Pith. sign in

REVIEW 2 major objections 4 minor 51 references

When Detection Fails: The Power of Fine-Tuned Models to Generate Human-Like Social Media Text

T0 review · 2 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 Fine-tuned AI text fools automated detectors and human readers alike.

desk verdict Fine-tuning LLMs on social media text makes them much harder to detect; this is a careful, useful study, but the headline 'realistic attacker' numbers hinge on a preprocessing assumption that deserves more probing. read the letter →

arxiv 2506.09975 v2 pith:O4CLT5F3 submitted 2025-06-11 cs.CL

Hillary Dawkins , Kathleen C. Fraser , Svetlana Kiritchenko This is my paper
classification cs.CL
keywords AI-generatedtextdetectionfine-tuningsocialmediaastroturfinglargelanguagemodelshumanevaluationbenchmarkdataset
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 asks whether a realistic attacker could use large language models to post social-media content that looks human-written. The answer it gives is yes: fine-tuning a language model on ordinary tweets makes its output dramatically harder to detect than output from the same model without fine-tuning. Under idealized conditions where the defender knows the generating model and has training data, a RoBERTa-based detector reaches 99.9% accuracy on base GPT-4o posts but only 71.6% on posts from fine-tuned GPT-4o. When knowledge of the generator is removed, off-the-shelf detectors fall to near chance, and human readers identify fine-tuned AI posts at 53.9% accuracy. The findings bear directly on astroturfing and influence campaigns, where the deception is the inauthenticity of a manufactured crowd.

What carries the argument

The central object is the fine-tuned language model: each base model (GPT-4o, GPT-4o-mini, Llama-3-8B, Llama-3.2-1B) is adapted to social-media style by supervised fine-tuning on 800 or 2000 training pairs of (topic description, human tweet), using OpenAI's fine-tuning interface for the closed models and QLoRA for the open ones. The evaluation machinery is a layered set of detection scenarios: an idealized 'complete knowledge' setting where the generating model is used as the measurement model and thresholds are calibrated in-distribution; an off-the-shelf setting with no access to the generator; and intermediate settings that relax one resource at a time. Detection is carried out with the MGTBench suite, including metric-based methods (log-likelihood, entropy, rank, log-rank, log-likelihood-log-rank ratio, Fast-DetectGPT, Binoculars) and PLM-based classifiers (OpenAI's RoBERTa detector, ChatGPT detector, GPTZero). The preprocessing step that removes @mentions and links from all tweets is part of the machinery, motivated by the assumption that a realistic attacker would insert such features separately.

What would settle it

Fine-tune a model on tweet data that has not been stripped of mentions and links, generate posts with that model, and run the off-the-shelf detectors exactly as the paper does; if accuracy stays at chance rather than rising toward the 79% the paper reports for the unprocessed condition, the claim that preprocessing drives the drop would be called into question.

Watch

Extended reading notes

Core claim

The paper's central claim is that fine-tuning is a cheap and generally applicable way to make AI-generated text evade detection, and that most existing detection benchmarks overstate real-world performance because they assume defenders have knowledge of the generating model. The authors construct a dataset of 505,159 AI-generated posts from four base models and their fine-tuned variants, across eleven controversial topics, and evaluate eight detection methods under progressively more realistic threat models. They find that while base-model text is nearly perfectly detectable in the idealized scenario (up to 99.9% accuracy), fine-tuned text drops to 71.6% accuracy for the strongest detector, and that even closely related fine-tuned models do not substitute well as resources for training or calibration. A human annotation study with 250 participants confirms the result: accuracy falls from 61.2% on base-model text to 53.9% on fine-tuned text, a statistically significant difference. The authors conclude that existing detectors are not as effective in the real world as in the research lab, since fine-tuning is a realistic and inexpensive step for an attacker.

Load-bearing premise

The headline numbers for the off-the-shelf detectors assume an attacker who generates the post text separately from any @mentions and links, so the dataset strips those features from both classes before evaluation; without that preprocessing, the fine-tuned text is noticeably easier to detect (for example, 79.2% vs 71.6% for the fine-tuned PLM detector).

