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

In the paper

“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].”

What the sources say

• ?Not matched

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

  cites Gagiano et al., 2021
• Supported

  Macko et al. (2024) is cited for the homoglyph-attack portion of the sentence. The Text edits section describes its homoglyph attacks, and Section 4.1 reports that these attacks often evade machine-generated-text detection across the tested languages. The paper also notes that preprocessing can mitigate them. (LLM, conf=0.94)

  cites Macko et al., 2024 source
• Supported

  The Introduction reports that adding a single space before a comma substantially reduces AI-text detection rates. Section 3, “Space Infiltration,” describes this whitespace-insertion attack as a way to bypass detectors. The other attack types in the citing sentence are attributed to other sources, not this paper. (LLM, conf=0.94)

  cites Cai and Cui, 2023 source

In the paper

“Techniques like slanted triangular learning rates and gradual unfreezing can be used to avoid catastrophic forgetting during fine-tuning [Howard and Ruder, 2018].”

What the sources say

• Supported

  The Introduction explicitly identifies slanted triangular learning rates and gradual unfreezing as techniques to retain previous knowledge and avoid catastrophic forgetting during fine-tuning. The Conclusion also says the proposed techniques, used together, prevent catastrophic forgetting. (LLM, conf=0.95)

  cites Howard and Ruder, 2018 source

In the paper

“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].”

What the sources say

• Supported

  Section 5 (p. 6) describes detecting machine-generated news by classifying articles as Human or Machine. Section 5.1 (p. 6) explicitly requires the discriminator to assign one of those two labels to each article in its unpaired evaluation, supporting the binary-classification characterization. (LLM, conf=0.88)

  cites Zellers et al., 2019 source
• Supported

  The “Automated ML-based detection” section describes classifiers that distinguish language-model outputs from a true-text distribution. “Our Work” describes a detector trained on samples equally likely to be real or fake, supporting the characterization of detection as a binary classification task. (LLM, conf=0.89)

  cites Solaiman et al., 2019 source
• Supported

  The Introduction describes the commonly used approach as a classification task distinguishing machine-generated from human-written text. Those are two classes, matching the claim that detection is typically approached as binary classification. No page number is provided in the supplied passage. (LLM, conf=0.93)

  cites Jawahar et al., 2020b source

In the paper

“Rivera-Soto et al. [2021] concluded that topic diversity and dataset size are crucial for effective cross-domain transfer.”

What the sources say

• Partly supported

  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)

  cites Rivera-Soto et al., 2021 sourceExperimental — accuracy not yet measured

In the paper

“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].”

What the sources say

• Partly supported

  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)

  cites Rivera-Soto et al., 2021 sourceExperimental — accuracy not yet measured
• Partly supported

  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)

  cites Huertas-Tato et al., 2022 sourceExperimental — accuracy not yet measured

In the paper

“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].”

What the sources say

• Supported

  The Introduction reports that LLMs can perform authorship attribution and use subtle stylistic and linguistic patterns to distinguish authors. The Conclusion describes applying end-to-end methods directly to authorship attribution while bypassing intermediate feature extraction. (LLM, conf=0.93)

  cites Huang et al., 2024 source

In the paper

“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].”

What the sources say

• Partly supported

  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)

  cites Kojima et al., 2022 sourceExperimental — accuracy not yet measured

In the paper

“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].”

What the sources say

• ?Not checked — no full text of the sourcecites Khosla et al., 2020 source
• Partly supported

  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)

  cites Huertas-Tato et al., 2022 sourceExperimental — accuracy not yet measured

In the paper

“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].”

What the sources say

• ?Not matched

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

  cites Cutler et al., 2021
• Supported

  Section 3.3, “Boundary Identification,” defines a task in which a human begins a text, an LLM continues it, and a detector identifies the boundary. The Introduction also lists human-written-to-machine-generated change-point detection as Task 3. This supports the cited description of a variation involving boundary detection. (LLM, conf=0.95)

  cites Wang et al., 2024 source

In the paper

“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].”

What the sources say

• ?Not checked — no full text of the sourcecites Guo et al., 2023 source
• Partly supported

  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)

  cites Dugan et al., 2023 sourceExperimental — accuracy not yet measured
• ?Not checked — no full text of the sourcecites Zhang et al., 2024 source

In the paper

“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].”

What the sources say

• Supported

  Section 2, “Automatic versus human evaluation,” explicitly says that human evaluation remains the gold standard for assessing overall NLG system quality. The Introduction describes human evaluation as crucial to assessing system quality, supporting the claim’s broader account of its role. The paper makes this point about NLG generally rather than singling out open-domain systems. (LLM, conf=0.91)

  cites van der Lee et al., 2019 source
• ?Could not verify

  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)

  cites Gehrmann et al., 2019a sourceExperimental — accuracy not yet measured

In the paper

“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].”

What the sources say

• ?Not matched

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

  cites Yang et al., 2023c
• Partly supported

  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)

  cites Crothers et al., 2023 sourceExperimental — accuracy not yet measured

In the paper

“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].”

