CoChat
Sign Up
Cited by 42Authorship Attribution and ProfilingAdvanced Steganography and Watermarking TechniquesTopic Modeling

Authorship Attribution in the Era of LLMs: Problems, Methodologies, and Challenges

Open access
Huang, Baixiang, Chen, Canyu, Shu, Kai·2025·ACM SIGKDD Explorations Newsletter
DOI 10.1145/3715073.3715076PMC PMC12019761
Open this paper in CoChat — free accountOpen PDF

Abstract

A literature review of human-written, LLM-generated, attributed, and co-authored text detection problems, methods, and challenges.

AI snapshot

Written from the paper’s full text.

TL;DR

This survey maps authorship attribution in the LLM era into four problem types and reviews methods, datasets, metrics, and open challenges. It argues that LLMs blur human–machine authorship boundaries and that progress requires better generalization, explainability, standardized benchmarks, and interdisciplinary work.

Key findings

  • The paper organizes the field into four representative problems: human-written text attribution, LLM-generated text detection, LLM-generated text attribution, and human-LLM co-authored text attribution, with complexity increasing across them.
  • Detection approaches are grouped into metric-based and model-based families, including feature-based, neural network-based, zero-shot, and watermarking methods; neural detectors often outperform metric-based ones but are less explainable, need retraining for new LLMs, and are vulnerable to adversarial or poisoning attacks.
  • The survey compares 21 benchmarks/datasets: all support LLM-generated text detection, but fewer support LLM-generated text attribution or human-LLM co-authored attribution, and many treat human-written text as one class, making them unsuitable for identifying individual human authors.
  • Human judgment is weak: untrained reviewers distinguish GPT-3 text from human text at roughly chance level, and experienced faculty/researchers identify GPT-generated academic writing correctly only about half the time. Commercial detectors often claim over 99% accuracy, yet few are tested on shared benchmarks and many have high false-positive rates.

Methods

  • Design: a literature review/survey that synthesizes existing research rather than presenting new experiments.
  • Data/sample: cited prior studies, 21 authorship attribution datasets/benchmarks, and a table of commercial and open-source LLM-text detectors.
  • Analysis: qualitative thematic categorization of four attribution problems, comparison of methodologies, datasets, evaluation metrics, and discussion of future directions.

Limitations

  • Stated: fewer studies address human-LLM co-authored attribution because it is difficult, and existing work often simplifies the task; many datasets are unsuitable for individual human authorship attribution; detectors lack robustness to decoding strategies, prompts, adversarial attacks, unseen domains, and unseen models.
  • Stated: LLM-based attribution becomes less effective as the number of candidate authors grows due to context-length limits, and training-data biases can produce inaccurate attributions for marginalized groups and languages.
  • Evident: as a survey, it provides no new empirical validation or standardized benchmark results, and its comparisons draw on heterogeneous prior datasets and metrics, so reported performance is not directly comparable; the text also does not describe a systematic search or inclusion protocol.

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

Integrity signals

Registry

Crossref · 10.1145/3715073.3715076

Cites

212 of 212 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

63 of 161 cited sources checked · 38 supported · 0 flagged · accuracy not yet measured

  • Premisecitation

    Detection algorithms frequently outperform humans in distinguishing LLM-generated from human-written text.

    “In contrast, detection algorithms frequently outperform the human in this task [Ippolito et al., 2020].”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites Ippolito et al., 2020 source
  • Premisecitation

    Untrained human reviewers are often unable to distinguish GPT-3-generated text from human-written text, identifying it correctly only at a rate consistent with random chance.

    “untrained human reviewers are often unable to distinguish GPT-3-generated text from humanwritten text, identifying it correctly only at a rate consistent with random chance [Clark et al., 2021].”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites Clark et al., 2021 source
  • Premisecitation

    The advent of LLMs has complicated authorship attribution, making it increasingly difficult to distinguish between LLM-generated texts and human-written texts.

    “The rapid advancements of Large Language Models (LLMs) have blurred the lines between human and machine authorship, posing significant challenges for traditional methods.”
    • Also stated in: § 1. INTRODUCTION · plain
      “The advent of LLMs has complicated authorship attribution, making it increasingly difficult to distinguish between LLM-generated texts and human-written texts [Clark et al., 2021;Sadasivan et al., 2023].”
    Quote in the paperSubstantial claim2 of 2 sources checked
    • cites Clark et al., 2021 source
    • cites Sadasivan et al., 2023 source
  • Premisecitation

    Existing detectors lack robustness to input sequence length.

    “Existing detectors also lack robustness to various factors, such as alternative decoding strategies [Ippolito et al., 2020], input sequence length [Solaiman et al., 2019], different prompts [Kumarage et al., 2023;Lu et al., 2023], and repetition penalties [Fishchuk and Braun, 2023].”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites Solaiman et al., 2019 source
  • Premisecitation

    The detection problem is becoming increasingly difficult.

    “Sadasivan et al. [2023] provided theoretical insights indicating that the detection problem is becoming increasingly difficult.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites Sadasivan et al., 2023 source
  • May combine two claims. 38
  • Classified as probably not a substantial claim. 42

  • May combine two claims.

    Premisecitation

    Detectors are vulnerable to adversarial attacks including homoglyph attacks, whitespace insertion, syntactic perturbations, synonym replacement, and paraphrasing.

    “Additionally, detectors are vulnerable to adversarial attacks, including homoglyph attacks [Gagiano et al., 2021;Macko et al., 2024], whitespace insertion [Cai and Cui, 2023], syntactic perturbations [Bhat and Parthasarathy, 2020a], synonym replacement [Kulkarni et al., 2023], and paraphrasing [Krishna et al., 2024b;Shi et al., 2024;Becker et al., 2023].”
    Quote in the paperSubstantial claim2 of 3 sources checked
    • cites Gagiano et al., 2021

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

    • cites Macko et al., 2024 source
    • cites Cai and Cui, 2023 source

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

    Premisecitation

    Techniques like slanted triangular learning rates and gradual unfreezing can be used to avoid catastrophic forgetting during fine-tuning.

    “Techniques like slanted triangular learning rates and gradual unfreezing can be used to avoid catastrophic forgetting during fine-tuning [Howard and Ruder, 2018].”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites Howard and Ruder, 2018 source

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

    Premisecitation

    The task of detecting LLM-generated text is typically approached as a binary classification problem.

