Trends and Challenges in Authorship Analysis: A Review of ML, DL, and LLM Approaches

Elisa Barney; Marcus Liwicki; Nudrat Habib; Tosin Adewumi

REVIEW 3 major objections 4 minor 54 references

Trends and Challenges in Authorship Analysis: A Review of ML, DL, and LLM Approaches

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This systematic review of 93 studies from 2015 to 2024 maps authorship attribution and verification, and identifies low-resource languages, multilingual adaptation, cross-domain generalization, and AI-generated text detection as the…

desk verdict A useful but over-labeled survey of AA/AV methods through 2024; the 'comprehensive systematic' claim outruns the first-10-pages search method, and one internal timeline inconsistency needs fixing, but it is a fair starting map for newcomers. read the letter →

arxiv 2505.15422 v1 pith:ZXCMEIDT submitted 2025-05-21 cs.CL

Nudrat Habib , Tosin Adewumi , Marcus Liwicki , Elisa Barney This is my paper · ORCID

classification cs.CL

keywords authorshipattributionverificationsystematicliteraturereviewstylometrylargelanguagemodelsmachine-generatedtextdetectionlow-resourcelanguagesmultilingualanalysis

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper is a systematic literature review of authorship analysis, focused on two subtasks: author attribution (deciding who wrote a text) and author verification (deciding whether two texts share an author). It synthesizes 93 studies from 2015 to 2024 that use traditional machine learning, deep learning, or large language models, and organizes them by sub-task, technique, feature extraction, dataset, and language. The review's central claim is that the field has moved from statistical and feature-based methods through deep learning to an early but growing use of LLMs, while remaining heavily English-centric. It identifies low-resource language processing, multilingual adaptation, cross-domain generalization, explainability, and human-versus-machine text detection as the principal gaps. A sympathetic reader would take this as a current map of the field and a checklist for future research.

What carries the argument

The carrying mechanism is the review's classification scheme: 93 articles grouped by sub-task (author attribution, author verification, or both) and by technique family (traditional ML, deep learning, LLMs, and unsupervised methods), with each article summarized by model, dataset, language, and feature extraction approach. Supporting this are the documented search queries, the inclusion criteria (peer-reviewed, English-language, textual data, 2015-2024, first ten pages of results), and a series of summary tables that turn individual study reports into cross-study patterns. The scheme is what lets the authors derive their trend claims, such as the dominance of English and the recency of LLM use.

What would settle it

A reader can reproduce the documented search protocol but go beyond the first ten pages; if additional qualifying studies on LLM-based attribution, non-English verification, or cross-domain transfer appear there, the review's gap analysis would shrink or shift. A narrower probe is to count qualifying verification studies in languages other than English: the paper reports one, and a single missed study would double the total.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is a structured account of where authorship analysis stands between 2015 and 2024. From 93 included articles, 59 address attribution only, 22 verification only, and 12 both; support vector machines and n-gram features still compete with CNNs, BERT-family models, and a small but rising number of LLM-based studies. The review reports that LLM use appears only from 2021 onward and jumps to 7 of 29 articles in 2024, yet most such work uses pre-trained models in zero-shot settings without fine-tuning. It also reports that verification research is almost entirely English, with one Urdu study, while attribution has scattered efforts in Arabic, Bangla, Urdu, Romanian, and a handful of other languages. The authors conclude that the critical gaps are multilingual adaptation, cross-domain generalization, explainability, and detection of machine-generated text.

Load-bearing premise

The load-bearing assumption is that the first ten pages of search results, screened manually by one author, capture enough of the field that the absence of work on a language or technique can be read as a real gap.

Editorial extensions

If this is right

If the review's map is right, new authorship-analysis research can be positioned against a known landscape: SVM and n-gram baselines remain competitive for long texts, while BERT variants dominate deep-learning work and LLMs are only beginning to be probed.
A researcher entering verification would see that almost all work is English-only; claims about other languages rest on a single Urdu study, so low-resource verification is essentially open territory.
The trend data imply that LLM-based attribution and verification have grown from 1 of 14 articles in 2021 to 7 of 29 in 2024, and that zero-shot prompting without fine-tuning is the common but underexplored mode.
The review's finding that performance degrades as the number of authors grows points to open-set and incremental learning as concrete directions, with two unsupervised clustering studies reporting an advantage at scale.
The treatment of machine-generated text as an author class in attribution and as a human-versus-machine verification task is presented as the newest and least settled front.

