CoChat
Sign Up
PreprintCited by 15Academic integrity and plagiarismAcademic Publishing and Open AccessAcademic Writing and Publishing

Have AI-Generated Texts from LLM Infiltrated the Realm of Scientific Writing? A Large-Scale Analysis of Preprint Platforms

Open access
Cheng, Huzi, Sheng, Bin, Lee, Aaron, Chaudary, Varun, Atanasov, Atanas G., Liu, Nan, Qiu, Yue, Wong, Tien Yin +2 more·2024·bioRxiv (Cold Spring Harbor Laboratory)
DOI 10.1101/2024.03.25.586710
Open this paper in CoChat — free accountOpen PDF

Abstract

Since the release of ChatGPT in 2022, AI-generated texts have inevitably permeated various types of writing, sparking debates about the quality and quantity of content produced by such large language models (LLM). This study investigates a critical question: Have AI-generated texts from LLM infiltrated the realm of scientific writing, and if so, to what extent and in what setting? By analyzing a dataset comprised of preprint manuscripts uploaded to arXiv, bioRxiv, and medRxiv over the past two years, we confirmed and quantified the widespread influence of AI-generated texts in scientific publications using the latest LLM-text detection technique, the Binoculars LLM-detector. Further analyses with this tool reveal that: (1) the AI influence correlates with the trend of ChatGPT web searches; (2) it is widespread across many scientific domains but exhibits distinct impacts within them (highest: computer science, engineering sciences); (3) the influence varies with authors who have different language speaking backgrounds and geographic regions according to the location of their affiliations (Italy, China, etc.); (4) AI-generated texts are used in various content types in manuscripts (most significant: hypothesis formulation, conclusion summarization); (5) AI usage has a positive influence on paper’s impact, measured by its citation numbers. Based on these findings, suggestions about the advantages and regulation of AI-augmented scientific writing are discussed.

AI snapshot

Written from the paper’s full text.

TL;DR

This large-scale analysis of 45,129 preprints from arXiv, bioRxiv, and medRxiv finds that Binoculars LLM-detector scores dropped after ChatGPT’s release, correlating with ChatGPT Google Trends and suggesting widespread AI-assisted text in scientific writing. A calibrated AI-content index estimates that about 5% of manuscripts contained AI-generated content after release, with variation by domain, country/language, and content type.

Key findings

  • Across 45,129 manuscripts, mean and minimum Binoculars scores decreased after ChatGPT’s release on 2022.11.30, while variance increased; these indices correlated with weekly and daily ChatGPT Google Trends. Mean scores remained above 1.01, still higher than the approximately 0.9 threshold from Hans et al., suggesting AI was not predominantly used to generate extensive text.
  • Domain differences were observed: biological sciences, computer science, and engineering showed the largest drops in minimum Binoculars values after release; engineering and computer science also had significant mean decreases. All domains showed a decrease in either mean or minimum Binoculars scores.
  • Country/language differences were observed: almost all countries/regions showed decreased minimum Binoculars values, with China, Italy, and India showing larger gaps in both mean and minimum drops. Countries/regions where English is an official language had higher overall mean and minimum Binoculars values.
  • Content-type analysis found literature reviews had low Binoculars scores and no significant drop after release, while hypothesis formulation, conclusion summarization, phenomenon description, and future work suggestions showed the largest drops. A logistic regression model reached an AUC of 0.69, and the calibrated AI-content index reached around 5% after release, with computer science showing higher average AI usage and China, Italy, and India showing stronger tendencies.