    “Typically, this task is approached as a binary classification problem [Zellers et al., 2019;Solaiman et al., 2019;Jawahar et al., 2020b;Fagni et al., 2021;Mitchell et al., 2023].”
    Quote in the paperNot a substantial claim3 of 3 sources checked
    • cites Zellers et al., 2019 source
    • cites Solaiman et al., 2019 source
    • cites Jawahar et al., 2020b source

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

    Premisecitation

    Topic diversity and dataset size are crucial for effective cross-domain transfer.

    “Rivera-Soto et al. [2021] concluded that topic diversity and dataset size are crucial for effective cross-domain transfer.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites Rivera-Soto et al., 2021 source

      The Conclusion attributes Reddit’s strong cross-domain transfer to both topic diversity and dataset size. But §4.2 says size only partially explains Reddit’s performance and that increasing the number of Amazon authors yielded marginal gains. Calling both factors “crucial” for cross-domain transfer generally is stronger than the paper’s finding. (LLM, conf=0.82)

      Experimental — accuracy not yet measured

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

    Premisecitation

    Authorship attribution methods are typically divided into classification-based methods for a small set of candidate authors and similarity-based ranking methods for larger numbers of authors.

    “Authorship attribution methods are typically divided into classification-based methods for a small set of candidate authors and similarity-based ranking methods for larger numbers of authors [Rivera-Soto et al., 2021;Huertas-Tato et al., 2022].”
    Quote in the paperNot a substantial claim2 of 2 sources checked
    • cites Rivera-Soto et al., 2021 source

      The Introduction (page unavailable) contrasts a prior closed-world study of 21 authors with this paper’s open-world setting involving far more authors. The Experiments metrics passage (page unavailable) describes ranking candidates by cosine similarity. The paper does not say that authorship attribution methods are typically divided into classification-based and similarity-based categories, or that classification is generally used for small candidate sets. (LLM, conf=0.84)

      Experimental — accuracy not yet measured
    • cites Huertas-Tato et al., 2022 source

      The paper describes classifier-based authorship approaches in §2.1 (p. 3) and evaluates its own similarity-based attribution from a pool of authors in §5.2 (p. 8). It does not say that methods are typically divided into these two categories or associate classification with small candidate sets and similarity-based ranking with larger ones. (LLM, conf=0.78)

      Experimental — accuracy not yet measured

  • May combine two claims.

    Premisecitation

    LLMs possess the ability to identify the author of unknown text based on nuanced linguistic features, making it possible to conduct end-to-end authorship attribution.

    “Beyond feature extraction, LLMs possess the ability to identify the author of unknown text based on nuanced linguistic features, making it possible to conduct end-to-end authorship attribution [Huang et al., 2024].”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites Huang et al., 2024 source

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

    Premisecitation

    LLMs possess significant capabilities in reasoning and problem-solving, particularly in zero-shot learning within resource-limited domains.

    “Authorship attribution is a complex reasoning task, and LLMs possess significant capabilities in reasoning and problem-solving, particularly in zero-shot learning within resource-limited domains [Kojima et al., 2022].”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites Kojima et al., 2022 source

      The Abstract (p. 1) reports substantial gains on diverse reasoning tasks using zero-shot prompting, and the Discussion (p. 9) describes LLMs as decent zero-shot reasoners with potential for generic problem solving. The paper does not establish these capabilities specifically within resource-limited domains; its zero-shot method avoids handcrafted examples and time-consuming fine-tuning, which is a narrower point. (LLM, conf=0.84)

      Experimental — accuracy not yet measured

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

    Premisecitation

    Contrastive learning is often used with pre-trained language models to enhance stylistic representation.

    “Contrastive learning [Khosla et al., 2020] is often used with pre-trained language models to enhance stylistic representation by maximizing similarity between texts written by the same authors and minimizing it with texts from different authors [Huertas-Tato et al., 2022].”
    Quote in the paperNot a substantial claim1 of 2 sources checked
    • cites Khosla et al., 2020 source
    • cites Huertas-Tato et al., 2022 source

      Section 2.3 (p. 5) says the authors pair pre-trained transformers with contrastive learning to build representations of writing style. The Introduction (p. 2) says texts by the same author should have higher similarity than texts by different authors. The paper supports this approach, but its discussion of recent efforts does not establish that it is often used for stylistic representation. (LLM, conf=0.82)

      Experimental — accuracy not yet measured

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

    Premisecitation

    One variation of the human-LLM co-authored text attribution task is to detect the boundary between human-written and LLM-generated text.

    “One variation of this task is to detect the boundary between human-written and LLM-generated text [Cutler et al., 2021;Wang et al., 2024].”
    Quote in the paperNot a substantial claim1 of 2 sources checked
    • cites Cutler et al., 2021

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

    • cites Wang et al., 2024 source

  • May combine two claims.

    Premisecitation

    Human-authored texts tend to be more coherent and exhibit greater lexical diversity compared to LLM-generated texts.

    “Human-authored texts tend to be more coherent and exhibit greater lexical diversity compared to LLM-generated texts [Guo et al., 2023;Dugan et al., 2023;Zhang et al., 2024].”
    Quote in the paperNot a substantial claim1 of 3 sources checked
    • cites Guo et al., 2023 source
    • cites Dugan et al., 2023 source

      Section 6, “RoFT-Specific Insights” (page unavailable), reports that irrelevant and contradictory sentences were among the most reliable signs participants used to detect generated text. This lends support to the coherence comparison, but the paper does not establish that human-authored text has greater lexical diversity; its sentence-level analysis examines length, parts of speech, and named entities instead. (LLM, conf=0.83)

      Experimental — accuracy not yet measured
    • cites Zhang et al., 2024 source

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

    Premisecitation

    The evaluation of the quality of machine-generated excerpts has traditionally relied on human judgement, considered the gold standard for open-domain generation systems.

