Simple techniques to bypass GenAI text detectors: implications for inclusive education
Open accessPublished version of a preprint
The DOI matches its Crossref record; 68 of the 68 references it cites were found in a registry or an index; 13 of the 40 sources its claims cite were checked for support.
Abstract
This study investigates the efficacy of six major Generative AI (GenAI) text detectors when confronted with machine-generated content modified to evade detection (n = 805). We compare these detectors to assess their reliability in identifying AI-generated text in educational settings, where they are increasingly used to address academic integrity concerns. Results show significant reductions in detector accuracy (17.4%) when faced with simple techniques to manipulate the AI generated content. The varying performances of GenAI tools and detectors indicate they cannot currently be recommended for determining academic integrity violations due to accuracy limitations and the potential for false accusation which undermines inclusive and fair assessment practices. However, these tools may support learning and academic integrity when used non-punitively. This study aims to guide educators and institutions in the critical implementation of AI text detectors in higher education, highlighting the importance of exploring alternatives to maintain inclusivity in the face of emerging technologies.
Overview
CoChat Summary
TL;DR
This experimental study tested seven AI text detectors on 114 samples, including 15 AI-generated texts, 89 adversarially modified AI texts, and 10 human controls. The detectors performed poorly overall and became substantially less accurate after simple adversarial techniques, raising concerns about fairness, false accusations, and the use of such tools in inclusive educational assessment.
Key findings
- Across unmanipulated AI-generated content, the mean detector accuracy was 39.5%; for human-written control samples, accuracy was 67%.
- After adversarial techniques were applied, mean detector accuracy fell to 22.14%; the paper also reports an overall 17.4% mean drop in accuracy.
- Some techniques were especially effective: adding spelling errors and increasing burstiness were described as highly effective, with manipulated-output accuracies of 12.9% and 15.9%, respectively; increase complexity had only a 2% drop in detectability and was the least effective.
- Detector performance varied widely: Copyleaks was the most accurate at 73.9% on non-manipulated output and 58.7% on manipulated output, while GPT Kit was the least accurate at 6% and 4.5%, respectively; Turnitin had the largest accuracy drop at 42.1%.
- Error analysis found that four of seven detectors did not misclassify any human-written samples, but Copyleaks misidentified 50% of human samples as AI-written; for undetected machine-generated cases, six of seven detectors had a UDC ratio above 50%, and Copyleaks still failed to identify 39% of AI-generated cases.
Methods
- Experimental design using three GenAI tools (GPT-4, Claude 2, Bard) to create 15 short text samples from five standardized prompts; those samples were then modified with six adversarial techniques to produce 89 additional AI-generated samples, alongside 10 human-written controls.
- All 114 samples were tested against seven AI text detectors: Turnitin, GPTZero, ZeroGPT, Copyleaks, Crossplag, GPT-2 Output Detector, and GPTKit; there were 805 tests in total and 797 valid tests.
- Accuracy was calculated using three methods—binary, semi-binary, and logarithmic—and the paper also conducted error analysis for false accusations and undetected cases.
Limitations
- The paper explicitly states that the study used a relatively small number of samples, only three GenAI tools and seven detectors, so the findings do not cover the full range of possible tools, styles, or manipulations.
- The authors say the results are only a snapshot in time because both GenAI tools and detectors evolve quickly; they may not remain stable for long.
- The paper notes that 8 samples could not be tested with some tools because of proprietary software issues; it also says some failures may have been due to text length limits, formatting incompatibilities, or temporary server issues.
- Evident from the design: the study used short, controlled samples and did not mirror the iterative real-world writing/editing process, which the authors say may mean real-world detectability could be even lower.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
10 Valid · 2 Not valid · 20 Not checked · 91 Own results and interpretations
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Mitchell et al. (2023) identified that DetectGPT correctly detected 70.3% of model-generated sequences from GPT2-XL.”
