Robustness of generative AI detection: adversarial attacks on black-box neural text detectors
Open accessAbstract
The increased quality and human-likeness of AI generated texts has resulted in a rising demand for neural text detectors, i.e. software that is able to detect whether a text was written by a human or generated by an AI. Such tools are often used in contexts where the use of AI is restricted or completely prohibited, e.g. in educational contexts. It is, therefore, important for the effectiveness of such tools that they are robust towards deliberate attempts to hide the fact that a text was generated by an AI. In this article, we investigate a broad range of adversarial attacks in English texts with six different neural text detectors, including commercial and research tools. While the results show that no detector is completely invulnerable to adversarial attacks, the latest generation of commercial detectors proved to be very robust and not significantly influenced by most of the evaluated attack strategies.
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
Integrity signals
Registry
Cites
Statistics
Claims the paper makes
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
Even minor textual alterations, such as synonym replacement, sentence reconfiguration, or translation cycles, can markedly diminish detection accuracy.
“Studies further illustrate that even minor textual alterations, such as synonym replacement, sentence reconfiguration, or translation cycles, can markedly diminish detection accuracy.”
From Harms and Ethical Risks of AI Detection Tools in Education: False Positives, Bias, Surveillance, and Student Rights (2026) · cites this paper as Fishchuk & Braun, 2024
Section 4.8 reports that paraphrasing sharply lowered scores for the GPT-2 Output Detector, but Copyleaks and Radar resisted it. Section 4.7 says translation had limited effects on most detectors, though some Japanese and Russian translations lowered GPT-2 detector scores. The paper does not separately establish the claimed effect for synonym replacement, so the sentence overgeneralizes across alterations and detectors. (LLM, conf=0.87)
Experimental — accuracy not yet measuredA recent study showed that no detector is completely invulnerable to adversarial attacks, but a newer version of Copyleaks resisted most types of attacks more often than not.
“Although no detector is completely invulnerable to adversarial attacks, the authors show that a newer version of Copyleaks (n.d.) resisted most types of attacks more often than not but lacked proper statistics.”
From AI for scientific integrity: detecting ethical breaches, errors, and misconduct in manuscripts (2025) · cites this paper as Fishchuk and Braun, 2024
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
The self-check needs the paper’s full text.
Cites
Explaining and Harnessing Adversarial Examples
Goodfellow, I. J., Shlens, J., & Szegedy, C. (2014) · arXiv · doi:10.48550/arxiv.1412.6572 · Cited by 10,381 · see citing works
Advances in Neural Information Processing Systems 14
(2002) · The MIT Press eBooks · doi:10.7551/mitpress/1120.001.0001 · Cited by 9,184 · see citing works
Intriguing properties of neural networks
Szegedy, C., Zaremba, W., Sutskever, I., Bruna, J., Erhan, D., Goodfellow, I., & Fergus, R. (2013) · arXiv · doi:10.48550/arxiv.1312.6199 · Cited by 8,143 · see citing works
Jin, D., Jin, Z., Zhou, J. T., & Szolovits, P. (2020) · Proceedings of the AAAI Conference on Artificial Intelligence · doi:10.1609/aaai.v34i05.6311 · Cited by 1,108 · see citing works
HotFlip: White-Box Adversarial Examples for Text Classification
Ebrahimi, J., Rao, A., Lowd, D., & Dou, D. (2018) · Proceedings of the conference - Association for Computational Linguistics. Meeting · doi:10.18653/v1/p18-2006 · Cited by 948 · see citing works
Adversarial Attacks and Defenses in Images, Graphs and Text: A Review
Xu, H., Ma, Y., Liu, H., Deb, D., Liu, H., Tang, J., & Jain, A. K. (2020) · Machine Intelligence Research · doi:10.1007/s11633-019-1211-x · Cited by 716 · see citing works
Black-Box Generation of Adversarial Text Sequences to Evade Deep Learning Classifiers
Gao, J., Lanchantin, J., Soffa, M. L., & Qi, Y. (2018) · 2018 IEEE Security and Privacy Workshops (SPW) · doi:10.1109/spw.2018.00016 · Cited by 693 · see citing works
GPT detectors are biased against non-native English writers
Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., & Zou, J. (2023) · Patterns · doi:10.1016/j.patter.2023.100779 · Cited by 673 · see citing works
Testing of detection tools for AI-generated text
Weber-Wulff, D., Anohina-Naumeca, A., Bjelobaba, S., Foltýnek, T., Guerrero-Dib, J., Popoola, O., Šigut, P., & Waddington, L. (2023) · International Journal for Educational Integrity · doi:10.1007/s40979-023-00146-z · Cited by 582 · cited 5× · see citing works
“Habibzadeh (2023) reports an accuracy of 0.80 for the detector, while Weber-Wulff et al. (2023) report an accuracy of only 54% for the very same tool.”