    “The evaluation of the quality of machine-generated excerpts has traditionally relied on human judgement, which is considered the gold standard for open-domain generation systems [van der Lee et al., 2019;Gehrmann et al., 2019a].”
    Quote in the paperNot a substantial claim2 of 2 sources checked
    • cites van der Lee et al., 2019 source
    • cites Gehrmann et al., 2019a source

      The Abstract reports a human-subjects study of detecting generated text, and the Method describes statistical tests for that detection task. Neither passage says that human judgment has traditionally been the gold standard for evaluating the quality of open-domain generation; searches for human evaluation and quality found no passage making that point. (LLM, conf=0.73)

      Experimental — accuracy not yet measured

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

    Premisecitation

    Some studies categorize any text generated, modified, or extended by a machine as LLM-generated, simplifying the task to LLM-generated Text Detection or LLM-generated Text Attribution.

    “Some studies categorize any text that is generated, modified, or extended by a machine as LLM-generated.This simplifies the task to either LLM-generated Text Detection or LLMgenerated Text Attribution [Yang et al., 2023c;Crothers et al., 2023].”
    Quote in the paperNot a substantial claim1 of 2 sources checked
    • cites Yang et al., 2023c

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

    • cites Crothers et al., 2023 source

      The Definition and Scope section defines machine-generated text as text produced, modified, or extended by a machine. Section IV describes detection as distinguishing machine-generated from human text, and Section V-B discusses attribution to a generative model. The paper uses the broader category “machine-generated text,” however; it does not categorize all such text as LLM-generated. (LLM, conf=0.86)

      Experimental — accuracy not yet measured

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

    Premisecitation

    Some studies have formulated a seven-class classification that includes one human class and six LLM classes.

    “Some studies have formulated a seven-class classification that includes one human class and six LLM classes [He et al., 2023;Wang et al., 2024].”
    Quote in the paperNot a substantial claim1 of 2 sources checked
    • cites He et al., 2023

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

    • cites Wang et al., 2024 source

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

    Premisecitation

    Models like DNA-GPT and DetectGPT utilized the T5 model to simulate scenarios in which humans modify LLM-generated texts.

    “Models like DNA-GPT [Yang et al., 2023a] and DetectGPT [Mitchell et al., 2023] utilized the T5 model [Raffel et al., 2020] to simulate scenarios in which humans modify LLMgenerated texts.”
    Quote in the paperNot a substantial claim1 of 3 sources checked
    • cites Yang et al., 2023a

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

    • cites Mitchell et al., 2023 source
    • cites Raffel et al., 2020

      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).

    Premisecitation

    In human authorship attribution, where there are a large number of candidate authors, retrieval metrics such as Mean Reciprocal Rank (MRR) and recall-at-k are used.

    “In human authorship attribution, where there are a large number of candidate authors, retrieval metrics such as Mean Reciprocal Rank (MRR) and recall-at-k are used [Rivera-Soto et al., 2021].”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites Rivera-Soto et al., 2021 source

  • May combine two claims.

    Premisecitation

    The increasingly large context length of LLMs enables in-context learning (ICL) by incorporating more documents, enhancing the model's ability to capture intricate writing patterns.

    “The increasingly large context length of LLMs enables in-context learning (ICL) [Brown et al., 2020] by incorporating more documents, enhancing the model's ability to capture intricate writing patterns.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites Brown et al., 2020 source

      The Introduction (p. 5) describes in-context learning with as many demonstrations as fit in the model’s context window, and §3.7 (p. 20) reports improved performance with more examples in context. These passages support the role of context capacity in ICL, but they do not establish that incorporating more documents enhances the model’s ability to capture intricate writing patterns. (LLM, conf=0.85)

      Experimental — accuracy not yet measured

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

    Premisecitation

    The challenge remains to provide higher-level explainability that aligns with human cognitive processes.

    “The challenge remains to provide higher-level explainability that aligns with human cognitive processes [Rudin, 2019].”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites Rudin, 2019 source

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

    Interpretationcitation

    Advanced techniques are required to dissect these integrations, identifying where and how the contributions of LLM intersect with human input.

    “Advanced techniques are required to dissect these integrations, identifying where and how the contributions of LLM intersect with human input [Wang et al., 2024].”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites Wang et al., 2024 source

  • May combine two claims.

    Premisecitation

    Models like BERT and RoBERTa have shown limitations in cross-domain tasks, and adapting models to new domains remains difficult due to factors like dataset size variability and writing styles.

    “Models like BERT and RoBERTa have shown limitations in cross-domain tasks [Barlas and Stamatatos, 2020b;Huertas-Tato et al., 2022], and adapting models to new domains remains difficult due to factors like dataset size variability and writing styles.”
    Quote in the paperNot a substantial claim1 of 2 sources checked
    • cites Barlas and Stamatatos, 2020b

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

    • cites Huertas-Tato et al., 2022 source

      Section 2.1 (p. 3) says adaptation to new domains is usually hard and identifies variation in document size and style; Section 5.2 (p. 8) shows the RoBERTa baseline underperforming other models on out-of-domain authorship tasks. The paper does not establish a cross-domain limitation for BERT specifically, and its discussion of varying size concerns documents rather than explicitly identifying variability in dataset size as a cause of adaptation difficulty. (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

    The Linguistic Inquiry and Word Count (LIWC) tool uses more than 100 psychological dimensions to analyze word use, reflecting distinct language variations among different groups in specific genres and languages.

    “The Linguistic Inquiry and Word Count (LIWC) tool exemplifies how automated text analysis can use more than 100 psychological dimensions to analyze word use, reflecting distinct language variations among different groups in specific genres and languages [Goldstein-Stewart et al., 2009;Pennebaker et al., 2015;Dudău and Sava, 2021].”
    Quote in the paperNot a substantial claim1 of 3 sources checked
    • cites Goldstein-Stewart et al., 2009 source
    • cites Pennebaker et al., 2015

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

    • cites Dudău and Sava, 2021 source

      The Introduction describes LIWC2015 as an automated word-counting tool with categories that include psychological constructs, and the Discussion reports differences among the languages studied. But the Introduction says it has approximately 90 categories in total—not more than 100 psychological dimensions—and the cited passages do not establish the claim about variations among groups in specific genres. (LLM, conf=0.88)

      Experimental — accuracy not yet measured

  • May combine two claims.

    Premisecitation

    The challenge in social media stems from the brief nature of posts and a large pool of potential authors, making the attribution of short documents particularly difficult.

