turbo regenerations of 2021 arXiv sections versus the original human texts. PELT runs on the cumulative sum of those word scores, and the paper's key quantity is the penalty threshold multiplier: the smallest variance-normalized penalty at which no changepoints remain. That multiplier quantifies how tightly LLM-associated words are clustered; the more clustered, the more noise would be needed to hide the segmentation. Z-score normalization within 25 length bins removes the strong length confound (length correlates -0.72 with log-odds and 0.88 with threshold in the raw data). A validation set—human texts with an LLM paragraph inserted (about 19.4% of length)—shows a significantly higher threshold multiplier than either pure human or fully regenerated texts, confirming the metric can detect partial insertion when it exists.

What would settle it

One concrete check is to apply the same pipeline to a corpus where authors disclose exactly how they used LLMs, such as papers from a journal requiring AI-disclosure statements; if papers disclosed as partially LLM-generated do not show higher length-normalized PELT thresholds than editing-only papers, the central claim is falsified, and if they do, the method's null result in the wild would be contradicted.

Watch

Extended reading notes

Core claim

The central claim is that when researchers use LLMs on arXiv preprints, they apply them uniformly across the paper. The paper measures each word's log-odds of appearing in GPT-3.5 Turbo-regenerated text versus human text, sums those per-word scores over each document, and runs PELT changepoint detection on the cumulative sum. The PELT threshold multiplier—the amount of noise needed to hide any segmentation—captures whether LLM-associated words are clumped or spread out. Across 2,408 preprints from 2023-2025, the raw correlation between total LLM log-odds and this threshold disappears after Z-score normalization over 25 length bins (r = -0.0233, p > 0.2). The paper takes this as evidence that LLM use is not concentrated in particular sections, so the partial-generation use case is 'highly uncommon,' and almost all LLM use is editing or translation. The authors acknowledge that the method cannot by itself separate uniform editing from complete LLM generation.

Load-bearing premise

The argument depends on the classifier's word scores actually measuring LLM authorship rather than what a paper is about or how formally it is written; if those scores mostly track topic or formatting, the null correlation cannot show that LLM use is uniform.

Editorial extensions

If this is right

  • If the finding holds, detection and journal policy should focus on telling complete LLM generation apart from editing and translation, rather than on hunting for partially generated sections.
  • The hallucination risk commonly tied to LLM-assisted writing shifts accordingly: the pattern with the highest risk of fabricated content is wholesale fabrication, because editing and translation preserve the human author's substantive claims.
  • For researchers who are not native English speakers, the result implies that LLM polishing of a manuscript does not imply machine-authored content, supporting policies that permit disclosed LLM editing.
  • At the population level, a whole-paper LLM signal is sufficient for monitoring arXiv; the high section-to-section correlations mean there is little additional information in treating each section as an independent unit.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper's own caveat leaves the biggest alternative open: complete LLM generation also produces stylistically uniform text, so the null result cannot quantify how much of the observed LLM signal is editing versus wholesale fabrication; the reassuring conclusion assumes the latter is rare.
  • Because the classifier was trained on GPT-3.5 Turbo regenerations of 2021 texts, the uniform-signal conclusion may be model-specific; other LLM families with different stylistic fingerprints could produce segmentation that this method misses.
  • A testable check follows from the classifier's word list: terms like 'efficiency' and 'configuration' are also topic words, so rerunning the analysis with discipline-matched controls would show whether the null correlation is a behavior or an artifact of subject vocabulary.
  • An extension would be to use journals that require AI-disclosure statements as ground truth, calibrating the PELT threshold against disclosed use type to measure how much partial generation actually occurs outside the uniform sample.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes a method to distinguish between LLM editing/translation and partial LLM generation in academic preprints. The authors train a naive Bayesian classifier on 2021 arXiv sections and GPT-3.5 Turbo regenerations, compute per-word log-odds, and apply PELT changepoint detection to the cumulative log-odds of each paper. They report that after Z-score normalization for text length, there is no correlation between total LLM log-odds and the PELT threshold multiplier (r = -0.0233, p > 0.2), and that section-level log-odds are highly correlated. They conclude that LLM use is uniformly applied and that 'almost all LLM use is for editing or translation,' with partial generation being 'highly uncommon.'

