Pith. sign in

REVIEW 2 major objections 5 minor 3 cited by

DAMAGE: Detecting Adversarially Modified AI Generated Text

T0 review · 2 major / 5 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 A detector trained on humanized text catches 98.26% of humanized AI essays at a 5% false-positive rate.

desk verdict A useful empirical study of humanizers with a solid detector result, but the 'detector-specific' attack is mislabeled and the unseen-humanizer claim lacks direct evidence. read the letter →

arxiv 2501.03437 v1 pith:AWBYDBEE submitted 2025-01-06 cs.CL

Elyas Masrour , Bradley Emi , Max Spero This is my paper · ORCID
classification cs.CL
keywords AItextdetectionhumanizeradversarialrobustnessdataaugmentationparaphraseLLM-generatedsyntheticmirrorsLoRAfine-tuning
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 studies 19 AI humanizer tools—websites that rewrite AI text to evade detection—and shows that popular detectors such as GPTZero and Binoculars lose much of their accuracy on humanized academic essays. It then builds DAMAGE, a classifier that detects humanized AI text at a 98.26% true-positive rate with a 5% false-positive rate, while keeping near-perfect accuracy on unmodified AI text. The central trick is data-centric: humanized examples, including human documents passed through humanizers and labeled as human, are added to training so that the model learns humanization as an invariance rather than a separate domain. The paper also shows that even after a humanizer is fine-tuned against DAMAGE's own predictions, the detector still catches 93.2% of the adversarially humanized essays at its default threshold.

What carries the argument

The central machinery is a training pipeline rather than a single identity. Synthetic mirror prompts generate an AI version of each human document; humanized samples are created by passing both human and AI documents through L1 humanizers, chunking them to roughly 300 words, oversampling them, and labeling humanized human documents as human. A 12-billion-parameter autoregressive language model is fine-tuned with LoRA adapters and a weighted cross-entropy loss, followed by hard-negative mining on false positives. The key design choice is to treat humanization as an input transform the model must become invariant to, not as a separate output class.

What would settle it

Hold out one of the 19 humanizers from the training set, then evaluate DAMAGE on text passed through that held-out humanizer; if the true-positive rate drops to the 95.7% no-humanizer baseline or lower, the cross-humanizer generalization claim fails.

Watch

Extended reading notes

Core claim

The paper claims that a deep-learning AI text detector can be made robust to humanizer paraphrasing by training on a small volume of humanized text from a curated set of high-quality (L1) humanizers, with humanized human documents included in the training data and labeled as human. This learned invariance lets DAMAGE maintain high recall on humanized text without inflating false positives. In the authors' benchmark, the resulting model detects 98.26% of humanized AI academic text at a 5% false-positive rate, versus 60.04% for GPTZero and 28.23% for Binoculars, and it also outperforms its own no-humanizer baseline (95.74%). When a humanizer is fine-tuned against DAMAGE's own predictions, detection remains at 100% at a 5% false-positive threshold and at 93.2% at the default threshold, which the authors attribute to residual detectable patterns in the humanizer's language model and to the detector's learned invariance.

Load-bearing premise

The central argument assumes that the paper's three quality tiers are stable, so training only on the best-tier humanizers teaches the model to detect all humanizers, including ones never seen in training.

Editorial extensions

If this is right

  • If DAMAGE's robustness generalizes, educational and publishing settings can deploy a detector that catches humanized AI text without substantially increasing false accusations against student writers, including ESL writers.
  • Existing perplexity-based detectors such as GPTZero and Binoculars are not sufficient against even modest humanizers; any serious detector needs training data that includes paraphrased or humanized text.
  • Watermarking schemes like SynthID are not a standalone defense: the paper's paraphrase experiment drops SynthID watermark detection from 87.6% to 5.4% true-positive rate at a 5% false-positive rate.
  • Detector-specific adversarial humanizers can reduce detection but not eliminate it, suggesting that robust detectors force attackers into a costlier fine-tuning loop; the paper's own adversarially fine-tuned humanizer produces only L2-level fluency.