Editorial extensions

If this is right

  • Existing benchmarks for AI-generated text detection, which assume access to the generating model, overestimate real-world detection performance.
  • Off-the-shelf detectors such as GPTZero and Binoculars operate near random levels on fine-tuned social-media text, suggesting commercial detection tools are not reliable against this threat.
  • Training data from a closely related but not identical fine-tuned model is a poor substitute: detection accuracy drops substantially when the training generator differs from the target, even when they share the same base model and fine-tuning data.
  • At the low false-positive rates likely required in practice (FPR = 0.01), the true positive rate of the strongest detector falls to as low as 0.129 on fine-tuned text.

Reading between the lines

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

  • If fine-tuning is as cheap and effective as reported (3 cents to $26 USD), the same recipe likely transfers beyond social media to any domain with abundant human-written examples, such as reviews, comments, or forum posts; the authors note the general applicability but do not demonstrate it.
  • The mention/link-stripping preprocessing is the main boundary condition: the near-chance off-the-shelf results may not hold against an attacker who does not separate content generation from mention/link insertion, and a detector that explicitly models those features could recover some signal.
  • A testable extension is to train detectors on a mixture of many fine-tuned variants of the same base model; the paper's ablation suggests this would not restore accuracy, but the mixture condition was not directly tested.
  • The human study's low accuracy on fine-tuned text (53.9%) suggests that human moderation alone cannot be the backstop for platform authenticity, and that network- or account-level signals (which the paper lists as out of scope) may be necessary.
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.

Referee Report

2 major / 4 minor

Summary. The paper constructs a large dataset of 505,159 AI-generated social media posts from open-source and closed-source LLMs, including fine-tuned variants, across 11 controversial topics. It reports that fine-tuned models produce text that is much harder to detect than base models under both an idealized in-distribution detection scenario and an off-the-shelf scenario, and confirms the difficulty with a human study in which fine-tuned GPT-4o text is detected at only 53.9% accuracy. The central claim is that under a 'realistic attacker' assumption—namely, that the attacker does not release the fine-tuned model and can insert mentions/links as a separate step—detectability drops dramatically for both machines and humans.

Significance. If the results hold, this is a timely and important contribution. The dataset release is a substantial asset for the community, and the experimental design is careful: topic-matched pairs, balanced classes, multiple detector families (metric-based, PLM, commercial), in-distribution threshold calibration where appropriate, and a human annotation study with quality checks. The fine-tuning cost table (Table 9) convincingly shows that the attacker capability is cheap and accessible. The paper also goes beyond a single headline number by systematically varying prompt strategy, fine-tuning data size, and knowledge of the generating model (Tables 2–4, 15–24), giving a rich picture of detector brittleness. The qualitative conclusion that fine-tuning reduces detectability is well supported. However, the headline magnitude of the 'dramatic drop' is partly conditional on a preprocessing decision (removal of mentions/links) whose effect is quantified for only one detector, as detailed below.

major comments (2)
  1. [Appendix B.5, Table 18] The central 'dramatic drop' claim for off-the-shelf detectors is measured only after removing all mentions and links from both classes. Table 18 shows that this removal is not neutral for fine-tuned generators: the in-distribution PLM accuracy on FT_GPT4o-large rises from 71.6% (processed) to 79.2% (unprocessed). Yet the paper does not report the unprocessed accuracies for the off-the-shelf detectors (Binoculars, OpenAI detector, ChatGPT detector, GPTZero) that carry the realistic-defender conclusion in Table 2 (right side). Without those numbers, the magnitude of the drop for a defender facing an attacker who does not perfectly perform the separate mention/link insertion step is unknown. Please provide the unprocessed off-the-shelf results, or explicitly state and justify the attacker model as an assumption and present the processed results as a sensitivity condition rather than the headline scenario.
  2. [Section 3.3.4 / Appendix B.5] The assumption that 'any reasonably sophisticated actor could insert the desired mentions and relevant links as a separate step' is presented without supporting evidence or a concrete model for how that separate step would be executed. Since the effect of removing mentions/links is substantial for fine-tuned models (Table 18: differences of 0.032–0.080 in PLM accuracy), the realism of this assumption is load-bearing for the quantitative conclusions. I am not asking for a full behavioral study, but the paper should at least (a) acknowledge that this is an untested assumption and (b) report the range of detection accuracies under both processing conditions so that readers can see how much the conclusion depends on it.
minor comments (4)
  1. [Section 4.2.1 / Table 2 caption] In Table 2 and the text, 'fine-tuned GTP4o model' appears to be a typo for 'fine-tuned GPT-4o model'; please correct.
  2. [Appendix D] The paper reports results from a single train–test split for most classification experiments, and notes that variance was low based on three detectors and three splits. It would be clearer to state explicitly in the main text that these are single-split results, and to include the variance information for at least the key fine-tuned conditions (e.g., FT_GPT4o-large).
  3. [Table 18 / Appendix D] The column headers 'Processed' and 'Unprocessed (@'s and links)' are slightly ambiguous; consider 'Mentions/links removed' and 'Mentions/links retained' to make the direction of the difference immediately obvious.
  4. [Section 6 / Limitations] The limitations section appropriately notes the short-text, English-only, single-platform scope, but it does not revisit the mention/link-removal preprocessing as a limitation. Adding a sentence acknowledging that the headline numbers assume a two-stage generation pipeline would be helpful for readers.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the paper's central claims are direct empirical measurements, with only a non-load-bearing background self-citation.