What the sources say

• ?Not matched

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

  cites He et al., 2023
• Supported

  Section 5.3, “Multi-way Detection,” explicitly describes Task 2 as seven-label classification and lists one human class and six LLM classes. This matches the classification attributed to Wang et al. (2024). (LLM, conf=0.95)

  cites Wang et al., 2024 source

In the paper

“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.”

What the sources say

• ?Not matched

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

  cites Yang et al., 2023a
• Supported

  Section 5.2 states that DetectGPT's authors simulated human revision of machine-generated text by replacing word spans with T5-3B samples. This supports the claim as it pertains to DetectGPT; the separate citation to DNA-GPT is not assessed here. (LLM, conf=0.95)

  cites Mitchell et al., 2023 source
• ?Not matched

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

  cites Raffel et al., 2020

In the paper

“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].”

What the sources say

• Supported

  The Experiments metrics passage, following §3.3, explicitly says the paper assesses performance with recall-at-8 and mean reciprocal rank (MRR). The Introduction describes its setting as open-world, with several orders of magnitude more authors than a 21-author study, supporting the large-candidate context. (LLM, conf=0.94)

  cites Rivera-Soto et al., 2021 source

In the paper

“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.”

What the sources say

• Partly supported

  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)

  cites Brown et al., 2020 sourceExperimental — accuracy not yet measured

In the paper

“The challenge remains to provide higher-level explainability that aligns with human cognitive processes [Rudin, 2019].”

What the sources say

• Supported

  In ‘Challenge #3’ (page unavailable), Rudin frames interpretability around how people explain their reasoning in visual classification and calls for models that reason about image parts so their explanations reflect the decision process. The Introduction also relates interpretable model design to limits on human cognitive capacity. This supports the claim’s broad point, though Rudin advocates inherently interpretable models rather than post-hoc explanations of black boxes. (LLM, conf=0.85)

  cites Rudin, 2019 source

In the paper

“Advanced techniques are required to dissect these integrations, identifying where and how the contributions of LLM intersect with human input [Wang et al., 2024].”

What the sources say

• Supported

  Section 3.3, “Boundary Identification,” describes detecting where human-written text changes to LLM-generated text. The Limitations section says real mixed texts can be more complicated and calls for methods that detect mixed text and identify all changing points, supporting the claim’s need for more capable techniques. (LLM, conf=0.88)

  cites Wang et al., 2024 source

In the paper

“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.”

What the sources say

• ?Not matched

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

  cites Barlas and Stamatatos, 2020b
• Partly supported

  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)

  cites Huertas-Tato et al., 2022 sourceExperimental — accuracy not yet measured

In the paper

“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].”

What the sources say

• ?Not checked — no full text of the sourcecites Goldstein-Stewart et al., 2009 source
• ?Not matched

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

  cites Pennebaker et al., 2015
• Partly supported

  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)

  cites Dudău and Sava, 2021 sourceExperimental — accuracy not yet measured

In the paper

“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].”

What the sources say

• ?Not checked — no full text of the sourcecites Aborisade and Anwar, 2018 source
• Supported

  The Abstract states that attributing online texts such as blogs and e-mails is more challenging than attributing traditional texts because they tend to be short and have more candidate authors. The Related Work section likewise identifies many candidate authors and short e-mails as challenges. (LLM, conf=0.94)

  cites Seroussi et al., 2014 source
• ?Not checked — no full text of the sourcecites Theophilo et al., 2021 source

In the paper

“Misattribution can lead to wrongful accusations or misinterpretation of an author's intent [Lund et al., 2023].”

What the sources say

• ?Could not verify

  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)

  cites Lund et al., 2023 sourceExperimental — accuracy not yet measured

In the paper

“For instance, Kowalczyk et al. [2022] detected GPT-2-generated fake reviews using Shapley Additive Explanations (SHAP) [Lundberg and Lee, 2017].”

What the sources say

• ?Not checked — no full text of the sourcecites Kowalczyk et al., 2022 source
• Supported

  The Abstract (p. 1) identifies SHAP as “SHapley Additive exPlanations” and describes it as a framework for interpreting model predictions. The citation to Lundberg and Lee identifies the method named in the sentence; the fake-review study is attributed separately to Kowalczyk et al. (LLM, conf=0.94)

  cites Lundberg and Lee, 2017 source

In the paper

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

In the paper

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

In the paper

“However, fewer support LLM-generated text attribution (Problem 3) and Human-LLM Co-authored Text Attribution (Problem 4).”

In the paper

“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].”

What the sources say

• ?Not matched

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

  cites He et al., 2023
• ?Not checked — no full text of the sourcecites Zhang et al., 2024 source

In the paper

“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.”

What the sources say

• ?Not matched

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

  cites Chakraborty et al., 2023a

In the paper

“LLM-generated texts are typically less emotional and more objective than human-written texts [Guo et al., 2023].”

What the sources say

• ?Not checked — no full text of the sourcecites Guo et al., 2023 source

In the paper

“whereas deep learning methods demand significant computational resources and extensive labeled data, with the risk of catastrophic forgetting [Ramasesh et al., 2022].”