Methods

  • Data: PDFs from arXiv, bioRxiv, and medRxiv from 2022.01.01 to 2024.03.01, with at most 1,000 random manuscripts per month per platform; after cleaning, 45,129 manuscripts were analyzed across nine domains.
  • Detection: manuscripts were split into 512-token chunks and scored with the Binoculars detector using Falcon-7b and Falcon-7b-instruct; manuscript-level mean, variance, and minimum Binoculars scores were computed and compared with weekly and daily Google Trends for “ChatGPT,” including a 30-day rolling average.
  • Analysis: OLS regression included domain, whether English is an official language, and after-ChatGPT status; zero-shot text classification used Meta’s BART-large-mnli model for 12 content types; a logistic regression trained on 1,000 pre-release manuscripts (500 revised by GPT-3.5, 80/20 train/test split) produced the calibrated AI-content index.

Limitations

  • The paper states it is impossible to completely determine whether text is AI-generated; Binoculars relies on statistical patterns and may be less reliable under improper use or attacks, and other tools used (zero-shot classifier, regression model) make similar errors.
  • The paper states preprint platforms do not cover all scientific papers and domains differ in preprint use, so the dataset cannot represent the entire picture.
  • The paper states there is no direct access to authors’ country/region/native language information; nationality inference introduces errors, and multilingual contributions make country/region analysis imprecise. Evident from the text: the AI-content index depends on a classifier with AUC 0.69 and calibration assuming 90% pre-release prediction accuracy, so the 5% estimate is approximate.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1101/2024.03.25.586710

Cites

Statistics

Nothing to recompute.

Claims the paper makes

0 of 15 cited sources checked · 0 supported · 0 flagged

  • Findingown data

    After 2022.11.30, the average of overall Binoculars scores significantly decreased.

    “We found that after 2022.11.30, the date ChatGPT was launched, the average of overall Binoculars scores significantly decreased.”
    Quote in the paperSubstantial claim
  • Findingown data

    Contents containing novel information, such as data presentation and phenomenon description, have the highest average Binoculars scores.

    “We found that the average Binoculars scores of different content types matched our intuition: literature review content has very low Binoculars scores, while contents containing novel information, such as data presentation and phenomenon description, have the highest average Binoculars scores.”
    Quote in the paperNot a substantial claim
  • Findingown data

    The average mean Binoculars value is higher before the release of ChatGPT than after.

    “The average mean and minimum Binoculars values are higher before the release of ChatGPT, while the variance is higher after the release.”
    Quote in the paperSubstantial claim
  • Findingown data

    The subsequent fluctuation of Binoculars indices correlates with ChatGPT Google Trends data.

    “Moreover, the subsequent fluctuation of these indices correlates with ChatGPT Google Trends data.”
    Quote in the paperNot a substantial claim
  • Findingown data

    Content type distributions in high and low Binoculars score collections are relatively stable.

    “Additionally, the content type distributions in high and low Binoculars score collections are relatively stable, as the portion shifts are small.”
    Quote in the paperNot a substantial claim
  • Classified as probably not a substantial claim. 26
  • May combine two claims. 9
  • The quote only partly states this claim. 1

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    The search trend for ChatGPT rises after its release on 2022.11.30.

    “As shown by the gray lines in the left column of Fig. 1, the search trend for ChatGPT rises after its release on 2022.11.30.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The mean Binoculars values significantly dropped after the introduction of ChatGPT, but remained above 1.01, which is still much higher than the threshold of approximately 0.9 set by Hans et al. (2024).

    “Firstly, although the mean Binoculars values have significantly dropped after the introduction of ChatGPT, remaining above 1.01, they are still much higher than the threshold of approximately 0.9 set by (Hans et al., 2024).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Hans et al., 2024

      Not checked in the free run.

  • May combine two claims.

    Findingown data

    The decrease in mean Binoculars values is present but not as pronounced.

    “From Fig. 3, it is evident that almost all countries exhibit a decrease in minimum Binoculars values, while the decrease in mean Binoculars values is present but not as pronounced.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Daily-level correlation between ChatGPT Google Trends and Binoculars indices persists and is consistent with the weekly level analysis.

    “The results on the right side of Fig. 1D indicate that the correlation persists and is consistent with the weekly level analysis.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingreasoning

    Compared with Canada and the UK, the US also exhibits similar trends in AI usage.

