Pith. sign in

REVIEW 2 major objections 4 minor 2 cited by

GenAI Content Detection Task 3: Cross-Domain Machine-Generated Text Detection Challenge

T0 review · 2 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Trained detectors can flag machine-written text from a fixed set of LLMs and domains at over 99% true positive rate while keeping false positives at 5%.

desk verdict A useful shared-task report whose 'robust detection' headline is undercut by formatting confounds the authors themselves document. read the letter →

arxiv 2501.08913 v1 pith:WYZEK4RZ submitted 2025-01-15 cs.CL cs.LG

Liam Dugan , Andrew Zhu , Firoj Alam , Preslav Nakov , Marianna Apidianaki , Chris Callison-Burch This is my paper · ORCID
classification cs.CLcs.LG
keywords machine-generatedtextdetectionsharedtaskRAIDbenchmarkcross-domainadversarialrobustnessTPRatFPRLLMclassification
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 reports a three-month shared task in which 9 teams submitted 23 detectors for machine-generated text, all trained and evaluated on the RAID benchmark's fixed set of 11 LLMs, 8 domains, 4 decoding strategies, and 11 adversarial attacks. Its central claim is that a single detector can reach over 99% true positive rate at a 5% false positive rate on ordinary generated text, and 97.7% on text altered by any of 11 attacks, when every generator and domain is known in advance. The result matters because real-world detectors typically face a known, relatively small set of models and domains, so the task asks whether robustness within that finite universe is achievable rather than whether detectors generalize to unseen generators. The paper's answer is yes, with the explicit caveat that this in-distribution success does not extend to models or domains never seen during training.

What carries the argument

The load-bearing object is the RAID benchmark, a dataset of roughly 10 million documents built by sampling about 2000 human-written documents from each of 8 domains, generating a machine version of each from 11 LLMs under 4 decoding strategies, and then applying 11 adversarial attacks. Its test set holds out 200 human documents per domain plus all generations based on them, with every model and domain seen during training. The evaluation metric is TPR@FPR=5%, a domain-adjusted true positive rate measured at a fixed 5% false positive rate, found by a threshold search over each detector's scores; this metric makes near-99% numbers comparable across systems. The argument runs through the gap between the best submissions (99.4% and 99.3%) and the provided baselines (Binoculars at 79.0%): the shared-task systems, using preprocessing and hard-example mining, close most of that gap.

What would settle it

Run the same shared-task setup on a version of RAID where all formatting and layout cues are stripped: unify whitespace and newlines, convert numbered lists to prose, remove punctuation artifacts, and standardize case. If TPR@FPR=5% for the top detectors falls substantially toward or below the 89.67% level of the paper's cleaning experiment, that would show the 99% scores were carried by superficial confounds rather than by detection of machine-generated content. A complementary test is to add human-written texts that mimic LLM formatting conventions, such as numbered steps and no newlines, and observe whether false positives rise.

Watch

Extended reading notes

Core claim

On the RAID benchmark, where every LLM, domain, decoding strategy, and adversarial attack is known at training time and only specific articles are held out, a single trained classifier can separate machine-generated from human-written text at near-ceiling accuracy. The winning subtask A system reaches 99.4% TPR at 5% FPR averaged over all eight domains, the runner-up 99.3%, and under the 11 adversarial attacks the two top systems both reach 97.7%. The paper takes this as evidence that detectors can be built for a large yet fixed set of generators and domains, and that many attacks (whitespace insertion, article deletion, zero-width spaces) become easy once the defender knows the attack code, while paraphrase, synonym swap, and homoglyph attacks remain the hardest. It also reports that the most common successful ingredients were aggressive text normalization or preprocessing and hard-example mining during training.

Load-bearing premise

The result rests on the assumption that the near-perfect scores reflect detection of machine-authored content itself, not accidental surface differences such as numbered recipe steps or line breaks appearing only in one class; the paper's own cleaning experiment drops a RoBERTa detector only from 92.67% to 89.67% TPR, so this premise is not fully confirmed.