What the sources say
- Partly supportedcites Mitchell et al. (2023) sourceExperimental — accuracy not yet measured
Table 1 reports that DetectGPT detects GPT-2-generated samples, but its results are AUROC values, not the percentage of generated sequences correctly detected. The cited paper’s results do not substantiate the specific 70.3% figure. (LLM, conf=0.82)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“These include 'recursive paraphrasing attacks' , which utilise automated, network-based paraphrasing tools to reduce the accuracy of detectors on watermarked texts”
What the sources say
- Could not verifycites Solaiman et al. (2019) sourceExperimental — accuracy not yet measured
The paper’s ‘4.3 Detecting Synthetic Text’ section discusses adversaries using better generation systems to challenge detectors, but does not address recursive paraphrasing attacks. Searches for recursive paraphrasing and for watermark-related detector accuracy found no passage about automated paraphrasing tools reducing detection accuracy on watermarked text. (LLM, conf=0.73)
- Supportedcites Sadasivan et al. (2023) source
The Introduction (p. 3) describes an automated recursive attack using neural network-based paraphrasing and reports that it lowers detection rates for watermarked text. Section 2.3 (p. 7) likewise reports that repeated paraphrasing further degrades the watermark detector’s detection rate. (LLM, conf=0.94)
- Not matchedcites Ippolito etal. (2020)
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Could not verify
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“However, claims of accuracy are contradicted by studies which demonstrate the varied levels of the detectors' ability to distinguish accurately between AI and human-generated content. (Chaka, 2023a; Gao etal., 2022; Krishna etal., 2023; Orenstrakh etal., 2023; Perkins etal., 2023; Walters, 2023; Weber-Wulff etal., 2023).”
What the sources say
- Supportedcites Chaka, 2023a source
The Copyleaks AI Content Detector section reports the provider’s claim of 99.12% detection accuracy. The Discussion reports substantially varied results across the five detectors, including misclassifications and a detector that correctly classified only one text; the Implications and limitations section says they could not clearly distinguish AI-generated from human-produced texts. No page numbers are available in our copy. (LLM, conf=0.92)
- Not matchedcites Gao etal., 2022
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Not matchedcites Krishna etal., 2023
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
boosted- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“However, after altering the content using an Automated Paraphrasing tool (APT), the detection rate decreased to 4.6%.”
What the sources say
- Partly supportedcites Mitchell et al. (2023) sourceExperimental — accuracy not yet measured
Section 5.2 reports that detection performance generally declines as machine-generated text is increasingly revised. It simulates revisions by replacing text spans with T5-3B output and reports AUROC, not a detection rate of 4.6% after use of an Automated Paraphrasing Tool. No page number is provided in our copy. (LLM, conf=0.86)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“we noted that AI detectors performed significantly worse in our study, even with unedited text.”
What the sources say
- Supportedcites Weber-Wulff (2023) source
The Variations in accuracy section (p. 16) reports 74% accuracy for AI-generated texts with no further modifications, providing the cited study’s unedited-text benchmark. The statement that detectors performed worse in the citing authors’ own study is their comparison, not a result attributed to Weber-Wulff et al. (LLM, conf=0.87)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
53 more claims · 1 with their cited sources checked
Create a free account to see all 123 claims and to verify the 1 reference still open.
- Probably not a substantial claim. 41
- It may state two separate claims as one. 38
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?
The source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“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 Perkins et al., 2024
Experimental — accuracy not yet measuredThe cited source says what the claim attributes to it.
In the citing paper
“However, in the current state, reliable detection procedures are scarce (Elkhatat et al., 2023;Kofinas et al., 2025;Perkins et al., 2024).”
