CoChat

A Survey on LLM-Generated Text Detection: Necessity, Methods, and Future Directions

Open access
Wu, Junchao, Yang, Shu, Zhan, Runzhe, Yuan, Yulin, Chao, Lidia Sam, Wong, Derek Fai · 2025 · Computational Linguistics

The DOI matches its Crossref record; 308 of the 310 references it cites were found in a registry or an index; 71 of the 197 sources its claims cite were checked for support.

Abstract

The remarkable ability of large language models (LLMs) to comprehend, interpret, and generate complex language has rapidly integrated LLM-generated text into various aspects of daily life, where users increasingly accept it. However, the growing reliance on LLMs underscores the urgent need for effective detection mechanisms to identify LLM-generated text. Such mechanisms are critical to mitigating misuse and safeguarding domains like artistic expression and social networks from potential negative consequences. LLM-generated text detection, conceptualized as a binary classification task, seeks to determine whether an LLM produced a given text. Recent advances in this field stem from innovations in watermarking techniques, statistics-based detectors, and neural-based detectors. Human-assisted methods also play a crucial role. In this survey, we consolidate recent research breakthroughs in this field, emphasizing the urgent need to strengthen detector research. Additionally, we review existing datasets, highlighting their limitations and developmental requirements. Furthermore, we examine various LLM-generated text detection paradigms, shedding light on challenges like out-of-distribution problems, potential attacks, real-world data issues, and ineffective evaluation frameworks. Finally, we outline intriguing directions for future research in LLM-generated text detection to advance responsible artificial intelligence. This survey aims to provide a clear and comprehensive introduction for newcomers while offering seasoned researchers valuable updates in the field.1

Overview

CoChat Summary

TL;DR

This survey reviews LLM-generated text detection, organizing methods into watermarking, statistics-based, neural-based, and human-assisted approaches, and argues robust detectors are urgently needed. It synthesizes datasets, benchmarks, evaluation metrics, and open challenges, concluding with future research directions.

Key findings

  • The survey’s systematic literature review screened work up to November 2023 and identified 83 pertinent pieces of literature; most relevant research appeared in 2023.
  • Detection methods are categorized into watermarking (data-driven, model-driven, post-processing), statistics-based (linguistic features, white-box, black-box), neural-based (feature classifiers, pre-trained classifiers, LLMs as detectors), and human-assisted methods.
  • It reports that current detectors face out-of-distribution problems (cross-domain, cross-lingual, cross-LLMs), attacks (paraphrase, adversarial, prompt, training threat models), real-world data issues (mixed/human-edited text, data ambiguity), model-size effects, and a lack of effective evaluation framework.
  • It cites selected performance figures from reviewed work: Hamed and Wu’s bigram-based system identified 98 of 100 LLM-written academic papers; LLMDet achieved 98.54% classification accuracy; GHOSTBUSTER achieved an average F1 of 99.0; fine-tuned BERT-family models averaged about 95% in-domain accuracy.

Methods

  • Design: a systematic literature review following Kitchenham’s SLR framework, guided by the research question of prevailing LLM-generated text detection methods and their main challenges.
  • Data: searches across Google Scholar, arXiv, Scopus, Web of Science, IEEE Xplore, Springer Link, ACL Anthology, and ACM Digital Library using Boolean search terms; focused on publicly accessible articles before November 2023; after de-duplication and manual screening, 83 works were included.
  • Analysis: synthesizes datasets and benchmarks, detector paradigms, evaluation metrics, important issues, and future directions; no primary experiments are conducted.

Limitations

  • Stated: current datasets and benchmarks are limited in attack types, domains, tasks, LLMs, and languages, and suffer from temporal staleness; there is a lack of effective evaluation framework; no detection technique is acknowledged as infallible; data ambiguity is not formally addressed by existing studies.
  • Stated: the paper argues prior reviews are insufficient in depth and comprehensiveness, but it does not state limitations of its own survey.
  • Evident: as a survey, it presents no new empirical validation, and its coverage is bounded by the search terms, databases, and the November 2023 cutoff with 83 included works.

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

Integrity signals

Registry

Crossref · 10.1162/coli_a_00549

Cites

308 of 310 references verified0 citation errors · 2 not checkable

Statistics

Nothing to recompute.

Figures

Figure 1 Toy picture of LLM-generated text detection task. This task is a binary classification task that detects whether the provided text is generated by LLMs or written by humans.
Figure 1 Toy picture of LLM-generated text detection task. This task is a binary classification task that detects whether the provided text is generated by LLMs or written by humans.CC-BY-NC-ND · From the PDF
Figure 2 The most critical reasons why LLM-generated text detection is needed urgently. We discussed it from five perspectives: Regulation, Users, Developments, Science, and Human Society.
Figure 2 The most critical reasons why LLM-generated text detection is needed urgently. We discussed it from five perspectives: Regulation, Users, Developments, Science, and Human Society.CC-BY-NC-ND · From the PDF
Figure 3 The distribution by year of the last 5 years of literature obtained from the screening is plotted. The number of published articles obtain significant attention in 2023.
Figure 3 The distribution by year of the last 5 years of literature obtained from the screening is plotted. The number of published articles obtain significant attention in 2023.CC-BY-NC-ND · From the PDF
Image 4
CC-BY-NC-ND · From the PDF

Verification

Claims the paper makes

50 Valid · 21 Not valid · 106 Not checked · 68 Own results and interpretations

64 more claims · 20 with their cited sources checked

Create a free account to see all 245 claims and to verify the 16 references still open.