Editorial extensions

If this is right

  • For a fixed, known set of LLMs and domains, a single detector can exceed 99% TPR at 5% FPR, so deployment in closed settings such as a known API or a curated corpus is feasible at near-perfect accuracy.
  • Once a defender knows the attack code, most simple adversarial attacks (whitespace, zero-width, case swapping, article deletion) can be neutralized with text normalization, but paraphrase, synonym swap, and homoglyph attacks remain the hard cases.
  • The top shared-task systems beat the strongest evaluated baseline, Binoculars (79.0%), by about 20 points, showing that task-specific training on RAID data is what drives the gain.
  • High in-distribution accuracy does not carry over to unseen models or domains; the paper explicitly warns that detectors still generalize poorly out of distribution.
  • Future shared tasks should standardize a preprocessing script to strip formatting confounds, include more prompt and paraphrase variation, and supply more human-written text to balance the dataset.

Reading between the lines

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

  • The paper's own Table 6 shows that all generated recipes are numbered while human recipes are prose, and that human abstracts contain newlines that generated abstracts lack; a fair inference is that a meaningful share of the 99% reflects such formatting artifacts, and a cleaned re-run would likely lower the ceiling.
  • If formatting cues are the dominant signal, then detectors trained on cleaned data would be more vulnerable to 'humanized' LLM outputs, so real-world robustness of these systems could be markedly lower than the headline figure suggests.
  • Given near-ceiling accuracy on a fixed generator set, a single trained classifier may be competitive with per-model watermarking in closed deployments, since it achieves the same discrimination without modifying the generation process.
  • A natural extension of this evaluation protocol is to measure how accuracy degrades as the set of known domains and models grows or becomes partially open, which would map the boundary between in-distribution and out-of-distribution detection.
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. This paper reports the organization and outcomes of GenAI Content Detection Task 3, a shared task on cross-domain machine-generated text detection using the RAID benchmark. The task covered 8 domains, 11 generator models, 4 decoding strategies, and 11 adversarial attacks, with all models and domains seen during training. Over three months, 9 teams submitted 23 detectors. The winning systems achieve over 99% TPR at a 5% FPR on non-adversarial data and around 97.7% on adversarially attacked data. The paper interprets this as evidence that detectors can robustly identify machine-generated text from a fixed, known set of domains and models, and discusses trends in preprocessing, hard-example mining, and approach diversity, while also listing limitations including potential test-data leakage and formatting confounds.

Significance. If the headline result survives the confound concerns, the paper makes a useful empirical contribution: it shows that, under a fixed known generator set, fine-tuned classifiers can separate human and machine text at very low false positive rates, and it documents which adversarial attacks remain difficult even with access to attack code. The shared task is reasonably well structured, the leaderboard is public, and the inclusion of multiple independent participant systems strengthens the empirical grounding. The paper is also honest in acknowledging several limitations, particularly the formatting confounds in Section 8 and the oracle-style threshold search in Section 4.2. However, the central claim that detectors 'robustly detect' machine-generated text is not yet fully separated from the documented surface-form artifacts, so the significance of the result for real deployment remains conditional.

major comments (2)
  1. [Section 6.1 and Limitations] The paper's central claim that multiple detectors 'robustly detect text from many domains and models' is not yet cleanly separated from the formatting confounds documented in Table 6 and the Limitations section. The cleaning experiment with RoBERTa-base (92.67 to 89.67 TPR) only evaluates a single baseline detector, not the winning submissions that achieve >99%, and the paper itself states that there is not enough evidence to conclude whether the confounds drive the high performance. Since the headline result depends on the detectors exploiting content rather than surface formatting, this is a load-bearing gap; the authors should either evaluate the top systems on the cleaned data or substantially weaken the abstract's conclusion to reflect the unresolved confound.
  2. [Section 4.2] The official metric searches for the decision threshold using human-written texts from the test set itself, per domain and detector. This makes the reported TPR@FPR=5% an oracle-threshold number, not a rate that a participant could obtain without access to test-set human examples. This is particularly concerning because the paper simultaneously documents systematic formatting differences between human and generated text in the test set (Table 6): a threshold tuned on test human texts can exploit any such cue, including trivial ones. The authors should report an additional result with thresholds fixed on a held-out validation split, or explicitly discuss the oracle nature of the metric and its consequences for the robustness interpretation.
minor comments (4)
  1. [Abstract] The abstract and Introduction refer to 'accuracies of over 99% on machine-generated text,' but the reported metric is TPR at a fixed 5% FPR, not overall accuracy. I recommend using 'true positive rate' throughout for precision.
  2. [Limitations] The Limitations section states that 'all human-written recipes were written without numbered lists of steps whereas all generated recipes included numbered lists of steps,' but Table 6 only shows two examples. If this is a corpus-wide property, a quantitative verification (e.g., numbering frequency across the full dataset) should be reported; otherwise the wording should be softened to 'the inspected examples show.'
  3. [Section 6.1] The discussion of domain difficulty would benefit from reporting the per-domain human sample sizes in the test set (200 per domain in Table 3) and noting that the Poetry domain's lower average TPR (58.1%) comes with high variability across systems, which is visible in Table 4 but not discussed.
  4. [Section 5] Team USTC-BUPT is described without a system-paper citation, unlike the other teams; adding a citation or explicitly stating that no description paper was submitted would improve consistency.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper reports an empirical shared-task evaluation with independent participant systems on a public benchmark; the acknowledged confounds are a validity concern, not a circular reduction.

