pith. sign in

arxiv: 2606.25152 · v1 · pith:RY42N3LNnew · submitted 2026-06-23 · 💻 cs.CL · cs.AI

Hitting a Moving Target: Test-Time Adaptation for AI Text Detection under Continual Distribution Shift

Kevin Ren , Manish Raghavan , Nikhil Garg This is my paper

Pith reviewed 2026-06-25 23:29 UTC · model grok-4.3

classification 💻 cs.CL cs.AI
keywords AI text detectiontest-time adaptationdistribution shiftsemi-supervised learningLLM detectionadversarial humanizationtemporal drift
0
0 comments X

The pith

Test-time adaptation with semi-supervised learning on unlabeled samples maintains robust AI text detection under continual shifts.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that standard supervised AI text detectors, which require labeled data at training time, break down under three ongoing post-deployment shifts: adversarial humanization of AI output, release of new LLMs, and natural changes in human writing styles over time. It proposes adapting the detector at test time by applying semi-supervised learning to batches of unlabeled samples, using the fact that those samples tend to come from the same source and therefore share homogeneity. This matters because labeled data is rarely available in real deployments, yet the adapted detector stays effective where fixed models collapse. A sympathetic reader sees the work as showing a practical path to keep detectors functional without constant retraining.

Core claim

Deployed AI text detectors fail under continual distribution shifts for which labeled data is unavailable. A test-time adaptation method using semi-supervised learning adapts by leveraging homogeneity among unlabeled samples observed at inference time. State-of-the-art supervised detectors systematically fail on both adversarial and natural shifts in AI-generated and human text, while the test-time approach remains largely robust, detecting 90.5 percent of adversarial AI-generated text compared to 24.1 percent for the commercial model Pangram.

What carries the argument

Test-time adaptation with semi-supervised learning that exploits inference-time homogeneity among unlabeled samples to adapt the detector without new labels.

If this is right

  • Supervised detectors fail on adversarial humanization, new LLMs, and temporal drift in human writing.
  • Test-time adaptation recovers high accuracy on the same shifted distributions without requiring labels.
  • A commercial detector reaches only 24.1 percent detection on adversarial AI text while the adapted method reaches 90.5 percent.
  • The framework supports ongoing detection in the wild after deployment.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Detectors could run continuously in production and update themselves from the unlabeled traffic they already see.
  • The same homogeneity signal might be tested on other generated-media tasks such as image or audio detection.
  • Smaller inference batches or faster shifts could be measured to find the practical limits of the adaptation step.

Load-bearing premise

Unlabeled samples observed together at inference time are homogeneous enough to supply a reliable signal that semi-supervised learning can use to adapt to the current shift.

What would settle it

A batch of inference-time samples drawn from the same distribution that are too heterogeneous for the semi-supervised step to improve, or even to maintain, detection accuracy on held-out test cases from that distribution.

Figures

Figures reproduced from arXiv: 2606.25152 by Kevin Ren, Manish Raghavan, Nikhil Garg.