Reading between the lines

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

Taking the paper at face value, the deepest omission may be in its own corpus: because only the first ten pages of each search were kept, very recent LLM-based work is likely undercounted, which would only strengthen the review's characterization of LLM authorship analysis as early-stage.
Perplexity-based human-versus-machine verification, which the review highlights, is plausibly fragile against paraphrasing or style-mimicking attacks; a natural test would be to rerun the GPT-2 perplexity classifiers on paraphrased machine text.
A concrete extension of the gap analysis is to build a multilingual verification benchmark with matched cross-domain pairs, since no such resource appears among the 93 studies.
The review's pairing of explainability techniques with forensic applications suggests that courts, not just accuracy benchmarks, are the implied end-use; that motivates annotation of decisions in legal terms.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

You should know two things about this one. First, it is a genuinely useful survey for someone entering authorship analysis: it covers 93 papers from 2015 to 2024, including the recent LLM-based work and human-vs-machine verification, and it organizes the literature into readable tables by task, model, dataset, and language. Second, the paper calls itself a 'comprehensive systematic review,' and that label is not really supported by the method described. The search stops at the first 10 pages of Google Scholar and Scopus, selection rests on one author's manual screening, and there is no PRISMA-style flow or inter-rater check. That does not destroy the value of the survey, but it should not be sold as exhaustive.

The novelty is real but modest. It extends Tyo et al. (2022) and He et al. (2024) by covering LLM approaches and AI-text detection, which those earlier surveys missed. The gap analysis—low-resource languages, multilingual adaptation, cross-domain generalization, explainability—is sensible and clearly grounded in the reviewed papers. The authors also deserve credit for being transparent about their inclusion criteria and for printing their search strings in Appendix B.

On the soft spots: the reader's report flags a technique-count discrepancy (97 vs 93), but the paper already addresses that in Appendix C—the total is 97 because four articles use both LLM and ML/DL techniques and are counted twice. So scratch that one. The real inconsistency is Section 3.1.2, which says 'Research involving LLMs for AA started in 2024,' while Table 5 lists Barlas and Stamatatos (2021) as an LLM paper using GPT-2, and Section 4.4 correctly says 2021 was the first LLM year. That is a plain internal contradiction and should be fixed. A related conceptual issue is that the paper lumps BERT, ELMo, and GPT-2 together as 'LLMs.' BERT is an encoder, not a generative LLM, and ELMo is a contextual embedding model. The grouping is too loose and could mislead newcomers.

Who is this for? A master's student or a researcher new to the area who wants a quick map of methods, datasets, and open problems. It is not a rigorous meta-analysis and breaks no new empirical ground. The flaws are correctable and do not sink the core content.

My recommendation: send it to peer review, but ask the authors to soften the 'comprehensive systematic' claim, fix the 2021/2024 LLM timeline contradiction, and tighten the 'LLM' terminology. With those changes it is a solid reference survey.

Referee Report

3 major / 4 minor

Summary. This paper presents a systematic literature review of author attribution (AA) and author verification (AV) covering ML, DL, and LLM methods from 2015 to 2024. The authors report screening 93 articles from Google Scholar and Scopus, summarizing models, feature extraction techniques, datasets, languages, and sub-tasks, and identifying research gaps in low-resource languages, multilingual adaptation, cross-domain generalization, explainability, and AI-generated text detection. The review aims to serve as a current reference map and to guide future research.

Significance. If its coverage claims held, this review would be a valuable reference map for the AA/AV community: it documents a reproducible search string protocol (Appendix B), extracts method/dataset/language details in tabular form across 93 studies, and covers the recent influx of LLM-based work including human-versus-machine verification. The paper is weaker as a systematic review than its abstract claims, however, and several quantitative statements about the evolution of the field are internally inconsistent. The gap analysis is useful as a starting point, but it should be treated with caution until the corpus and counts are reconciled.

major comments (3)