Reading between the lines

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

  • Editorial extension: The L1-only training choice suggests that a humanizer deliberately designed to mimic L2 or L3 degradation could exploit the precision trade-off reported in the paper, since including all tiers in training raised the false-positive rate.
  • Editorial extension: The 19 surveyed tools are a snapshot; future humanizers trained on detector feedback, or trained with more data and compute, may close the remaining 6.8% gap, so the 93.2% adversarial result is not a stability guarantee.
  • Editorial extension: The invariance-based approach implies that AI detectors should be evaluated on held-out humanizers the way image classifiers are evaluated on unseen corruptions; building a tiered humanizer leaderboard would make cross-humanizer claims directly testable.
  • Editorial extension: The watermark-removal result implies that watermarking must be paired with paraphrase detection to be useful for provenance, rather than used as a standalone defense.
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 / 5 minor

Summary. The paper studies AI-text humanizer tools. It audits 19 tools, categorizes them into three fluency tiers, shows that watermarking (SynthID) is defeated by DIPPER paraphrasing, and evaluates existing detectors on humanized academic essays. It then introduces DAMAGE, a 12B-parameter Mistral-NeMo-based detector fine-tuned with LoRA on synthetic mirror-prompt data augmented with humanizer outputs, and reports high true-positive rates at 5% FPR on humanized text and on RAID paraphrase/synonym subsets. Finally, it describes a 'detector-specific' adversarial humanizer built by fine-tuning GPT-4o on false-negative examples and reports that DAMAGE still detects 93.2% of such text at the default threshold.

Significance. If the detection results are reproducible, the paper makes a useful practical contribution: a robust detector for humanized AI text, an ablation study showing which design choices matter, and a public audit of 19 humanizer tools with tier labels. The bootstrapped confidence intervals in Table 3 and the decomposition in Table 8 are strengths, as is the evaluation on held-out student essay datasets. However, the two headline claims--generalization to unseen humanizers and robustness to a detector-specific attack--are not fully supported by the evidence as presented, which tempers the significance. The core Table 3 result is solid and meaningful, but the adversarial-robustness claim needs either reframing or a genuinely detector-optimized attack.

major comments (2)
  1. [§7.1 and Table 5] The adversarial humanizer described in §7.1 is not directly optimized against the detector as the abstract and §7 claim. The procedure selects L1 humanizer outputs that the detector already labels as human (false negatives) and fine-tunes GPT-4o to imitate those outputs; the detector is used only as a filter to choose training pairs, and no detector loss, gradient, or iterative feedback enters the fine-tuning. Table 5 therefore supports robustness to one fine-tuned paraphraser that mimics a particular humanizer, but it does not test a white-box attack that explicitly minimizes the detector's score. The claim that the detector 'remains robust to a detector-specific fine-tuning attack' (abstract, §7) should either be rephrased to describe the actual imitation-based attack, or supported by an additional experiment using a detector-aware reward such as DPO/RL as in Nicks et al. (2024).
  2. [§5.4, Table 3, Table 8] The claim of robustness to humanizers 'unseen during training' (abstract; §1) is not directly evidenced. Training uses only L1 humanizers, but the evaluation in Table 3 appears to aggregate all humanizers, and no table or appendix reports performance separately by humanizer, by tier, or by seen versus unseen status. Because L2 and L3 humanizers were excluded from training and are lower quality, their inclusion in the aggregate could mask weaker performance on held-out L1 humanizers. A per-humanizer breakdown is needed to determine whether the 98.26% TPR at 5% FPR reflects true cross-humanizer generalization or is driven by easy, low-quality humanizers; the paper should also state whether any L1 humanizers are held out of training.
minor comments (5)
  1. [Table 4] The RAID results are reported as 'TPR @ FPR=5% for Academic Text,' but RAID contains multiple domains; please clarify whether the numbers are on the full RAID benchmark or a domain subset, and specify how the 5% FPR threshold is calibrated for each attack split.
  2. [§5.3] The sentence 'the hidden state from the final token in the sequence is used as the input to both classification heads' appears to be a typo; there is only one classification head described. Also, 'train to convergence ... for 1 epoch' is contradictory; please specify the actual stopping criterion.
  3. [§3.4 and Table 2] The claim that humanizers remove watermarks is supported by only one humanizer (DIPPER) and one generator (Gemma-2B-IT); either soften the generalization or add results for additional humanizers.
  4. [§5.5] Please specify explicitly that the RAID evaluation uses the held-out test split, since §5.5 says a small amount of RAID train data was incorporated into the final training run.
  5. [Table 8] The ablation table would be more informative if it reported TPR separately for L1, L2, and L3 humanized test data, not just the aggregate, since the choice to exclude L2/L3 is a central design decision.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the central detection results are measured on held-out essay corpora and external RAID benchmarks, and the self-citations to the authors' earlier methods are not load-bearing.

