The Science of Detecting LLM-Generated Text
Open accessAbstract
While many detection methods have been proposed, understanding the challenges is far more daunting.
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
Integrity signals
Registry
Cites
Statistics
Claims the paper makes
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The ground-truth setup can introduce biases because it requires assumptions about which models scientists use and how they prompt them.
“This setup can introduce biases (16), as it requires assumptions on which models scientists use for their LLM-assisted writing, and how exactly they prompt them.”
From Delving into LLM-assisted writing in biomedical publications through excess vocabulary (2025) · cites this paper as [16]
The “Limitations of black-box detection” section says data collection can introduce biases that harm a detector’s generalization. “Data acquisition” also says researchers must choose generation models and prompts when collecting LLM-generated text. The paper does not specifically discuss assumptions about which models scientists use for LLM-assisted writing or how those scientists prompt them, so that application is an extension of its point. (LLM, conf=0.85)
Experimental — accuracy not yet measuredQuantitative comparisons have looked mainly at basic grammar and syntax or features such as word choice, punctuation, and sentence length, finding evidence of some differences between human- and LLM-written text.
“So far, quantitative comparisons have looked mainly at basic grammar and syntax (5) or features such as word choice, punctuation, sentence length, and so on, finding evidence of some differences between human- and LLM-written text (10–14).”
From Do LLMs write like humans? Variation in grammatical and rhetorical styles (2025) · cites this paper as [10]
The “Linguistic patterns” section describes comparisons of vocabulary, part of speech, punctuation, and grammar in human- and LLM-written text; “Statistical disparities” also reports differences in word-use patterns. These passages support the existence of measurable differences, but they do not establish that quantitative comparisons have looked mainly at those basic features or specifically at sentence length. No page number is available for these passages in our copy. (LLM, conf=0.82)
Experimental — accuracy not yet measuredOne of the methods to determine whether a text containing a world fact is generated by an LLM (and hence might require vetting) is to identify psycholinguistic patterns known to pervade 'factual' LLM language, such as lower levels of emotional expression.
“One of the methods to determine whether a text containing a world fact is generated by an LLM (and hence might require vetting) is to identify psycholinguistic patterns known to pervade ‘factual’ LLM language, such as lower levels of emotional expression62.”
From A psychometric framework for evaluating and shaping personality traits in large language models (2025) · cites this paper as [62]
The Linguistic patterns section describes sentiment as a feature for detecting LLM-generated text and reports that it is generally less emotional than human-authored text. The paper does not establish this pattern specifically for text containing world facts, and the White-Box Detection section cautions that prompts can change an LLM’s sentiment and style. (LLM, conf=0.82)
Experimental — accuracy not yet measuredDatasets of ground-truth LLM-generated and human-written scientific texts can easily be biased, with no guarantee that the corpus of LLM-generated texts is representative of all LLM use cases occurring in actual scholarly practice.
“Such datasets can easily be biased (16), with no guarantee that the corpus of LLM-generated texts is representative of all LLM use cases occurring in actual scholarly practice.”
From Delving into LLM-assisted writing in biomedical publications through excess vocabulary (2025) · cites this paper as [16]
There are structural differences between human- and LLM-written text.
“Though not definitive, these results suggest there are indeed structural differences between human- and LLM-written text.”
From Do LLMs write like humans? Variation in grammatical and rhetorical styles (2025) · cites this paper as [10]
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
The self-check needs the paper’s full text.