Significance. The methodological combination of PELT with a Bayesian classifier to measure stylistic segmentation is a creative contribution, and the paper is transparent about its data and code. The null result, if valid, would be informative for the community studying LLM use in scientific writing. However, the central interpretive claim—that LLM use is predominantly editorial rather than generative—is not supported by the experimental design, because the null correlation is equally compatible with uniform complete LLM generation. The validation of the classifier is also entirely synthetic, using the same GPT-3.5 Turbo pipeline for training and testing, and the paper does not report classifier accuracy numerically.

major comments (4)
  1. [Abstract; Sections 3.2 and 5] The central claim that 'almost all LLM use is for editing or translation' is not supported by the null correlation between Z-score-normalized LLM log-odds and the PELT threshold (r = -0.0233, p > 0.2). As the paper itself states in the Introduction, a uniform LLM signal would occur 'as would occur when an author uses LLMs for general editing or translation, or when a manuscript is completely fabricated.' No independent estimate of the prevalence of complete generation is provided, so the data cannot separate case 2 (editing/translation) from case 4 (complete generation). The abstract's assertion that hallucination risk is reduced therefore overreaches the evidence.
  2. [Section 2.4 and Figures 1-2] The classifier's accuracy is never reported numerically. Figures 1 and 2 are described as a histogram and confusion matrix, but the text gives no true-positive rate, false-positive rate, or AUC. The only quantitative validation (Tables 1-2) concerns the PELT threshold on the synthetic segmented dataset, not the classifier's discrimination. Because the training data, validation data, and LLM-associated word lists all come from the same GPT-3.5 Turbo regeneration pipeline, the metric's validity for other LLMs and for real 2023-2025 papers is not established.
  3. [Section 3.2 and Table 5] The Z-score length-normalization procedure is under-specified. The paper reports raw correlations of r = -0.72 between length and log-odds and r = 0.88 between length and threshold, but after 25-bin Z-score normalization the correlation drops to r = -0.0233 with p > 0.2. No partial correlation is reported, and there is no sensitivity analysis for the number of bins or the choice of bin boundaries. Given that the main conclusion rests entirely on this null result, the adequacy of the length correction is load-bearing and needs stronger justification.
  4. [Section 2.1] The sample restriction to papers with Abstract, Introduction, and Conclusion sections each at least 500 characters reduces the 2023-2025 sample from n = 8000 to n = 2408 and the 2021 sample from n = 550 to n = 205. This selection rule may systematically exclude short, unstructured, or formulaic papers, which could be more likely to be completely LLM-generated. The Limitations section does not discuss this potential bias, and it could affect the external validity of the uniform-use conclusion.
minor comments (5)
  1. [Section 2.2] Equation (1) defines LogOdds as a ratio of averaged frequency rates, but the text calls these 'odds'; please use consistent terminology, since a frequency-rate ratio is not an odds ratio.
  2. [Section 3.2] The sentence 'The total log odds has a magnitude that scales with text length as well' is difficult to reconcile with the negative correlation reported in Table 5 (Length vs. Log Odds r = -0.7204); clarify whether 'total log-odds' is a sum or an average over words.
  3. [Section 2.3] The binary search is described as producing a threshold with 'a margin of error of 10^-2 for a threshold M,' but it is unclear whether this is an absolute or relative error; please define the stopping criterion precisely.
  4. [Table 3] The table header contains a typo ('W ord'), and the entry 'delf (delve)' should be cleaned up; also consider explaining why 'delf' is listed as an LLM-associated word.
  5. [Figure 3] The caption 'The distribution difference of PELT thresholds for the segmented dataset' is vague; specify which groups are being compared and what kernel or density estimate is used.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the null correlation is an empirical result, and the main weakness is an inferential gap rather than a circular derivation.

