CoChat
Sign Up
Cited by 30Academic integrity and plagiarismMisinformation and Its ImpactsAuthorship Attribution and Profiling

AI for scientific integrity: detecting ethical breaches, errors, and misconduct in manuscripts

Open access
Pellegrina, Diogo, Helmy, Mohamed·2025·Frontiers in Artificial Intelligence
DOI 10.3389/frai.2025.1644098PMC PMC12436494
Open this paper in CoChat — free accountOpen PDF

Abstract

The use of Generative AI (GenAI) in scientific writing has grown rapidly, offering tools for manuscript drafting, literature summarization, and data analysis. However, these benefits are accompanied by risks, including undisclosed AI authorship, manipulated content, and the emergence of papermills. This perspective examines two key strategies for maintaining research integrity in the GenAI era: (1) detecting unethical or inappropriate use of GenAI in scientific manuscripts and (2) using AI tools to identify mistakes in scientific literature, such as statistical errors, image manipulation, and incorrect citations. We reviewed the capabilities and limitations of existing AI detectors designed to differentiate human-written (HWT) from machine-generated text (MGT), highlighting performance gaps, genre sensitivity, and vulnerability to adversarial attacks. We also investigate emerging AI-powered systems aimed at identifying errors in published research, including tools for statistical verification, citation validation, and image manipulation detection. Additionally, we discuss recent publishing industry initiatives to combat AI-driven papermills. Our investigation shows that these developments are not yet sufficiently accurate or reliable yet for use in academic assessment, they mark an early but promising steps toward scalable, AI-assisted quality control in scholarly publishing.

AI snapshot

Written from the paper’s full text.

TL;DR

This perspective reviews AI’s dual role in scientific integrity: detecting unethical GenAI use in manuscripts and using AI to find errors, image manipulation, and misconduct in published literature. It concludes that current detectors and error-detection tools are promising but not yet accurate or reliable enough for automated academic assessment, so human oversight remains essential.

Key findings

  • AI-text detectors vary in accuracy and are vulnerable to evasion: an early study reported false negative rates of 32% for GPTZero and 49% for OpenAI’s Text Classifier, rising to 95% and 98% after QuillBot paraphrasing; OpenAI removed its classifier in July 2023 citing low accuracy.
  • Detector performance is sensitive to genre and source model: MGTBench used 1,000 human-written and 1,000 machine-generated texts per genre; ConDA’s F1 dropped from 0.99 to 0.67 when trained on news and tested on essays, and to 0.00 when trained on StableLM and tested on Claude; DEMASQ dropped from 0.92 to 0.71 across LLM changes and to 0.23 across genre changes.
  • Bias findings conflict: an earlier study claimed higher false positive rates for non-native English writers, but a later GRE-writing study found non-native texts had the highest perplexity and machine-generated texts much lower perplexity, reporting 99.9% accuracy in detecting machine-generated texts using perplexity alone.
  • In the authors’ own test, DeepSeek generated 10 fake computer science abstracts; GPTZero and Copyleaks confidently detected 9 of 10 as machine-generated, but after AIUndetect obfuscation GPTZero considered 9 highly confident human and Copyleaks rated all as 0% AI. The paper also reviews AI error-detection tools such as Statcheck, Retraction Watch, LLM-based citation checking, Imagetwin, Proofig AI, PubPeer, The Black Spatula Project, and YesNoError, plus publisher initiatives from Wiley, Springer Nature/Slimmer AI, and the STM Integrity Hub.

Methods

  • Perspective/review design: the authors synthesize prior studies and examples on GenAI detection, AI detection of mistakes in scientific literature, and AI detection of papermills.
  • Illustrative experiment: in May 2025 they prompted DeepSeek to generate 10 made-up computer science abstracts of 100–250 words, evaluated them with GPTZero and Copyleaks, then obfuscated them with AIUndetect using Manual/Balance/Academic settings and retested; they also tried humanizing texts via ChatGPT using prior abstracts.
  • Main analysis is narrative and descriptive, comparing reported detector performance, error-detection tools, and publisher initiatives rather than conducting a systematic review or primary validation study.