Cites
RoBERTa: A Robustly Optimized BERT Pretraining Approach
Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L., & Stoyanov, V. (2019) · arXiv · doi:10.48550/arxiv.1907.11692 · Cited by 16,515 · see citing works
GPT-3: Its Nature, Scope, Limits, and Consequences
Floridi, L., & Chiriatti, M. (2020) · Minds and Machines · doi:10.1007/s11023-020-09548-1 · Cited by 2,519 · see citing works
The Curious Case of Neural Text Degeneration
Holtzman, A., Buys, J., Du, L., Forbes, M., & Choi, Y. (2019) · arXiv · doi:10.48550/arxiv.1904.09751 · Cited by 1,304 · see citing works
Zipf’s word frequency law in natural language: A critical review and future directions
Piantadosi, S. T. (2014) · Psychonomic Bulletin & Review · doi:10.3758/s13423-014-0585-6 · Cited by 894 · see citing works
Electronic marking and identification techniques to discourage document copying
Brassil, J., Low, S. H., Maxemchuk, N. F., et al. (1995) · IEEE Journal on Selected Areas in Communications · doi:10.1109/49.464718 · Cited by 373 · see citing works
An estimate of an upper bound for the entropy of English
Brown, P. F., Pietra, V. J. D., Mercer, R. L., Pietra, S. A. D., & Lai, J. C. (1992) · OLAC - Open Language Archives Community · doi:10.5555/146680.146685 · Cited by 356 · see citing works
How Close is ChatGPT to Human Experts? Comparison Corpus, Evaluation, and Detection
Guo, B., Zhang, X., Wang, Z., Jiang, M., Nie, J., Ding, Y., Yue, J., & Wu, Y. (2023) · arXiv · doi:10.48550/arxiv.2301.07597 · Cited by 318 · see citing works
Release Strategies and the Social Impacts of Language Models
Solaiman, I., Brundage, M., Clark, J., Askell, A., Herbert-Voss, A., Wu, J., Radford, A., Krueger, G., Kim, J. W., Kreps, S., et al. (2019) · arXiv · doi:10.48550/arxiv.1908.09203 · Cited by 292 · see citing works
All That’s ‘Human’ Is Not Gold: Evaluating Human Evaluation of Generated Text
Clark, E., August, T., Serrano, S., Haduong, N., Gururangan, S., & Smith, N. A. (2021) · Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers) · doi:10.18653/v1/2021.acl-long.565 · Cited by 290 · see citing works
All That's 'Human' Is Not Gold: Evaluating Human Evaluation of Generated Text
Clark, E., August, T., Serrano, S., Haduong, N., Gururangan, S., & Smith, N. A. (2021) · arXiv · doi:10.48550/arxiv.2107.00061 · Cited by 290
Word Reordering and a Dynamic Programming Beam Search Algorithm for Statistical Machine Translation
Tillmann, C., & Ney, H. (2003) · Computational Linguistics · doi:10.1162/089120103321337458 · Cited by 258 · see citing works
Natural Language Watermarking: Design, Analysis, and a Proof-of-Concept Implementation
Atallah, M. J., Raskin, V., Crogan, M., Hempelmann, C. F., Kerschbaum, F., Mohamed, D., & Naik, S. (2001) · Lecture notes in computer science · doi:10.1007/3-540-45496-9_14 · Cited by 201 · see citing works
A Watermark for Large Language Models
Kirchenbauer, J., Geiping, J., Wen, Y., Katz, J., Miers, I., & Goldstein, T. (2023) · arXiv · doi:10.48550/arxiv.2301.10226 · Cited by 194 · see citing works
TweepFake: about detecting deepfake tweets
Fagni, T., Falchi, F., Gambini, M., Martella, A., & Tesconi, M. (2021) · PLOS ONE · doi:10.1371/journal.pone.0251415 · Cited by 185 · see citing works
Can AI-Generated Text be Reliably Detected?
Sadasivan, V. S., Kumar, A., Balasubramanian, S., Wang, W., & Feizi, S. (2023) · arXiv · doi:10.48550/arxiv.2303.11156 · Cited by 161 · see citing works
The hiding virtues of ambiguity
Topkara, U., Topkara, M., & Atallah, M. J. (2006) · ACM workshop on Multimedia and security · doi:10.1145/1161366.1161397 · Cited by 152 · see citing works
Dugan, L., Ippolito, D., Kirubarajan, A., et al. (2022) · AAAI Conference on Artificial Intelligence · doi:10.48550/arxiv.2212.12672 · Cited by 138 · see citing works
Feature-based detection of automated language models: tackling GPT-2, GPT-3 and Grover
Fröhling, L., & Zubiaga, A. (2021) · PeerJ Computer Science · doi:10.7717/peerj-cs.443 · Cited by 106 · see citing works
TURINGBENCH: A Benchmark Environment for Turing Test in the Age of Neural Text Generation
Uchendu, A., Ma, Z., Le, T., Zhang, R., & Lee, D. (2021) · Findings of the Association for Computational Linguistics: EMNLP 2021 · doi:10.18653/v1/2021.findings-emnlp.172 · Cited by 82 · see citing works
How Decoding Strategies Affect the Verifiability of Generated Text
Massarelli, L., Petroni, F., Piktus, A., Ott, M., Rocktäschel, T., Plachouras, V., Silvestri, F., & Riedel, S. (2020) · Findings of the Association for Computational Linguistics: EMNLP (Findings of EMNLP) · doi:10.18653/v1/2020.findings-emnlp.22 · Cited by 40 · see citing works
Cross-domain detection of GPT-2-generated technical text
Rodriguez, J., Hay, T., Gros, D., Shamsi, Z., & Srinivasan, R. (2022) · Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies · doi:10.18653/v1/2022.naacl-main.88 · Cited by 40 · see citing works
Adversarial Robustness of Neural-Statistical Features in Detection of Generative Transformers
Crothers, E., Japkowicz, N., Viktor, H. L., & Branco, P. (2022) · International Joint Conference on Neural Networks (IJCNN) · doi:10.1109/ijcnn55064.2022.9892269 · Cited by 30 · see citing works
Unsupervised and Distributional Detection of Machine-Generated TextPreprint
Gallé, M., Rozen, J., Kruszewski, G., & Elsahar, H. (2021) · arXiv (Cornell University) · doi:10.48550/arxiv.2111.02878 · Cited by 27 · see citing works
Neural Deepfake Detection with Factual Structure of TextPreprint
Zhong, W., Tang, D., Xu, Z., et al. (2020) · arXiv (Cornell University) · doi:10.48550/arxiv.2010.07475 · Cited by 4 · see citing works
Flagged first, then by citations
Research works printed but not linked: 12
- Annual Meeting of the ACL, 2020, 1808-1822.