From Policy analysis of higher education institutes in Ontario, Canada: A focus on artificial intelligence (2025) · cites this paper as Perkins et al., 2024
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
“| Total/Average | 9 | 414 | 15 | 65.7 |”
Reported 414 · recomputed 503
Citations
Cites
A Survey on Evaluation of Large Language Models
Chang, Y., Wang, X., Wang, J., Wu, Y., Yang, L., Zhu, K., Chen, H., Yi, X., Wang, C., Wang, Y., et al. (2024) · ACM Transactions on Intelligent Systems and Technology · doi:10.1145/3641289 · Cited by 4,254
A Survey on Evaluation of Large Language Models
Chang, Y., Wang, X., Wang, J., Wu, Y., Yang, L., Zhu, K., Chen, H., Yi, X., Wang, C., Wang, Y., et al. (2023) · arXiv · doi:10.48550/arxiv.2307.03109 · Cited by 3,159 · cited 1×
“Despite these achievements, Claude 2 lags in scientific writing and quantitative accuracy compared to other models (Chang et al., 2023; Lozić & Štular, 2023; Z. Wu et al., 2023), primarily because of the ability of GPT-4 and Bard to access the internet for information gathering.”
MethodChatting and cheating: Ensuring academic integrity in the era of ChatGPT
Cotton, D. R. E., Cotton, P. A., & Shipway, J. R. (2024) · Innovations in Education and Teaching International · doi:10.1080/14703297.2023.2190148 · Cited by 3,087 · cited 1×
“…on problematising text-detection software from a perspective of inclusiv-ity, too much focus on GenAI as an emerging risk to academic integrity (Cotton et al., 2023; Perkins, 2023) may lead us to forget that extant threats to inclusivity and educational equity, such as contract cheating or…”
Chatting and Cheating. Ensuring academic integrity in the era of ChatGPT
Cotton, D., Cotton, P., & Shipway, J. R. (2023) · doi:10.35542/osf.io/mrz8h · Cited by 3,079
A comprehensive AI policy education framework for university teaching and learning
Chan, C. K. Y. (2023) · International Journal of Educational Technology in Higher Education · doi:10.1186/s41239-023-00408-3 · Cited by 1,577 · cited 1×
“…which accepts the use of GenAI in crafting manuscripts, albeit with certain restrictions (Per-kins & Roe, 2024a), and studies have shown the rising popularity of using GenAI tools to support academic writing and research (Bedington et al., 2024; Chan, 2023; Perkins & Roe, 2024b; Sobaih, 2024).”
Gao, C. A., Howard, F. M., Markov, N. S., Dyer, E. C., Ramesh, S., Luo, Y., & Pearson, A. T. (2022) · doi:10.1101/2022.12.23.521610 · Cited by 1,099 · cited 3×
“All samples created were tested against seven AI text detectors which had been previously identified in literature (Chaka, 2023a; Gao et al., 2022; Krishna et al., 2023; Orenstrakh et al., 2023; Perkins, Roe, et al., 2023; Weber-Wulff et al., 2023)as being somewhat effective at successfully…”
MethodGao, C. A., Howard, F. M., Markov, N. S., Dyer, E. C., Ramesh, S., Luo, Y., & Pearson, A. T. (2023) · npj Digital Medicine · doi:10.1038/s41746-023-00819-6 · Cited by 1,099
Perkins, M. (2023) · Journal of University Teaching and Learning Practice · doi:10.53761/1.20.02.07 · Cited by 839
Self-citationCan 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 810 · cited 2×
“…evade detection as ‘adversarial techniques’ in reference to the combination of methods that have been referred to as adversarial attacks in the extant literature related to attacks on machine learning based tools (Alsmadi et al., 2021; Qiu et al., 2019; Sadasivan et al., 2023; Wang et al., 2023).”
BackgroundMethodGPT 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 · cited 8×
“Detection tool biases against NNES have been highlighted as a potential issue of bias by Fröhling and Zubiaga (2021) and Liang et al. (2023) because of their reliance on standardised linguistic metrics such as perplexity and burstiness, which can disadvantage NNES.”
