Pith. sign in

REVIEW 5 major objections 5 minor 27 references

DeBERTa-Sentinel: Toward Transparent and Trustworthy Detection of AI-Generated Text

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read DeBERTa-Sentinel reports AI-text detection at 97.5% accuracy

desk verdict A useful DeBERTa fine-tuning experiment for AI-text detection, but the UID-level split leak and inconsistent numbers mean the headline results should not be trusted until the authors re-split by UID, run multiple seeds, and fix the arithmetic. read the letter →

arxiv 2608.01046 v1 pith:WQIOC2K4 submitted 2026-08-02 cs.CL cs.AI

Muhammad Yousaf Rehman , Muhammad Islam This is my paper
classification cs.CLcs.AI
keywords AI-generatedtextdetectiondisentangledattentionDeBERTa-v3transformerclassifiersparaphrasedexplainabilitycross-generatorgeneralization
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

DeBERTa-Sentinel is a claim that AI-generated text detection improves when the classifier can separate what a token means from where it appears. The paper builds on GPT-Sentinel, replacing its frozen RoBERTa encoder with a fully fine-tuned DeBERTa-v3-small, and trains on GLC-AIText, a new dataset of 28,057 paraphrases from GPT-3.5, LLaMA, and Claude paired with 29,142 human OpenWebText-Final samples. It reports 97.53% test accuracy, 99.33% recall, 99.53% ROC-AUC, and a 0.67% false-negative rate, beating the RoBERTa-Sentinel baseline's 95.3% accuracy. The paper's distinctive promise is transparency: the model emits token-level attributions, so detection decisions come with evidence a human can audit. If the numbers and the architecture hold up, a small task-tuned model with disentangled attention can detect synthetic text more reliably across generators, while giving educators and journalists a basis to challenge individual verdicts.

What carries the argument

The mechanism that carries the argument is DeBERTa's disentangled self-attention, which splits the attention computation into content-to-content, content-to-position, and position-to-content terms using relative position offsets $\delta(i,j) = \mathrm{clip}(j-i, -k, k)$. This is what the paper says lets the model pick up structural regularities in synthetic text instead of conflating them with lexical content. The supporting machinery is the GLC-AIText dataset itself: paired human/AI paraphrases from three generator families, which the authors use both to train and to stage a held-out-generator test of generalization.

What would settle it

Run the identical training and evaluation but split the GLC-AIText pairs by UID, ensuring that no human original and its AI paraphrase fall on opposite sides of the train/test boundary; if accuracy drops materially below 97.5%, the reported separation is inflated by content overlap. Separately, re-run the Claude held-out test with the Claude paraphrases' paired human originals removed from training, and check whether 100% recall on Claude persists.

Watch

Extended reading notes

Core claim

The paper claims that DeBERTa-Sentinel—a DeBERTa-v3-small sequence classifier fine-tuned end to end—detects AI-generated text by exploiting DeBERTa's disentangled attention, which computes separate content-to-content ($A^{c2c}$), content-to-position ($A^{c2p}$), and position-to-content ($A^{p2c}$) attention terms. This decomposition, the authors argue, lets the model capture structural markers of machine text—formulaic transitions, templated sentence openings, systematic hedging—that a fused content+position attention (RoBERTa) blurs. Trained on the GLC-AIText dataset (paraphrases from gpt-3.5-turbo, LLaMA, and Claude, each sharing a UID with a human OpenWebText-Final original), the model re

Load-bearing premise

The reported scores assume the random 60/20/20 split does not leak information between training and test through the paired human originals; because each AI paraphrase shares a UID with its human text, a sample-level split can put one member of a pair in training and the other in test, letting the model score by recognizing document content instead of synthetic-text cues.

Editorial extensions

If this is right

  • A fully fine-tuned DeBERTa-v3-small can replace frozen RoBERTa encoders in Sentinel-style detectors, improving accuracy by about 2.2 percentage points and cutting false negatives from 3.7% to 0.67%.
  • Training on paraphrases from several LLM families may make detectors ready for a new generator without retraining; the paper's held-out Claude test is evidence, not proof.
  • Token-level attribution gives human reviewers something concrete to check, which is a step toward using detectors as decision-support rather than as proof of AI authorship.
  • The zero-shot collapse to 50% accuracy shows the pretrained backbone alone is not sufficient; the gain comes from fine-tuning on diverse paraphrase data, so detector quality depends on data curation as much as architecture.