full rationale

The paper's central claim (TPR@FPR=5% >99% on RAID) is an empirical measurement of independent participant systems, not a quantity derived from the authors' own assumptions. RAID (Dugan et al., 2024) is cited as the benchmark; it is a public, code-reproduced dataset with a withheld test split, and 9 outside teams trained their own detectors, so the measurement is externally grounded. No fitted parameter is renamed a prediction: the threshold search (Section 4.2) enforces FPR=5% on human texts and then reports TPR on machine texts, which is not forced by construction. The passage most relevant to circularity is the Limitations section, where the authors admit 'Artifacts such as these may trivialize the detection task' and report that cleaning the confounds only drops a RoBERTa-base baseline from 92.67 to 89.67. This is an acknowledged validity threat to generalization, not a circular step: the paper explicitly does not claim to have controlled for confounds, and the cleaning ablation evaluates a baseline rather than the winning detectors. Because every load-bearing step is an externally measured result rather than a definitional identity or a self-citation chain used to forbid alternatives, there is no significant circularity. Score 0.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new theoretical entities, particles, forces, or conserved quantities. The only hand-chosen parameter affecting the reported metric is the threshold-search tolerance, which is a standard calibration detail rather than a fitted parameter in a derivation.

free parameters (1)
  • Threshold-search tolerance epsilon = 0.0005; 50 max iterations
    Used in Section 4.2 to stop threshold search when FPR is within 0.0005 of the target. Chosen by the authors; small effect on reported TPR values.
assumptions (4)
  • domain assumption The RAID test set accurately represents the fixed distribution of domains and models used in the task.
    The paper's conclusion is about RAID; if the test set were unbalanced or leaky, the TPR numbers would not support the claim. The paper discusses leakage risk in Limitations.
  • domain assumption Participants did not have access to the hidden test set.
    Section 1 and Section 3 state the test set is not released; the shared task protocol assumes this holds.
  • domain assumption Selecting the operating threshold on the test set human data yields a valid measure of detection performance.
    Section 4.2 describes threshold search to achieve exactly 5% FPR on test humans; this makes the metric well-defined but leaks label information into the evaluation.
  • domain assumption The manually cleaned dataset in the Limitations section removes the main confounds, and the remaining drop in accuracy is small and not fully explained.
    The cleaning experiment is preliminary; the paper states it lacks enough evidence to conclude whether artifacts are the source of high performance.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GenAI Content Detection Task 3: Cross-Domain Machine-Generated Text Detection Challenge." pith.science (2026). https://pith.science/paper/WYZEK4RZ

@misc{pith2026250108913,
  author       = {Pith},
  title        = {Pith review of: GenAI Content Detection Task 3: Cross-Domain Machine-Generated Text Detection Challenge},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WYZEK4RZ}},
  note         = {Machine review of arXiv:2501.08913}
}
read the original abstract

Recently there have been many shared tasks targeting the detection of generated text from Large Language Models (LLMs). However, these shared tasks tend to focus either on cases where text is limited to one particular domain or cases where text can be from many domains, some of which may not be seen during test time. In this shared task, using the newly released RAID benchmark, we aim to answer whether or not models can detect generated text from a large, yet fixed, number of domains and LLMs, all of which are seen during training. Over the course of three months, our task was attempted by 9 teams with 23 detector submissions. We find that multiple participants were able to obtain accuracies of over 99% on machine-generated text from RAID while maintaining a 5% False Positive Rate -- suggesting that detectors are able to robustly detect text from many domains and models simultaneously. We discuss potential interpretations of this result and provide directions for future research.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Stress-testing Machine Generated Text Detection: Shifting Language Models Writing Style to Fool Detectors

    cs.CL 2025-05 conditional novelty 6.0 of 10

    Fine-tuning LLMs with DPO to push generated news and abstracts toward human style substantially reduces the F1 scores of state-of-the-art machine-generated text detectors.