[Section 3.1.2 / Table 5 / Section 4.4] The claim that "Research involving LLMs for AA started in 2024" (Section 3.1.2) is directly contradicted by Table 5, which lists Barlas and Stamatatos (2021) using BERT, GPT-2, ELMo, and ULMFiT for AA, and by Section 4.4, which states that 2021 was the first year of LLM application (1 of 14 articles). Specify whether BERT/ELMo/GPT-2 count as LLMs in this review, and correct the text or the table so that the timeline is consistent.
[Appendix C, Table C3 / Section 4.4] The technique counts in Table C3 sum to 97 for 93 articles, and the appendix explains this as overlap of 4 articles counted in both ML/DL and LLM. However, the per-row sums imply more overlaps: 2021 has 14 ML/DL + 1 LLM for 13 articles, 2023 has 15+2 for 15 articles, and 2024 has 22+7 for 28 articles, giving 5 row-level overlaps. Moreover, Section 4.4 uses technique counts ("1 out of 14 articles", "2 out of 17 articles") instead of distinct article counts. Please reconcile the table and base the evolution analysis on consistent article counts.
[Section 2, inclusion criterion 5] The search protocol limits inclusion to the first 10 pages of Google Scholar and Scopus results and relies on a single author's manual screening (steps 8-11) without a PRISMA-style flow diagram or inter-rater reliability check. As a result, the "comprehensive systematic" claim in the Abstract and Section 1.2 is stronger than the described methodology supports. Please add a limitations paragraph, report the number of records screened/excluded at each stage, and either extend the search or soften the "comprehensive" claim to a scoping review.

minor comments (4)

[Table 1] The column headings "Venue / Articles / Venue count" are ambiguous; please define "Venue count" and explain the relationship between the 93 article total and the 61 venue-count total.
[Section 4, opening paragraph] "the the landscape" is a typo; also "At the end of the this" in Section 2, step 11 should be corrected.
[Table 6, Halvani et al. row] The entry "12 approches from PAN competition" contains a typo ("approches" should be "approaches").
[Section 4.4] The statement that no articles from 2015 and 2016 made it into the review is not explained; one sentence on exclusion reasons would improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is a literature review whose claims are summaries of a surveyed corpus; the only self-citation appears as a reviewed item and is not load-bearing.

full rationale

This manuscript is a systematic literature review of author attribution and author verification. It does not present a derivation, a fitted model, or a prediction that reduces to its inputs. The central claims are descriptive: the review synthesizes 93 selected studies, groups them by sub-task and technique, and identifies research gaps. No equation is derived from an assumption that already contains the conclusion, and no fitted parameter is renamed as a prediction. The methodology in Section 2 is a manual search and screening protocol, with transparent inclusion and exclusion criteria, including the page-10 cutoff; that cutoff is a limitation on coverage and representativeness, not a circular step. The only self-citation, Adewumi et al. (2024), is one of the 93 surveyed articles (Table 5, LLM-based AA research), and it is reported as a study finding limitations of LLMs for attribution. It is not invoked as an authority to justify the review's framework, its gap analysis, or any methodological choice, so it is not load-bearing. The internal inconsistency noted in Appendix C, where technique counts exceed the number of articles due to overlap, is an editing or arithmetic issue that affects the quantitative evolution analysis (Figure 3 and Table C3) but does not constitute circular reasoning. Likewise, the statement in Section 3.1.2 that 'Research involving LLMs for AA started in 2024' is contradicted elsewhere by the inclusion of Barlas and Stamatatos (2021) and by Section 4.4's statement that 2021 was the first LLM year; this is an internal consistency error, not circularity. In sum, the review's claims are self-contained summaries of its corpus, and any weakness lies in corpus coverage and internal consistency, which are outside the definition of circularity used in this pass.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

The review introduces no free parameters, fitting, or new entities. Its claims rest on search and screening assumptions: first-10-pages cutoff, manual single-author screening, and the representativeness of the 93 selected papers.

assumptions (3)

domain assumption The first 10 pages of search results from Google Scholar and Scopus contain the most relevant literature for AA and AV
Stated in inclusion/exclusion criterion 5 (Section 2). The review assumes that relevance drops after page 10, which bounds the search but may exclude relevant work.
domain assumption Manual screening of titles and abstracts by a single author is sufficient to identify all relevant articles
Section 2 step 4 describes skimming titles and abstracts; no inter-rater reliability or second screening pass is reported, so a subjective selection is baked into the corpus.
domain assumption The 93 selected articles are representative of the field's methods and gaps between 2015 and 2024
The gap analysis in Sections 4 and 5 is inferred from the selected corpus only; if the corpus is not representative, the identified gaps could change.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Trends and Challenges in Authorship Analysis: A Review of ML, DL, and LLM Approaches." pith.science (2026). https://pith.science/paper/ZXCMEIDT