Limitations

  • The paper states that current AI-generated text detectors vary in accuracy and remain vulnerable to paraphrasing and genre-shifting adversarial attacks; tools for detecting scientific errors or misconduct require further validation before high-stakes use; training-data biases and discipline variation challenge generalizability.
  • The paper states these technologies are in early stages and should not yet be relied upon to automate evaluation or judgment of researchers’ work; human oversight remains essential.
  • Evident from the text: the authors’ own detector test is small (10 abstracts), uses only two accessible detectors, and is an illustrative check rather than a systematic validation; the article is a perspective, not a systematic review or primary empirical study.

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

Integrity signals

Registry

Crossref · 10.3389/frai.2025.1644098

Cites

16 of 16 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

7 of 26 cited sources checked · 3 supported · 1 flagged · accuracy not yet measured

  • Premisecitation

    Bias in the training datasets used by AI detectors can impact their ability to classify some types of human-written text.

    “Another factor that contributes to the low accuracy in MGT detection is bias in the training datasets used by AI detectors, which can impact their ability to classify some types of HWT.”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites Liang et al., 2023 source

      The section “GPT detectors exhibit bias against non-native English authors” reports that detectors misclassified more than half of the human-written TOEFL essays while accurately classifying the US student essays. The Discussion says details of most detectors’ training data are unavailable; it does not establish that bias in their training datasets caused the misclassification. (LLM, conf=0.87)

      Experimental — accuracy not yet measured
  • Premisecitation

    LLMs can detect incorrect or misattributed citations with limited context.

    “A recent study demonstrated that LLMs can detect incorrect or misattributed citations with limited context, offering a valuable layer of quality control for reference accuracy, an area often overlooked during peer review (Zhang and Abernethy, 2024).”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites Zhang and Abernethy, 2024 source
  • Findingown data

    Humanizing texts from within ChatGPT by giving it samples of previous abstracts and asking it explicitly to create texts that look human did not disrupt either detector.

    “We also tried humanizing the texts from within ChatGPT, by giving it samples of our previous abstracts and by asking it explicitly to create texts that look human, both detectors were not disrupted by this approach.”
    Quote in the paperNot a substantial claim
  • Interpretationreasoning

    These developments are not yet sufficiently accurate or reliable for use in academic assessment.

    “Our investigation shows that these developments are not yet sufficiently accurate or reliable yet for use in academic assessment, they mark an early but promising steps toward scalable, AI-assisted quality control in scholarly publishing.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The use of Generative AI in scientific writing has grown rapidly.
  • Premisecitation

    The Black Spatula Project and YesNoError aim to identify a broader range of mistakes in published literature, including mathematical inconsistencies, incorrect units, and flawed experimental logic.

    “In addition, newer initiatives such as The Black Spatula Project (n.d.) and YesNoError (n.d.) aim to identify a broader range of mistakes in published literature, including mathematical inconsistencies, incorrect units, and flawed experimental logic (Gibney, 2025).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Gibney, 2025 source
  • May combine two claims. 21
  • Classified as probably not a substantial claim. 17

  • May combine two claims.

    Premisecitation

    There is a trade-off: detectors robust against genre changes (e.g., ConDA) are poor at detecting MGT created with a model different from the one they were trained on, while detectors robust against changes in LLM (e.g., DEMASQ) fail when there is a change in genre.

    “There is a trade-off where detectors that are robust against genre changes like ConDA (Bhattacharjee et al., 2023) (F1-score when trained with news dropped from 0.99 to 0.67 when testing essays) are poor at detecting MGT created with a model different than the one it was trained on, when trained with StableLM (StableLMTuned-Alpha, n.d.) the F1-score testing Claude drops to 0.00. On the other hand,”
    Quote in the paperSubstantial claim2 of 3 sources checked
    • cites Bhattacharjee et al., 2023 source

      Split into 2 atoms — 0 confirmed, 1 partially supported, 1 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

      Checked by a reasoning model — may be wrong.