    “The challenge in social media stems from the brief nature of posts and a large pool of potential authors, making the attribution of short documents particularly difficult [Aborisade and Anwar, 2018;Seroussi et al., 2014;Theophilo et al., 2021].”
    Quote in the paperNot a substantial claim1 of 3 sources checked
    • cites Aborisade and Anwar, 2018 source
    • cites Seroussi et al., 2014 source
    • cites Theophilo et al., 2021 source

  • May combine two claims.

    Premisecitation

    Misattribution can lead to wrongful accusations or misinterpretation of an author's intent.

    “Misattribution can lead to wrongful accusations or misinterpretation of an author's intent [Lund et al., 2023].”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites Lund et al., 2023 source

      The 'Authorship attribution is a major issue' section discusses ownership, plagiarism, and the difficulty of tracing sources in AI-generated text (page unavailable). Searches for wrongful accusations and misinterpretation of an author's intent found no passage addressing either consequence. (LLM, conf=0.72)

      Experimental — accuracy not yet measured

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

    Premisecitation

    Kowalczyk et al. detected GPT-2-generated fake reviews using Shapley Additive Explanations (SHAP).

    “For instance, Kowalczyk et al. [2022] detected GPT-2-generated fake reviews using Shapley Additive Explanations (SHAP) [Lundberg and Lee, 2017].”
    Quote in the paperNot a substantial claim1 of 2 sources checked
    • cites Kowalczyk et al., 2022 source
    • cites Lundberg and Lee, 2017 source

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

    Findingnone given

    Although many detectors claim over 99% accuracy in detecting LLM-generated text, few are tested on shared benchmark datasets.

    “Although many of these detectors claim over 99% accuracy in detecting LLM-generated text, few are tested on shared benchmark datasets.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Despite their high accuracy claims, many detectors suffer from high false positive rates, which could falsely accuse individuals of plagiarism and undermine credibility and trust in genuine authors.

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

    Findingnone given

    These detectors are primarily designed for LLM-generated text detection (Problem 2).

    “These detectors are primarily designed for LLM-generated text detection (Problem 2).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingnone given

    Fewer listed datasets support LLM-generated text attribution (Problem 3) and Human-LLM Co-authored Text Attribution (Problem 4).

    “However, fewer support LLM-generated text attribution (Problem 3) and Human-LLM Co-authored Text Attribution (Problem 4).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Neural network-based detectors generally outperform metric-based methods in both human authorship attribution and LLM-generated text detection.

    “Neural network-based detectors generally outperform metric-based methods in both human authorship attribution and LLM-generated text detection problems [He et al., 2023;Zhang et al., 2024].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites He et al., 2023

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

    • cites Zhang et al., 2024 source

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

    Premisecitation

    Detecting LLM-generated text is nearly always feasible with the collection of multiple samples, with precise sample complexity bounds established.

    “Chakraborty et al. [2023a] employed theoretical analysis to argue that detecting LLM-generated text is nearly always feasible with the collection of multiple samples, and they established precise sample complexity bounds for this detection.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Chakraborty et al., 2023a

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

  • May combine two claims.

    Premisecitation

    LLM-generated texts are typically less emotional and more objective than human-written texts.

    “LLM-generated texts are typically less emotional and more objective than human-written texts [Guo et al., 2023].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Guo et al., 2023 source

  • May combine two claims.

    Premisecitation

    Deep learning methods demand significant computational resources and extensive labeled data, with the risk of catastrophic forgetting.

    “whereas deep learning methods demand significant computational resources and extensive labeled data, with the risk of catastrophic forgetting [Ramasesh et al., 2022].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Ramasesh et al., 2022

      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).

    Interpretationnone given

    Neural network approaches often offer less explainability compared to metric-based counterparts.

    “However, these neural network approaches often offer less explainability compared to their metric-based counterparts.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    The rapid evolution of LLMs complicates detection, as newer versions exhibit different stylistic characteristics, making detection models quickly obsolete.

    “The rapid evolution of LLMs further complicates detection, as newer versions exhibit different stylistic characteristics, making detection models quickly obsolete [Chakraborty et al., 2023b].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Chakraborty et al., 2023b

      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

    BERT performed well with large vocabularies, outperforming multi-headed RNNs.

    “Barlas and Stamatatos [2020a] found that BERT performed well with large vocabularies, outperforming multi-headed RNNs.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Barlas and Stamatatos, 2020a

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

  • May combine two claims.

    Premisecitation

    LLMs can generalize across various domains without fine-tuning, mitigating issues related to domain specificity.

    “Unlike BERT-based models, which require computationally expensive fine-tuning and large amounts of domain-specific data for optimal performance, LLMs can generalize across various domains without fine-tuning, thereby mitigating issues related to domain specificity [Barlas and Stamatatos, 2020a].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Barlas and Stamatatos, 2020a

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

  • May combine two claims.

    Premisecitation

    Collaborative texts blend multiple writing styles, masking individual contributions and obscuring distinct authorial signals.

    “Collaborative texts, which blend multiple writing styles, further mask individual contributions and obscure distinct authorial signals [Dauber et al., 2017].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Dauber et al., 2017 source

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

    Premisecitation

    Traditionally, LLMs have been employed mainly for auxiliary tasks such as feature extraction and data annotation.

    “Traditionally, LLMs have been employed mainly for auxiliary tasks such as feature extraction and data annotation [Patel et al., 2023].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Patel et al., 2023 source

  • May combine two claims.

    Interpretationnone given

    Pre-trained language model methods do not require hand-crafted features but require substantial training time and domain-specific labeled data, struggling with cross-domain generalization and explainability.

    “These methods do not require hand-crafted features but require substantial training time and domain-specific labeled data, struggling with cross-domain generalization and explainability.”
    Quote in the paperNot a substantial claim

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

    Interpretationnone given

    The natural language understanding capability allows LLMs to recognize nuances, styles, and patterns in language, which are crucial for distinguishing between authors.

    “The natural language understanding capability allows LLMs to recognize nuances, styles, and patterns in language, which are crucial for distinguishing between authors.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    LLMs are effective with shorter texts, reducing the necessity for long inputs to derive meaningful representations.

    “LLMs are also effective with shorter texts, reducing the necessity for long inputs to derive meaningful representations [Eder, 2015].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Eder, 2015 source

  • May combine two claims.