@misc{pith2026250515422,
  author       = {Pith},
  title        = {Pith review of: Trends and Challenges in Authorship Analysis: A Review of ML, DL, and LLM Approaches},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZXCMEIDT}},
  note         = {Machine review of arXiv:2505.15422}
}

read the original abstract

Authorship analysis plays an important role in diverse domains, including forensic linguistics, academia, cybersecurity, and digital content authentication. This paper presents a systematic literature review on two key sub-tasks of authorship analysis; Author Attribution and Author Verification. The review explores SOTA methodologies, ranging from traditional ML approaches to DL models and LLMs, highlighting their evolution, strengths, and limitations, based on studies conducted from 2015 to 2024. Key contributions include a comprehensive analysis of methods, techniques, their corresponding feature extraction techniques, datasets used, and emerging challenges in authorship analysis. The study highlights critical research gaps, particularly in low-resource language processing, multilingual adaptation, cross-domain generalization, and AI-generated text detection. This review aims to help researchers by giving an overview of the latest trends and challenges in authorship analysis. It also points out possible areas for future study. The goal is to support the development of better, more reliable, and accurate authorship analysis system in diverse textual domain.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

54 extracted references · 47 canonical work pages

[1]

Table 5 shows the summary of the research using LLMs

s based Research: Leveraging LLMs for AA is relatively new with only 3 articles exploring LLMs for Attribution, highlighting the research gap. Table 5 shows the summary of the research using LLMs. AA using the GPT-2 model and perplexity score was done by W. Huang et al. (2024) and they argued about the 12 superiority of perplexity over stylometry for the ...

work page 2024
[2]

There are significantly fewer AV-based research articles compared to AA

Verify if the text is written by a human or machine (Guo et al., 2024). There are significantly fewer AV-based research articles compared to AA. The methods and techniques employed for AV span from traditional ML models to more advanced encoder-decoder architectures and transformer-based models. With the rise of LLMs, their potential for this task has sta...

work page 2024
[3]

Studies that focused solely on traditional ML techniques include Castillo et al

Traditional ML and DL: Research in both traditional ML and DL often involves the use of multiple models simultaneously, reporting the best-performing model for specific settings. Studies that focused solely on traditional ML techniques include Castillo et al. (2019) using 2 variants of SVM classifier, namely C-SVC and Nu-SVC and Weeras- inghe et al. (2021...

work page 2019
[4]

It includes the top 50 authors based on the total size of their articles, all of whom have written at least one article in the CCAT (corporate/industrial) category

Reuters-50: This dataset is a subset of the Reuters Corpus Volume 1 (RCV1) (Lewis et al., 2004), which has already been used in author identification studies. It includes the top 50 authors based on the total size of their articles, all of whom have written at least one article in the CCAT (corporate/industrial) category. The training set contains 2,500 t...

work page 2004
[5]

Large Language Model (LLM): The use of LLM for AV is relatively new, with only 4 out of 22 articles leveraging LLMs for this task. In X. Liu and Kong (2024), the authors used a strided sliding 14 Table 6: A summary of research on AV # Article Models Used Dataset Language 1 W. Huang and Grieve (2024) GPT2, SVM PAN 2024 English 2 Guo et al. (2024) BERT, BiL...

work page 2024
[6]

For instance, Khan et al

Some researchers did experiments using both traditional ML and DL techniques and reported the results of which technique pro- duced the best results for their data or experimental setting. For instance, Khan et al. (2023) employed RF, SVM and CNN while Giorgi et al. (2020) used SVM, DT, RF, AdaBoost, SGD, RNN and CNN. Jiménez et al. (2023) tested LR, SVM ...

work page 2023
[7]

Shao et al

Traditional ML and DL: The Ensembler approach using Deep Forest was employed by S. Shao et al. (2020) for internet relay chat (IRC) by monitoring online channels. Different channels like darkscience, computer, hak5 were monitored for data along with IRC public datasets provided by the AZSecure-data project. Profile-based and instance-based features were u...

work page 2020
[8]

(2024) used LLMs for Latin AA and AV, focusing on historic language rather than modern language

LLMs: Schmidt et al. (2024) used LLMs for Latin AA and AV, focusing on historic language rather than modern language. They used LLMs such as GPT-4o, Gemini, Mistral, and Claude. GPT-4 performed best for AV so only that was used for the AA task. Other studies using LLMs include B. Huang et al. (2024b) using Mistral 7B, Llama 2, GPT 3.5 turbo and GPT 4 Turb...

work page 2024

Show all 54 references

[9]

IMDB62: A dataset comprising movie reviews from 62 prolific authors on IMDb, with over 1,000 reviews per author (Seroussi et al., 2014)

2014
[10]

It has many datasets of different sub-tasks

PAN AA: PAN is a series of scientific events and shared tasks focused on digital text forensics and writing style analysis (stylometry). It has many datasets of different sub-tasks. For AA it provides both close set (the author of disputed text belongs to known candidates) and...