      Experimental — accuracy not yet measured
    • cites Kumari et al., 2023 source

      Split into 2 atoms — 0 confirmed, 2 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

      Experimental — accuracy not yet measured
    • cites StableLMTuned-Alpha, n.d.

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • May combine two claims.

    Premisecitation

    GenAI tools are used in scientific writing for manuscript drafting, literature summarization, and data analysis.

    “The use of Generative AI (GenAI) in scientific writing has grown rapidly, offering tools for manuscript drafting, literature summarization, and data analysis.”
    • Also stated in: § Introduction · plain
      “These tools assist in generating introductions, summarizing findings, aligning content with journal guidelines, and automating literature reviews (Gauckler and Werner, 2024; Kim, 2023).”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites Gauckler and Werner, 2024 source
    • cites Kim, 2023 source

      In “Using ChatGPT for language editing in scientific articles” (page unavailable), Kim describes ChatGPT editing prose for scientific articles and notes recent attention to AI chatbots. The paper does not establish rapid growth in use or describe manuscript drafting, literature summarization, and data analysis as uses of these tools. (LLM, conf=0.84)

      Experimental — accuracy not yet measured

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

    Premisecitation

    An earlier study claimed that several detectors exhibited higher false positive rates on texts written by non-native English speakers, attributed to non-native writing showing lower text perplexity.

    “An earlier study claimed that several detectors exhibited higher false positive rates (FPR) on texts written by non-native English speakers (Liang et al., 2023). This was attributed to non-native writing showing lower text perplexity, making it appear, paradoxically, more machine-like.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites Liang et al., 2023 source

  • May combine two claims.

    Premisecitation

    A recent study showed that no detector is completely invulnerable to adversarial attacks, but a newer version of Copyleaks resisted most types of attacks more often than not.

    “Although no detector is completely invulnerable to adversarial attacks, the authors show that a newer version of Copyleaks (n.d.) resisted most types of attacks more often than not but lacked proper statistics.”
    Quote in the paperNot a substantial claim1 of 2 sources checked
    • cites Fishchuk and Braun, 2024 source
    • cites Copyleaks, n.d.

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • May combine two claims.

    Findingown data

    In the authors' test, DeepSeek generated 10 abstracts for made-up computer science papers; GPTZero and Copyleaks confidently detected 9 out of 10 as MGT, GPTZero was uncertain about one abstract, and Copyleaks evaluated another one as 0% AI.

    “The resulting texts were evaluated by the two easily accessible AI detectors, GPTZero and Copyleaks. Both detected confidently that 9 out of 10 abstracts were MGT, GPTZero was uncertain about one abstract, and Copyleaks evaluated another one as 0% AI.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    After obfuscation by AI Undetect, GPTZero was moderately confident one abstract was AI while 9 were considered highly confident to be human; Copyleaks evaluated all of them as 0% AI.

    “After each abstract was obfuscated by AI Undetect (n.d.) (using the settings Manual/Balance/Academic), and the tests were repeated and GPTZero was moderately confident one abstract was AI while 9 were considered highly confident to be human. Copyleaks evaluated all of them as 0% AI (Figure 2; Supplementary File 1).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites AI Undetect, n.d.

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • May combine two claims.

    Findingown data

    Although the detectors had acceptable accuracy on the unedited MGTs, AIUndetect was able to fool them almost all times.

    “Although the detectors had acceptable accuracy on the unedited MGTs, AIUndetect was able to fool them almost all times.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites AI Undetect, n.d.

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • May combine two claims.

    Interpretationcitation

    Existing AI detectors designed to differentiate human-written from machine-generated text have performance gaps, genre sensitivity, and vulnerability to adversarial attacks.

    “We reviewed the capabilities and limitations of existing AI detectors designed to differentiate human-written (HWT) from machine-generated text (MGT), highlighting performance gaps, genre sensitivity, and vulnerability to adversarial attacks.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The use of Generative AI in scientific writing has grown rapidly.

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

    Premisecitation

    OpenAI removed its text classifier tool in July 2023, citing low accuracy concerns.