full rationale

I walked the paper's derivation chain and found no step in which a claimed prediction or first-principles result reduces to its own inputs by construction. The Bayesian classifier is trained on human-written 2021 text and GPT-3.5 Turbo regenerations of that text, and the PELT threshold is measured on 2023-2025 arXiv papers; the headline finding is an observed absence of correlation after length normalization. This is not a fitted parameter renamed as a prediction: the classifier and PELT procedure are validated on held-out synthetic data, but the final correlation is computed on a separate real-world corpus rather than being forced by the training labels. The paper's self-citations (Bockting et al. 2023 and Van Dis et al. 2023, both involving author Johan Bollen) are background policy statements and are not load-bearing for the method or the null result. The genuine weakness is that the null correlation is logically compatible with complete LLM generation, uniform editing, or a mix, and the paper explicitly concedes in Section 5 that the method 'does not distinguish between a text which is generated entirely by an LLM and one which is edited by one.' That is an overclaim or underdetermination issue, not a circularity: the conclusion does not equal the input by definition, nor does it rely on the authors' prior work to make the inference go through. The same-pipeline training and validation data raise a generalization risk for the classifier's transfer to real usage, but that is an external-validity limitation rather than a circular reduction. Under the seven enumerated circularity patterns, no specific reduction can be quoted, so the appropriate finding is no significant circularity.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The paper's conclusion rests on a chain of assumptions about the 2021 baseline being LLM-free, GPT-3.5 Turbo style being representative, the word-frequency classifier being valid across disciplines, and the PELT threshold being sensitive to real partial generation. None of these is externally benchmarked, and the main interpretive claim depends on all of them.

free parameters (5)
  • LogOdds pseudocount (10^-4) = 10^-4
    Added to both numerator and denominator in the LogOdds formula (Section 2.2) to avoid log(0); chosen by hand, it affects low-frequency word scores.
  • Z-score length-bin count (25) = 25
    Z-score normalization groups papers into 25 length bins (Section 3.2); the bin count is arbitrary and no sensitivity analysis is reported, although the null correlation depends on this normalization.
  • Minimum section length (500 chars) = 500 characters
    Inclusion criterion for Abstract, Introduction, and Conclusion sections (Section 2.1); selecting only papers with all three sections and each at least 500 characters drops n from 8000 to 2408 and may bias the sample.
  • PELT threshold margin (10^-2) = 10^-2
    Binary search precision for the penalty multiplier threshold (Section 2.3); a precision parameter, not fitted, but affects the threshold value.
  • PELT penalty variance-scale factor = 1 (linear scaling)
    Penalty constant is normalized to scale linearly with per-document variance of word log-odds (Section 2.3); this scaling rule determines the meaning of the threshold and was not varied.
assumptions (6)
  • domain assumption 2021 arXiv text contains no LLM-generated content.
    The baseline training set is sampled from May-September 2021, pre-ChatGPT (Section 2.1).
  • domain assumption GPT-3.5 Turbo regenerations are a representative sample of LLM writing style.
    The classifier and validation data derive entirely from this one model (Sections 2.1, 2.4); the paper notes in Section 4 that other LLMs may have different linguistic biases, yet the analysis assumes transfer.
  • domain assumption Word-frequency log-odds are a valid per-word proxy for LLM authorship in arbitrary academic text.
    The naive Bayesian classifier averages word frequency rates across documents (Section 2.2); its accuracy is referenced via Figures 1-2 but never numerically reported.
  • domain assumption PELT changepoints in the cumulative log-odds series correspond to meaningful stylistic segmentation.
    PELT is applied to the cumulative sum of word log-odds (Section 2.3); the validation only tests a single 19.4% block insertion, not smaller or interleaved edits.
  • domain assumption The absence of correlation after Z-score normalization is not an artifact of the binning scheme.
    The 25-bin Z-score normalization removes the raw correlation (Section 3.2); no alternative bin counts or null-model simulations are reported.
  • domain assumption Sections (Abstract, Introduction, Conclusion) are sufficient to represent a paper's LLM usage.
    The analysis concatenates only three sections per paper (Section 2.1), excluding methods and results, based on the claim that these sections see higher LLM use [Liang et al. 2024b].