Figure 1
Figure 1. Figure 1: Comparing AI text detection with supervised learning (top) and test-time adaptation with [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: (a) We simulate the back-and-forth between detectors and evaders. PU + TTA and PNU + [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Detectors trained on one LLM have low recall on the outputs of other LLMs. TTA performs [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: (a) Due to temporal distribution shift in human writing, models trained in 2010 degrade in [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Plotting the same experiment as in Figure [PITH_FULL_IMAGE:figures/full_fig_p030_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Plotting the same experiment as in Figure [PITH_FULL_IMAGE:figures/full_fig_p031_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Plotting the same experiment as in Figure [PITH_FULL_IMAGE:figures/full_fig_p032_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Sensitivity analysis for introducing varied amounts of sentences from naive AI-generated [PITH_FULL_IMAGE:figures/full_fig_p033_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: (9a) Distribution of hallucination scores and omission scores per-prompt. We generally preserve the content that was originally in the human abstract. (9b) The scores do not correlate with the predictions of our detection models (same detection model-prompt mapping as in [PITH_FULL_IMAGE:figures/full_fig_p035_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Plotting the same experiment as in Figure [PITH_FULL_IMAGE:figures/full_fig_p036_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: The same figure as in Figure [PITH_FULL_IMAGE:figures/full_fig_p037_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: We plot heatmaps of all collected metrics for models trained on five different Gemini [PITH_FULL_IMAGE:figures/full_fig_p038_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: The same figure as the AI recall subplot in Figure [PITH_FULL_IMAGE:figures/full_fig_p039_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Linear Discriminant Analysis as in Figure [PITH_FULL_IMAGE:figures/full_fig_p039_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Error rates of logistic regressions trained and evaluated in-sample on abstracts from two [PITH_FULL_IMAGE:figures/full_fig_p041_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Scatter plots, for each pairwise combination of LLMs, for the error rates as in Figure [PITH_FULL_IMAGE:figures/full_fig_p041_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: We find that Pangram performs significantly better on some LLMs than others. Notably, [PITH_FULL_IMAGE:figures/full_fig_p042_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Grouped bar plots as in Figure [PITH_FULL_IMAGE:figures/full_fig_p043_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Demonstrating that Figure 4a is robust on multiple metrics. While PU + TTA performs poorly on AI-generated text compared to supervised methods, we see that, due to temporal shift in human writing from 2010 to 2020, supervised methods gradually predict worse on human writing (lower AUC, balanced accuracy, and continuous predictions on human writing; higher balanced cross-entropy and bias on prevalence esti… view at source ↗
Figure 20
Figure 20. Figure 20: We plot CDFs of the distribution of predictions on held-out AI-generated writing from [PITH_FULL_IMAGE:figures/full_fig_p044_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: We plot PDFs (histograms) of the distribution of predictions on held-out human writing [PITH_FULL_IMAGE:figures/full_fig_p045_21.png] view at source ↗
read the original abstract

Deployed approaches for AI text detection often rely on training-time access to labeled datasets of both human-written and AI-generated text. This approach is vulnerable to three types of distribution shifts that occur continually post-deployment, and for which labeled data is often unavailable: adversarial humanization, new LLMs being released, and temporal drift in human writing. Simultaneously, existing approaches do not leverage a key signal of LLM usage: inference-time homogeneity. We propose a test-time adaptation (TTA) approach, using semi-supervised learning, that adapts to distribution shifts by leveraging homogeneity among unlabeled samples observed at inference time. Empirically, we find that state-of-the-art supervised detectors systematically fail when they encounter distribution shifts in AI-generated and human writing, both adversarial and natural, while test-time adaptation with semi-supervised learning is largely robust; e.g., the commercial model Pangram detects just 24.1% of our adversarial AI-generated text, compared to 90.5% for our test-time approach. We establish that test-time adaptation is a promising framework for AI text detection in the wild. We publicly release our code (which includes code for model training, evaluation, and plots) at https://github.com/kkr36/llm_detection.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 1 minor

Summary. The paper claims that supervised AI text detectors are vulnerable to three continual post-deployment distribution shifts (adversarial humanization, new LLMs, temporal drift in human writing) because labeled data is unavailable, while a proposed test-time adaptation (TTA) method using semi-supervised learning can adapt by exploiting homogeneity among unlabeled inference-time samples. It reports that this yields substantially higher robustness, e.g., 90.5% detection on adversarial AI-generated text versus 24.1% for the commercial Pangram detector, and concludes that TTA is a promising framework for detection in the wild. Code for training, evaluation, and plots is released.

Significance. If the empirical claims hold, the work would be significant for practical AI text detection because it directly targets the post-deployment shift problem without requiring new labeled data. The use of inference-time homogeneity as an adaptation signal is a concrete contribution, and the public code release (including training/evaluation scripts and plots) is a clear strength that supports reproducibility.

major comments (1)
  1. [Abstract] Abstract: The central claim that TTA with semi-supervised learning is 'largely robust' to the three shifts rests on the assumption that homogeneity among unlabeled inference-time samples supplies a usable signal. The reported numbers (90.5% vs. 24.1%) are given only for controlled batches; no experiments or controls are described for mixed human/AI batches or streaming multi-distribution regimes that would remove the homogeneity signal, which is load-bearing for the 'in the wild' robustness conclusion.
minor comments (1)
  1. The abstract states that 'state-of-the-art supervised detectors systematically fail' under the shifts but provides no quantitative baseline numbers or dataset details for the non-adversarial cases.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for highlighting the centrality of the homogeneity assumption and the need to clarify the experimental scope for 'in the wild' claims. We address the point directly below and propose targeted revisions.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central claim that TTA with semi-supervised learning is 'largely robust' to the three shifts rests on the assumption that homogeneity among unlabeled inference-time samples supplies a usable signal. The reported numbers (90.5% vs. 24.1%) are given only for controlled batches; no experiments or controls are described for mixed human/AI batches or streaming multi-distribution regimes that would remove the homogeneity signal, which is load-bearing for the 'in the wild' robustness conclusion.