    Premisecitation

    Zero-shot detection methods are generally statistics-based, enabling the detection of LLM-generated text without additional training.

    “Zero-shot detection methods are generally statistics-based, enabling the detection of LLM-generated text without additional training [Su et al., 2023b].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Su et al., 2023b

      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.

    Interpretationnone given

    This attribution task extends beyond binary classification to handle multiple classes, increasing the complexity of LLM-generated text detection.

    “This attribution task extends beyond binary classification to handle multiple classes, increasing the complexity of LLMgenerated text detection.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    LLM-generated texts typically exhibit consistent syntax and a broader vocabulary.

    “In contrast, LLMgenerated texts typically exhibit consistent syntax and a broader vocabulary [Guo et al., 2023].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Guo et al., 2023 source

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

    Interpretationnone given

    LLMs may inadvertently reproduce snippets of their training data.

    “Additionally, LLMs may inadvertently reproduce snippets of their training data.”
    Quote in the paperNot a substantial claim

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

    Interpretationnone given

    Feature-based methods are more explainable but have drawbacks, such as poor generalizability of certain features across different domains and sampling methods.

    “Feature-based methods are more explainable but have drawbacks, such as poor generalizability of certain features across different domains and sampling methods.”
    Quote in the paperNot a substantial claim

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

    Premisecitation

    Some approaches consider multiple human classes, e.g., a 10-class classification with seven human classes and three LLM classes.

    “For instance, a 10-class classification might include seven human classes and three LLM classes [Tripto et al., 2023].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Tripto et al., 2023 source

  • May combine two claims.

    Premisecitation

    Human-LLM co-authored texts may begin as human drafts that are later extended or revised by LLMs, or vice versa, resulting in an integration of styles that obscures authorial markers.

    “These texts may begin as human drafts that are later extended or revised by LLMs, or vice versa, resulting in an integration of styles that obscures authorial markers [Liu et al., 2023b].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Liu et al., 2023b

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

  • Classified as probably not a substantial claim.

    Interpretationcitation

    Attributing texts to LLMs versus human writers involves recognizing inherent differences in their text generation capabilities.

    “Attributing texts to LLMs versus human writers involves recognizing inherent differences in their text generation capabilities.”
    Quote in the paperSubstantial claim

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

    Premisereasoning

    The detection of LLM-generated text has garnered significant attention.

    “As a result, the detection of LLM-generated text has garnered significant attention [Kumarage and Liu, 2023;Tang et al., 2023a;Wu et al., 2023b;Yang et al., 2023c].”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites Kumarage and Liu, 2023 source
    • cites Tang et al., 2023a

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

    • cites Wu et al., 2023b

      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).

    Interpretationnone given

    Transformer-based models, such as BERT and RoBERTa, are fine-tuned on datasets containing both human-written and LLM-generated texts to conduct attribution.

    “Transformer-based models, such as BERT and RoBERTa, are fine-tuned on datasets containing both human-written and LLM-generated texts to conduct the attribution.”
    Quote in the paperNot a substantial claim

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

    Interpretationnone given

    LLMs can be misused to generate content that conceals true authorship by mimicking others or using LLMs to alter their work.

    “Additionally, LLMs can be misused to generate content that conceals true authorship by mimicking others or using LLMs to alter their work.”
    Quote in the paperNot a substantial claim

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

    Interpretationreasoning

    This variability complicates authorship attribution, necessitating the identification of consistent and unique stylistic markers.

    “This variability complicates authorship attribution, necessitating the identification of consistent and unique stylistic markers.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    Variations in training data, model architecture, and fine-tuning processes contribute to the distinctive outputs of different LLMs, though these differences are often subtle and difficult to detect.

    “Variations in training data, model architecture, and fine-tuning processes contribute to the distinctive outputs of different LLMs, though these differences are often subtle and difficult to detect [Uchendu et al., 2021].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Uchendu et al., 2021 source

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

    Premisecitation

    LLM-generated text detectors can be categorized into metric-based and model-based methods, further divided into feature-based, neural network-based, zero-shot-based, and watermark-based methods.

    “LLM-generated text detectors can be categorized into metricbased and model-based methods [He et al., 2023], which are further divided into feature-based, neural network-based, zero-shot-based, and watermark-based methods.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites He et al., 2023

      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). May combine two claims.

    Interpretationnone given

    Watermarking-based methods typically fall under the white-box detection category, while proprietary models are restricted to black-box methods.

    “Watermarking-based methods typically fall under the white-box detection category, while proprietary models are restricted to black-box methods.”
    Quote in the paperNot a substantial claim

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

    Premisecitation

    MIXSET offers a more comprehensive dataset that includes text refined by LLMs through polishing, completion, and rewriting operations.

    “MIXSET [Zhang et al., 2024] offers a more comprehensive dataset that includes text refined by LLMs through polishing, completion, and rewriting operations.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Zhang et al., 2024 source

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

    Interpretationnone given

    LLM-generated text attribution builds upon techniques for LLM-generated text detection.

    “LLM-generated text attribution builds upon techniques for LLM-generated text detection.”
    Quote in the paperNot a substantial claim

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

    Premisecitation

    These detectors are classified as white-box or black-box, depending on their access to the LLM weights.

    “These detectors are also classified as white-box or black-box, depending on their access to the LLM weights [Tang et al., 2023b;Yang et al., 2023c].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites Tang et al., 2023b

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

    • cites Yang et al., 2023c

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

  • May combine two claims.

    Premisecitation

    Leveraging LLMs for data augmentation and synthetic data generation can create diverse training datasets, which in turn improves the generalization of attribution models.

    “Moreover, leveraging LLMs for data augmentation and synthetic data generation can create diverse training datasets, which in turn improves the generalization of attribution models [Albalak et al., 2024].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Albalak et al., 2024 source

  • May combine two claims.

    Premisecitation

    Factors influencing the performance of existing authorship attribution models include the size of the training text, the number of candidate authors, and the imbalanced distribution of training texts among candidate authors.

    “Factors influencing the performance of existing authorship attribution models include the size of the training text [Hirst and Feiguina, 2007;Marton et al., 2005], the number of candidate authors [Koppel et al., 2006], and the imbalanced distribution of training texts among candidate authors [Stamatatos, 2008].”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites Hirst and Feiguina, 2007 source
    • cites Marton et al., 2005 source
    • cites Koppel et al., 2006 source

  • May combine two claims.