- Jalil, Z. and Mirza, A.M. A review of digital watermarking techniques for text documents. In Proceedings of IEEE 2009 Intern. Conf. Information and Multimedia Technology. IEEE Computer Society, Washington D.C., U.S., 230-234.
- Kantchelian, A. et al. Better malware ground truth: Techniques for weighting anti-virus vendor labels. In Proceedings of the 8 th ACM Workshop on AI and Security. ACM, Denver, CO, USA, 2015, 45-56.
- Pagnoni, A., Graciarena, M. and Tsvetkov, Y. Threat scenarios and best practices to detect neural fake news. In Proceedings of the 29 th Intern. Conf. Computational Linguistics.(Gyeongju, Republic of Korea, 2022), 1233-1249.
- Pagnoni, A., Graciarena, M. and Tsvetkov, Y. Threat scenarios and best practices to detect neural fake news. In Proceedings of the 29 th Intern. Conf. Computational Linguistics. 2022, 1233-1249.
- Zellers, R. et al. Defending against neural fake news. Advances in Neural Information Processing Systems 32 (2019), 9054-9065.
- Abdelnabi, S. and Fritz, M. Adversarial watermarking transformer: Towards tracing text provenance with data hiding. In Proceedings of the 2021 IEEE Symp. Security and Privacy. IEEE, 121-140.
- Bhatt, P. and Rios, A. Detecting bot-generated text by characterizing linguistic accommodation in human-bot interactions. Findings of the Association for Computational Linguistics. ACL, Bangkok, Thailand, 2021, 3235-3247.
- Dugan, L., Ippolito, D., Kirubarajan, A. and Callison- Burch, C. Roft: A tool for evaluating human detection of machine-generated text. In Proceedings of the 2020 Conf. Empirical Methods in NLP: System Demonstrations. ACL Online, 2020, 189-196.
- Fan, A. et al. Eli5: Long form question answering. In Proceedings of the 57 th Annual Meeting of the ACL (Florence, Italy, 2019), 3558-3567.
- Gehrmann, S., Strobelt, H. and Rush, A.M. GLTR: Statistical detection and visualization of generated text. In Proceedings of the 57 th Annual Meeting of the ACL: System Demonstrations, 2019, 111-116.
- Ippolito, D., Duckworth, D., Callison-Burch, C. and Eck, D. Automatic detection of generated text is easiest when humans are fooled. In Proceedings of the 58 th
Cited by
Sorted by citations
Delving into LLM-assisted writing in biomedical publications through excess vocabulary
Kobak, D., González-Márquez, R., Horvát, E.-Á., & Lause, J. (2024) · arXiv · doi:10.48550/arxiv.2406.07016 · Cited by 214
Delving into LLM-assisted writing in biomedical publications through excess vocabulary
Kobak, D., González-Márquez, R., Horvát, E.-Á., & Lause, J. (2025) · Science Advances · doi:10.1126/sciadv.adt3813 · Cited by 213 · see citing works
“However, LLMs are infamous for making up references (Walters and Wilder, 2023), providing inaccurate summaries (Tang et al., 2024a; Kim et al., 2024), and making false claims that sound authoritative and convincing (Mittelstadt et al., 2023; Ji et al., 2023; Zhang et al., 2023; Zheng and Zhan,…”
A Survey on LLM-Generated Text Detection: Necessity, Methods, and Future Directions
Wu, J., Yang, S., Zhan, R., Yuan, Y., Chao, L. S., & Wong, D. F. (2025) · Computational Linguistics · doi:10.1162/coli_a_00549 · Cited by 212 · see citing works
Contrasting Linguistic Patterns in Human and LLM-Generated News Text
Muñoz-Ortiz, A., Gómez-Rodríguez, C., & Vilares, D. (2024) · Artificial Intelligence Review · doi:10.1007/s10462-024-10903-2 · Cited by 94 · see citing works
Anatomy of an AI-powered malicious social botnet
Yang, K., & Menczer, F. (2024) · Journal of Quantitative Description: Digital Media · doi:10.51685/jqd.2024.icwsm.7 · Cited by 71 · see citing works
Are Large Language Models a Threat to Digital Public Goods? Evidence from Activity on Stack Overflow
del Rio-Chanona, M., Laurentsyeva, N., & Wachs, J. (2023) · arXiv · doi:10.48550/arxiv.2307.07367 · Cited by 62 · see citing works
Do LLMs write like humans? Variation in grammatical and rhetorical styles