Reading between the lines

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

  • Because each AI paraphrase shares a UID with its human original and the split is only described at the sample level, the reported test accuracy is likely to overstate real-world performance; a UID-grouped split is the natural check and could be done with the same data.
  • The held-out Claude result should be read with the same caveat: if the human counterparts of the Claude test paragraphs appear in training, 100% recall may reflect memorization of content rather than generator-generalizable cues.
  • The paper's explainability analysis is presented as averaged attribution scores; without user studies it is not yet shown that these attributions help journalists or educators make better decisions, a gap the authors themselves flag as future work.
  • If the disentangled-attention advantage replicates, it suggests detector architecture—not just model size—is the deciding factor for forensic tasks, pointing toward small, transparent models rather than ever-larger black boxes.
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

5 major / 5 minor

Summary. The paper introduces DeBERTa-Sentinel, a DeBERTa-v3-small classifier fine-tuned for detecting AI-generated text. It uses a dataset (GLC-AIText) of paraphrases from GPT-3.5, LLaMA, and Claude paired with human OpenWebText passages, and claims a 60/20/20 split, 97.53% test accuracy, 99.33% recall, 99.53% ROC-AUC, and F1 0.976, outperforming a RoBERTa-Sentinel baseline. It also reports 98.46% accuracy on a held-out Claude test set, and provides token-level explainability examples. The central claim is that DeBERTa's disentangled attention plus multi-generator paraphrase data yields better, more transparent detection.

Significance. If the results hold, the paper would make a useful empirical contribution to AI-text detection: it combines a well-motivated architectural change (DeBERTa disentangled attention), multi-generator paraphrase training, a held-out generator evaluation, and an explicit transparency component. The authors report detailed hyperparameters, a fixed seed, and a public code/data link, which are strengths for reproducibility. However, several load-bearing issues prevent me from crediting the headline numbers as stated: dataset counts are internally inconsistent, the train/test split does not appear to be UID-grouped, and the held-out Claude experiment has an unexplained test size. The comparison with the RoBERTa-Sentinel baseline also lacks sufficient methodological detail. These concerns are fixable, but they are not purely presentational.