how reviews work

0 comments
Cite this review

Pith. "Pith review of GPT Editors, Not Authors: The Stylistic Footprint of LLMs in Academic Preprints." pith.science (2026). https://pith.science/paper/LLKCJQ5D

@misc{pith2026250517327,
  author       = {Pith},
  title        = {Pith review of: GPT Editors, Not Authors: The Stylistic Footprint of LLMs in Academic Preprints},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LLKCJQ5D}},
  note         = {Machine review of arXiv:2505.17327}
}
read the original abstract

The proliferation of Large Language Models (LLMs) in late 2022 has impacted academic writing, threatening credibility, and causing institutional uncertainty. We seek to determine the degree to which LLMs are used to generate critical text as opposed to being used for editing, such as checking for grammar errors or inappropriate phrasing. In our study, we analyze arXiv papers for stylistic segmentation, which we measure by varying a PELT threshold against a Bayesian classifier trained on GPT-regenerated text. We find that LLM-attributed language is not predictive of stylistic segmentation, suggesting that when authors use LLMs, they do so uniformly, reducing the risk of hallucinations being introduced into academic preprints.

Figures

Figures reproduced from arXiv: 2505.17327 by the authors.

Figure 1
Figure 1. Histogram of classifier results on GPT￾regenerated text [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 3
Figure 3. The distribution difference of PELT thresholds for the segmented dataset. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. The original data. Observe length as a confounding variable. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: After normalization with Z-score. 9 [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

26 extracted references · 8 canonical work pages

  1. [2]

    doi: 10.1007/s00256-023-04340-5

    ISSN 1432-2161. doi: 10.1007/s00256-023-04340-5. arXiv.org submitters. arxiv dataset,

    work page doi:10.1007/s00256-023-04340-5

  2. [4]

    doi: 10.1038/d41586-023-03266-1

    ISSN 0028-0836, 1476-4687. doi: 10.1038/d41586-023-03266-1. URL https://www.nature.com/ articles/d41586-023-03266-1. Huzi Cheng, Bin Sheng, Aaron Lee, Varun Chaudary, Atanas G. Atanasov, Nan Liu, Yue Qiu, Tien Yin Wong, Yih-Chung Tham, and Yingfeng Zheng. Have AI-Generated Texts from LLM Infiltrated the Realm of Scientific Writing? A Large-Scale Analysis ...

    work page doi:10.1038/d41586-023-03266-1

  3. [5]

    Pages: 2024.03.25.586710 Section: New Results

    URLhttps: //www.biorxiv.org/content/10.1101/2024.03.25.586710v2. Pages: 2024.03.25.586710 Section: New Results. Tjibbe Donker. The dangers of using large language models for peer review.The Lancet Infectious Diseases, 23(7):781, July

    work page doi:10.1101/2024.03.25.586710v2 2024

  4. [6]

    contamination

    ISSN 1473-3099, 1474-4457. doi: 10.1016/S1473-3099(23)00290-6. URL https: //www.thelancet.com/journals/laninf/article/PIIS1473-3099(23)00290-6/fulltext. Publisher: Elsevier. Andrew Gray. ChatGPT "contamination": estimating the prevalence of LLMs in the scholarly literature, March

    doi:10.1016/s1473-3099(23)00290-6

  5. [7]

    arXiv:2403.16887 [cs]

    URLhttp://arxiv.org/abs/2403.16887. arXiv:2403.16887 [cs]. AbhimanyuHans, AviSchwarzschild, ValeriiaCherepanova, HamidKazemi, AniruddhaSaha, MicahGoldblum, Jonas Geiping, and Tom Goldstein. Spotting LLMs With Binoculars: Zero-Shot Detection of Machine- Generated Text, October

    arXiv

  6. [8]

    arXiv:2401.12070 [cs]