Reinhart, A., Markey, B., Laudenbach, M., Pantusen, K., Yurko, R., Weinberg, G., & Brown, D. W. (2025) · Proceedings of the National Academy of Sciences · doi:10.1073/pnas.2422455122 · Cited by 59 · see citing works
A psychometric framework for evaluating and shaping personality traits in large language models
Serapio-García, G., Safdari, M., Crepy, C., Sun, L., Fitz, S., Romero, P., Abdulhai, M., Faust, A., & Matarić, M. (2025) · Nature Machine Intelligence · doi:10.1038/s42256-025-01115-6 · Cited by 30 · see citing works
Generative artificial intelligence: implications for biomedical and health professions education
Hersh, W. (2025) · Annual Review of Biomedical Data Science · doi:10.1146/annurev-biodatasci-103123-094756 · Cited by 28 · see citing works
M4: Multi-generator, Multi-domain, and Multi-lingual Black-Box Machine-Generated Text Detection
Wang, Y., Mansurov, J., Ivanov, P., Su, J., Shelmanov, A., Tsvigun, A., Whitehouse, C., Afzal, O. M., Mahmoud, T., Sasaki, T., et al. (2023) · arXiv · doi:10.48550/arxiv.2305.14902 · Cited by 19 · see citing works
“Features can be categorized into three types: statistical distributions (e.g., GLTR-like word rankings), linguistic patterns (such as vocabulary features, part-of-speech tags, dependency parsing, sentiment analysis, and stylistic features) and fact-verification features (Tang et al., 2023).”
BackgroundYang, Z., Feng, Z., Huo, R., Lin, H., Zheng, H., Nie, R., & Chen, H. (2025) · Expert Systems with Applications · doi:10.1016/j.eswa.2025.126694 · Cited by 17 · see citing works
Sparks of Large Audio Models: A Survey and Outlook
Latif, S., Shoukat, M., Shamshad, F., Usama, M., Ren, Y., Cuayáhuitl, H., Wang, W., Zhang, X., Togneri, R., Cambria, E., et al. (2023) · arXiv · doi:10.48550/arxiv.2308.12792 · Cited by 17 · see citing works
“Tang et al. [62] focus on techniques for detecting text generated by LLMs, while Chang et al. [29] have examined the various ways to evaluate LLMs. Additionally, there are a number of surveys dedicated to investigating the specialised applications of Large Models in various fields such as vision…”
BackgroundHuman-LLM Coevolution: Evidence from Academic Writing
Geng, M., & Trotta, R. (2025) · Findings of the Association for Computational Linguistics: ACL 2025 · doi:10.18653/v1/2025.findings-acl.657 · Cited by 8 · see citing works
Towards AI-Generated Essay Classification Using Numerical Text Representation
Krawczyk, N., Probierz, B., & Kozak, J. (2024) · Applied Sciences · doi:10.3390/app14219795 · Cited by 4 · see citing works
Cheng, Z., Zhou, L., Jiang, F., Wang, B., & Li, H. (2025) · Proceedings of the ACM on Web Conference 2025 · doi:10.1145/3696410.3714770 · Cited by 3 · see citing works
“However, this method not only requires pre-editing that is not applicable to open-source models [42] but also affects the quality of model generation due to the insertion of watermarks [40, 48].”
Artificial Writing and Automated Detection
Jabarian, B., & Imas, A. (2025) · doi:10.2139/ssrn.5407424 · Cited by 3 · see citing works
- Authors
- Tang, R., Chuang, Y.-N., & Hu, X.
- Affiliations
- Rice University (US)
- Venue
- Communications of the ACM
- Published
- 2024-03-25
- Volume
- 67
- Issue
- 4
- Pages
- 50-59
- ISSN
- 0001-0782 (ISSN-L)
- 1557-7317
- OpenAlex
- W4392563948
- Cited by
- 125
- References
- 24
- Citations by year
- 2026: 32 · 2025: 66 · 2024: 26 · 2023: 1
- FWCI
- 18.30
- Type
- journal article
- Language
- en
- Open access
- bronze
- doi.org
- Locations
- Communications of the ACM (publishedVersion)
- Funders
- NSF (IIS-2310260, IIS-1849085, IIS-2224843)
- 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 › Adversarial Robustness in Machine Learning
- Keywords
- text detection, generative AI detection, AI-generated text
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.
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