full rationale

The paper's central claims are empirical rather than derivational: it constructs a dataset of 505,159 generated posts, measures detector accuracies under several access scenarios, and runs a human annotation study. The headline numbers are direct measurements on held-out data: the fine-tuned PLM detector drops from 99.9% accuracy on base GPT-4o to 71.6% on fine-tuned GPT-4o (Table 2), and human accuracy is 53.9% for fine-tuned text (Section 4.3). No fitted parameter is renamed as a prediction: the detectors were not trained on the human-study test items, and the off-the-shelf detectors were used with their published or default configurations. The only self-citation is Fraser et al. (2025), cited in the background as one of several survey references; it is not load-bearing for any result. Appendix B.5 strips @mentions and links from both classes before detection, on the explicit assumption that a sophisticated attacker could insert them as a separate step; this is a stated, quantified scope assumption rather than a hidden reduction, and Table 18 discloses the sensitivity of the fine-tuned PLM detector to that preprocessing (e.g., FT_GPT4o-large rises from 71.6% processed to 79.2% unprocessed). Even in the unprocessed condition, the qualitative conclusion that fine-tuning reduces detectability remains an external measurement, not an identity. GPT-4o is used to extract topic descriptions for generation, but this is a generation tool, not an argument that entails detectability. No equation or definition in the paper makes a claimed output equal to its input by construction, and no load-bearing step is justified solely by a self-citation. The observed conditional dependence on the mention/link preprocessing is a limitation on generalizability, not circularity.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The central claim depends on domain assumptions about attacker behavior and data representativeness, not on fitted parameters or invented entities. The most consequential assumption is the removal of mentions and links, which strengthens the reported detection difficulty.

assumptions (4)
  • domain assumption The 11,000 human-written tweets sampled from public datasets are representative of human social media style on controversial topics.
    The paper relies on this to compare generated text against a trusted human baseline; Section 3.1.
  • domain assumption A realistic threat actor will fine-tune on a benign, off-topic tweet sample and will insert @mentions and links separately.
    This justifies removing all mentions and links from the data before detection experiments, which strongly affects measured detectability (Appendix B.5, Table 18).
  • domain assumption The 'Generate from Topic' prompting strategy is a realistic approximation of how attackers would mass-produce astroturfing posts.
    Section 3.3.3; this is the main scenario for fine-tuned models.
  • domain assumption The two model families (GPT-4o/4o-mini, Llama-3-8B/3.2-1B) and the specific fine-tuning setups (QLoRA rank 64, 5 epochs; OpenAI interface) represent the space of realistic attackers.
    Limitations section acknowledges other models and settings may differ.

how reviews work

0 comments
Cite this review

Pith. "Pith review of When Detection Fails: The Power of Fine-Tuned Models to Generate Human-Like Social Media Text." pith.science (2026). https://pith.science/paper/O4CLT5F3

@misc{pith2026250609975,
  author       = {Pith},
  title        = {Pith review of: When Detection Fails: The Power of Fine-Tuned Models to Generate Human-Like Social Media Text},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/O4CLT5F3}},
  note         = {Machine review of arXiv:2506.09975}
}
read the original abstract

Detecting AI-generated text is a difficult problem to begin with; detecting AI-generated text on social media is made even more difficult due to the short text length and informal, idiosyncratic language of the internet. It is nonetheless important to tackle this problem, as social media represents a significant attack vector in online influence campaigns, which may be bolstered through the use of mass-produced AI-generated posts supporting (or opposing) particular policies, decisions, or events. We approach this problem with the mindset and resources of a reasonably sophisticated threat actor, and create a dataset of 505,159 AI-generated social media posts from a combination of open-source, closed-source, and fine-tuned LLMs, covering 11 different controversial topics. We show that while the posts can be detected under typical research assumptions about knowledge of and access to the generating models, under the more realistic assumption that an attacker will not release their fine-tuned model to the public, detectability drops dramatically. This result is confirmed with a human study. Ablation experiments highlight the vulnerability of various detection algorithms to fine-tuned LLMs. This result has implications across all detection domains, since fine-tuning is a generally applicable and realistic LLM use case.