    Authors: We agree that the homogeneity signal is load-bearing and that our primary quantitative results (including the 90.5% vs. 24.1% comparison) are obtained on controlled batches constructed to preserve that signal. This design choice reflects the intended use case: TTA is applied to batches or streams where inference-time samples share distributional properties (e.g., repeated queries on similar topics or from the same user base), which is common in deployed detectors. We do not claim robustness for arbitrary mixed human/AI batches or fully heterogeneous streaming regimes, where the semi-supervised signal would degrade. The manuscript text motivates the method under the homogeneity assumption but does not include explicit ablation on mixed or multi-distribution streaming settings. To address this, we will revise the abstract and add a dedicated limitations paragraph that explicitly states the operating regime, together with a new experiment quantifying performance under controlled partial mixing. This constitutes a partial revision; the core empirical claims remain valid within the stated scope. revision: partial

Circularity Check

0 steps flagged

No derivation chain present; results are empirical comparisons

full rationale

The paper advances an empirical claim that test-time adaptation via semi-supervised learning remains robust to distribution shifts (adversarial humanization, new LLMs, temporal drift) while supervised detectors degrade, supported by reported accuracy numbers on controlled test batches. No equations, uniqueness theorems, fitted parameters renamed as predictions, or self-citation chains appear in the abstract or described method; the homogeneity assumption is stated as a modeling choice whose validity is tested experimentally rather than derived from prior self-work. The central performance numbers (e.g., 24.1% vs 90.5%) are direct measurements, not quantities forced by construction from the inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Review is based on abstract only; no explicit free parameters, axioms, or invented entities are described. The approach implicitly rests on the domain assumption that inference-time homogeneity is exploitable, but this is captured in the weakest_assumption field rather than as a new invented entity.

pith-pipeline@v0.9.1-grok · 5754 in / 1141 out tokens · 31450 ms · 2026-06-25T23:29:19.406211+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

112 extracted references · 8 canonical work pages

  1. [1]

    K., Bai, Y., Baker, B., Bao, H., et al

    Agarwal, S., Ahmad, L., Ai, J., Altman, S., Applebaum, A., Arbus, E., Arora, R. K., Bai, Y., Baker, B., Bao, H., et al. gpt-oss-120b & gpt-oss-20b model card. arXiv preprint arXiv:2508.10925, 2025

    Pith/arXiv arXiv 2025

  2. [2]

    Fast- DetectGPT : Efficient zero-shot detection of machine-generated text via conditional probability curvature

    Bao, G., Zhao, Y., Teng, Z., Yang, L., and Zhang, Y. Fast- DetectGPT : Efficient zero-shot detection of machine-generated text via conditional probability curvature. arXiv preprint arXiv:2310.05130, 2023

    arXiv 2023

  3. [3]

    and Davis, J

    Bekker, J. and Davis, J. Estimating the class prior in positive and unlabeled data through decision tree induction. In Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018

    2018

  4. [4]

    Conda: Contrastive domain adaptation for ai-generated text detection

    Bhattacharjee, A., Kumarage, T., Moraffah, R., and Liu, H. Conda: Contrastive domain adaptation for ai-generated text detection. arXiv preprint arXiv:2309.03992, 2023

    arXiv 2023

  5. [5]

    C., and Dhillon, P

    Chakrabarty, T., Ginsburg, J. C., and Dhillon, P. Readers prefer outputs of AI trained on copyrighted books over expert human writers. arXiv preprint arXiv:2510.13939, 2025 a

    arXiv 2025

  6. [6]

    AI -slop to AI -polish? aligning language models through edit-based writing rewards and test-time computation

    Chakrabarty, T., Laban, P., and Wu, C.-S. AI -slop to AI -polish? aligning language models through edit-based writing rewards and test-time computation. arXiv preprint arXiv:2504.07532, 2025 b

    arXiv 2025

  7. [7]

    arxiv dataset

    Cornell-University. arxiv dataset. Kaggle Dataset, 2020. URL https://www.kaggle.com/datasets/Cornell-University/arxiv