major comments (5)
  1. [§3.1, §4, Table 1] The dataset arithmetic is inconsistent. Section 4 states 28,057 AI-generated samples + 29,142 human samples = 58,537 total, but 28,057 + 29,142 = 57,199. The stated split sizes (35,121 + 11,708 + 11,708 = 58,537) imply the human count should be 30,480, not 29,142. This error propagates to class-balance assumptions and to every reported accuracy/F1 number. Please correct the counts and report the exact per-split class distributions.
  2. [§3.1, §4, §5.10] The split likely leaks content-level information. Section 3.1 says each AI paraphrase shares the same UID as its human original, but Section 4 describes only a 60/20/20 random split with seed 42, with no UID-level grouping. If the split is at the sample level, a human original can be in training while its AI paraphrase is in test (or vice versa). The model could then exploit topic/entity overlap rather than generalizable detection cues. This also affects the held-out Claude experiment: the paired human originals may be in the training set, potentially explaining the 100% recall. Please report UID-overlap statistics and, ideally, a UID-grouped split as a robustness check.
  3. [§5.10, Table 8] The held-out Claude test set size is unexplained and inconsistent with the dataset description. Table 1 lists 8,283 Claude-generated samples, and each is paired with a human original. A balanced paired test set would contain up to 16,566 samples, yet Table 8 reports 11,658 test samples. If a subset was selected, the selection procedure must be described. Otherwise the 98.46% accuracy and 100% recall are not reproducible.
  4. [§5.6, Table 3] The ablation claimed in Table 3 does not isolate the contribution of multi-LLM data diversity. The only comparison is zero-shot DeBERTa-v3 vs fine-tuned DeBERTa-Sentinel, which isolates fine-tuning, not data diversity. The held-out Claude experiment tests generalization but is not an ablation of the multi-LLM training set. To support the claim that multi-generator data is essential, a single-generator (e.g., GPT-3.5-only) fine-tuned model should be compared on the same test set.
  5. [§5.5.1, §5.5.2, Tables 4/7] The operating threshold and reported metrics are not mutually consistent. Section 5.5.1 reports an optimal threshold of 0.984 (Youden's J) with TPR 98.58% and FPR 2.28%, while Table 7 reports recall 99.33% and FPR 4.28%. If the default threshold is 0.5, this should be stated; if the metrics use a different threshold, the confusion matrix should be provided. Also, Table 7 lists F1 as 97.58% while Tables 3 and 4 list it as 0.976; the percentage values should be reconciled.
minor comments (5)
  1. [Abstract] The phrase "28,057 human and LLM-generated samples" is misleading; 28,057 is the number of AI-generated paraphrases. The human count is separate.
  2. [§5.7, Table 5] The RoBERTa-Sentinel baseline (credited to HULLMI [10]) is said to have been "evaluated on the GLC-AIText test set for this comparison," but no details are given about whether this is a fine-tuned checkpoint, the training split, or hyperparameters. Please document this baseline evaluation to make the comparison reproducible.
  3. [§5.8] The explainability analysis is described as "attention-based" and then interpreted as feature importance. Attention weights are not necessarily faithful explanations; please soften the causal language or add a grounding/faithfulness check.
  4. [§3.2] There is a spacing typo: "theOpenGPTTextdataset" should be "the OpenGPTText dataset."
  5. [References] Reference [10] is titled "HULLMI: Human vs LLM Identification with Explainability," yet the text calls this method "RoBERTa-Sentinel." Please clarify the relationship between HULLMI and RoBERTa-Sentinel, or rename the baseline consistently.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: reported metrics are measured on a held-out test set; the UID-split overlap issue is a validity concern, not a circular derivation.

full rationale

Walked the claimed derivation chain: data construction (Section 3), architecture and fine-tuning (Section 4), and evaluation (Section 5). No prediction or first-principles result is defined in terms of its own target, and no fitted parameter is renamed as a prediction. The central metrics (97.53% test accuracy, 99.53% ROC-AUC, 98.46% held-out Claude accuracy) are measured on a 20% test split after validation-based model selection; test labels are not used to fit anything. The RoBERTa-Sentinel baseline in Table 5 is attributed to HULLMI [10] and was evaluated on the same GLC-AIText test set; comparing on an identical test set is a standard experimental practice and not circular. The most serious limitation is that Section 3.1 states each AI paraphrase shares a UID with its human original, while Section 4 describes only a random 60/20/20 split with seed 42 and no UID-level grouping, so paired human/AI samples may straddle the train/test boundary. That is a data-leakage or experimental-design threat that could inflate accuracy through content memorization, but it is not a circular derivation: no equation, fitted parameter, or cited uniqueness result is equivalent by construction to the reported outcome. There is no load-bearing self-citation chain, as [10] is not authored by the present authors. Therefore no circular steps are identified; the appropriate circularity score is 0.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The paper invents no new physical or formal objects. The 'Sentinel' naming is branding of a standard classifier. The C2C/C2P/P2C interpretations ('formulaic sentence-opening transitions', 'templating behaviors') are rhetorical glosses placed on DeBERTa's standard attention components, not new entities with independent handles. The scientific content is a fine-tuning recipe plus a dataset artifact, resting on three domain assumptions (what counts as human text, what counts as AI text, and whether the split leaks paired content) and on the standard DeBERTa equations. The largest unexamined input is the pairing structure of the data, a design choice that can bias every reported number.