    “Compared with Canada and the UK, the US, as an English-speaking country, also exhibits similar trends.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Findingnone given

    The statistical results are stable for major domains and countries/regions.

    “Therefore, the dataset we used for analysis cannot represent the entire picture, though the statistical results are stable for major domains and countries/regions.”
    Quote in the paperNot a substantial claim

  • The quote only partly states this claim.

    Findingreasoning

    The results still reveal stable patterns at the collective level despite limitations in country/region analysis.

    “However, the results still reveal stable patterns at the collective level.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The logistic regression model achieved an AUC of 0.69 on the test set, using 80% of the 1,000 ground-truth manuscripts for training and 20% for testing.

    “From the dataset of 1,000 manuscripts with ground truth, 80% was used for training and 20% for testing, achieving an AUC of 0.69.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Some LLM detectors tend to recognize texts written by non-native English speakers as LLM-generated.

    “This finding aligns with some previous studies indicating that some LLM detectors tend to recognize texts written by non-native English speakers as LLM-generated (Liang et al., 2023).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Liang et al., 2023

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationreasoning

    The zero-shot text classification model and the regression model also make similar errors.

    “Other statistical tools we used for analysis, like the zero-shot text classification model and the regression model, also make similar errors.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The native language spoken by authors may influence the use of ChatGPT.

    “Another important factor that may influence the use of ChatGPT is the native language spoken by the authors of the paper (El kah et al., 2023;Hwang et al., 2023).”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites El kah et al., 2023

      Not checked in the free run.

    • cites Hwang et al., 2023

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    When the Binoculars score is high, the input text is more likely generated by humans; when it is lower than a certain threshold, the text is more suspicious of containing LLM-generated content.

    “When Binoculars is high, it indicates that the input text is more likely generated by humans.When Binoculars is lower than a certain threshold, the text is more suspicious of containing LLM-generated content (Fig. 1B).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Hans et al., 2024

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    An increasing number of authors tend to upload their manuscripts to preprint platforms.

    “Second, although an increasing number of authors tend to upload their manuscripts to preprint platforms (Piwowar et al., 2018), these platforms still do not cover all scientific papers, not to mention that different domains also have varying tendencies in using preprints as a distribution channel.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Piwowar et al., 2018

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing). May combine two claims.

    Interpretationreasoning

    Not only will article types such as literature reviews be affected by AI, but the creation of new knowledge will likely be influenced as well.

    “It suggests that not only will article types such as literature reviews be affected by AI (Aydın & Karaarslan, 2022), but the creation of new knowledge will likely be influenced as well.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Aydın & Karaarslan, 2022

      Not checked in the free run.

  • May combine two claims.

    Interpretationown data

    The use of AI in scientific writing may be widespread, but it is not predominantly for generating extensive texts.

    “This suggests that while the use of AI in scientific writing may be widespread, it is not predominantly for generating extensive texts.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationnone given

    Different domains have varying tendencies in using preprints as a distribution channel.

    “Second, although an increasing number of authors tend to upload their manuscripts to preprint platforms (Piwowar et al., 2018), these platforms still do not cover all scientific papers, not to mention that different domains also have varying tendencies in using preprints as a distribution channel.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Interpretationreasoning

    The heterogeneous use of AI in countries without English as their official languages indirectly confirmed that AI can be used to bridge the communication gap caused by region/language barriers.

    “The heterogeneous use of AI in countries without English as their official languages indirectly confirmed this point.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationnone given

    The low average Binoculars score in engineering and computer science before release may be attributed to the size of the corpus of these domains in the training dataset of LLMs like ChatGPT.

    “However, these domains also exhibit a relatively low average Binoculars score even before the release, which may be attributed to the size of the corpus of these domains in the training dataset of LLMs like ChatGPT.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Premisecitation

    A higher Binoculars score suggests human authorship and a lower score indicates potential AI generation.