    Premisecitation

    Some detectors like GPTZero can also detect human-LLM co-authored text and identify portions likely to be LLM-generated, addressing the boundary detection sub-problem of Problem 4.

    “Some detectors like GPTZero [Tian and Cui, 2023] can also detect human-LLM co-authored text and identify portions likely to be LLM-generated, addressing the boundary detection sub-problem of Problem 4 [Cutler et al., 2021].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites Tian and Cui, 2023

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

    • cites Cutler et al., 2021

      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).

    Interpretationreasoning

    Detecting subtle shifts in style and contextual cues that indicate the extent of LLM involvement allows for accurate segmentation and attribution of authorship within hybrid documents.

    “This involves detecting subtle shifts in style and contextual cues that indicate the extent of LLM involvement, allowing for accurate segmentation and attribution of authorship within hybrid documents.”
    Quote in the paperNot a substantial claim

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

    Premisecitation

    Authorship attribution is inherently multidisciplinary, encompassing elements of linguistics, computer science, forensic science, and psychology.

    “Authorship attribution is inherently multidisciplinary, encompassing elements of linguistics, computer science, forensic science, and psychology [Stamatatos, 2009].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Stamatatos, 2009 source

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

    Premisecitation

    Self-BLEU are useful metrics, with a lower score indicating higher textual diversity.

    “Additionally, Self-BLEU are useful metrics, with a lower score indicating higher textual diversity [Zhang et al., 2024].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Zhang et al., 2024 source

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

    Premisecitation

    Tweets are limited to 280 characters, whereas legal judgment documents contain thousands of words.

    “For example, tweets are limited to 280 characters, whereas legal judgment documents contain thousands of words [Seroussi et al., 2011].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Seroussi et al., 2011

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

  • May combine two claims.

    Interpretationreasoning

    Since Problems 2, 3, and 4 frequently treat human-written text as a single category for simplicity, rather than identifying individual authors as in Problem 1, many of these datasets are unsuitable for Problem 1, which requires unique author identification.

    “Since Problems 2, 3, and 4 frequently treat human-written text as a single category for simplicity, rather than identifying individual authors as in Problem 1, many of these datasets are unsuitable for Problem 1, which requires unique author identification.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    Combining interdisciplinary perspectives enhances our ability to distinguish between human- and machine-generated texts, addressing the emerging challenges posed by sophisticated language models.

    “Combining these interdisciplinary perspectives enhances our ability to distinguish between human-and machine-generated texts, addressing the emerging challenges posed by sophisticated language models.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    These detectors often lack robustness against variations in sampling strategies, adversarial attacks, and unseen domains and language models.

    “Additionally, these detectors often lack robustness against variations in sampling strategies, adversarial attacks, and unseen domains and language models [Dugan et al., 2024].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Dugan et al., 2024 source
67 of 67

  • The paper relies on this without support.

    The proprietary nature of many LLMs restricts access to comparative data, posing significant hurdles.

    “However, the proprietary nature of many LLMs restricts access to comparative data, posing significant hurdles.”

  • The paper relies on this without support.

    The high degree of stylistic overlap among LLMs, especially those with similar architectures or trained on overlapping datasets, further complicates accurate classification.

    “Additionally, the high degree of stylistic overlap among LLMs, especially those with similar architectures or trained on overlapping datasets, further complicates accurate classification.”

  • The paper relies on this without support.

    Hybrid texts may originate as human-written content, with LLMs employed for conditional generation, making it difficult to clearly distinguish between the stylistic features of human and machine contributions.

    “Hybrid texts may originate as human-written content, with LLMs employed for conditional generation, making it difficult to clearly distinguish between the stylistic features of human and machine contributions.”

  • The paper relies on this without support.

    There are fewer studies on human-LLM co-authored text attribution due to its difficulty, and existing research often makes simplifications.

    “There are fewer studies on this task due to its difficulty, and existing research often makes simplifications.”

  • The paper relies on this without support.

    The complexity of human-LLM co-authored text attribution increases as users may employ multiple LLMs to compose different sections of an article, further blurring the lines between human and machine-generated content.

    “The complexity of this task increases as users may employ multiple LLMs to compose different sections of an article, further blurring the lines between human and machine-generated content.”

  • The paper relies on this without support.

    Analyzing and classifying texts co-authored by humans and LLMs presents a significant challenge due to the blending of human and machine stylistic features.

    “Analyzing and classifying texts co-authored by humans and LLMs presents a significant challenge due to the blending of human and machine stylistic features.”

  • The paper relies on this without support.

    Human-LLM co-authored texts often start as human drafts and are later extended or revised by LLMs, or the process might occur in reverse.

    “These texts often start as human drafts and are later extended or revised by LLMs, or the process might occur in reverse.”

  • The paper relies on this without support.

    The integration of styles in human-LLM co-authored texts creates a hybrid form that makes it difficult to distinguish distinct authorial markers, thereby complicating the attribution process.

    “This integration of styles creates a hybrid form that makes it difficult to distinguish distinct authorial markers, thereby complicating the attribution process.”

  • The paper relies on this without support.

    Human-LLM co-authored texts pose a more intricate challenge due to the blending of stylistic and linguistic elements from both human authors and LLMs.

    “Human-LLM co-authored texts pose a more intricate challenge due to the blending of stylistic and linguistic elements from both human authors and LLMs.”

  • The paper relies on this without support.

    Manually collecting authorship data is time-consuming and costly, motivating researchers to utilize web data sources such as Wikipedia and Reddit.

    “Manually collecting data is time-consuming and costly, motivating researchers to utilize web data sources such as Wikipedia and Reddit.”

  • The paper relies on this without support.

    Custom datasets for LLM-generated text are easier and more affordable to create, and are often built alongside human-written text to maintain a similar domain and format.

    “In contrast, custom datasets for LLM-generated text are easier and more affordable to create, and are often built alongside human-written text to maintain a similar domain and format.”

  • The paper relies on this without support.

    LLM-generated and human-written texts are often disproportionate, making class imbalance a crucial issue.

    “Addressing class imbalance is crucial, as LLM-generated and human-written texts are often disproportionate.”