full rationale

The paper's main claim is an empirical one: a detector trained with humanizer-augmented data achieves high TPR on humanized AI text. The evaluation uses held-out student essay datasets (PERSUADE, ELLIPSE, BAWE, etc.) and the external RAID benchmark, so the headline numbers are not derived from the training labels by construction. The data-centric augmentation uses outputs from real third-party humanizer tools, not from the detector itself, and the ablation study (Table 8) isolates the contribution of humanizer data. The self-citations to Emi and Spero (2024) concern mirror-prompt synthetic data generation and active learning; both procedures are described in sufficient detail in the present paper, and the detector's robustness is not justified solely by that citation. Section 7's adversarial humanizer is built by fine-tuning GPT-4o on detector false negatives rather than by optimizing a detector loss, so the phrase 'optimized against our detector's predictions' overstates the attack's strength; however, this is a threat-model limitation or correctness concern, not circularity, because the reported 93.2% TPR is an independent empirical measurement and is not forced by the training procedure. Similarly, the claim of generalization to unseen humanizers is not directly supported by a seen-versus-unseen split, but the absence of that evidence does not make the derivation circular. Overall, no step in the paper reduces to its own inputs by definition or by fitted-parameter renaming.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central result rests on several domain assumptions about humanizer behavior and data labeling that are not validated against independent evidence. The main free parameters are training choices (oversampling, context length, L1-only selection) that were set by hand and not ablated for sensitivity. No new entities are postulated.

free parameters (3)
  • oversample_factor = 18
    Humanized text is oversampled by 18x to compensate for its small volume (0.68% of the final dataset); no sensitivity analysis is reported.
  • context_window = 512 tokens
    Inputs are cropped to 512 tokens to constrain the model to short-range features; this truncation choice affects what the classifier can see.
  • l1_only_training = True
    The final model is trained only on L1 humanizer outputs; L2/L3 are excluded because they compromise precision (FPR). This is a post-hoc selection based on validation performance and underpins the generalization claim.
assumptions (4)
  • domain assumption Humanizers marketed at students work best on student writing, so Fineweb-EDU is an adequate proxy for student essays.
    Section 5.4 uses Fineweb-EDU documents as the base corpus for humanizer augmentation without validating that humanizer behavior on this corpus matches student essays.
  • domain assumption Labeling humanized human text as 'human' in training teaches the desired invariance rather than a spurious humanizer signal.
    Section 5.4; the ablation shows this choice improves FPR, but the mechanism is not proven.
  • domain assumption The L1/L2/L3 tiering is a stable property of each humanizer and does not drift over time.
    Sections 3 and 4 tier humanizers based on a small qualitative sample; the robustness evaluation assumes these tiers remain valid for the dataset/eval.
  • domain assumption GPT-4o fine-tuning on the detector's false negatives is a representative 'detector-specific adversarial humanizer.'
    Section 7; the attack is a proof of concept with a weak optimization signal, yet conclusions about robustness are drawn from it.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DAMAGE: Detecting Adversarially Modified AI Generated Text." pith.science (2026). https://pith.science/paper/AWBYDBEE

@misc{pith2026250103437,
  author       = {Pith},
  title        = {Pith review of: DAMAGE: Detecting Adversarially Modified AI Generated Text},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AWBYDBEE}},
  note         = {Machine review of arXiv:2501.03437}
}
read the original abstract

AI humanizers are a new class of online software tools meant to paraphrase and rewrite AI-generated text in a way that allows them to evade AI detection software. We study 19 AI humanizer and paraphrasing tools and qualitatively assess their effects and faithfulness in preserving the meaning of the original text. We show that many existing AI detectors fail to detect humanized text. Finally, we demonstrate a robust model that can detect humanized AI text while maintaining a low false positive rate using a data-centric augmentation approach. We attack our own detector, training our own fine-tuned model optimized against our detector's predictions, and show that our detector's cross-humanizer generalization is sufficient to remain robust to this attack.