2012
[12]

Liu, 2006)

and includes 5,000 financial news articles written by 50 authors, with each author having 50 training documents and 50 testing documents (Z. Liu, 2006). 4.1.2 AV

2006
[13]

The DTs used are Essays, Emails, Text messages and Business memos

PAN22 Authorship Analysis: AV: This AV dataset is based on a new corpus in English which provides cross-Discourse type (DTs) verification cases. The DTs used are Essays, Emails, Text messages and Business memos. The focus of the dataset is on more challenging scenarios where e...

2022
[14]

It was collected by the Federal Energy Regulatory Commission during their investigation into Enron’s collapse (Cohen, 2015)

Enron Email Dataset: This dataset includes 500,000 emails written by employ- ees of the Enron Corporation. It was collected by the Federal Energy Regulatory Commission during their investigation into Enron’s collapse (Cohen, 2015)

2015
[15]

It includes news articles from Reuters written between 1996 and

RCV1: The RCV1 dataset is a well-known benchmark for text categorization. It includes news articles from Reuters written between 1996 and

1996
[17]

The dataset consists of articles obtained from Internet news portals, is compiled into a dataset in the Comma Separated Value (CSV) document format

Indonesian News Article dataset: This dataset was created by Etania and Riccosan (2023) and was also utilized by Saputra and Riccosan (2024). The dataset consists of articles obtained from Internet news portals, is compiled into a dataset in the Comma Separated Value (CSV) doc...

2023
[18]

(2023) and contains over 2 million articles

arXiv dataset: This is a large dataset created from research papers available on arXiv, created by Bauersfeld et al. (2023) and contains over 2 million articles. Various trimmed versions of the dataset were also created for analysis. For example 19 in one version they limited ...

2023
[19]

The dataset comprises a total of 10 authors and 582 samples

AAALitCorpus: This dataset was created by Misini, Kadriu, and Canhasi (2024) and it consists of a collection of literary texts authored by prominent Albanian writers. The dataset comprises a total of 10 authors and 582 samples

2024
[20]

UNAAC-20: The Urdu News AA Corpus was developed by Nazir et al. (2021). It contains 26,118,475 words with a vocabulary of 291,728 unique tokens, and includes 21,938 articles written by 94 authors. The authors claim this is the largest AA dataset ever created for a low-resource...

2021
[21]

It was created by Khan et al

UACV-22: This dataset was built by gathering articles from Urdu newspapers to create a benchmark corpus for AV in the Urdu language. It was created by Khan et al. (2023) and was also utilized by Khan et al. (2024). The dataset comprises of 15 authors each with 400 articles. Th...

2023
[23]

From Figure 3a it is clear that from year 2017 to 2020, research predominantly utilized ML and DL techniques, with no reported use of LLMs

Figure 3a shows the articles per year and what technique they used whereas 3b displays the distribution with respect to sub-task. From Figure 3a it is clear that from year 2017 to 2020, research predominantly utilized ML and DL techniques, with no reported use of LLMs. The use...

2017
[24]

Does a text retain its original authorship when it undergoes numerous paraphrasing iterations

This indicates a growing interest in integrating LLMs for AA and AV. Prior to 2024, deep learning models dominated authorship analysis research. Inter- estingly even with the advent of LLMs and their powerful nature, the use of traditional ML and DL has dominated the research....

2023
[25]

It operates by recursively dividing the dataset based on feature values, creating a tree-like structure

Traditional ML based models: • Decision Tree (DT): A DT is a supervised learning algorithm utilized for clas- sification and regression tasks. It operates by recursively dividing the dataset based on feature values, creating a tree-like structure. In this structure, an interna...

1995
[26]

DL-based: Besides ML, DL models have also been utilized for the task of AA and AV. The models that have been used and included in this review are: • Convolutional Neural Network (CNN): A CNN is a special type of feed- forward neural network (also called a multi-layer perceptro...

1997
[27]

It is capable of generating coherent text and performing tasks such as translation and summarization without task-specific training (Radford et al., 2019)

LLM Based: • GPT-2: GPT-2 is a transformer-based language model developed by OpenAI, containing 1.5 billion parameters. It is capable of generating coherent text and performing tasks such as translation and summarization without task-specific training (Radford et al., 2019). •...