    URLhttp://arxiv.org/abs/2401.12070. arXiv:2401.12070 [cs]. Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Yejin Bang, Delong Chen, Wenliang Dai, Ho Shu Chan, Andrea Madotto, and Pascale Fung. Survey of Hallucination in Natural Language Generation.ACM Computing Surveys, 55(12):1–38, December

    arXiv

  7. [9]

    doi: 10.1145/3571730

    ISSN 0360-0300, 1557-7341. doi: 10.1145/3571730. URLhttp://arxiv.org/abs/2202.03629. arXiv:2202.03629 [cs]. 11 R. Killick, P. Fearnhead, and I. A. Eckley. Optimal detection of changepoints with a linear computational cost. Journal of the American Statistical Association, 107(500):1590–1598, December

    arXiv

  8. [11]

    arXiv:2404.01261 [cs]

    URLhttp://arxiv.org/abs/2404.01261. arXiv:2404.01261 [cs]. Dmitry Kobak, Rita González-Márquez, Emőke Ágnes Horvát, and Jan Lause. Delving into ChatGPT usage in academic writing through excess vocabulary, February

    arXiv

Show all 26 references
  1. [13]

    org/abs/2303.13408

    URLhttp://arxiv. org/abs/2303.13408. arXiv:2303.13408 [cs]. Tharindu Kumarage, Paras Sheth, Raha Moraffah, Joshua Garland, and Huan Liu. How Reliable Are AI-Generated-Text Detectors? An Assessment Framework Using Evasive Soft Prompts. In Houda Bouamor, Juan Pino, and Kalika Ba...

    2023 arXiv

  2. [14]

    doi: 10.18653/v1/2023.findings-emnlp.94

    Association for Computational Linguistics. doi: 10.18653/v1/2023.findings-emnlp.94. URLhttps://aclanthology.org/2023.findings-emnlp.94/. Giuseppe Russo Latona, Manoel Horta Ribeiro, Tim R. Davidson, Veniamin Veselovsky, and Robert West. The AI Review Lottery: Widespread AI-Ass...

    2023 doi

  3. [15]

    arXiv:2405.02150 [cs]

    URLhttp://arxiv.org/abs/2405.02150. arXiv:2405.02150 [cs]. Mariana Lenharo. SCIENCE’S LANGUAGE BARRIER: THE COST FOR NON-NATIVE SPEAKERS. Weixin Liang, Mert Yuksekgonul, Yining Mao, Eric Wu, and James Zou. GPT detectors are biased against non- native English writers.Patterns, ...

    arXiv

  4. [16]

    doi: 10.1016/j.patter.2023.100779

    ISSN 2666-3899. doi: 10.1016/j.patter.2023.100779. URLhttps://www.sciencedirect.com/science/article/pii/S2666389923001307. Weixin Liang, Zachary Izzo, Yaohui Zhang, Haley Lepp, Hancheng Cao, Xuandong Zhao, Lingjiao Chen, Haotian Ye, Sheng Liu, Zhi Huang, Daniel A. McFarland, a...

    2023

  5. [17]

    doi: 10.1038/s41551-024-01185-8

    ISSN 2157-846X. doi: 10.1038/s41551-024-01185-8. URLhttps: //www.nature.com/articles/s41551-024-01185-8. Publisher: Nature Publishing Group. Ning Lu, Shengcai Liu, Rui He, Qi Wang, Yew-Soon Ong, and Ke Tang. Large Language Models can be Guided to Evade AI-Generated Text Detection, May

    doi

  6. [18]

    arXiv:2305.10847 [cs]

    URLhttp://arxiv.org/abs/2305.10847. arXiv:2305.10847 [cs]. Eric Mitchell, Yoonho Lee, Alexander Khazatsky, Christopher D. Manning, and Chelsea Finn. DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature. InProceedings of the 40th International Confe...

    arXiv

  7. [19]

    doi: 10.1007/s10489-024-05298-0

    ISSN 1573-7497. doi: 10.1007/s10489-024-05298-0. URL https://doi.org/10.1007/s10489-024-05298-0. Vinu Sankar Sadasivan, Aounon Kumar, Sriram Balasubramanian, Wenxiao Wang, and Soheil Feizi. Can AI-Generated Text be Reliably Detected?, January

    doi

  8. [20]

    arXiv:2303.11156 [cs]

    URL http://arxiv.org/abs/2303.11156. arXiv:2303.11156 [cs]. Liyan Tang, Igor Shalyminov, Amy Wong, Jon Burnsky, Jake Vincent, Yu’an Yang, Siffi Singh, Song Feng, Hwanjun Song, Hang Su, Lijia Sun, Yi Zhang, Saab Mansour, and Kathleen McKeown. TofuEval: Evaluating Hallucinations...