  • The paper relies on this without support.

    The rapid growth of online communication has led to shifts in writing behavior, resulting in shorter, fragmented, and less coherent social media tweets and text messages.

    “However, the rapid growth of online communication has also led to shifts in writing behavior, resulting in shorter, fragmented, and less coherent social media tweets and text messages.”

  • The paper relies on this without support.

    Despite their high accuracy claims, many detectors suffer from high false positive rates, which could falsely accuse individuals of plagiarism and undermine credibility and trust in genuine authors.

    “Despite their high accuracy claims, many detectors suffer from high false positive rates, which could falsely accuse individuals of plagiarism and undermine credibility and trust in genuine authors.”

  • The paper relies on this without support.

    Current authorship attribution methods face limitations when handling a more extensive range of candidate human authors or LLMs.

    “Current authorship attribution methods face limitations when handling a more extensive range of candidate human authors or LLMs, presenting opportunities for future research.”

  • The paper relies on this without support.

    Existing approaches for LLM-generated Text Attribution typically manage only a limited number of authors or models, which restricts their applicability in real-world scenarios where the pool of potential authors or models can be vast.

    “For instance, existing approaches for LLM-generated Text Attribution typically manage only a limited number of authors or models, which restricts their applicability in real-world scenarios where the pool of potential authors or models can be vast.”

  • The paper relies on this without support.

    Domain generalization poses significant challenges due to variability in vocabularies, syntax, and styles across different subjects, complicating accurate authorship attribution.

    “Domain generalization poses significant challenges due to variability in vocabularies, syntax, and styles across different subjects, complicating accurate authorship attribution.”

  • The paper relies on this without support.

    Authors' adaptation to various genres dilutes their identifiable stylistic traits, complicating attribution.

    “Authors' adaptation to various genres dilutes their identifiable stylistic traits, complicating attribution.”

  • The paper relies on this without support.

    Traditional attribution methods struggle to distinguish between human-authored texts and those generated by LLMs, which adeptly replicate these features.

    “Traditional attribution methods, which rely on stylistic and linguistic features to identify an author, struggle to distinguish between human-authored texts and those generated by LLMs, which adeptly replicate these features.”

  • The paper relies on this without support.

    The diversity in datasets and evaluation metrics currently hinders the comparability and generalizability of different authorship attribution methods.

    “The diversity in datasets and evaluation metrics currently hinders the comparability and generalizability of different authorship attribution methods.”

  • The paper relies on this without support.

    Authorship attribution techniques are essential in digital forensics, cybersecurity, and plagiarism detection.

    “Authorship attribution techniques are essential in digital forensics, cybersecurity, and plagiarism detection.”

  • The paper relies on this without support.

    The attribution of text to specific authors or models raises significant ethical and privacy issues.

    “The attribution of text to specific authors or models raises significant ethical and privacy issues.”

  • The paper relies on this without support.

    The potential to reveal the identities of anonymous authors presents significant ethical challenges.

    “However, the potential to reveal the identities of anonymous authors presents significant ethical challenges.”

  • The paper relies on this without support.

    Applications such as linking user accounts across platforms and identifying compromised accounts raise privacy concerns and ethical questions about surveillance and profiling individuals based on their writing style.

    “Applications such as linking user accounts across platforms and identifying compromised accounts raise privacy concerns and ethical questions about surveillance and profiling individuals based on their writing style.”

  • The paper relies on this without support.

    Despite existing measures to prohibit the unethical use of LLMs, these restrictions could be evaded through prompt engineering and jail-breaking, posing risks of phishing and fraud scams.

    “Despite existing measures to prohibit the unethical use of LLMs, these restrictions could be evaded through prompt engineering and jail-breaking, posing risks of phishing and fraud scams.”

  • The paper relies on this without support.

    The increasing difficulty in distinguishing between human and LLM-generated content raises concerns about intellectual property, plagiarism, and accountability.

    “Furthermore, the increasing difficulty in distinguishing between human and LLM-generated content raises concerns about intellectual property, plagiarism, and accountability.”
26 of 26

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.

  • Opposing claims about LLM vocabulary diversity

    Minor

    One section says LLM-generated text has greater vocabulary breadth than human writing, while another says humans have greater lexical diversity than LLMs. The text does not explain how these opposing comparisons fit together.

  • Corpus citation points to a different work

    Minor

    The entry cited for the Aston 100 Idiolects Corpus is titled as a paper about GPTZero detection, not the described corpus.