Figures

Figures reproduced from arXiv: 2506.09975 by the authors.

Figure 1
Figure 1. Length of text in characters (top), Type-Token Ratio or TTR (middle), and number of hashtags (bottom), [PITH_FULL_IMAGE:figures/full_fig_p018_1.png] view at source ↗
Figure 2
Figure 2. Frequency of occurrence of the most frequent “private verbs” by Llama-3-8B and humans. [PITH_FULL_IMAGE:figures/full_fig_p025_2.png] view at source ↗
Figure 3
Figure 3. Proportion of tweets containing at least one profanity. [PITH_FULL_IMAGE:figures/full_fig_p026_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Proportion of tweets containing at least one slang word or abbreviation. [PITH_FULL_IMAGE:figures/full_fig_p026_4.png]
Figure 5
Figure 5. Figure 5: The task instructions and one group of tweets for annotation in the human study. [PITH_FULL_IMAGE:figures/full_fig_p027_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

51 extracted references · 25 canonical work pages

  1. [1]

    Guangsheng Bao, Yanbin Zhao, Zhiyang Teng, Linyi Yang, and Yue Zhang. 2024. https://openreview.net/forum?id=Bpcgcr8E8Z Fast-DetectGPT : E fficient zero-shot detection of machine-generated text via conditional probability curvature . In Proceedings of the Twelfth International Conference on Learning Representations

    work page 2024

  2. [2]

    Douglas Biber. 1993. https://aclanthology.org/J93-2001/ Using register-diversified corpora for general language studies . Computational Linguistics, 19(2):219--241

    work page 1993

  3. [3]

    Jovy Chan. 2024. https://journals.sagepub.com/doi/full/10.1177/01914537221108467 Online astroturfing: A problem beyond disinformation . Philosophy & Social Criticism, 50(3):507--528

    work page doi:10.1177/01914537221108467 2024

  4. [4]

    Stefano Cresci. 2020. https://dl.acm.org/doi/abs/10.1145/3409116 A decade of social bot detection . Communications of the ACM, 63(10):72--83

    doi:10.1145/3409116 2020

  5. [5]

    Evan Crothers, Nathalie Japkowicz, and Herna L Viktor. 2023. https://ieeexplore.ieee.org/abstract/document/10177704 Machine-generated text: A comprehensive survey of threat models and detection methods . IEEE Access

    work page arXiv 2023

  6. [6]

    Limeng Cui and Dongwon Lee. 2020. https://arxiv.org/abs/2006.00885 CoAID : COVID-19 healthcare misinformation dataset . Preprint, arXiv:2006.00885

    arXiv 2020

  7. [7]

    Wanyun Cui, Linqiu Zhang, Qianle Wang, and Shuyang Cai. 2023. Who said that? B enchmarking social media AI detection. arXiv preprint arXiv:2310.08240

    arXiv 2023

  8. [8]

    Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. 2024. https://proceedings.neurips.cc/paper_files/paper/2023/hash/1feb87871436031bdc0f2beaa62a049b-Abstract-Conference.html QLORA : E fficient finetuning of quantized LLM s . In Proceedings of the 37th International Conference on Neural Information Processing Systems, NIPS '23, Red Hook, N...

    work page 2024

Show all 51 references
  1. [9]

    Devereaux, Alexandre Bergeron-Guyard, Bruce Forrester, and Marc-Andre Labrie

    Zachary P. Devereaux, Alexandre Bergeron-Guyard, Bruce Forrester, and Marc-Andre Labrie. 2025. https://cdainstitute.ca/publications/category/on-track/ AI and information warfare: T he M ockingbird prototype . On Track, 35:47--56

    2025

  2. [10]

    Tiziano Fagni, Fabrizio Falchi, Margherita Gambini, Antonio Martella, and Maurizio Tesconi. 2021. https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0251415 TweepFake : About detecting deepfake tweets . Plos one, 16(5):e0251415