    “A higher Binoculars score suggests human authorship and a lower score indicates potential AI generation.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Hans et al., 2024

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    arXiv provides only a list of author names for each paper.

    “arXiv provides only a list of author names for each paper.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing). May combine two claims.

    Premisecitation

    Scientific writing has traditionally been viewed as a means of disseminating ideas and new knowledge to humanity, raising concerns within the community about ethical issues more than just plagiarism.

    “However, unlike scenarios where students employ AI to complete assignments, scientific writing has traditionally been viewed as a means of disseminating ideas and new knowledge to humanity, raising concerns within the community about ethical issues more than just plagiarism (Salvagno et al., 2023).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Salvagno et al., 2023

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationreasoning

    The introduction of nationality inference services inevitably leads to errors in specific papers.

    “The introduction of nationality inference services inevitably leads to errors in specific papers.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductreasoning

    Due to limitations of platforms like arXiv, the study has no direct access to the authors' country/region/native language information.

    “Lastly, due to the limitations of platforms like arXiv, we have no direct access to the authors' country/region/native language information.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    The bias was quantified to ensure 90% accuracy of the model's predictions pre-ChatGPT release.

    “In this step, the bias was quantified to ensure 90% accuracy of the model's predictions pre-ChatGPT release.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductown data

    The logistic regression model predicts the presence or absence of AI-revised content in a manuscript.

    “This model predicts the presence or absence of AI-revised content in a manuscript.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

    Interpretationreasoning

    Our conclusions may not remain valid indefinitely for future scientific writings, considering the ongoing development of new models and the continual addition of contemporary texts to training corpora.

    “Lastly, considering the ongoing development of new models and the continual addition of contemporary texts to training corpora, our conclusions may not remain valid indefinitely for future scientific writings.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationreasoning

    The lower Binoculars score in computer science compared to other domains may reflect the abundance of computer-related content online.

    “(for instance, the lower Binoculars score in computer science compared to other domains may reflect the abundance of computer-related content online).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning

    A manuscript may contain contributions from people speaking different languages, making the country/region analysis imprecise.

    “Moreover, as discussed above, a manuscript may contain contributions from people speaking different languages, making the country/region analysis imprecise.”
    Quote in the paperNot a substantial claim
    • The authors state this as an assumption: The corresponding author largely determines the content and writing style of the manuscript.

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    All platforms do not provide author country/region information directly.

    “All platforms do not provide author country/region information directly.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    bioRxiv and medRxiv provide the corresponding author name and institution information.

    “bioRxiv and medRxiv provide the corresponding author name and institution information.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationreasoning

    As new content is incorporated into an AI's training corpus, it becomes more likely to be recognized as AI-generated in subsequent analyses.

    “As new content is incorporated into an AI's training corpus, it becomes more likely to be recognized as AI-generated in subsequent analyses”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductreasoning

    The ratios were calibrated as an AI-content index, measuring the extent of AI usage in scientific writing.

    “these ratios were then calibrated as an AI-content index, measuring the extent of AI usage in scientific writing.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductreasoning

    The step constructed a dataset with ground truth labels of AI revision.

    “This step thus constructed a dataset with ground truth labels of AI revision.”
    Quote in the paperSubstantial claim
33 of 33

  • The authors state this as an assumption.

    Current ChatGPT usage may be reflected in manuscripts submitted in the near future rather than at the same moment, and a 30-day average lag is assumed.

    “assuming the current usage of ChatGPT may be reflected in the manuscripts submitted in the near future rather than at the same moment, given that a manuscript usually takes a relatively long time to finish (a 30-day average lag is assumed).”

  • The paper relies on this without support.

    The study is the first endeavor to reveal the trend of AI's footprint in contemporary scientific writing through quantitative, large-scale analysis.

    “Despite the limitations outlined above, this study represents, to our knowledge, the first endeavor to reveal the trend of AI's footprint in contemporary scientific writing activities through quantitative, large-scale analysis.”