    2020

  8. [8]

    Introducing AI to an online petition platform changed outputs but not outcomes

    Corpus, I., Gilbert, E., Koenecke, A., and Naaman, M. Introducing AI to an online petition platform changed outputs but not outcomes. arXiv preprint arXiv:2511.13949, 2025

    arXiv 2025

  9. [9]

    Adversarial classification

    Dalvi, N., Domingos, P., Mausam, Sanghai, S., and Verma, D. Adversarial classification. In Proceedings of the tenth ACM SIGKDD international conference on Knowledge discovery and data mining, pp.\ 99--108, 2004

    2004

  10. [10]

    C., and Kiritchenko, S

    Dawkins, H., Fraser, K. C., and Kiritchenko, S. When detection fails: The power of fine-tuned models to generate human-like social media text. arXiv preprint arXiv:2506.09975, 2025

    arXiv 2025

  11. [11]

    Doshi, A. R. and Hauser, O. P. Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science advances, 10 0 (28): 0 eadn5290, 2024

    2024

  12. [12]

    C., Niu, G., and Sugiyama, M

    Du Plessis, M. C., Niu, G., and Sugiyama, M. Analysis of learning from positive and unlabeled data. Advances in neural information processing systems, 27, 2014

    2014

  13. [13]

    RAID : A Shared Benchmark for Robust Evaluation of Machine-Generated Text Detectors

    Dugan, L., Hwang, A., Trhl \'i k, F., Zhu, A., Ludan, J. M., Xu, H., Ippolito, D., and Callison-Burch, C. RAID : A shared benchmark for robust evaluation of machine-generated text detectors. In Ku, L.-W., Martins, A., and Srikumar, V. (eds.), Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p...

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

  14. [14]

    and Antoniak, M

    Elazar, Y. and Antoniak, M. LLM -generated or human-written? comparing review and non-review papers on arxiv. arXiv preprint arXiv:2601.17036, 2026

    arXiv 2026

  15. [15]

    and Noto, K

    Elkan, C. and Noto, K. Learning classifiers from only positive and unlabeled data. In Proceedings of the 14th ACM SIGKDD international conference on Knowledge discovery and data mining, pp.\ 213--220, 2008

    2008

  16. [16]

    and Spero, M

    Emi, B. and Spero, M. Technical report on the Pangram AI -generated text classifier. arXiv preprint arXiv:2402.14873, 2024

    arXiv 2024

  17. [17]

    Unsupervised and distributional detection of machine-generated text

    Gall \'e , M., Rozen, J., Kruszewski, G., and Elsahar, H. Unsupervised and distributional detection of machine-generated text. arXiv preprint arXiv:2111.02878, 2021

    arXiv 2021

  18. [18]

    J., Balakrishnan, S., and Lipton, Z

    Garg, S., Wu, Y., Smola, A. J., Balakrishnan, S., and Lipton, Z. Mixture proportion estimation and pu learning: A modern approach. Advances in Neural Information Processing Systems, 34: 0 8532--8544, 2021

    2021

  19. [19]

    Gehrmann, S., Strobelt, H., and Rush, A. M. Gltr: Statistical detection and visualization of generated text. arXiv preprint arXiv:1906.04043, 2019

    Pith/arXiv arXiv 1906

  20. [20]

    A., Chandra, K

    Goel, S., Str \"u ber, J., Auzina, I. A., Chandra, K. K., Kumaraguru, P., Kiela, D., Prabhu, A., Bethge, M., and Geiping, J. Great models think alike and this undermines AI oversight. In Forty-second International Conference on Machine Learning, 2025

    2025

  21. [21]

    Note: Robust continual test-time adaptation against temporal correlation

    Gong, T., Jeong, J., Kim, T., Kim, Y., Shin, J., and Lee, S.-J. Note: Robust continual test-time adaptation against temporal correlation. Advances in Neural Information Processing Systems, 35: 0 27253--27266, 2022

    2022

  22. [22]

    Free AI humanizer: Humanize AI text

    Grammarly . Free AI humanizer: Humanize AI text. https://www.grammarly.com/ai-humanizer, 2026. Accessed: 2026-04-17

    2026

  23. [23]

    The Llama 3 herd of models

    Grattafiori, A., Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Letman, A., Mathur, A., Schelten, A., Vaughan, A., et al. The Llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024