    “Since July 2023, OpenAI removed its texts classifier tool, citing low accuracy concerns (OpenAI, n.d.) and has not released a new version.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites OpenAI, n.d.

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

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

    Premisecitation

    Retraction Watch and the accompanying Retraction Database track retracted papers and the reasons behind them.

    “A second early example comes from Retraction Watch (n.d.) and the accompanying Retraction Database, which tracks retracted papers and the reasons behind them.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Retraction Watch, n.d.

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • May combine two claims.

    Premisecitation

    A more recent study found that non-native texts had the highest perplexity, while machine-generated texts consistently had much lower perplexity, and using this feature alone achieved 99.9% accuracy in detecting MGTs.

    “Contrary to the earlier claims, this analysis showed that non-native texts had the highest perplexity, while MGTs consistently had much lower perplexity. Using this feature alone, the authors reported 99.9% accuracy in detecting MGTs.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Jiang et al., 2024 source

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

    Interpretationreasoning

    The conflicting findings on non-native bias may be explained by differences in dataset composition, detector models, or evaluation design.

    “These conflicting findings may be explained by differences in dataset composition, detector models, or evaluation design.”
    Quote in the paperSubstantial claim

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

    Premisenone given

    Errors in scientific literature can mislead subsequent research, influence policy decisions, and erode public trust in science.

    “Regardless of their origin, such errors can have a lasting impact, misleading subsequent research, influencing policy decisions, and eroding public trust in science.”
    Quote in the paperSubstantial claim

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

    Premisereasoning

    Mistakes in scientific literature are an enduring challenge, and errors are an inevitable part of the process.

    “Mistakes in scientific literature are an enduring challenge in research. Since science is a human endeavor, errors are an inevitable part of the process.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Risks of GenAI in scientific writing include undisclosed AI authorship, manipulated content, and papermills.

    “However, these benefits are accompanied by risks, including undisclosed AI authorship, manipulated content, and the emergence of papermills.”
    • Also stated in: § Introduction · plain
      “Among the most pressing are the unethical use of GenAI in writing, such as undisclosed AI usage, manipulation of scientific content or the proliferation of papermills (Pérez-Neri et al., 2022), which risk eroding trust in scholarly communication.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Pérez-Neri et al., 2022 source

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

    Premisenone given

    These image manipulation detection tools have contributed to the exposure of widespread image duplication, leading to retractions and corrections in prominent journals.

    “These tools have contributed to the exposure of widespread image duplication, leading to retractions and corrections in prominent journals.”
    Quote in the paperNot a substantial claim

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

    Premisecitation

    Some mistakes stem from honest miscalculations, misinterpretations, or technical oversights, while others arise from fraudulent practices including data fabrication, image manipulation, and methodological misreporting.

    “Some mistakes stem from honest miscalculations, misinterpretations, or technical oversights, while others arise from fraudulent practices, including data fabrication, image manipulation, and methodological misreporting (Conroy, 2025).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Conroy, 2025 source

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

    Premisenone given

    AI presents opportunities for strengthening research integrity by identifying mistakes in manuscripts, including factual inaccuracies, statistical errors, and subtle inconsistencies that peer review may overlook.

    “At the same time, AI presents promising opportunities for strengthening research integrity by identifying mistakes in manuscripts, including factual inaccuracies, statistical errors, and subtle inconsistencies that peer review may overlook.”
    Quote in the paperNot a substantial claim

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

    Interpretationreasoning

    These systems continue to face significant limitations in terms of accuracy, consistency, and contextual understanding.

    “However, they continue to face significant limitations in terms of accuracy, consistency, and contextual understanding.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Springer Nature, in collaboration with Slimmer AI’s Science division, has developed two AI tools focused on detecting fraudulent submissions.
    • Rests on an assumption: The STM Integrity Hub offers a centralized cloud-based platform for publishers to share intelligence and detect papermill-generated manuscripts through automated screening applications.

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    Tools designed to detect scientific errors or misconduct require further validation before they can be reliably applied in high-stakes settings such as manuscript screening or academic evaluations.