Figures

Figures reproduced from arXiv: 2501.03437 by the authors.

Figure 1
Figure 1. Example of an AI humanizer tool from large language models without getting caught. Other humanizers target their product towards SEO marketers, who may generate hundreds of blog posts using AI and apply humanizers to evade AI detection by search engine algorithms. In this work, we attempt to comprehensively study these AI humanizers: what they are doing, and whether it is possible to identify humanized AI-generated … view at source ↗
Figure 2
Figure 2. Augmenting the training set with high quality humanizer data improves robustness. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Two out of the four most popular Writing [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: This paraphraser performs a very close paraphrase, only replacing individual words and phrases rather [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: We segment humanizers into three tiers, based on their fluency. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: This humanizer is an LLM that is instructed to paraphrase "in a conversational tone". [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: This humanizer is an LLM that is instructed to "write at a college level" and "asked to rephrase the [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 3 Pith papers

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

  1. T5-CSBoost: Adversarial Perturbation Resistant LLM Fingerprinting

    cs.CL 2026-05 conditional novelty 6.0 of 10

    Adding a margin-based triplet loss to T5-Sentinel's decoder embeddings improves LLM source attribution robustness to word/character edits, paraphrasing, and unseen models/domains.

  2. Pangram 4 Technical Report

    cs.CL 2026-07 conditional novelty 5.0 of 10

    Pangram 4 is a commercial MoE-based detector claiming 0.9916 AUROC, 0.0041% FPR, 0.3396% FNR, plus tokenwise human/AI-assisted/AI-generated labels and humanizer detection.

  3. Can You Detect the Difference?

    cs.CL 2025-07 reject novelty 4.0 of 10

    A 2,000-sample comparison finds diffusion-generated LLaDA text can match human perplexity and burstiness when rephrasing, while LLaMA text is more predictable and easier to flag.

Reference graph

Works this paper leans on

37 extracted references · 13 canonical work pages · cited by 3 Pith papers

  1. [1]

    Navid Ayoobi, Lily Knab, Wen Cheng, David Pantoja, Hamidreza Alikhani, Sylvain Flamant, Jin Kim, and Arjun Mukherjee. 2024. https://arxiv.org/abs/2409.14285 Esperanto: Evaluating synthesized phrases to enhance robustness in ai detection for text origination . Preprint, arXiv:2409.14285

    arXiv 2024

  2. [2]

    Guangsheng Bao, Yanbin Zhao, Zhiyang Teng, Linyi Yang, and Yue Zhang. 2024. https://arxiv.org/abs/2310.05130 Fast-detectgpt: Efficient zero-shot detection of machine-generated text via conditional probability curvature . Preprint, arXiv:2310.05130

    arXiv 2024

  3. [3]

    Counter Turing Test CT^2: AI-Generated Text Detection is Not as Easy as You May Think -- Introducing AI Detectability Index

    Megha Chakraborty, S. M Towhidul Islam Tonmoy, S M Mehedi Zaman, Krish Sharma, Niyar R Barman, Chandan Gupta, Shreya Gautam, Tanay Kumar, Vinija Jain, Aman Chadha, Amit P. Sheth, and Amitava Das. 2023. https://arxiv.org/abs/2310.05030 Counter turing test ct\^2: Ai-generated text detection is not as easy as you may think -- introducing ai detectability ind...

    work page Pith review arXiv 2023

  4. [4]

    Crossley, Perpetual Baffour, Yingqi Tian, Abigail Franklin, Margaret Benner, and Ulrich Boser

    Scott A. Crossley, Perpetual Baffour, Yingqi Tian, Abigail Franklin, Margaret Benner, and Ulrich Boser. 2024. A large-scale corpus for assessing written argumentation: PERSUADE 2.0. Assessing Writing, 61. Dataset comprising over 25,000 argumentative essays from 6th-12th grade students, released under CC BY-NC-SA 4.0

    work page 2024

  5. [5]