InfluentialElkhatat, A. M., Elsaid, K., & Almeer, S. (2023) · International Journal for Educational Integrity · doi:10.1007/s40979-023-00140-5 · Cited by 392 · cited 2× · see citing works
“In the study by Elkhatat et al. (2023), the best-performing detector achieves a recall of 93% and a precision of 80%, significantly higher than the results reported by Weber-Wulff et al. (2023).”
Advances in Adversarial Attacks and Defenses in Computer Vision: A Survey
Akhtar, N., Mian, A. S., Kardan, N., & Shah, M. (2021) · IEEE Access · doi:10.1109/access.2021.3127960 · Cited by 319 · see citing works
A Watermark for Large Language Models
Kirchenbauer, J., Geiping, J., Wen, Y., Katz, J., Miers, I., & Goldstein, T. (2023) · arXiv · doi:10.48550/arxiv.2301.10226 · Cited by 194 · see citing works
Can AI-Generated Text be Reliably Detected?
Sadasivan, V. S., Kumar, A., Balasubramanian, S., Wang, W., & Feizi, S. (2023) · arXiv · doi:10.48550/arxiv.2303.11156 · Cited by 161 · see citing works
Automatic Detection of Machine Generated Text: A Critical Survey
Jawahar, G., Abdul-Mageed, M., & Lakshmanan, V. S. (2020) · Proceedings of the 28th International Conference on Computational Linguistics · doi:10.18653/v1/2020.coling-main.208 · Cited by 156 · see citing works
Walters, W. H. (2023) · Open Information Science · doi:10.1515/opis-2022-0158 · Cited by 109 · see citing works
A Survey on Adversarial Attack in the Age of Artificial Intelligence
Kong, Z., Xue, J., Wang, Y., Huang, L., Niu, Z., & Li, F. (2021) · Wireless Communications and Mobile Computing · doi:10.1155/2021/4907754 · Cited by 83 · see citing works
Habibzadeh, F. (2023) · Journal of Korean Medical Science · doi:10.3346/jkms.2023.38.e319 · Cited by 68 · see citing works
RADAR: Robust AI-Text Detection via Adversarial Learning
Hu, X., Chen, P.-Y., & Ho, T.-Y. (2023) · Advances in Neural Information Processing Systems 36 · doi:10.52202/075280-0662 · Cited by 62 · see citing works
Untargeted, Targeted and Universal Adversarial Attacks and Defenses on Time Series
Rathore, P., Basak, A., Nistala, S. H., & Runkana, V. (2020) · International Joint Conference on Neural Networks (IJCNN) · doi:10.1109/ijcnn48605.2020.9207272 · Cited by 53 · see citing works
Yoo, K., Kim, J., Jang, J., & Kwak, N. (2022) · Findings of the Association for Computational Linguistics: ACL 2022 · doi:10.18653/v1/2022.findings-acl.289 · Cited by 33 · see citing works
Adversarial Robustness of Neural-Statistical Features in Detection of Generative Transformers
Crothers, E., Japkowicz, N., Viktor, H. L., & Branco, P. (2022) · International Joint Conference on Neural Networks (IJCNN) · doi:10.1109/ijcnn55064.2022.9892269 · Cited by 30 · see citing works
An Empirical Study of AI Generated Text Detection Tools
Akram, A. (2023) · arXiv · doi:10.48550/arxiv.2310.01423 · Cited by 22 · see citing works
Attacking Neural Text Detectors
Wolff, M., & Wolff, S. (2020) · arXiv · doi:10.48550/arxiv.2002.11768 · Cited by 18 · see citing works
AI tools vs AI text: Detecting AI-generated writing in foot and ankle surgery
Cooperman, S. R., & Brandão, R. A. (2024) · Foot & Ankle Surgery Techniques Reports & Cases · doi:10.1016/j.fastrc.2024.100367 · Cited by 10 · see citing works
Enhancing Neural Text Detector Robustness with μAttacking and RR-Training
Liang, G., Guerrero, J., Zheng, F., & Alsmadi, I. M. (2023) · Electronics · doi:10.3390/electronics12081948 · Cited by 8 · see citing works
Efficient text-based evolution algorithm to hard-label adversarial attacks on text
Peng, H., Wang, Z., Zhao, D., Wu, Y., Han, J., Guo, S., Ji, S., & Zhong, M. J. (2023) · Journal of King Saud University - Computer and Information Sciences · doi:10.1016/j.jksuci.2023.03.017 · Cited by 7 · see citing works
Efficient Black-Box Adversarial Attacks on Neural Text Detectors
Fishchuk, V., & Braun, D. (2023) · arXiv · doi:10.48550/arxiv.2311.01873 · Cited by 3 · see citing works
Self-citationTextDefense: Adversarial Text Detection based on Word Importance EntropyPreprint
Shen, L., Zhang, X., Ji, S., Pu, Y., Ge, C., Yang, X., & Feng, Y. (2023) · arXiv (Cornell University) · doi:10.48550/arxiv.2302.05892 · Cited by 3 · see citing works
Flagged first, then by citations
Research works printed but not linked: 4
- Emi, B., & Spero, M. (2024). Technical report on the Checkfor.ai AI- generated text classifier. arXiv preprint arXiv: 2402. 14873
- Liang, G., Guerrero, J., & Alsmadi, I. (2023a). Mutation-based adver- sarial attacks on neural text detectors. arXiv preprint arXiv: 2302. 05794
- Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., & Zou, J. (2023c). GPT detectors are biased against non-native english writers. In ICLR 2023 workshop on trustworthy and reliable large-scale machine learning models.