2019
[29]

The method used was K-means clustering

Unsupervised techniques: Very few articles used unsupervised techniques for classification purposes. The method used was K-means clustering. • K-means clustering is an unsupervised learning algorithm that divides a dataset into K groups by minimizing the distance between point...

1967
[30]

Combination of the above techniques. Appendix B Evidence of methodology 29 Table B1: Primary Search Terms Used in the Systematic Review Keyword ID Search Term K1 Large Language Models K2 Machine Learning K3 Deep Learning K4 Author Attribution K5 Author Verification K6 Text Tab...

2014
[35]

G., Bradley, H., OBrien, K., Hallahan, E., Khan, M

Biderman, S., Schoelkopf, H., Anthony, Q. G., Bradley, H., OBrien, K., Hallahan, E., Khan, M. A., Purohit, S., Prashanth, U. S., Raff, E., et al. (2023). Pythia: A suite for analyzing large language models across training and scaling. International Conference on Machine Learni...

2023
[37]

Jafariakinabad, F., & Hua, K. A. (2022). A self-supervised representation learn- ing of sentence structure for authorship attribution. ACM Transactions on Knowledge Discovery from Data (TKDD), 16 (4), 1–16. Jalal Yousef Zaidieh, A. (2024). Combatting cybersecurity threats on s...

2022 arXiv
[39]

Fourkioti, O., Symeonidis, S., & Arampatzis, A. (2019). Language models and fusion for authorship attribution. Information Processing & Management, 56 (6), 102061. Freund, Y., & Schapire, R. E. (1997). A decision-theoretic generalization of on-line learning and an application ...

2019
[41]

LaValley, M. P. (2008). Logistic regression. Circulation, 117 (18), 2395–2399. Lewis, D. D., Yang, Y., Rose, T. G., & Li, F. (2004). Rcv1: A new benchmark collection for text categorization research. Journal of machine learning research, 5 (Apr), 361–397. Linardatos, P., Papas...

2008
[42]

H., DGama, R., & Yu, H

Hu, X., Ou, W., Acharya, S., Ding, S. H., DGama, R., & Yu, H. (2023). TDRLM: Stylometric learning for authorship verification by Topic-Debiasing. Expert Systems with Applications, 233, 120745. Huang, B., Chen, C., & Shu, K. (2024a). Authorship attribution in the era of LLMs: P...

2023 arXiv
[44]

F., Anwar, W., Arshad, H., & Abbas, S

Khan, T. F., Anwar, W., Arshad, H., & Abbas, S. N. (2023). An Empirical Study on Authorship Verification for Low Resource Language using Hyper-Tuned CNN Approach. IEEE Access. Khan, T. F., Sabir, M., Malik, M. H., Ghous, H., Ijaz, H. M., Nadeem, A., & Ejaz, A. (2024). Comparat...

2023
[47]

Liu, X., & Kong, L. (2024). AI text detection method based on perplexity features with strided sliding window. Working notes of clef. Liu, Z. (2006). Reuter5050 [DOI: https://doi.org/10.24432/C5DS42]. 36 MacQueen, J. (1967). Some methods for classification and analysis of mult...

2024 doi
[50]

B., Tabassum, M., Feng, D., Das, A., Rahgooy, T., Dozier, G., & Seals, C

Rahgouy, M., Giglou, H. B., Tabassum, M., Feng, D., Das, A., Rahgooy, T., Dozier, G., & Seals, C. D. (2024). Towards effective authorship attribution: Integrat- ing class-incremental learning. 2024 IEEE 6th International Conference on Cognitive Machine Intelligence (CogMI), 56...

2024 arXiv
[52]

Romanov, A., Shelupanov, A., Kurtukova, A., & Fedotova, A. (2024). Integrated tech- nique of natural and artificial texts authorship verification in the academic environment. Rosen-Zvi, M., Chemudugunta, C., Griffiths, T., Smyth, P., & Steyvers, M. (2010). Learning author-topi...

2024 arXiv
[93]

Table C3: Distribution over the years Year No of articles Technique Sub-tasks ML / DL LLM AA AV Both 2017 2 2 0 2 0 0 2018 9 9 0 7 2 0 2019 6 6 0 3 3 0 2020 10 10 0 5 2 3 2021 13 14 1 9 4 1 2022 9 9 0 8 0 1 2023 15 15 2 8 5 2 2024 28 22 7 17 6 5 Total 93 87 10 59 22 12 31 Refe...