free parameters (2)
  • Operating threshold on P(AI) = 0.984 (Youden optimal) for one reported point; headline metrics use an unstated threshold
    Section 5.5.1 reports 98.58% TPR and 2.28% FPR at 0.984, while Table 7 reports 99.33% recall and 4.28% FPR; the threshold producing the abstract's headline numbers is not given.
  • Max token length = 256
    Chosen instead of the standard 512 for computational efficiency; asserted 'empirically validated' without a supporting experiment in Section 4.
assumptions (4)
  • domain assumption OpenWebText 2019 web content is valid ground-truth 'human writing'
    Section 3.2 justifies the negative class by corpus age (2019, pre-LLM) and Reddit sourcing. This defines the task, and makes the negative class informal web text rather than the academic or formal text the paper later flags as hard cases.
  • domain assumption The paraphrase prompt 'Rephrase the following paragraph by paragraph' plus fluency filtering yields representative 'AI-generated text' across the three models
    Section 3.3. The positive class is paraphrase-augmented web text, not typical LLM usage; the paper's own related work [15, 26] says paraphrase settings are a distinct, harder regime, so results on GLC-AIText are defined by this choice.
  • domain assumption A sample-level random split without UID grouping yields a contamination-free test set
    Section 4 states only a seed-42 three-way split; since Section 3.1 couples each AI sample to a human UID, the split's integrity is assumed, never shown. This is the paper's most fragile premise.
  • standard math DeBERTa disentangled attention equations (1)-(3) as specified in [12]
    Section 4.1.1 reproduces DeBERTa's three-way attention decomposition; treated as given from the cited paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DeBERTa-Sentinel: Toward Transparent and Trustworthy Detection of AI-Generated Text." pith.science (2026). https://pith.science/paper/WQIOC2K4

@misc{pith2026260801046,
  author       = {Pith},
  title        = {Pith review of: DeBERTa-Sentinel: Toward Transparent and Trustworthy Detection of AI-Generated Text},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WQIOC2K4}},
  note         = {Machine review of arXiv:2608.01046}
}
read the original abstract

The rapid spread of large language models (LLMs) across the web raises concerns about misinformation, academic integrity, automated content manipulation, and risks to vulnerable online communities. Existing transformer-based detectors, such as GPT-Sentinel, show promise but struggle to generalize to diverse model outputs and paraphrasing attacks, limiting their role in building trustworthy web ecosystems. This work introduces DeBERTa-Sentinel, a responsible AI-generated text detection framework leveraging DeBERTa-v3's disentangled attention to capture subtle structural irregularities in synthetic content. A central design principle is transparency: unlike black-box commercial detectors, DeBERTa-Sentinel exposes token-level explanations of its decisions, enabling affected stakeholders journalists, educators, and platform trust and safety teams to audit, challenge, and contextualize detection outcomes. Using the GLC-AIText dataset of 28,057 human and LLM-generated samples (GPT, LLaMA, and Claude) with a 60-20-20 split, DeBERTa-Sentinel achieves 98.21\% validation accuracy and surpasses the RoBERTa-Sentinel baseline from NeurIPS 2025, achieving 97.53\% test accuracy, 95.89\% precision, 99.33\% recall, and 99.53\% ROC-AUC, and maintaining a 0.665\% false negative rate. The model's interpretability reveals linguistic markers such as academic phrasing and formal transitions associated with synthetic text, directly supporting stakeholder needs for verifiable, auditable content-authenticity decisions. By advancing responsible detection methods that reduce bias and enhance explainability, DeBERTa-Sentinel promotes trustworthy, ethical, and human-centric AI systems. Code and data are available at https://github.com/Galileo-Galili/HUMAN-VS-AI-TEXT-DETECTION.

Figures

Figures reproduced from arXiv: 2608.01046 by the authors.

Figure 1
Figure 1. The diagram illustrates the DeBERTa-Sentinel architecture. The input sequence (e.g., “This is AI Generated...”) is embedded and processed through 12 layers of disentangled attention in Step 3 (our proposed approach). The final [CLS] token representation is used for classification via the internal feedforward layer. The DeBERTa-Sentinel architecture is fully end-to-end fine-tuned with gradients backpropagating throug… view at source ↗
Figure 2
Figure 2. DeBERTa’s disentangled self-attention architecture as applied in DeBERTa-Sentinel. Standard attention (top) combines content and position into a single query/key/value operation. DeBERTa decomposes this into three independent components: (1) Content-to-Content (C2C), capturing semantic co-occurrence patterns; (2) Content-to-Position (C2P), modelling how a token’s semantic role varies with structural location; and (3… view at source ↗
Figure 3
Figure 3. [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: ROC-AUC comparison across baseline models. DeBERTa-v3-small achieves 99.53% AUC (0.995), substantially outperforming traditional base￾lines including TF-IDF + LogReg (93.6% AUC) and random/majority class baselines (50.0% AUC). The results demonstrate the effectiveness …
Figure 6
Figure 6. Figure 6: presents the analysis of feature importance aggregated across multiple text samples, revealing the linguistic patterns that DeBERTa-Sentinel has learned to distinguish AI from human text. The model demonstrates sophisticated pattern recognition, identifying formal voca…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