    Pith/arXiv arXiv 2024

  24. [24]

    A survey on LLM -as-a-judge

    Gu, J., Jiang, X., Shi, Z., Tan, H., Zhai, X., Xu, C., Li, W., Shen, Y., Ma, S., Liu, H., et al. A survey on LLM -as-a-judge. The Innovation, 2024

    2024

  25. [25]

    Detective: Detecting AI -generated text via multi-level contrastive learning

    Guo, X., Zhang, S., He, Y., Zhang, T., Feng, W., Huang, H., and Ma, C. Detective: Detecting AI -generated text via multi-level contrastive learning. Advances in Neural Information Processing Systems, 37: 0 88320--88347, 2024

    2024

  26. [26]

    Spotting LLM s with binoculars: Zero-shot detection of machine-generated text

    Hans, A., Schwarzschild, A., Cherepanova, V., Kazemi, H., Saha, A., Goldblum, M., Geiping, J., and Goldstein, T. Spotting LLM s with binoculars: Zero-shot detection of machine-generated text. arXiv preprint arXiv:2401.12070, 2024

    arXiv 2024

  27. [27]

    Strategic classification

    Hardt, M., Megiddo, N., Papadimitriou, C., and Wootters, M. Strategic classification. In Proceedings of the 2016 ACM conference on innovations in theoretical computer science, pp.\ 111--122, 2016

    2016

  28. [28]

    Detree: Detecting human- AI collaborative texts via tree-structured hierarchical representation learning

    He, Y., Zhang, S., Cao, Y., Ma, L., and Luo, P. Detree: Detecting human- AI collaborative texts via tree-structured hierarchical representation learning. arXiv preprint arXiv:2510.17489, 2025

    arXiv 2025

  29. [29]

    Hu, L., Immorlica, N., and Vaughan, J. W. The disparate effects of strategic manipulation. In Proceedings of the Conference on Fairness, Accountability, and Transparency, pp.\ 259--268, 2019

    2019

  30. [30]

    RADAR : Robust AI -text detection via adversarial learning

    Hu, X., Chen, P.-Y., and Ho, T.-Y. RADAR : Robust AI -text detection via adversarial learning. Advances in neural information processing systems, 36: 0 15077--15095, 2023

    2023

  31. [31]

    Humanize AI text with the smartest AI humanizer

    Humanize AI . Humanize AI text with the smartest AI humanizer. https://www.humanizeai.pro/, 2026. Accessed: 2026-04-17

    2026

  32. [32]

    Dedpul: Difference-of-estimated-densities-based positive-unlabeled learning

    Ivanov, D. Dedpul: Difference-of-estimated-densities-based positive-unlabeled learning. In 2020 19th IEEE International Conference on Machine Learning and Applications (ICMLA), pp.\ 782--790. IEEE, 2020

    2020

  33. [33]

    and Imas, A

    Jabarian, B. and Imas, A. Artificial writing and automated detection. Technical report, National Bureau of Economic Research, 2025

    2025

  34. [34]

    T., and Naaman, M

    Jakesch, M., Hancock, J. T., and Naaman, M. Human heuristics for AI -generated language are flawed. Proceedings of the National Academy of Sciences, 120 0 (11): 0 e2208839120, 2023

    2023

  35. [35]

    Automatic detection of machine generated text: A critical survey

    Jawahar, G., Abdul-Mageed, M., and Laks Lakshmanan, V. Automatic detection of machine generated text: A critical survey. In Proceedings of the 28th international conference on computational linguistics, pp.\ 2296--2309, 2020

    2020

  36. [36]

    The subjectivity of monoculture

    Jo, N., Garg, N., and Raghavan, M. The subjectivity of monoculture. arXiv preprint arXiv:2602.24086, 2026

    arXiv 2026

  37. [37]

    Generative AI and perceptual harms: Who's suspected of using LLM s? In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, pp.\ 1--17, 2025

    Kadoma, K., Metaxa, D., and Naaman, M. Generative AI and perceptual harms: Who's suspected of using LLM s? In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, pp.\ 1--17, 2025

    2025

  38. [38]

    autoresearch: AI agents running research on single- GPU nanochat training automatically

    Karpathy, A. autoresearch: AI agents running research on single- GPU nanochat training automatically. https://github.com/karpathy/autoresearch, March 2026. GitHub repository

    2026

  39. [39]