2024 arXiv
[131]

M., & Mimno, D

Hicke, R. M., & Mimno, D. (2023). T5 meets Tybalt: Author Attribution in Early Modern English Drama Using Large Language Models. arXiv preprint arXiv:2310.18454. 34 Hochreiter, S., & Schmidhuber, J. (1997). Long short-term memory. Neural computa- tion, 9 (8), 1735–1780. Hossai...

2023 arXiv
[190]

Kuo, C.-C. J. (2016). Understanding convolutional neural networks with a mathemat- ical model. Journal of Visual Communication and Image Representation, 41, 406–413. Lastowka, G. (2007). Digital attribution: Copyright and the right to credit. BUL Rev., 87,

2016
[297]

Cover, T., & Hart, P. (1967). Nearest neighbor pattern classification. IEEE transac- tions on information theory, 13 (1), 21–27. Custódio, J. E., & Paraboni, I. (2021). Stacked authorship attribution of digital texts. Expert Systems with Applications, 176, 114866. da Rocha, M....

1967
[348]

Avram, S.-M., & Oltean, M. (2022). A comparison of several AI techniques for authorship attribution on Romanian texts. Mathematics, 10 (23),

2022
[435]

Reisi, E., & Mahboob Farimani, H. (2020). Authorship attribution in historical and lit- erary texts by a deep learning classifier. Journal of Applied Intelligent Systems and Information Sciences, 1 (2), 118–127. Ribeiro, R., Carvalho, J. P., & Coheur, L. (2024). Leveraging fuz...

2020
[497]

Fang, Y., Yang, Y., & Huang, C. (2020). Emaildetective: An email authorship identification and verification model. The Computer Journal, 63 (11), 1775–

2020
[838]

(2022, March)

Stamatatos, E., Kredens, K., Pezik, P., Heini, A., Kestemont, M., Bevendorff, J., Pot- thast, M., & Stein, B. (2022, March). Pan22 authorship analysis: Authorship verification (Version 0.1.1). Zenodo. https://doi.org/10.5281/zenodo.6337151 39 Suman, C., Raj, A., Saha, S., & Bh...

2021 arXiv
[1787]

Fedotova, A., Romanov, A., Kurtukova, A., & Shelupanov, A. (2021). Authorship attribution of social media and literary Russian-language texts using machine learning methods and feature selection. Future Internet, 14 (1),

2021
[1997]

The dataset has 804,414 documents that were manually labeled and classified by industry, topic, and region using three controlled vocabularies (Lewis et al., 2004)

2004
[2004]

It includes 681,288 posts with over 140 million words, averaging about 35 posts and 7,250 words per blogger (Schler et al., 2006). 18

2006
[2023]

• Gemini: Gemini, developed by Google DeepMind and launched on December 6, 2023, is a multi-modal large language model

The latest iteration, Llama 3.2, released in September 2024, is a multi-modal model capable of processing both text and images, enhancing its versatility across various applications (Touvron et al., 2023). • Gemini: Gemini, developed by Google DeepMind and launched on December...

2023
[2024]

Due to inclusion and exclusion criteria mentioned in Section 2, no articles from 2015 and 2016 made it into the review as shown in Table C3 and Figure

Table C3 in Appendix C presents the distribution of methodologies used across the surveyed literature. Due to inclusion and exclusion criteria mentioned in Section 2, no articles from 2015 and 2016 made it into the review as shown in Table C3 and Figure

2015
[2207]

R., Cohen, T., & McGill, S

Zhang, O. R., Cohen, T., & McGill, S. (2018). Did Gaius Julius Caesar Write De Bello Hispaniensi? A Computational Study of Latin Classics Authorship. Human IT: Journal for Information Technology Studies as a Human Science, 14 (1), 28–58. Ziegel, E. R. (2003). The elements of s...

2018
[2700]

G., & Kertész, G

Oldal, L. G., & Kertész, G. (2022). Evaluation of Deep Learning-based Authorship Attribution Methods on Hungarian Texts. 2022 IEEE 10th Jubilee Interna- tional Conference on Computational Cybernetics and Cyber-Medical Systems (ICCC), 000161–000166. OpenAI. (2023). GPT-4 techni...

2022 arXiv
[4589]

Barlas, G., & Stamatatos, E. (2021). A transfer learning approach to cross-domain authorship attribution. Evolving Systems, 12 (3), 625–643. Bauersfeld, L., Romero, A., Muglikar, M., & Scaramuzza, D. (2023). Cracking double- blind review: Authorship attribution with deep learn...