What the sources say

• ?Not matched

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

  cites Ramasesh et al., 2022

In the paper

“However, these neural network approaches often offer less explainability compared to their metric-based counterparts.”

In the paper

“The rapid evolution of LLMs further complicates detection, as newer versions exhibit different stylistic characteristics, making detection models quickly obsolete [Chakraborty et al., 2023b].”

What the sources say

• ?Not matched

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

  cites Chakraborty et al., 2023b

In the paper

“Barlas and Stamatatos [2020a] found that BERT performed well with large vocabularies, outperforming multi-headed RNNs.”

What the sources say

• ?Not matched

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

  cites Barlas and Stamatatos, 2020a

In the paper

“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].”

What the sources say

• ?Not matched

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

  cites Barlas and Stamatatos, 2020a

In the paper

“Collaborative texts, which blend multiple writing styles, further mask individual contributions and obscure distinct authorial signals [Dauber et al., 2017].”

What the sources say

• ?Not checked — no full text of the sourcecites Dauber et al., 2017 source

In the paper

“Traditionally, LLMs have been employed mainly for auxiliary tasks such as feature extraction and data annotation [Patel et al., 2023].”

What the sources say

• ?Not checked — no full text of the sourcecites Patel et al., 2023 source

In the paper

“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.”

In the paper

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

In the paper

“LLMs are also effective with shorter texts, reducing the necessity for long inputs to derive meaningful representations [Eder, 2015].”

What the sources say

• ?Not checked — no full text of the sourcecites Eder, 2015 source

In the paper

“Zero-shot detection methods are generally statistics-based, enabling the detection of LLM-generated text without additional training [Su et al., 2023b].”

What the sources say

• ?Not matched

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

  cites Su et al., 2023b

In the paper

“This attribution task extends beyond binary classification to handle multiple classes, increasing the complexity of LLMgenerated text detection.”

In the paper

“In contrast, LLMgenerated texts typically exhibit consistent syntax and a broader vocabulary [Guo et al., 2023].”

What the sources say

• ?Not checked — no full text of the sourcecites Guo et al., 2023 source

In the paper

“Additionally, LLMs may inadvertently reproduce snippets of their training data.”

In the paper

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

In the paper

“For instance, a 10-class classification might include seven human classes and three LLM classes [Tripto et al., 2023].”

What the sources say

• ?Not checked — no full text of the sourcecites Tripto et al., 2023 source

In the paper

“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].”

What the sources say

• ?Not matched

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

  cites Liu et al., 2023b

In the paper

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

In the paper

“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].”

What the sources say

• ?Not checked — no full text of the sourcecites Kumarage and Liu, 2023 source
• ?Not matched

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

  cites Tang et al., 2023a
• ?Not matched

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

  cites Wu et al., 2023b

In the paper

“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.”

In the paper

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

In the paper

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

In the paper

“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].”

What the sources say

• ?Not checked — no full text of the sourcecites Uchendu et al., 2021 source

In the paper

“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.”

What the sources say

• ?Not matched

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

  cites He et al., 2023

In the paper

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

In the paper

“MIXSET [Zhang et al., 2024] offers a more comprehensive dataset that includes text refined by LLMs through polishing, completion, and rewriting operations.”

What the sources say

• ?Not checked — no full text of the sourcecites Zhang et al., 2024 source

In the paper

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

In the paper

“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].”

What the sources say

• ?Not matched

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

  cites Tang et al., 2023b
• ?Not matched

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

  cites Yang et al., 2023c

In the paper

“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].”

What the sources say

• ?Not checked — no full text of the sourcecites Albalak et al., 2024 source

In the paper

“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].”

What the sources say

• ?Not checked — no full text of the sourcecites Hirst and Feiguina, 2007 source
• ?Not checked — no full text of the sourcecites Marton et al., 2005 source
• ?Not checked — no full text of the sourcecites Koppel et al., 2006 source

In the paper

“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].”

What the sources say

• ?Not matched

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

  cites Tian and Cui, 2023
• ?Not matched

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

  cites Cutler et al., 2021

In the paper

“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.”

In the paper

“Authorship attribution is inherently multidisciplinary, encompassing elements of linguistics, computer science, forensic science, and psychology [Stamatatos, 2009].”

What the sources say

• ?Not checked — no full text of the sourcecites Stamatatos, 2009 source

In the paper

“Additionally, Self-BLEU are useful metrics, with a lower score indicating higher textual diversity [Zhang et al., 2024].”

What the sources say

• ?Not checked — no full text of the sourcecites Zhang et al., 2024 source

In the paper

“For example, tweets are limited to 280 characters, whereas legal judgment documents contain thousands of words [Seroussi et al., 2011].”

What the sources say

• ?Not matched

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

  cites Seroussi et al., 2011

In the paper

“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.”

In the paper

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

In the paper

“Additionally, these detectors often lack robustness against variations in sampling strategies, adversarial attacks, and unseen domains and language models [Dugan et al., 2024].”

What the sources say

• ?Not checked — no full text of the sourcecites Dugan et al., 2024 source