    2021 doi

  3. [11]

    Bruce Forrester, Akiva Bacovcin, Zachary Devereaux, and Stefany Bedoya. 2019. https://www.sto.nato.int/publications/STO Technical report, NATO Science & Technology Organization

    2019

  4. [12]

    Kathleen C Fraser, Hillary Dawkins, and Svetlana Kiritchenko. 2025. https://www.jair.org/index.php/jair/article/view/16665 Detecting AI -generated text: F actors influencing detectability with current methods . Journal of Artificial Intelligence Research, 82:2233--2278

    2025

  5. [13]

    Sebastian Gehrmann, Hendrik Strobelt, and Alexander Rush. 2019. https://doi.org/10.18653/v1/P19-3019 GLTR : Statistical detection and visualization of generated text . In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics: System Demonstrat...

    2019 doi

  6. [14]

    Soumya Suvra Ghosal, Souradip Chakraborty, Jonas Geiping, Furong Huang, Dinesh Manocha, and Amrit Bedi. 2023. https://openreview.net/forum?id=AXtFeYjboj A survey on the possibilities & impossibilities of AI -generated text detection . Transactions on Machine Learning Research

    2023

  7. [15]

    Biyang Guo, Xin Zhang, Ziyuan Wang, Minqi Jiang, Jinran Nie, Yuxuan Ding, Jianwei Yue, and Yupeng Wu. 2023. https://arxiv.org/abs/2301.07597 How close is ChatGPT to human experts? C omparison corpus, evaluation, and detection . arXiv preprint arXiv:2301.07597

    2023 arXiv

  8. [16]

    Abhimanyu Hans, Avi Schwarzschild, Valeriia Cherepanova, Hamid Kazemi, Aniruddha Saha, Micah Goldblum, Jonas Geiping, and Tom Goldstein. 2024. https://openreview.net/forum?id=axl3FAkpik Spotting llms with binoculars: Zero-shot detection of machine-generated text . In Proceedin...

    2024

  9. [18]

    Steffen Herbold, Annette Hautli-Janisz, Ute Heuer, Zlata Kikteva, and Alexander Trautsch. 2023. A large-scale comparison of human-written versus ChatGPT -generated essays. Scientific reports, 13(1):18617

    2023

  10. [19]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen - Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen - Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen. 2021. https://arxiv.org/abs/2106.09685 LoRA : L ow-rank adaptation of large language models . CoRR, abs/2106.09685

    2021 arXiv

  11. [20]

    Kalpesh Krishna, Yixiao Song, Marzena Karpinska, John Wieting, and Mohit Iyyer. 2024. https://proceedings.neurips.cc/paper_files/paper/2023/hash/575c450013d0e99e4b0ecf82bd1afaa4-Abstract-Conference.html Paraphrasing evades detectors of AI -generated text, but retrieval is an e...

    2024

  12. [21]

    Tharindu Kumarage, Joshua Garland, Amrita Bhattacharjee, Kirill Trapeznikov, Scott Ruston, and Huan Liu. 2023 a . https://arxiv.org/abs/2303.03697 Stylometric detection of AI -generated text in T witter timelines . arXiv preprint arXiv:2303.03697

    2023 arXiv

  13. [22]

    Tharindu Kumarage, Paras Sheth, Raha Moraffah, Joshua Garland, and Huan Liu. 2023 b . https://doi.org/10.18653/v1/2023.findings-emnlp.94 How reliable are AI -generated-text detectors? an assessment framework using evasive soft prompts . In Findings of the Association for Compu...

    2023 doi

  14. [23]

    Linyang Li, Pengyu Wang, Ke Ren, Tianxiang Sun, and Xipeng Qiu. 2023. https://arxiv.org/abs/2304.14072 Origin tracing and detecting of LLM s . arXiv preprint arXiv:2304.14072

    2023 arXiv

  15. [24]

    Yafu Li, Qintong Li, Leyang Cui, Wei Bi, Zhilin Wang, Longyue Wang, Linyi Yang, Shuming Shi, and Yue Zhang. 2024. https://doi.org/10.18653/v1/2024.acl-long.3 MAGE : Machine-generated text detection in the wild . In Proceedings of the 62nd Annual Meeting of the Association for ...