    Crossley, Yuan Tian, Perpetual Baffour, Alicia Franklin, Yeoil Kim, William Morris, Brooke Benner, Auriane Picou, and Ulrich Boser

    Scott A. Crossley, Yuan Tian, Perpetual Baffour, Alicia Franklin, Yeoil Kim, William Morris, Brooke Benner, Auriane Picou, and Ulrich Boser. 2023. Measuring second language proficiency using the English Language Learner Insight , Proficiency and Skills Evaluation ( ELLIPSE ) Corpus . International Journal of Learner Corpus Research, 9(2):248--269

    work page 2023

  6. [6]

    Liam Dugan, Alyssa Hwang, Filip Trhlik, Josh Magnus Ludan, Andrew Zhu, Hainiu Xu, Daphne Ippolito, and Chris Callison-Burch. 2024. https://arxiv.org/abs/2405.07940 Raid: A shared benchmark for robust evaluation of machine-generated text detectors . Preprint, arXiv:2405.07940

    arXiv 2024

  7. [7]

    Bradley Emi and Max Spero. 2024. https://arxiv.org/abs/2402.14873 Technical report on the pangram ai-generated text classifier . Preprint, arXiv:2402.14873

    arXiv 2024

  8. [8]

    Angela Fan, Yacine Jernite, Ethan Perez, David Grangier, Jason Weston, and Michael Auli. 2019. https://arxiv.org/abs/1907.09190 Eli5: Long form question answering . Preprint, arXiv:1907.09190

    arXiv 2019

Show all 37 references
  1. [9]

    Alex Franklin, asiegel, HCL-Jevster, Jules King, julianmante, Maggie, Perpetual Baffour, Ryan Holbrook, and Scott Crossley. 2023. Commonlit - evaluate student summaries. https://kaggle.com/competitions/commonlit-evaluate-student-summaries. Kaggle

    2023

  2. [10]

    Google DeepMind . 2024. https://deepmind.google/technologies/synthid/ Synthid: Identifying ai-generated content with digital watermarks . Accessed: 2024-11-14

    2024

  3. [11]

    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

  4. [12]

    Langdon Holmes, Scott Crossley, Perpetual Baffour, Jules King, Lauryn Burleigh, Maggie Demkin, Ryan Holbrook, Walter Reade, and Addison Howard. 2024. The learning agency lab - pii data detection. https://kaggle.com/competitions/pii-detection-removal-from-educational-data. Kaggle

    2024

  5. [13]

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685

    2022 arXiv

  6. [14]

    Xiaomeng Hu, Pin-Yu Chen, and Tsung-Yi Ho. 2023. https://arxiv.org/abs/2307.03838 Radar: Robust ai-text detection via adversarial learning . Preprint, arXiv:2307.03838

    2023 arXiv

  7. [15]

    Shin'ichiro Ishikawa. 2023. The ICNALE Guide: An Introduction to a Learner Corpus Study on Asian Learners' L2 English . Routledge

    2023

  8. [16]

    Alan Juffs, Na-Rae Han, and Ben Naismith. 2020. https://doi.org/10.5281/zenodo.3991977 The University of Pittsburgh English Language Corpus ( PELIC )

    2020 doi

  9. [17]

    John Kirchenbauer, Jonas Geiping, Yuxin Wen, Jonathan Katz, Ian Miers, and Tom Goldstein. 2023. https://arxiv.org/abs/2301.10226 A watermark for large language models . Preprint, arXiv:2301.10226

    2023 arXiv

  10. [18]

    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

  11. [19]

    Tharindu Kumarage, Paras Sheth, Raha Moraffah, Joshua Garland, and Huan Liu. 2023. https://arxiv.org/abs/2310.05095 How reliable are ai-generated-text detectors? an assessment framework using evasive soft prompts . Preprint, arXiv:2310.05095

    2023 arXiv

  12. [20]

    Laida Kushnareva, Tatiana Gaintseva, German Magai, Serguei Barannikov, Dmitry Abulkhanov, Kristian Kuznetsov, Eduard Tulchinskii, Irina Piontkovskaya, and Sergey Nikolenko. 2024. https://arxiv.org/abs/2311.08349 Ai-generated text boundary detection with roft . Preprint, arXiv:...