- Solaiman, I., Brundage, M., Clark, J., Askell, A., Herbert-Voss, A., & Wu, J. (2019). Release strategies and the social impacts of lan- guage models. arXiv preprint arXiv: 1908. 09203
Cited by
Sorted by citations
Fiedler, A., & Döpke, J. (2025) · International Review of Economics Education · doi:10.1016/j.iree.2025.100321 · Cited by 40 · see citing works
AI for scientific integrity: detecting ethical breaches, errors, and misconduct in manuscripts
Pellegrina, D., & Helmy, M. (2025) · Frontiers in Artificial Intelligence · doi:10.3389/frai.2025.1644098 · Cited by 30 · see citing works
Makhmutova, A., Sharimbayev, B., Amirzhanov, A., & Shalkarbay-uly, A. (2026) · Journal of Advances in Information Technology · doi:10.12720/jait.17.3.438-449 · Cited by 6 · see citing works
Aravind, D., Babulak, E., & Suresh, N. V. (2026) · Advances in computational intelligence and robotics book series · doi:10.4018/979-8-3373-7852-7.ch004 · Cited by 1 · see citing works
GİRAY, L., CANER, H. N., & AMOOZEGAR, A. (2026) · Journal of Narrative and Language Studies · doi:10.59045/nalans.2026.109
Evaluating Robustness of Neural Text Detectors in Generative AI Detection
Parasdeep, Verma, M. K., Mehra, R., Soni, K., & Kumar, M. (2025) · International Conference on Computing, Communication and Automation · doi:10.1109/iccca66364.2025.11325575
MDD-TD: A dataset of large language model text source and content authenticity detection
Li, S., & Ning, M. A. (2026) · China Scientific Data · doi:10.11922/11-6035.csd.2025.0191.zh
Reimagining Peer Review in the Age of AI: Lessons from a Graduate Education Experiment: Pre-Results
Freeshah, M., Osama, N., Masoud, A. R., Bediwy, A., & Abdelsamie, M. (2026) · IEEE Global Engineering Education Conference (EDUCON) · doi:10.1109/educon67543.2026.11574319
A Systematic Review and Comparative Analysis of AI-Generated Text Detection in Education
Daoudi, H., Riad, M., Salhi, I., Gouraguine, S., & Qbadou, M. (2026) · International Conference on Circuit, Systems and Communication (ICCSC) · doi:10.1109/iccsc71566.2026.11650035
- Authors
- Fishchuk, V., & Braun, D.
- Affiliations
- University of Twente (NL)
- Corresponding author
- Fishchuk, Vitalii
- Braun, Daniel
- ORCID
- Braun, Daniel
- Venue
- International Journal of Speech Technology
- Published
- 2024-10-16
- Volume
- 27
- Issue
- 4
- Pages
- 861-874
- ISSN
- 1381-2416 (ISSN-L)
- 1572-8110
- OpenAlex
- W4403457328
- Cited by
- 7
- References
- 28
- Citations by year
- 2026: 5 · 2025: 2
- FWCI
- 1.04
- Type
- journal article
- Language
- en
- Open access
- hybrid
- doi.org
- Locations
- International Journal of Speech Technology (publishedVersion, cc-by)
- University of Twente Research Information (submittedVersion, cc-by)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Adversarial Robustness in Machine Learning
- Physical Sciences › Computer Science › Artificial Intelligence › Authorship Attribution and Profiling
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Digital Media Forensic Detection
- Keywords
- AI-generated text detection, adversarial attacks, robustness, generative AI, English text
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