    2024 doi

  16. [25]

    Weixin Liang, Mert Yuksekgonul, Yining Mao, Eric Wu, and James Zou. 2023. GPT detectors are biased against non-native E nglish writers. Patterns, 4(7)

    2023

  17. [26]

    Zeyan Liu, Zijun Yao, Fengjun Li, and Bo Luo. 2024. https://dl.acm.org/doi/abs/10.1145/3658644.3670392 On the detectability of ChatGPT content: B enchmarking, methodology, and evaluation through the lens of academic writing . In Proceedings of the 2024 on ACM SIGSAC Conference...

    2024

  18. [27]

    Jason Lucas, Adaku Uchendu, Michiharu Yamashita, Jooyoung Lee, Shaurya Rohatgi, and Dongwon Lee. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.883 Fighting fire with fire: T he dual role of LLM s in crafting and detecting elusive disinformation . In Proceedings of the 2023...

    2023 doi

  19. [28]

    Dominik Macko, Jakub Kopal, Robert Moro, and Ivan Srba. 2024. https://arxiv.org/abs/2406.12549 M ulti S ocial: M ultilingual benchmark of machine-generated text detection of social-media texts . arXiv preprint arXiv:2406.12549

    2024 arXiv

  20. [29]

    Emi, and Max Spero

    Elyas Masrour, Bradley N. Emi, and Max Spero. 2025. https://aclanthology.org/2025.genaidetect-1.9/ DAMAGE : Detecting adversarially modified AI generated text . In Proceedings of the 1st Workshop on GenAI Content Detection (GenAIDetect), pages 120--133, Abu Dhabi, UAE. Interna...

    2025

  21. [30]

    Eric Mitchell, Yoonho Lee, Alexander Khazatsky, Christopher D Manning, and Chelsea Finn. 2023. https://proceedings.mlr.press/v202/mitchell23a.html DetectGPT : Z ero-shot machine-generated text detection using probability curvature . In Proceedings of the International Conferen...

    2023

  22. [31]

    Saif Mohammad, Svetlana Kiritchenko, Parinaz Sobhani, Xiaodan Zhu, and Colin Cherry. 2016. https://doi.org/10.18653/v1/S16-1003 S em E val-2016 task 6: Detecting stance in tweets . In Proceedings of the 10th International Workshop on Semantic Evaluation ( S em E val-2016) , pa...

    2016 doi

  23. [32]

    Jiameng Pu, Zain Sarwar, Sifat Muhammad Abdullah, Abdullah Rehman, Yoonjin Kim, Parantapa Bhattacharya, Mobin Javed, and Bimal Viswanath. 2023. https://ieeexplore.ieee.org/abstract/document/10179387 Deepfake text detection: L imitations and opportunities . In Proceedings of th...

    2023

  24. [33]

    Alex Reinhart, David West Brown, Ben Markey, Michael Laudenbach, Kachatad Pantusen, Ronald Yurko, and Gordon Weinberg. 2024. https://arxiv.org/abs/2410.16107 Do LLMs write like humans? variation in grammatical and rhetorical styles . arXiv preprint arXiv:2410.16107

    2024 arXiv

  25. [34]

    Vinu Sankar Sadasivan, Aounon Kumar, Sriram Balasubramanian, Wenxiao Wang, and Soheil Feizi. 2023. https://arxiv.org/abs/2303.11156 Can AI -generated text be reliably detected? arXiv preprint arXiv:2303.11156

    2023 arXiv

  26. [35]

    Tony Berber Sardinha. 2024. https://www.sciencedirect.com/science/article/pii/S2666799123000436 AI -generated vs human-authored texts: A multidimensional comparison . Applied Corpus Linguistics, 4(1):100083

    2024

  27. [36]

    Areg Mikael Sarvazyan, Jos \'e \'A ngel Gonz \'a lez, Marc Franco-Salvador, Francisco Rangel, Berta Chulvi, and Paolo Rosso. 2023. https://arxiv.org/abs/2309.11285 Overview of AuTexTification at IberLEF 2023 : D etection and attribution of machine-generated text in multiple do...

    2023 arXiv

  28. [37]

    Ruiting Shao, Ryan Schwarz, Christopher Clifton, and Edward Delp. 2024. https://aclanthology.org/2024.lrec-main.92/ A natural approach for synthetic short-form text analysis . In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Reso...