  2. Characterizing Human-Likeness in AI Generated Poetry: A Zero-shot Classification Study

    cs.CL 2026-07 conditional novelty 5.0 of 10

    Zero-shot Gemma 4 classification separates AI from human poems at 90% weighted F1, while human evaluators reach only 45% accuracy, and the study catalogs the linguistic attributes behind correct and incorrect detections.

Reference graph

Works this paper leans on

49 extracted references · 24 canonical work pages · cited by 2 Pith papers

  1. [1]

    Shifali Agrahari, Prabhat Mishra, and Sujit Kumar. 2025. Team Random at GenAI Detection Task 3: A Hybrid Approach to Cross-Domain Detection of Machine-Generated Text with Adversarial Attack Mitigation . In Proceedings of the 1st Workshop on GenAI Content Detection (GenAIDetect), Abu Dhabi, UAE

    work page 2025

  2. [2]

    Mazal Bethany, Athanasios Galiopoulos, Emet Bethany, Mohammad Bahrami Karkevandi, Nishant Vishwamitra, and Peyman Najafirad. 2024. https://arxiv.org/abs/2401.09727 Large Language Model Lateral Spear Phishing: A Comparative Study in Large-Scale Organizational Settings . Preprint, arXiv:2401.09727

    arXiv 2024

  3. [3]

    Janek Bevendorff, Matti Wiegmann, Jussi Karlgren, Luise D \"u rlich, Evangelia Gogoulou, Aarne Talman, Efstathios Stamatatos, Martin Potthast, and Benno Stein. 2024. Overview of the ``Voight-Kampff'' Generative AI Authorship Verification Task at PAN and ELOQUENT 2024 . In Working Notes of CLEF 2024 - Conference and Labs of the Evaluation Forum, CEUR Works...

    work page 2024

  4. [4]

    Meghana Moorthy Bhat and Srinivasan Parthasarathy. 2020. https://doi.org/10.18653/v1/2020.insights-1.7 How Effectively Can Machines Defend Against Machine-Generated Fake News? An Empirical Study . In Proceedings of the First Workshop on Insights from Negative Results in NLP, pages 48--53, Online. Association for Computational Linguistics

    doi:10.18653/v1/2020.insights-1.7 2020

  5. [5]

    Shuyang Cai and Wanyun Cui. 2023. https://arxiv.org/abs/2307.02599 Evade chatgpt detectors via a single space . Preprint, arXiv:2307.02599

    arXiv 2023

  6. [6]

    Shammur Absar Chowdhury, Hind Al-Merekhi, Mucahid Kutlu, Kaan Efe Keleş, Fatema Ahmad, Tasnim Mohiuddin, Georgios Mikros, and Firoj Alam. 2025. GenAI content detection task 2: AI vs. human – academic essay authenticity challenge. In Proceedings of the 1st Workshop on GenAI Content Detection (GenAIDetect), Abu Dhabi, UAE. International Conference on Comput...

    work page 2025

  7. [7]

    Matthieu Dubois, François Yvon, and Pablo Piantanida. 2025. MOSAIC at GenAI Content Detection Task 3: Cross-Domain Machine Generated Text Detection . In Proceedings of the 1st Workshop on GenAI Content Detection (GenAIDetect), Abu Dhabi, UAE

    work page 2025

  8. [8]

    Liam Dugan, Alyssa Hwang, Filip Trhl \' k, Andrew Zhu, Josh Magnus Ludan, Hainiu Xu, Daphne Ippolito, and Chris Callison-Burch. 2024. https://doi.org/10.18653/v1/2024.acl-long.674 RAID : A shared benchmark for robust evaluation of machine-generated text detectors . In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics ...

    doi:10.18653/v1/2024.acl-long.674 2024

Show all 49 references
  1. [9]

    Liam Dugan, Daphne Ippolito, Arun Kirubarajan, and Chris Callison-Burch. 2020. https://doi.org/10.18653/v1/2020.emnlp-demos.25 R o FT : A Tool for Evaluating Human Detection of Machine-Generated Text . In Proceedings of the 2020 Conference on Empirical Methods in Natural Langu...