2021
[7255]

Aouchiche, R. I. A., Boumahdi, F., Remmide, M. A., & Madani, A. (2024). Authorship attribution in twitter: A comparative study of machine learning and deep learning approaches. International Journal of Information Technology, 16 (5), 3303–3310. Apoorva, K., & Sangeetha, S. (20...

2024
[7518]

Modupe, A., Celik, T., Marivate, V., & Olugbara, O. O. (2023). Integrating bidi- rectional long short-term memory with subword embedding for authorship attribution. 2023 IEEE International Conference on Systems, Man, and Cybernetics (SMC), 1910–1917. Najafi, M., & Sadidpur, S....

2023
[9071]

https://doi.org/10.1109/TPAMI.2024.3415112 Guo, L., Yang, W., Ma, L., & Ruan, J. (2024). BLGAV: generative AI author verification model based on BERT and BiLSTM. Working Notes of CLEF. Halvani, O., Winter, C., & Graner, L. (2019). Assessing the applicability of author- ship ve...

2024

Pith tools

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

[1] [1]

Table 5 shows the summary of the research using LLMs

s based Research: Leveraging LLMs for AA is relatively new with only 3 articles exploring LLMs for Attribution, highlighting the research gap. Table 5 shows the summary of the research using LLMs. AA using the GPT-2 model and perplexity score was done by W. Huang et al. (2024) and they argued about the 12 superiority of perplexity over stylometry for the ...

work page 2024

[2] [2]

There are significantly fewer AV-based research articles compared to AA

Verify if the text is written by a human or machine (Guo et al., 2024). There are significantly fewer AV-based research articles compared to AA. The methods and techniques employed for AV span from traditional ML models to more advanced encoder-decoder architectures and transformer-based models. With the rise of LLMs, their potential for this task has sta...

work page 2024

[3] [3]

Studies that focused solely on traditional ML techniques include Castillo et al

Traditional ML and DL: Research in both traditional ML and DL often involves the use of multiple models simultaneously, reporting the best-performing model for specific settings. Studies that focused solely on traditional ML techniques include Castillo et al. (2019) using 2 variants of SVM classifier, namely C-SVC and Nu-SVC and Weeras- inghe et al. (2021...

work page 2019

[4] [4]

It includes the top 50 authors based on the total size of their articles, all of whom have written at least one article in the CCAT (corporate/industrial) category

Reuters-50: This dataset is a subset of the Reuters Corpus Volume 1 (RCV1) (Lewis et al., 2004), which has already been used in author identification studies. It includes the top 50 authors based on the total size of their articles, all of whom have written at least one article in the CCAT (corporate/industrial) category. The training set contains 2,500 t...

work page 2004

[5] [5]

Large Language Model (LLM): The use of LLM for AV is relatively new, with only 4 out of 22 articles leveraging LLMs for this task. In X. Liu and Kong (2024), the authors used a strided sliding 14 Table 6: A summary of research on AV # Article Models Used Dataset Language 1 W. Huang and Grieve (2024) GPT2, SVM PAN 2024 English 2 Guo et al. (2024) BERT, BiL...

work page 2024

[6] [6]

For instance, Khan et al

Some researchers did experiments using both traditional ML and DL techniques and reported the results of which technique pro- duced the best results for their data or experimental setting. For instance, Khan et al. (2023) employed RF, SVM and CNN while Giorgi et al. (2020) used SVM, DT, RF, AdaBoost, SGD, RNN and CNN. Jiménez et al. (2023) tested LR, SVM ...

work page 2023

[7] [7]

Shao et al

Traditional ML and DL: The Ensembler approach using Deep Forest was employed by S. Shao et al. (2020) for internet relay chat (IRC) by monitoring online channels. Different channels like darkscience, computer, hak5 were monitored for data along with IRC public datasets provided by the AZSecure-data project. Profile-based and instance-based features were u...

work page 2020

[8] [8]

(2024) used LLMs for Latin AA and AV, focusing on historic language rather than modern language

LLMs: Schmidt et al. (2024) used LLMs for Latin AA and AV, focusing on historic language rather than modern language. They used LLMs such as GPT-4o, Gemini, Mistral, and Claude. GPT-4 performed best for AV so only that was used for the AA task. Other studies using LLMs include B. Huang et al. (2024b) using Mistral 7B, Llama 2, GPT 3.5 turbo and GPT 4 Turb...

work page 2024