  • The paper relies on this without support.

    The most advanced AI tools are usually transformer-based LLMs, and ChatGPT has been the dominating choice among these LLMs at the time of writing.

    “As in the use case of text generation, the most advanced AI tools are usually transformer based LLMs, and ChatGPT has been the dominating choice among these LLMs at the time of writing.”

  • The authors state this as an assumption.

    The corresponding author largely determines the content and writing style of the manuscript.

    “First, we assume that corresponding author largely determines the content and writing style of the manuscript.”

  • The authors state this as an assumption.

    The last author provided by arXiv at most of the time can be treated as the corresponding author of that paper.

    “Second, we assume the last author provided by arXiv at most of the time, can be treated as the corresponding author of that paper.”

  • The paper relies on this without support.

    The content type list covers the majority of content found in a typical scientific paper.

    “This list covers the majority of content found in a typical scientific paper.”

  • The paper relies on this without support.

    The daily Google Trends data reconstruction approach allows the reconstruction of long-interval daily Google Trends data using the earliest month as the base.

    “This approach allows the reconstruction of long-interval daily Google Trends data using the earliest month as the base.”
7 of 7

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • Some studies have reported differences in estimated LLM usage between English- and non-English-speaking countries, academic fields, and publishing venues.

    “Some studies have reported differences in estimated LLM usage between English- and non-English–speaking countries (18–20), academic fields (17), and publishing venues.”

    From Delving into LLM-assisted writing in biomedical publications through excess vocabulary (2025) · cites this paper as [18]

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Minimum Binoculars scores are described as both decreasing and increasing

    Major

    The paper says average minimum scores were higher before ChatGPT's release, which means they decreased afterward. The next sentence instead cites an increase in minimum scores after the release.

1 of 1

The citation lists haven’t been fetched yet

References 22 · Cited by 15

Research works printed but not linked: 1

  1. Almazrouei, E. , Alobeidli, H. , Alshamsi, A. , Cappelli, A. , Cojocaru, R. , Debbah, M. , Goffinet, E. , Heslow, D. , Launay, J. , Malartic, Q. , et al. (2023). Falcon-40b: an open large language model with state-of-the-art performance. Findings of the Association for Computational Linguistics: ACL, 2023:10755–10773.

Authors
Cheng, H., Sheng, B., Lee, A., Chaudary, V., Atanasov, A. G., Liu, N., Qiu, Y., Wong, T. Y., Tham, Y.-C., & Zheng, Y.
Affiliations
  • Indiana University Bloomington (US)
  • Shanghai Jiao Tong University (CN)
  • University of Washington (US)
  • McMaster University (CA)
  • Ludwig Boltzmann Institute Digital Health and Patient Safety (AT)
  • National University of Singapore (SG)
  • Tsinghua University (CN)
  • Singapore National Eye Center (SG)
  • State Key Laboratory of Ophthalmology (CN)
Corresponding author
Zheng, Yingfeng
ORCID
Venue
bioRxiv (Cold Spring Harbor Laboratory)
Publisher
openRxiv
Published
2024-03-26
DOI
10.1101/2024.03.25.586710
OpenAlex
W4393190288
Cited by
15
References
22
Citations by year
2026: 3 · 2025: 7 · 2024: 5
Type
posted content
Language
en
Open access
Licence
cc-by-nc-nd
Locations
bioRxiv (Cold Spring Harbor Laboratory) (acceptedVersion, cc-by-nc-nd)
Topics
  • Social Sciences › Safety Research › Academic integrity and plagiarism
  • Social Sciences › Decision Sciences › Information Systems and Management › Academic Publishing and Open Access
  • Social Sciences › Arts and Humanities › History and Philosophy of Science › Academic Writing and Publishing
Keywords
AI-generated text, large language models, ChatGPT, scientific writing, arXiv, bioRxiv, medRxiv, citation impact, hypothesis generation

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-10.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-10.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review