    2020 doi

  2. [10]

    Liam Dugan, Daphne Ippolito, Arun Kirubarajan, Sherry Shi, and Chris Callison-Burch. 2023. https://doi.org/10.1609/aaai.v37i11.26501 Real or Fake Text? Investigating Human Ability to Detect Boundaries between Human-Written and Machine-Generated text . In Proceedings of the Thi...

    2023 doi

  3. [11]

    Abishek R Edikala, Gregorios A Katsios, Noelie V Creaghe, and Ning Yu. 2025. Leidos at GenAI Content Detection Task 3: A Weight-Balanced Transformer Approach for AI Generated Text Detection Across Domains . In Proceedings of the 1st Workshop on GenAI Content Detection (GenAIDe...

    2025

  4. [12]

    Bradley Emi, Elyas Masrour, and Max Spero. 2025. Pangram at GenAI Content Detection Task 3: An Active Learning Approach to Machine Generated Text Generation . In Proceedings of the 1st Workshop on GenAI Content Detection (GenAIDetect), Abu Dhabi, UAE

    2025

  5. [13]

    Pieter Fivez, Walter Daelemans, Tim Van de Cruys, Yury Kashnitsky, Savvas Chamezopoulos, Hadi Mohammadi, Anastasia Giachanou, Ayoub Bagheri, Wessel Poelman, Juraj Vladika, Esther Ploeger, Johannes Bjerva, Florian Matthes, and Hans van Halteren. 2024. https://clinjournal.org/cl...

    2024

  6. [14]

    Rinaldo Gagiano, Maria Myung-Hee Kim, Xiuzhen Zhang, and Jennifer Biggs. 2021. https://aclanthology.org/2021.alta-1.12 Robustness Analysis of Grover for Machine-Generated News Detection . In Proceedings of the 19th Annual Workshop of the Australasian Language Technology Associ...

    2021

  7. [15]

    Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, Fran c ois Laviolette, Mario March, and Victor Lempitsky. 2016. http://jmlr.org/papers/v17/15-239.html Domain-adversarial training of neural networks . Journal of Machine Learning Research, 17(59):1--35

    2016

  8. [16]

    Ji Gao, Jack Lanchantin, Mary Lou Soffa, and Yanjun Qi. 2018. https://arxiv.org/abs/1801.04354 Black-box Generation of Adversarial Text Sequences to Evade Deep Learning Classifiers . Preprint, arXiv:1801.04354

    2018 arXiv

  9. [17]

    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

  10. [18]

    German Gritsai, Anastasia Voznyuk, Andrey Grabovoy, and Yury Chekhovich. 2025. https://arxiv.org/abs/2410.14677 Are ai detectors good enough? a survey on quality of datasets with machine-generated texts . Preprint, arXiv:2410.14677

    2025 arXiv

  11. [19]

    Jesus Guerrero, Gongbo Liang, and Izzat Alsmadi. 2022. https://arxiv.org/abs/2212.11808 A Mutation-based Text Generation for Adversarial Machine Learning Applications . Preprint, arXiv:2212.11808

    2022 arXiv

  12. [20]

    Abhimanyu Hans, Avi Schwarzschild, Valeriia Cherepanova, Hamid Kazemi, Aniruddha Saha, Micah Goldblum, Jonas Geiping, and Tom Goldstein. 2024. https://arxiv.org/abs/2401.12070 Spotting LLMs With Binoculars: Zero-Shot Detection of Machine-Generated Text . Preprint, arXiv:2401.12070

    2024 arXiv

  13. [21]

    Xiaomeng Hu, Pin-Yu Chen, and Tsung-Yi Ho. 2023. RADAR: Robust AI-Text Detection via Adversarial Learning . In Proceedings of the 37th International Conference on Neural Information Processing Systems, NeurIPS 2023, Red Hook, NY, USA. Curran Associates Inc

    2023

  14. [22]

    Ram Mohan Rao Kadiyala. 2024. https://doi.org/10.18653/v1/2024.semeval-1.78 RK adiyala at S em E val-2024 Task 8: Black-Box Word-Level Text Boundary Detection in Partially Machine Generated Texts . In Proceedings of the 18th International Workshop on Semantic Evaluation (SemEv...