    2024 arXiv

  9. [21]

    doi: 10.18653/v1/2024.naacl-long.251

    Association for Computational Linguistics. doi: 10.18653/v1/2024.naacl-long.251. URLhttps://aclanthology.org/ 2024.naacl-long.251. Eva A. M. Van Dis, Johan Bollen, Willem Zuidema, Robert Van Rooij, and Claudi L. Bockting. ChatGPT: five priorities for research.Nature, 614(7947)...

    2024 doi

  10. [22]

    doi: 10.1038/d41586-023-00288-7

    ISSN 0028-0836, 1476-4687. doi: 10.1038/d41586-023-00288-7. URLhttps://www.nature.com/articles/d41586-023-00288-7. William H. Walters and Esther Isabelle Wilder. Fabrication and errors in the bibliographic citations generated by ChatGPT.Scientific Reports, 13(1):14045, September

    doi

  11. [23]

    ISSN 2045-2322. doi:

  12. [24]

    ISBN 9798350361643

    IEEE. ISBN 9798350361643. doi: 10.1109/ICETCI61221.2024.10594194. URLhttps://ieeexplore.ieee.org/document/10594194/. Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Yulong Chen, Longyue Wang, Anh Tuan Luu, Wei Bi, Freda Shi...

    2024

  13. [25]

    arXiv:2309.01219 [cs]

    URL http: //arxiv.org/abs/2309.01219. arXiv:2309.01219 [cs]. Haoyi Zheng and Huichun Zhan. ChatGPT in Scientific Writing: A Cautionary Tale.The American Journal of Medicine, 136(8):725–726.e6, August

    arXiv

  14. [26]

    doi: 10.1016/j.amjmed.2023.02.011

    ISSN 0002-9343, 1555-7162. doi: 10.1016/j.amjmed.2023.02.011. URLhttps://www.amjmed.com/article/S0002-9343(23)00159-6/fulltext. Publisher: Elsevier. 13

    2023 doi

  15. [2012]

    doi: 10.1080/01621459.2012.737745

    ISSN 0162-1459, 1537- 274X. doi: 10.1080/01621459.2012.737745. URL http://arxiv.org/abs/1101.1438. arXiv:1101.1438 [stat]. YekyungKim, YapeiChang, MarzenaKarpinska, AparnaGarimella, VarunManjunatha, KyleLo, TanyaGoyal, and Mohit Iyyer. FABLES: Evaluating faithfulness and conte...

    2012

  16. [2023]

    doi: 10.1371/journal.pbio.3002184

    ISSN 1545-7885. doi: 10.1371/journal.pbio.3002184. URL https://journals.plos.org/plosbiology/article?id=10.1371/journal.pbio.3002184. Publisher: Public Library of Science. Sisith Ariyaratne, Karthikeyan P. Iyengar, Neha Nischal, Naparla Chitti Babu, and Rajesh Botchu. A compar...

    doi

  17. [2024]

    Claudi L

    URLhttps://www.kaggle.com/dsv/7548853. Claudi L. Bockting, Eva A. M. Van Dis, Robert Van Rooij, Willem Zuidema, and Johan Bollen. Living guidelines for generative AI — why scientists must oversee its use.Nature, 622(7984):693–696, October

  18. [2025]

    arXiv:2406.07016 [cs]

    URLhttp://arxiv.org/abs/2406.07016. arXiv:2406.07016 [cs]. Kalpesh Krishna, Yixiao Song, Marzena Karpinska, John Wieting, and Mohit Iyyer. Paraphrasing evades detectors of AI-generated text, but retrieval is an effective defense, October

    arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.