    T., Moazam, H., et al

    Khattab, O., Singhvi, A., Maheshwari, P., Zhang, Z., Santhanam, K., Vardhamanan, S., Haq, S., Sharma, A., Joshi, T. T., Moazam, H., et al. DSPy : Compiling declarative language model calls into self-improving pipelines. arXiv preprint arXiv:2310.03714, 2023

    Pith/arXiv arXiv 2023

  40. [40]

    M., Garg, A., Peng, K., and Garg, N

    Kim, E. M., Garg, A., Peng, K., and Garg, N. Correlated errors in large language models. In Forty-second International Conference on Machine Learning, 2025

    2025

  41. [41]

    M., Lee, K., Zhu, P., Raheja, V., and Kang, D

    Kim, Z. M., Lee, K., Zhu, P., Raheja, V., and Kang, D. Threads of subtlety: Detecting machine-generated texts through discourse motifs. In Ku, L.-W., Martins, A., and Srikumar, V. (eds.), Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp.\ 5449--5474, Bangkok, Thailand, August 2024. Associa...

    work page doi:10.18653/v1/2024.acl-long.298 2024

  42. [42]

    C., and Sugiyama, M

    Kiryo, R., Niu, G., Du Plessis, M. C., and Sugiyama, M. Positive-unlabeled learning with non-negative risk estimator. Advances in neural information processing systems, 30, 2017

    2017

  43. [43]

    and Raghavan, M

    Kleinberg, J. and Raghavan, M. How do classifiers induce agents to invest effort strategically? ACM Transactions on Economics and Computation (TEAC), 8 0 (4): 0 1--23, 2020

    2020

  44. [44]

    and Raghavan, M

    Kleinberg, J. and Raghavan, M. Algorithmic monoculture and social welfare. Proceedings of the National Academy of Sciences, 118 0 (22): 0 e2018340118, 2021

    2021

  45. [45]

    OUTFOX : LLM -generated essay detection through in-context learning with adversarially generated examples

    Koike, R., Kaneko, M., and Okazaki, N. OUTFOX : LLM -generated essay detection through in-context learning with adversarially generated examples. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pp.\ 21258--21266, 2024

    2024

  46. [46]

    Paraphrasing evades detectors of ai-generated text, but retrieval is an effective defense

    Krishna, K., Song, Y., Karpinska, M., Wieting, J., and Iyyer, M. Paraphrasing evades detectors of ai-generated text, but retrieval is an effective defense. Advances in neural information processing systems, 36: 0 27469--27500, 2023

    2023

  47. [47]

    A survey of AI -generated text forensic systems: Detection, attribution, and characterization

    Kumarage, T., Agrawal, G., Sheth, P., Moraffah, R., Chadha, A., Garland, J., and Liu, H. A survey of AI -generated text forensic systems: Detection, attribution, and characterization. arXiv preprint arXiv:2403.01152, 2024

    arXiv 2024

  48. [48]

    Scientific production in the era of large language models

    Kusumegi, K., Yang, X., Ginsparg, P., de Vaan, M., Stuart, T., and Yin, Y. Scientific production in the era of large language models. Science, 390 0 (6779): 0 1240--1243, 2025

    2025

  49. [49]

    Is contrasting all you need? contrastive learning for the detection and attribution of ai-generated text

    La Cava, L., Costa, D., and Tagarelli, A. Is contrasting all you need? contrastive learning for the detection and attribution of ai-generated text. arXiv preprint arXiv:2407.09364, 2024

    arXiv 2024

  50. [50]

    R., McGregor, S., and Ovadya, A

    Leibowicz, C. R., McGregor, S., and Ovadya, A. The deepfake detection dilemma: a multistakeholder exploration of adversarial dynamics in synthetic media. In Proceedings of the 2021 AAAI/ACM Conference on AI, Ethics, and Society, pp.\ 736--744, 2021

    2021

  51. [51]

    The authoritative synthesis trap: Human- AI collaboration and the quality of science

    Leippold, M. The authoritative synthesis trap: Human- AI collaboration and the quality of science. Available at SSRN 6080627, 2026

    2026

  52. [52]