    2024 doi

  15. [23]

    Ram Mohan Rao Kadiyala, Siddartha Pullakhandam, Kanwal Mehreen, Ashay Srivastava, Subhasya TippaReddy, Arvind Reddy Bobbili, Drishti Sharma, Suraj Chandrashekhar, Modabbir Adeeb, and Srinadh Vura. 2024. https://doi.org/10.57967/hf/3194 mmgtd-corpus (v1)

    2024 doi

  16. [24]

    Hemanth Kandula, Chak Fai Li, Haoling Qiu, Damianos Karakos, Hieu Man Duc Trong, Thien Huu Nguyen, and Brian Ulicny. 2025. BBN-U.Oregon's ALERT system at GenAI Content Detection Task 3: Robust Authorship Style Representations for Cross-Domain Machine-Generated Text Detection ....

    2025

  17. [25]

    Yury Kashnitsky, Drahomira Herrmannova, Anita de Waard, George Tsatsaronis, Catriona Catriona Fennell, and Cyril Labbe. 2022. https://aclanthology.org/2022.sdp-1.26 Overview of the DAGP ap22 shared task on detecting automatically generated scientific papers . In Proceedings of...

    2022

  18. [26]

    Junseong Kim, Seolhwa Lee, Jihoon Kwon, Sangmo Gu, Yejin Kim, Minkyung Cho, Jy yong Sohn, and Chanyeol Choi. 2024. https://getlinq.com/blog/linq-embed-mistral/ Linq-embed-mistral:elevating text retrieval with improved gpt data through task-specific control and quality refineme...

    2024

  19. [27]

    Jules King, Perpetual Baffour, Scott Crossley, Ryan Holbrook, and Maggie Demkin. 2023. Llm - detect ai generated text. https://kaggle.com/competitions/llm-detect-ai-generated-text. Kaggle

    2023

  20. [28]

    Ryuto Koike, Masahiro Kaneko, and Naoaki Okazaki. 2024. https://aclanthology.org/2024.findings-emnlp.841 How you prompt matters! E ven task-oriented constraints in instructions affect LLM -generated text detection . In Findings of the Association for Computational Linguistics:...

    2024

  21. [29]

    Kalpesh Krishna, Yixiao Song, Marzena Karpinska, John Wieting, and Mohit Iyyer. 2023. https://arxiv.org/abs/2303.13408 Paraphrasing evades detectors of ai-generated text, but retrieval is an effective defense . Preprint, arXiv:2303.13408

    2023 arXiv

  22. [30]

    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

  23. [31]

    Gongbo Liang, Jesus Guerrero, and Izzat Alsmadi. 2023 a . https://arxiv.org/abs/2302.05794 Mutation-based adversarial attacks on neural text detectors . Preprint, arXiv:2302.05794

    2023 arXiv

  24. [32]

    Weixin Liang, Mert Yuksekgonul, Yining Mao, Eric Wu, and James Zou. 2023 b . GPT detectors are biased against non-native english writers. Patterns, 4(7):100779

    2023

  25. [33]

    Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. 2017. https://doi.org/10.1109/ICCV.2017.324 Focal Loss for Dense Object Detection . In 2017 IEEE International Conference on Computer Vision (ICCV), pages 2999--3007, Los Alamitos, CA, USA. IEEE Computer Society

    2017 doi

  26. [34]

    Brady D Lund, Ting Wang, Nishith Reddy Mannuru, Bing Nie, Somipam Shimray, and Ziang Wang. 2023. ChatGPT and a new academic reality: Artificial Intelligence-written research papers and the ethics of the large language models in scholarly publishing . Journal of the Association...

    2023

  27. [35]

    Md Kamrujjaman Mobin and Md Saiful Islam. 2025. LuxVeri at GenAI Content Detection Task 3: Cross-Domain Detection of AI-Generated Text Using Inverse Perplexity-Weighted Ensemble of Fine-Tuned Transformer Models . In Proceedings of the 1st Workshop on GenAI Content Detection (G...