BackgroundInfluential
Research works printed but not linked: 3
- Originality.AI. (2023). AI Content Detector Accuracy Review + Open Source Dataset and Research Tool – Originality.AI. https://originality.ai
- Roe, J. (2024). AI and the Anthropological Imagination: Rethinking Education in the Digital Age. Open Anthropology Research Repository. https://openanthroresearch.org/index.php/oarr/preprint/view/399
- Sweenor, D., & Ramanathan, K. (2023). The CIO’s Guide to Adopting Generative AI: Five Keys to Success. TinyTechMedia LLC.
Cited by
Talk is cheap: why structural assessment changes are needed for a time of GenAI
Corbin, T., Dawson, P., & Liu, D. (2025) · Assessment & Evaluation in Higher Education · doi:10.1080/02602938.2025.2503964 · Cited by 107
The AI Assessment Scale (AIAS) in action: A pilot implementation of GenAI supported assessment
Furze, L., Perkins, M., Roe, J., & MacVaugh, J. (2024) · Australasian Journal of Educational Technology · doi:10.14742/ajet.9434 · Cited by 92
Self-citationGenerative AI tools and empowerment in L2 academic writing
Moorhouse, B. L., Wan, Y., Wu, C., Wu, M., & Ho, T. Y. (2025) · System · doi:10.1016/j.system.2025.103779 · Cited by 63
Bjelobaba, S., Waddington, L., Perkins, M., Foltýnek, T., Bhattacharyya, S., & Weber-Wulff, D. (2025) · International Journal for Educational Integrity · doi:10.1007/s40979-025-00191-w · Cited by 120
“…is not recommended owing to inherent problems with interpretation of the results, accuracy, and bias (Liang, 2023, Weber-Wulff et al. 2023; Perkins et al. 2024a, b), it is worth noting that the reliability of the detectors significantly drops if machine translation is used on human…”
Self-citationGenerative artificial intelligence: implications for biomedical and health professions education
Hersh, W. (2025) · Annual Review of Biomedical Data Science · doi:10.1146/annurev-biodatasci-103123-094756 · Cited by 28
Metadata
- Authors
- Perkins, M., Roe, J., Vu, B. H., Postma, D., Hickerson, D., McGaughran, J., & Khuat, H. Q.
- Affiliations
- British University Vietnam (VN)
- James Cook University Singapore (SG)
- Corresponding author
- Perkins, Mike
- Venue
- International Journal of Educational Technology in Higher Education
- Published
- 2024-09-09
- Volume
- 21
- Issue
- 1
- ISSN
- 2365-9440 (ISSN-L)
- OpenAlex
- W4402340204
- Cited by
- 117
- References
- 53
- Citations by year
- 2026: 64 · 2025: 47 · 2024: 6
- FWCI
- 48.31
- Type
- journal article
- Language
- en
- Open access
- diamond
- doi.org
- Licence
- cc-by
- creativecommons.org/licenses/by/4.0 (tdm)
- creativecommons.org/licenses/by/4.0 (vor)
- Locations
- International Journal of Educational Technology in Higher Education (publishedVersion, cc-by)
- arXiv (Cornell University) (submittedVersion, cc-by-nc-nd)
- ResearchOnline - JCU (James Cook University) (submittedVersion)
- DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion)
- Durham Research Online (Durham University) (submittedVersion)
- arXiv (Cornell University)
- Topics
- Social Sciences › Education › Academic integrity and plagiarism
- Social Sciences › Education › Artificial Intelligence in Education
- Physical Sciences › Computer Science › Artificial Intelligence › Authorship Attribution and Profiling
- Keywords
- academic integrity, false accusation, inclusive assessment, higher education, assessment fairness
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.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol).
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
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: Simple techniques to bypass GenAI text detectors: implications for inclusive education. https://app.cochat.ai/papers/10.1186/s41239-024-00487-w
- The paper
- Perkins, M., Roe, J., Vu, B. H., Postma, D., Hickerson, D., McGaughran, J., & Khuat, H. Q. (2024). Simple techniques to bypass GenAI text detectors: implications for inclusive education. International Journal of Educational Technology in Higher Education. https://doi.org/10.1186/s41239-024-00487-w
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.