2 of 2

Cites

Flagged first, then by citations

Research works printed but not linked: 20

  1. [7]7. Shlomo Argamon. 2018. Computational forensic author- ship analysis: Promises and pitfalls. Language and Law/Linguagem e Direito 5, 2 (2018), 7-37.
  2. [16]16. Georgios Barlas and Efstathios Stamatatos. 2020b. Cross- domain authorship attribution using pre-trained language models. In Artificial Intelligence Applications and Innova- tions: 16th IFIP WG 12.5 International Conference, AIAI 2020, Neos Marmaras, Greece, June 5-7, 2020, Proceed- ings, Part I 16. Springer, 255-266.
  3. [54]54. Maël Fabien, Esau Villatoro-Tello, Petr Motlicek, and Shan- tipriya Parida. 2020. BertAA : BERT fine-tuning for Authorship Attribution. In Proceedings of the 17th In- ternational Conference on Natural Language Processing (ICON). NLP Association of India (NLPAI), Indian Insti- tute of Technology Patna, Patna, India, 127-137. https: //aclanthology.org/2020.icon-main.16
  4. [58]58. Rinaldo Gagiano, Maria Myung-Hee Kim, Xiuzhen Zhang, and Jennifer Biggs. 2021. Robustness Analysis of Grover for Machine-Generated News Detection. In Proceedings of the The 19th Annual Workshop of the Australasian Language Technology Association. Australasian Language Technology Association, Online, 119-127. https://aclanthology. org/2021.alta-1.12
  5. [90]90. Ganesh Jawahar, Muhammad Abdul-Mageed, and Laks Lakshmanan, V.S. 2020b. Automatic Detection of Ma- chine Generated Text: A Critical Survey. In Proceed- ings of the 28th International Conference on Computa- tional Linguistics. International Committee on Computa- tional Linguistics, Barcelona, Spain (Online), 2296-2309. https://doi.org/10.18653/v1/2020.coling-main.208 · DOI · Paper page
  6. [94]94. Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. 2020. Supervised Contrastive Learn- ing. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, Hugo Larochelle, Marc'Aurelio Ranzato, Raia Hadsell, Maria-Florina Balcan, and Hsuan-Tien Lin (Eds.). https://proceedings.neurips.cc/paper/2020/hash/ d89a66c7c80a29b1bdbab0f2a1a94af8-Abstract.html
  7. [99]99. Moshe Koppel, Jonathan Schler, and Eran Messeri. 2008. Au- thorship attribution in law enforcement scenarios. NATO Security Through Science Series D-Information and Com- munication Security 15 (2008), 111.
  8. [108]108. Thomas Lavergne, Tanguy Urvoy, and François Yvon. 2008. Detecting Fake Content with Relative Entropy Scoring. Pan 8, 27-31 (2008), 4.
  9. [123]123. David Madigan, Alexander Genkin, David D Lewis, Shlomo Argamon, Dmitriy Fradkin, and Li Ye. 2005b. Author identification on the large scale. In Proceedings of the 2005 Meeting of the Classification Society of North America (CSNA).
  10. [127]127. Rohith Menon and Yejin Choi. 2011. Domain Independent Authorship Attribution without Domain Adaptation. In Proceedings of the International Conference Recent Ad- vances in Natural Language Processing 2011. Association for Computational Linguistics, Hissar, Bulgaria, 309-315. https://aclanthology.org/R11-1043
  11. [145]145. Jiameng Pu, Zain Sarwar, Sifat Muhammad Abdullah, Ab- dullah Rehman, Yoonjin Kim, Parantapa Bhattacharya, Mobin Javed, and Bimal Viswanath. 2023a. Deepfake text detection: Limitations and opportunities. In 2023 IEEE Symposium on Security and Privacy (SP). IEEE, 1613-1630.
  12. [150]150. Vinay Venkatesh Ramasesh, Aitor Lewkowycz, and Ethan Dyer. 2022. Effect of scale on catastrophic forgetting in neural networks. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net. https://openreview. net/forum?id=GhVS8_yPeEa
  13. [155]155. Juan Rodriguez, Todd Hay, David Gros, Zain Shamsi, and Ravi Srinivasan. 2022b. Cross-Domain Detection of GPT-2-Generated Technical Text. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. Association for Computational Linguistics, Seattle, United States, 1213-1233. https: //doi.org/10.18653/v1/2022.naacl-main.88 · DOI · Paper page
  14. [163]163. Jonathan Schler, Moshe Koppel, Shlomo Argamon, and James W Pennebaker. 2006. Effects of age and gender on blogging.. In AAAI spring symposium: Computational approaches to analyzing weblogs, Vol. 6. 199-205.
  15. [168]168. Seroussi, Russell Smyth, and Ingrid Zukerman. 2011. Ghosts from the high court's past: Evidence from compu- tational linguistics for Dixon ghosting for Mctiernan and rich. University of New South Wales Law Journal, The 34, 3 (2011), 984-1005.
  16. [174]174. Prasha Shrestha, Sebastian Sierra, Fabio González, Manuel Montes, Paolo Rosso, and Thamar Solorio. 2017b. Con- volutional Neural Networks for Authorship Attribution of Short Texts. In Proceedings of the 15th Conference of the European Chapter of the Association for Computa- tional Linguistics: Volume 2, Short Papers. Association for Computational Linguistics, Valencia, Spain, 669-674. https://aclanthology.org/E17-2106
  17. [190]190. Jinyan Su, Terry Yue Zhuo, Di Wang, and Preslav Nakov. 2023c. Detectllm: Leveraging log rank information for zero- shot detection of machine-generated text. ArXiv preprint abs/2306.05540 (2023). https://arxiv.org/abs/2306. 05540 · arXiv
  18. [192]192. Kalaivani Sundararajan and Damon Woodard. 2018. What represents "style" in authorship attribution?. In Proceed- ings of the 27th International Conference on Computa- tional Linguistics. Association for Computational Linguis- tics, Santa Fe, New Mexico, USA, 2814-2822. https: //aclanthology.org/C18-1238
  19. [194]194. Ruixiang Tang, Yu-Neng Chuang, and Xia Hu. 2023b. The science of detecting llm-generated texts. ArXiv preprint abs/2303.07205 (2023). https://arxiv.org/abs/2303. 07205 · arXiv
  20. [216]216. Xianjun Yang, Wei Cheng, Yue Wu, Linda Petzold, William Yang Wang, and Haifeng Chen. 2023b. Dna-gpt: Divergent n-gram analysis for training-free detection of gpt-generated text. ArXiv preprint abs/2305.17359 (2023). https://arxiv.org/abs/2305.17359 · arXiv

Cited by

Sorted by citations

Related

Authors
Huang, B., Chen, C., & Shu, K.
Affiliations
  • Illinois Institute of Technology (US)
  • Emory University (US)
ORCID
Venue
ACM SIGKDD Explorations Newsletter
Publisher
Association for Computing Machinery (ACM)
Published
2025-01-21
Volume
26
Issue
2
Pages
21-43
ISSN
  • 1931-0145 (ISSN-L)
  • 1931-0153
DOI
10.1145/3715073.3715076
PMC
PMC12019761
OpenAlex
W4406939734
Cited by
42
References
209
Citations by year
2026: 28 · 2025: 14
FWCI
16.71
Type
journal article
Language
en
Open access
Licence
www.acm.org/publications/policies/copyright_policy#Background (vor)
Locations
  • ACM SIGKDD Explorations Newsletter (publishedVersion)
  • PubMed (publishedVersion)
  • PubMed Central (acceptedVersion)
  • Europe PMC (PubMed Central) (acceptedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • arXiv (Cornell University)
Funders
  • National Science Foundation (2346158, 2241068, 2339198)
  • U.S. Department of Homeland Security (17STQAC00001-07-04)
  • Cisco Systems
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Authorship Attribution and Profiling
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Steganography and Watermarking Techniques
  • Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
Keywords
authorship attribution, large language models, AI-generated text detection, misinformation, plagiarism detection, forensic investigation, explainability, domain generalization

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 2026-10-11.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-11.

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