    Who's your judge? on the detectability of LLM -generated judgments

    Li, D., Tan, Z., Zhao, C., Jiang, B., Huang, B., Ma, P., Alnaibari, A., Shu, K., and Liu, H. Who's your judge? on the detectability of LLM -generated judgments. arXiv preprint arXiv:2509.25154, 2025

    arXiv 2025

  53. [53]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers),

    Li, Y., Li, Q., Cui, L., Bi, W., Wang, Z., Wang, L., Yang, L., Shi, S., and Zhang, Y. MAGE : Machine-generated text detection in the wild. In Ku, L.-W., Martins, A., and Srikumar, V. (eds.), Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp.\ 36--53, Bangkok, Thailand, August 2024. Associat...

    work page doi:10.18653/v1/2024.acl-long.3 2024

  54. [54]

    A comprehensive survey on test-time adaptation under distribution shifts

    Liang, J., He, R., and Tan, T. A comprehensive survey on test-time adaptation under distribution shifts. International Journal of Computer Vision, 133 0 (1): 0 31--64, 2025 a

    2025

  55. [55]

    A., and Zou, J

    Liang, W., Izzo, Z., Zhang, Y., Lepp, H., Cao, H., Zhao, X., Chen, L., Ye, H., Liu, S., Huang, Z., McFarland, D. A., and Zou, J. Y. Monitoring AI -modified content at scale: a case study on the impact of ChatGPT on AI conference peer reviews. In Proceedings of the 41st International Conference on Machine Learning, ICML'24. JMLR.org, 2024 a

    2024

  56. [56]

    Mapping the increasing use of LLM s in scientific papers

    Liang, W., Zhang, Y., Wu, Z., Lepp, H., Ji, W., Zhao, X., Cao, H., Liu, S., He, S., Huang, Z., et al. Mapping the increasing use of LLM s in scientific papers. arXiv preprint arXiv:2404.01268, 2024 b

    arXiv 2024

  57. [57]

    The widespread adoption of large language model-assisted writing across society

    Liang, W., Zhang, Y., Codreanu, M., Wang, J., Cao, H., and Zou, J. The widespread adoption of large language model-assisted writing across society. arXiv preprint arXiv:2502.09747, 2025 b

    arXiv 2025

  58. [58]

    Quantifying large language model usage in scientific papers

    Liang, W., Zhang, Y., Wu, Z., Lepp, H., Ji, W., Zhao, X., Cao, H., Liu, S., He, S., Huang, Z., et al. Quantifying large language model usage in scientific papers. Nature Human Behaviour, pp.\ 1--11, 2025 c

    2025

  59. [59]

    S., Yu, P

    Liu, B., Lee, W. S., Yu, P. S., and Li, X. Partially supervised classification of text documents. In ICML, volume 2, pp.\ 387--394. Sydney, NSW, 2002

    2002

  60. [60]

    T., Garg, N., and Borgs, C

    Liu, L. T., Garg, N., and Borgs, C. Strategic ranking. In International Conference on Artificial Intelligence and Statistics, pp.\ 2489--2518. PMLR, 2022

    2022

  61. [61]

    Coco: Coherence-enhanced machine-generated text detection under low resource with contrastive learning

    Liu, X., Zhang, Z., Wang, Y., Pu, H., Lan, Y., and Shen, C. Coco: Coherence-enhanced machine-generated text detection under low resource with contrastive learning. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pp.\ 16167--16188, 2023

    2023

  62. [62]

    On the generalization and adaptation ability of machine-generated text detectors in academic writing

    Liu, Y., Zhong, Z., Liao, Y., Sun, Z., Zheng, J., Wei, J., Gong, Q., Tong, F., Chen, Y., Zhang, Y., et al. On the generalization and adaptation ability of machine-generated text detectors in academic writing. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2, pp.\ 5674--5685, 2025

    2025

  63. [63]

    and Ben-David, S

    Mansouri, F. and Ben-David, S. Learning from positive and unlabeled examples-finite size sample bounds. arXiv preprint arXiv:2507.07354, 2025

    arXiv 2025

  64. [64]

    Does pangram work on GPT-5.4 ? https://www.pangram.com/blog/does-pangram-work-on-gpt-5-4, March 2026

    Masrour, E. Does pangram work on GPT-5.4 ? https://www.pangram.com/blog/does-pangram-work-on-gpt-5-4, March 2026. Accessed: 2026-04-22

    2026

  65. [65]

    N., and Spero, M

    Masrour, E., Emi, B. N., and Spero, M. Damage: detecting adversarially modified AI generated text. In Proceedings of the 1stWorkshop on GenAI Content Detection (GenAIDetect), pp.\ 120--133, 2025