27 extracted references · 20 canonical work pages

  1. [1]

    GPT-3.5,

    OpenAI, “GPT-3.5, ” 2023. [Online]. Available: https://openai.com

    work page 2023

  2. [2]

    LLaMA: Open and Efficient Foundation Language Models,

    H. Touvron et al., “LLaMA: Open and Efficient Foundation Language Models, ”arXiv:2302.13971, 2023

    arXiv 2023

  3. [3]

    Claude: A Large Language Model,

    Anthropic, “Claude: A Large Language Model, ” 2023. Available: https://www.anthropic.com

    work page 2023

  4. [4]

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

    J. Wu, S. Yang, R. Zhan, Y. Yuan, L. S. Chao, and D. F. Wong, “A Survey on LLM-Generated Text Detection: Necessity, Methods, and Future Directions, ”Computational Linguistics, pp. 1–65, 2025

    work page 2025

  5. [5]

    Unified Large Language Models for Misinformation Detection in Low-Resource Linguistic Settings

    M. Islam, “Unified Large Language Models for Misinformation Detection in Low-Resource Lin- guistic Settings, ”arXiv:2506.01587, 2025

    work page Pith review arXiv 2025

  6. [6]

    GLTR: Statistical Detection and Visualization of Generated Text,

    S. Gehrmann, H. Strobelt, and A. Rush, “GLTR: Statistical Detection and Visualization of Generated Text, ” inProc. ACL System Demonstrations, 2019, pp. 111–116

    work page 2019

  7. [7]

    Automatic Detection of Generated Text is Easiest when Humans are Fooled,

    D. Ippolito, D. Duckworth, C. Callison-Burch, and D. Eck, “Automatic Detection of Generated Text is Easiest when Humans are Fooled, ” inProc. ACL, 2020, pp. 1808–1822

    work page 2020

  8. [8]

    Release Strategies and the Social Impacts of Language Models,

    I. Solaiman et al., “Release Strategies and the Social Impacts of Language Models, ”arXiv:1908.09203, 2019

    arXiv 1908

Show all 27 references
  1. [9]

    GPT-Sentinel: Distinguishing Human and ChatGPT Generated Content,

    Y. Chen, H. Kang, V. Zhai, L. Li, R. Singh, and B. Raj, “GPT-Sentinel: Distinguishing Human and ChatGPT Generated Content, ”arXiv:2305.07969, 2023

    2023 arXiv

  2. [10]

    HULLMI: Human vs LLM Identification with Explainability,

    P. D. Joshi, S. Pocker, R. A. Dandekar, R. Dandekar, and S. Panat, “HULLMI: Human vs LLM Identification with Explainability, ” inAdvances in Neural Information Processing Systems (NeurIPS),

  3. [11]

    Multimodal generative AI for human motion under- standing and generation: A survey and way forward,

    M. Islam, T. Huang, E. Ahn, and U. Naseem, “Multimodal generative AI for human motion under- standing and generation: A survey and way forward, ”Information Fusion, 2026

    2026

  4. [12]

    DeBERTa: Decoding-Enhanced BERT with Disentangled Attention,

    P. He, X. Liu, J. Gao, and W. Chen, “DeBERTa: Decoding-Enhanced BERT with Disentangled Attention, ” inICLR, 2021

    2021

  5. [13]

    Sarang at DEFACTIFY 4.0: Detecting AI-Generated Text Using Noised Data and an Ensemble of DeBERTa Models,

    A. Trivedi and S. Sivanesan, “Sarang at DEFACTIFY 4.0: Detecting AI-Generated Text Using Noised Data and an Ensemble of DeBERTa Models, ” inDe-Factify 4.0 Workshop at AAAI, 2025