See all claims — free account
  • Probably not a substantial claim. 119
  • It may state two separate claims as one. 92

Create a free account to see these claims.

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.

5 of 9 · the rest aren’t loaded

4 more findings

Create a free account to see every finding of this check.

See all findings — free account

Citations

Cites

Show all 310 references

Research works printed but not linked: 22

  1. Aich, 2022, Demystifying neural fake news via linguistic feature-based interpretation, 6586
  2. Christian, 2023, CNET secretly used AI on articles that didn’t disclose that fact, staff say, Futurism
  3. Cui, 2023, ChatLaw: Open-source legal large language model with integrated external knowledge bases, ArXiv preprint
  4. Deng, 2023, Efficient detection of LLM-generated texts with a Bayesian surrogate model, ArXiv preprint
  5. Dong, 2023, A survey for in-context learning, ArXiv preprint
  6. Giorgi, 2023, “I slept like a baby”: Using human traits to characterize deceptive ChatGPT and human text, 23
  7. Guo, 2020, Wiki-40B: Multilingual language model dataset, 2440
  8. He, 2023, Empirical analysis of beam search curse and search errors with model errors in neural machine translation, 91
  9. Kalinichenko, 2003, Digital libraries: Advanced methods and technologies, digital collections, D-Lib Magazine, 9, 1082
  10. Kitchenham, Barbara and StuartCharters. 2007. Guidelines for performing systematic literature reviews in software engineering. Technical Report, EBSE Technical Report EBSE-2007-01. pages 1–57. https://legacyfileshare.elsevier.com/promis_misc/525444systematicreviewsguide.pdf
  11. Kuditipudi, 2024, Robust distortion-free watermarks for language models, Transactions on Machine Learning Research
  12. Lambert, 2022, Illustrating reinforcement learning from human feedback (RLHF), HuggingFace Blog
  13. Liu, 2023, A private watermark for large language models, ArXiv preprint
  14. Liu, 2023, Check me if you can: Detecting ChatGPT-generated academic writing using CheckGPT, ArXiv preprint
  15. Ma, 2023, AI vs. human–differentiation analysis of scientific content generation, arXiv
  16. McCarthy, Philip M. 2005. An Assessment of the Range and Usefulness of Lexical Diversity Measures and the Potential of the Measure of Textual, Lexical Diversity (MTLD). Ph.D. thesis, The University of Memphis.
  17. Nicks, 2023, Language model detectors are easily optimized against
  18. Pagnoni, 2022, Threat scenarios and best practices to detect neural fake news, 1233
  19. Piccolo, 2023, Many bioinformatics programming tasks can be automated with ChatGPT, ArXiv preprint
  20. Radford, 2019, Language models are unsupervised multitask learners, OpenAI Blog, 1, 9
  21. Yan, 2023, Detection of AI-generated essays in writing assessment, Psychological Testing and Assessment Modeling, 65, 125
  22. Yu, 2023, GPT paternity test: GPT generated text detection with GPT genetic inheritance, ArXiv preprint

Cited by

Metadata

Authors
Wu, J., Yang, S., Zhan, R., Yuan, Y., Chao, L. S., & Wong, D. F.
Affiliations
  • University of Macau (MO)
  • Peking University (CN)
Venue
Computational Linguistics
Publisher
MIT Press
Published
2025
Volume
51
Issue
1
Pages
275-338
ISSN
  • 0891-2017 (ISSN-L)
  • 1530-9312
OpenAlex
W4406236959
Cited by
212
References
257
Citations by year
2026: 106 · 2025: 91
FWCI
79.02
Type
journal article
Language
en
Open access
Locations
  • Computational Linguistics (publishedVersion, cc-by-nc-nd)
  • arXiv (Cornell University) (submittedVersion, cc-by)
  • arXiv (Cornell University) (cc-by)
Funders
  • United Mitochondrial Disease Foundation
  • Universidade de Macau
  • Fundo para o Desenvolvimento das Ciências e da Tecnologia (FDCT/060/2022/AFJ, FDCT/0070/2022/AMJ)
  • State Administration for Science, Technology and Industry for National Defense
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Authorship Attribution and Profiling
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Digital Media Forensic Detection
  • Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
Keywords
LLM-generated text detection, large language models, watermarking, detection dataset, out-of-distribution detection, adversarial attacks, evaluation framework, responsible AI

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.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-26, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-26 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (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

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.