    2025

  66. [66]

    D., and Hardt, M

    Milli, S., Miller, J., Dragan, A. D., and Hardt, M. The social cost of strategic classification. In Proceedings of the conference on fairness, accountability, and transparency, pp.\ 230--239, 2019

    2019

  67. [67]

    D., and Finn, C

    Mitchell, E., Lee, Y., Khazatsky, A., Manning, C. D., and Finn, C. DetectGPT : Zero-shot machine-generated text detection using probability curvature. In International conference on machine learning, pp.\ 24950--24962. PMLR, 2023

    2023

  68. [68]

    Major AI conference flooded with peer reviews written fully by AI

    Naddaf, M. Major AI conference flooded with peer reviews written fully by AI . Nature, 648 0 (8093): 0 256--257, 2025

    2025

  69. [69]

    Nguyen, P., Le, T. M. V., and McAuley, J. Contrastive self-supervised learning for text incoherence detection. In 2025 IEEE International Conference on Big Data (BigData), pp.\ 1500--1507, 2025. doi:10.1109/BigData66926.2025.11401963

    work page doi:10.1109/bigdata66926.2025.11401963 2025

  70. [70]

    Codet-m4: Detecting machine-generated code in multi-lingual, multi-generator and multi-domain settings

    Orel, D., Azizov, D., and Nakov, P. Codet-m4: Detecting machine-generated code in multi-lingual, multi-generator and multi-domain settings. arXiv preprint arXiv:2503.13733, 2025

    arXiv 2025

  71. [71]

    and He, H

    Padmakumar, V. and He, H. Does writing with language models reduce content diversity? In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=Feiz5HtCD0

    2024

  72. [72]

    Enhancing domain generalization for robust machine-generated text detection

    Park, S., Han, S., and Cha, M. Enhancing domain generalization for robust machine-generated text detection. IEEE Transactions on Knowledge and Data Engineering, 2025

    2025

  73. [73]

    and Garg, N

    Peng, K. and Garg, N. Monoculture in matching markets. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. URL https://openreview.net/forum?id=P5yezHuMSS

    2024

  74. [74]

    Performative prediction

    Perdomo, J., Zrnic, T., Mendler-D \"u nner, C., and Hardt, M. Performative prediction. In International Conference on Machine Learning, pp.\ 7599--7609. PMLR, 2020

    2020

  75. [75]

    C., Bai, Y., Shao, E., and Wang, D

    Qian, Y., Wen, Z., Furnas, A. C., Bai, Y., Shao, E., and Wang, D. The rise of large language models and the direction and impact of us federal research funding. arXiv preprint arXiv:2601.15485, 2026

    Pith/arXiv arXiv 2026

  76. [76]

    Competition and diversity in generative AI

    Raghavan, M. Competition and diversity in generative AI . arXiv preprint arXiv:2412.08610, 2024

    Pith/arXiv arXiv 2024

  77. [77]

    Mixture proportion estimation via kernel embeddings of distributions

    Ramaswamy, H., Scott, C., and Tewari, A. Mixture proportion estimation via kernel embeddings of distributions. In International conference on machine learning, pp.\ 2052--2060. PMLR, 2016

    2052

  78. [78]

    AI use in american newspapers is widespread, uneven, and rarely disclosed

    Russell, J., Karpinska, M., Akinode, D., Thai, K., Emi, B., Spero, M., and Iyyer, M. AI use in american newspapers is widespread, uneven, and rarely disclosed. arXiv preprint arXiv:2510.18774, 2025 a

    Pith/arXiv arXiv 2025

  79. [79]

    People who frequently use chatgpt for writing tasks are accurate and robust detectors of AI -generated text

    Russell, J., Karpinska, M., and Iyyer, M. People who frequently use chatgpt for writing tasks are accurate and robust detectors of AI -generated text. arXiv preprint arXiv:2501.15654, 2025 b

    arXiv 2025

  80. [80]

    S., Kumar, A., Balasubramanian, S., Wang, W., and Feizi, S

    Sadasivan, V. S., Kumar, A., Balasubramanian, S., Wang, W., and Feizi, S. Can AI -generated text be reliably detected? arXiv preprint arXiv:2303.11156, 2023

    Pith/arXiv arXiv 2023

Showing first 80 references.