    2025

  28. [36]

    Jiameng Pu, Zain Sarwar, Sifat Muhammad Abdullah, Abdullah Rehman, Yoonjin Kim, Parantapa Bhattacharya, Mobin Javed, and Bimal Viswanath. 2023. https://doi.org/10.1109/SP46215.2023.10179387 Deepfake Text Detection: Limitations and Opportunities . In 2023 IEEE Symposium on Secu...

    2023

  29. [37]

    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? Preprint, arXiv:2303.11156

    2023 arXiv

  30. [38]

    Areg Mikael Sarvazyan, José Ángel González, Marc Franco-Salvador, Francisco Rangel, Berta Chulvi, and Paolo Rosso. 2023. https://arxiv.org/abs/2309.11285 Overview of autextification at iberlef 2023: Detection and attribution of machine-generated text in multiple domains . Prep...

    2023 arXiv

  31. [39]

    Tatiana Shamardina, Vladislav Mikhailov, Daniil Chernianskii, Alena Fenogenova, Marat Saidov, Anastasiya Valeeva, Tatiana Shavrina, Ivan Smurov, Elena Tutubalina, and Ekaterina Artemova. 2022. https://doi.org/10.28995/2075-7182-2022-21-497-511 Findings of the RuATD Shared Task...

    2022 doi

  32. [40]

    Filipo Sharevski, Jennifer Vander Loop, Peter Jachim, Amy Devine, and Emma Pieroni. 2023. Talking Abortion (Mis) information with ChatGPT on TikTok . In 2023 IEEE European Symposium on Security and Privacy Workshops (EuroS&PW), pages 594--608. IEEE

    2023

  33. [41]

    Irene Solaiman, Miles Brundage, Jack Clark, Amanda Askell, Ariel Herbert-Voss, Jeff Wu, Alec Radford, Gretchen Krueger, Jong Wook Kim, Sarah Kreps, Miles McCain, Alex Newhouse, Jason Blazakis, Kris McGuffie, and Jasmine Wang. 2019. https://doi.org/10.48550/ARXIV.1908.09203 Rel...

    2019 doi arXiv

  34. [42]

    Giovanni Spitale, Nikola Biller-Andorno, and Federico Germani. 2023. https://doi.org/10.1126/sciadv.adh1850 AI model GPT-3 (dis)informs us better than humans . Science Advances, 9(26):eadh1850

    2023 doi

  35. [43]

    L D M S Sai Teja, Annepaka Yadagiri, M Srikar Vardhan, and Partha Pakray. 2025. CNLP-NITS-PP at Task 3: Cross-Domain Machine-Generated Text Detection Using DistilBERT Techniques . In Proceedings of the 1st Workshop on GenAI Content Detection (GenAIDetect), Abu Dhabi, UAE

    2025

  36. [44]

    Yuxia Wang, Jonibek Mansurov, Petar Ivanov, Jinyan Su, Artem Shelmanov, Akim Tsvigun, Osama Mohammed Afzal, Tarek Mahmoud, Giovanni Puccetti, and Thomas Arnold. 2024. https://doi.org/10.18653/v1/2024.semeval-1.279 S em E val-2024 task 8: Multidomain, multimodel and multilingua...

    2024 doi

  37. [45]

    Yuxia Wang, Artem Shelmanov, Jonibek Mansurov, Akim Tsvigun, Vladislav Mikhailov, Rui Xing, Zhuohan Xie, Jiahui Geng, Giovanni Puccetti, Ekaterina Artemova, Jinyan Su, Minh Ngoc Ta, Mervat Abassy, Kareem Elozeiri, Saad El Dine Ahmed, Maiya Goloburda, Tarek Mahmoud, Raj Vardhan...

    2025

  38. [46]

    Max Weiss. 2019. https://techscience.org/a/2019121801/ Deepfake Bot Submissions to Federal Public Comment Websites Cannot Be Distinguished from Human Submissions. Technology Science, 2019121801

    2019

  39. [47]

    Max Wolff. 2020. https://arxiv.org/abs/2002.11768 Attacking Neural Text Detectors . Preprint, arXiv:2002.11768

    2020 arXiv

  40. [48]

    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...

  41. [49]

    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 10, 2026 · model on record in the stance chip above.