    2024 arXiv

  13. [21]

    Weixin Liang, Mert Yuksekgonul, Yining Mao, Eric Wu, and James Zou. 2023. https://arxiv.org/abs/2304.02819 Gpt detectors are biased against non-native english writers . Preprint, arXiv:2304.02819

    2023 arXiv

  14. [22]

    Anton Lozhkov, Loubna Ben Allal, Leandro von Werra, and Thomas Wolf. 2024. https://doi.org/10.57967/hf/2497 Fineweb-edu

    2024 doi

  15. [23]

    Mistral AI Team . 2024. https://mistral.ai/news/mistral-nemo/ Mistral NeMo . Released in collaboration with NVIDIA, July 18, 2024

    2024

  16. [24]

    Manning, and Chelsea Finn

    Eric Mitchell, Yoonho Lee, Alexander Khazatsky, Christopher D. Manning, and Chelsea Finn. 2023. https://arxiv.org/abs/2301.11305 Detectgpt: Zero-shot machine-generated text detection using probability curvature . arXiv preprint

    2023 arXiv

  17. [25]

    Hilary Nesi, Sheena Gardner, Paul Thompson, and Paul Wickens. 2008. http://ota.ox.ac.uk/desc/2539 The British Academic Written English ( BAWE ) Corpus . First version, 2008. Available from: http://ota.ox.ac.uk/desc/2539

    2008

  18. [26]

    Charlotte Nicks, Eric Mitchell, Rafael Rafailov, Archit Sharma, Christopher D Manning, Chelsea Finn, and Stefano Ermon. 2024. https://openreview.net/forum?id=4eJDMjYZZG Language model detectors are easily optimized against . In The Twelfth International Conference on Learning ...

    2024

  19. [27]

    OpenAI. 2023. https://arxiv.org/abs/2303.08774 Gpt-4 technical report . arXiv preprint arXiv:2303.08774

    2023 arXiv

  20. [28]

    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

  21. [29]

    Irene Solaiman, Miles Brundage, Jack Clark, Amanda Askell, Ariel Herbert-Voss, Jeff Wu, Alec Radford, Gretchen Krueger, Jong Wook Kim, Sarah Kreps, et al. 2019. Release strategies and the social impacts of language models. arXiv preprint arXiv:1908.09203

    2019 arXiv

  22. [30]

    Turnitin AI Technical Staff. 2024. https://www.turnitin.com/whitepapers/turnitin-ai-writing-detection-model-architecture-and-testing-protocol Turnitin ai writing detection: Model architecture and testing protocol . Technical report, Turnitin

    2024

  23. [31]

    Gemma Team, Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Rivière, Mihir Sanjay Kale, Juliette Love, Pouya Tafti, Léonard Hussenot, Pier Giuseppe Sessa, Aakanksha Chowdhery, Adam Roberts, Aditya Barua, Alex Botev, Alex...

    2024 arXiv

  24. [32]

    Edward Tian and Alexander Cui. 2023. https://gptzero.me Gptzero: Towards detection of ai-generated text using zero-shot and supervised methods

    2023

  25. [33]

    Vivek Verma, Eve Fleisig, Nicholas Tomlin, and Dan Klein. 2023. https://arxiv.org/abs/2305.15047 Ghostbuster: Detecting text ghostwritten by large language models . Preprint, arXiv:2305.15047

    2023 arXiv

  26. [34]

    Pengyu Wang, Linyang Li, Ke Ren, Botian Jiang, Dong Zhang, and Xipeng Qiu. 2023. https://arxiv.org/abs/2310.08903 Seqxgpt: Sentence-level ai-generated text detection . Preprint, arXiv:2310.08903

    2023 arXiv

  27. [35]

    Debora Weber-Wulff, Alla Anohina-Naumeca, Sonja Bjelobaba, Tomáš Foltýnek, Jean Guerrero-Dib, Olumide Popoola, Petr Šigut, and Lorna Waddington. 2023. https://doi.org/10.1007/s40979-023-00146-z Testing of detection tools for ai-generated text . International Journal for Educat...

    2023 doi

  28. [36]

    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...

  29. [37]

    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.