    2024

  29. [38]

    Zhouxing Shi, Yihan Wang, Fan Yin, Xiangning Chen, Kai-Wei Chang, and Cho-Jui Hsieh. 2024. https://direct.mit.edu/tacl/article/doi/10.1162/tacl_a_00639/119629 Red teaming language model detectors with language models . Transactions of the Association for Computational Linguist...

    2024 doi

  30. [39]

    Jinyan Su, Terry Zhuo, Di Wang, and Preslav Nakov. 2023. https://doi.org/10.18653/v1/2023.findings-emnlp.827 D etect LLM : Leveraging log rank information for zero-shot detection of machine-generated text . In Findings of the Association for Computational Linguistics: EMNLP 20...

    2023 doi

  31. [40]

    Zhen Sun, Zongmin Zhang, Xinyue Shen, Ziyi Zhang, Yule Liu, Michael Backes, Yang Zhang, and Xinlei He. 2024. Are we in the AI -generated text world already? Q uantifying and monitoring AIGT on social media. arXiv preprint arXiv:2412.18148

    2024 arXiv

  32. [41]

    Ruixiang Tang, Yu-Neng Chuang, and Xia Hu. 2024. https://doi.org/10.1145/3624725 The science of detecting LLM -generated texts . Communications of the ACM, 67:50--–59

    2024 doi

  33. [42]

    Cagri Toraman, Oguzhan Ozcelik, Furkan Sahinuc, and Fazli Can. 2024. https://aclanthology.org/2024.lrec-main.986 M i D e22: A n annotated multi-event tweet dataset for misinformation detection . In Proceedings of the 2024 Joint International Conference on Computational Linguis...

    2024

  34. [43]

    Brian Tufts, Xuandong Zhao, and Lei Li. 2024. https://arxiv.org/abs/2412.05139 A practical examination of AI -generated text detectors for large language models . arXiv preprint arXiv:2412.05139

    2024 arXiv

  35. [44]

    Adaku Uchendu, Thai Le, and Dongwon Lee. 2023. https://doi.org/10.1145/3606274.3606276 Attribution and obfuscation of neural text authorship: A data mining perspective . SIGKDD Explor. Newsl., 25(1):1–18

    2023

  36. [45]

    Vivek Verma, Eve Fleisig, Nicholas Tomlin, and Dan Klein. 2024. https://doi.org/10.18653/v1/2024.naacl-long.95 Ghostbuster: D etecting text ghostwritten by large language models . In Proceedings of the 2024 Conference of the North American Chapter of the Association for Comput...

    2024 doi

  37. [46]

    Pengyu Wang, Linyang Li, Ke Ren, Botian Jiang, Dong Zhang, and Xipeng Qiu. 2023. https://aclanthology.org/2023.emnlp-main.73/ SeqXGPT : S entence-level AI -generated text detection . In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pag...

    2023

  38. [47]

    Yichen Wang, Shangbin Feng, Abe Hou, Xiao Pu, Chao Shen, Xiaoming Liu, Yulia Tsvetkov, and Tianxing He. 2024. https://doi.org/10.18653/v1/2024.acl-long.160 Stumbling blocks: Stress testing the robustness of machine-generated text detectors under attacks . In Proceedings of the...

    2024 doi

  39. [48]

    Junchao Wu, Shu Yang, Runzhe Zhan, Yulin Yuan, Lidia Sam Chao, and Derek Fai Wong. 2025. https://direct.mit.edu/coli/article/doi/10.1162/coli_a_00549/127462/A-Survey-on-LLM-Generated-Text-Detection-Necessity A survey on LLM -generated text detection: Necessity, methods, and fu...

    2025 doi

  40. [49]

    Wong, Shu Yang, Xinyi Yang, Yulin Yuan, and Lidia S

    Junchao Wu, Runzhe Zhan, Derek F. Wong, Shu Yang, Xinyi Yang, Yulin Yuan, and Lidia S. Chao. 2024. https://openreview.net/forum?id=ZGMkOikEyv Detect RL : Benchmarking LLM -generated text detection in real-world scenarios . In Proceedings of the Thirty-eight Conference on Neura...

    2024

  41. [50]

    Xianjun Yang, Liangming Pan, Xuandong Zhao, Haifeng Chen, Linda Petzold, William Yang Wang, and Wei Cheng. 2023. https://arxiv.org/abs/2310.15654 A survey on detection of LLM s-generated content . arXiv preprint arXiv:2310.15654

    2023 arXiv

  42. [51]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  43. [52]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

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