    2025

  6. [14]

    DetectGPT: Zero-Shot Machine- Generated Text Detection using Probability Curvature,

    E. Mitchell, Y. Lee, A. Khazatsky, C. D. Manning, and C. Finn, “DetectGPT: Zero-Shot Machine- Generated Text Detection using Probability Curvature, ” inICML, 2023, pp. 24950–24962

    2023

  7. [15]

    Paraphrasing Evades Detectors of AI-Generated Text, but Retrieval is an Effective Defense,

    K. Krishna, Y. Song, M. Karpinska, J. Wieting, and M. Iyyer, “Paraphrasing Evades Detectors of AI-Generated Text, but Retrieval is an Effective Defense, ” inNeurIPS, vol. 36, 2023

    2023

  8. [16]

    OpenWebText Corpus,

    A. Gokaslan and V. Cohen, “OpenWebText Corpus, ” 2019. [Online]. Available: https://skylion007. github.io/OpenWebTextCorpus/

    2019

  9. [17]

    Language Models are Unsuper- vised Multitask Learners,

    A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, and I. Sutskever, “Language Models are Unsuper- vised Multitask Learners, ” OpenAI Technical Report, 2019

    2019

  10. [18]

    Detecting and Unmasking AI-Generated Texts through Explainable Artificial Intelligence using Stylistic Features,

    A. Shah et al., “Detecting and Unmasking AI-Generated Texts through Explainable Artificial Intelligence using Stylistic Features, ”IJACSA, vol. 14, no. 10, pp. 968–976, 2023

    2023

  11. [19]

    Testing of Detection Tools for AI-Generated Text,

    L. Weber-Wulff et al., “Testing of Detection Tools for AI-Generated Text, ”Int. J. Educational Integrity, vol. 19, no. 1, p. 26, 2023

    2023

  12. [20]

    Can AI-Generated Text be Reliably Detected?

    V. S. Sadasivan, A. Kumar, S. Balasubramanian, W. Wang, and S. Feizi, “Can AI-Generated Text be Reliably Detected?”arXiv:2303.11156, 2023

    2023 arXiv

  13. [21]

    GPT Detectors are Biased Against Non- Native English Writers,

    W. Liang, M. Yuksekgonul, Y. Mao, E. Wu, and J. Zou, “GPT Detectors are Biased Against Non- Native English Writers, ”Patterns, vol. 4, no. 7, p. 100779, 2023

    2023

  14. [22]

    A Survey on Detection of LLMs-Generated Content,

    X. Yang, Y. Cheng, W. Pang, and Y. Chen, “A Survey on Detection of LLMs-Generated Content, ” arXiv:2310.15654, 2023

    2023 arXiv

  15. [23]

    Language Models are Few-Shot Learners,

    T. B. Brown et al., “Language Models are Few-Shot Learners, ” inAdvances in Neural Information Processing Systems (NeurIPS), vol. 33, pp. 1877–1901, 2020

    1901

  16. [24]

    Training Language Models to Follow Instructions with Human Feedback,

    L. Ouyang, J. Wu, X. Jiang, D. Almeida, C. L. Wainwright, P. Mishkin, C. Zhang, S. Agarwal, K. Slama, A. Ray, et al., “Training Language Models to Follow Instructions with Human Feedback, ” inAdvances in Neural Information Processing Systems (NeurIPS), vol. 35, 2022. Preprint:...

    2022 arXiv

  17. [25]

    Constitutional AI: Harmlessness from AI Feedback,

    Y. Bai et al., “Constitutional AI: Harmlessness from AI Feedback, ”arXiv:2212.08073, 2022

    2022 arXiv

  18. [26]

    MAGE: Machine- Generated Text Detection in the Wild,

    Y. Li, Q. Li, L. Cui, W. Bi, Z. Wang, L. Wang, L. Yang, S. Shi, and Y. Zhang, “MAGE: Machine- Generated Text Detection in the Wild, ” inProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL), pp. 36–53, 2024

    2024

  19. [2025]

    Preprint: arXiv:2409.04808

    arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.