    “Similarly, tools designed to detect scientific errors or misconduct require further validation before they can be reliably applied in high-stakes settings such as manuscript screening or academic evaluations.”
    Quote in the paperSubstantial claim

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

    Premisenone given

    These platforms are not yet foolproof.

    “While not yet foolproof, they represent a shift toward scalable, AI-assisted post-publication review.”
    Quote in the paperNot a substantial claim

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

    Premisecitation

    Wiley announced a pilot of a new AI-powered Papermill Detection service, but the specific tools or technologies behind this effort have not been publicly disclosed.

    “For instance, Wiley recently announced a pilot of a new AI-powered Papermill Detection service, although the specific tools or technologies behind this effort have not been publicly disclosed (Rose, 2024).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Rose, 2024

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • May combine two claims.

    Interpretationreasoning

    Current AI-generated text detectors vary in accuracy and remain vulnerable to paraphrasing and genre-shifting adversarial attacks.

    “Current AI-generated text detectors vary in accuracy and remain vulnerable to paraphrasing and genre-shifting adversarial attacks.”
    Quote in the paperSubstantial claim

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

    Interpretationnone given

    These technologies are still in the early stages of integration into journal workflows, and their accuracy and reliability are being actively evaluated.

    “While promising, these technologies are still in the early stages of integration into journal workflows, and their accuracy and reliability are being actively evaluated.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Springer Nature, in collaboration with Slimmer AI’s Science division, has developed two AI tools focused on detecting fraudulent submissions.
    • Rests on an assumption: The STM Integrity Hub offers a centralized cloud-based platform for publishers to share intelligence and detect papermill-generated manuscripts through automated screening applications.

  • May combine two claims.

    Interpretationnone given

    Biases in training data and variation across disciplines pose challenges to the generalizability of AI integrity tools.

    “Biases in training data and variation across disciplines also pose challenges to their generalizability.”
    Quote in the paperNot a substantial claim

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

    Interpretationreasoning

    The broader deployment and refinement of these tools will take time, and their effectiveness in real-world editorial settings remains to be fully assessed.

    “The broader deployment and refinement of these tools will take time, and their effectiveness in real-world editorial settings remains to be fully assessed.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Springer Nature, in collaboration with Slimmer AI’s Science division, has developed two AI tools focused on detecting fraudulent submissions.
    • Rests on an assumption: The STM Integrity Hub offers a centralized cloud-based platform for publishers to share intelligence and detect papermill-generated manuscripts through automated screening applications.

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

    Premisecitation

    Proofig and Imagetwin apply machine learning to identify duplicated or altered figures, common hallmarks of papermill submissions.

    “Services like Proofig and Imagetwin (discussed above) apply machine learning to identify duplicated or altered figures, common hallmarks of papermill submissions.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites Proofig AI, n.d.

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites Imagetwin, n.d.

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

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

    Premisereasoning

    These systems can identify key indicators of papermill involvement, such as reused templates, duplicated phrases, or unnatural statistical patterns.

    “These systems can identify key indicators of papermill involvement, such as reused templates, duplicated phrases, or unnatural statistical patterns.”
    Quote in the paperNot a substantial claim
28 of 28

  • The paper relies on this without support.

    The use of Generative AI in scientific writing has grown rapidly.

    “The use of Generative AI (GenAI) in scientific writing has grown rapidly, offering tools for manuscript drafting, literature summarization, and data analysis.”

  • The paper relies on this without support.

    Springer Nature, in collaboration with Slimmer AI’s Science division, has developed two AI tools focused on detecting fraudulent submissions.

    “Springer Nature, in collaboration with Slimmer AI’s Science division, has developed two AI tools focused on detecting fraudulent submissions.”

  • The paper relies on this without support.

    The STM Integrity Hub offers a centralized cloud-based platform for publishers to share intelligence and detect papermill-generated manuscripts through automated screening applications.

    “Another industry-wide initiative, the STM Integrity Hub, offers a centralized cloud-based platform for publishers to share intelligence and detect papermill-generated manuscripts through automated screening applications.”
3 of 3

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?

  • Tools like LLMs, Black Spatula, YesNoError, and GRIM-Test improve accuracy, but they need good data and human checks.

    “Tools like LLMs, Black Spatula, YesNoError, and GRIM-Test improve accuracy, but they need good data and human checks.”

    From Artificial Intelligence in Detecting Statistical Errors: Implications for Authors, Reviewers, and Editors (2025) · cites this paper as [26]

    The Introduction’s ‘AI detection of mistakes in scientific literature’ section says LLMs can detect citation errors and that Black Spatula and YesNoError flag anomalies for expert review. The ‘GenAI detection tools’ section also describes sensitivity to training datasets. But our copy does not mention GRIM-Test or establish that all four named tools improve accuracy. (LLM, conf=0.87)

    Experimental — accuracy not yet measured

  • Recent studies show that AI is increasingly helpful in research, assisting in spotting errors in methodology, citations, and statistical analyses.

    “Recent studies show that AI is increasingly helpful in research, assisting in spotting errors in methodology, citations, and statistical analyses.”

    From Artificial Intelligence in Detecting Statistical Errors: Implications for Authors, Reviewers, and Editors (2025) · cites this paper as [26]

  • A 2025 article in the Frontiers explored the use and possible risks of AI tools in drafting paper manuscripts and in data analysis, and explored the use of LLMs in detection of reference errors, the Black Spatula Project and YesNoError tools in detection of mathematical and experimental logic errors.

    “Another 2025 article in the Frontiers explored the use and possible risks that accompany the use of AI tools in the drafting of paper manuscripts and in data analysis. It also explored the use of tools like large language models (LLMs) in the detection of reference errors, the Black Spatula Project and YesNoError tools in the detection of mathematical and experimental logic errors.26”

    From Artificial Intelligence in Detecting Statistical Errors: Implications for Authors, Reviewers, and Editors (2025) · cites this paper as [26]

Cites

Flagged first, then by citations

Research works printed but not linked: 8

  1. Undetectable AI, AI Rewriter, Rewording Tool.
  2. AI Detector—Free AI Checker for ChatGPT, GPT-4, Gemini & More.
  3. Beta Version for Detecting AI-Generated Images.
  4. Image Integrity Risks in Life Science Publications.
  5. PubPeer - PubPeer 2.0.
  6. QuillBot’s Paraphraser: The Best AI Paraphrasing Tool
  7. Tracking Retractions as a Window into the Scientific Process.
  8. Rose, 2024, Wiley Announces Pilot of New AI-Powered Papermill Detection Service

Cited by

Sorted by citations

Related

Authors
Pellegrina, D., & Helmy, M.
Affiliations
  • University of Saskatchewan (CA)
  • Idaho State University (US)
Corresponding author
Helmy, Mohamed
ORCID
Venue
Frontiers in Artificial Intelligence
Publisher
Frontiers Media SA
Published
2025-09-02
Volume
8
Pages
1644098-1644098
ISSN
2624-8212 (ISSN-L)
DOI
10.3389/frai.2025.1644098
PMC
PMC12436494
OpenAlex
W4413902283
Semantic Scholar
c0e48fde9744efa5b0761ed12288bd50296227b2
Cited by
30
References
13
Citations by year
2026: 28 · 2025: 2
FWCI
31.37
Type
journal article
Language
en
Update notices
correction (2025-11-19)
Open access
Licence
Locations
  • Frontiers in Artificial Intelligence (publishedVersion, cc-by)
  • Europe PMC (PubMed Central) (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • PubMed Central (publishedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
Innovation Saskatchewan
Topics
  • Social Sciences › Safety Research › Academic integrity and plagiarism
  • Social Sciences › Sociology and Political Science › Misinformation and Its Impacts
  • Physical Sciences › Computer Science › Artificial Intelligence › Authorship Attribution and Profiling
Keywords
research integrity, human-written vs machine-generated text, image forgery detection, citation validation

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 2026-10-29.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-29.

Cited sources are checked by gpt-6-sol.

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