REVIEW 3 major objections 6 minor 8 references
Cat and Mouse -- Can Fake Text Generation Outpace Detector Systems?
T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash
Pith's one-line read A half-version upgrade to Gemini made fake detective fiction far harder for statistical classifiers to spot, while newer GPT models gained no deceptive edge over GPT 3.5.
desk verdict Gemini 2.0's deception gain is real but length-confounded; the GPT plateau is too noisy to carry the architecture claim. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The carrying mechanism is the controlled rewrite-and-detect loop. Human text from out-of-copyright detective novels is cut into about 100-word excerpts; each LLM receives a fixed prompt asking for a same-length rewrite; generated texts are pruned so human and AI sets match in average length, median, and standard deviation; and each set is split 80/20 and scored by four classical classifiers—Support Vector Machine, Random Forest, Naive Bayes, and MLP. The central metric is the Deception Rate, $D = 1 - \mathrm{Recall}$, which isolates the rate at which AI-generated text is mistaken for human writing. The paper's conclusions hang on comparing that rate across versions of the same model family while keeping the prompt, genre, and balancing procedure fixed.
What would settle it
Re-run the experiment with many random 80/20 splits and bootstrap confidence intervals for accuracy and Deception Rate; the central claim collapses if the Gemini 2.0-versus-1.5 gap regularly drops below 10 percentage points or overlaps the GPT 3.5-versus-4.1 gap. An even sharper test is the paper's proposed self-hosted experiment: fix one architecture, increase parameter count incrementally, and see whether deceptive ability rises as it did for Gemini.
Extended reading notes
Core claim
The discovery is a divergence in how two model families improve with version: Gemini 2.0-flash rewrote detective-fiction excerpts into text that classical classifiers misclassified as human much more often than Gemini 1.5-flash did, with accuracy dropping from 95.25 to 82.25 (Random Forest), 96.25 to 85.50 (SVM), 97.00 to 83.00 (MLP), and 96.25 to 79.75 (Naive Bayes). GPT models did not show this trend: GPT 4.1 was no harder to detect than GPT 3.5, and the speed-optimized GPT 4o-mini was the easiest to catch, with the MLP classifier reaching almost 98% accuracy. The paper interprets this as a sign that GPT-style models may be hitting a plateau in deceptive ability despite a roughly tenfold growth in parameters, while the Gemini architecture is better able to convert scale into more human-like text. It defines the Deception Rate as $D = 1 - \mathrm{Recall}$, the proportion of AI text mistaken for human writing, and uses it to compare generators on the failure mode that matters for disinformation.
Load-bearing premise
The load-bearing premise is that a single 80/20 split of one small, curated corpus gives a stable measure of each model's deceptiveness, and that the two Gemini flash models differ mainly by version and architecture rather than by undisclosed changes in training, distillation, or prompt behavior.
Editorial extensions
If this is right
- If GPT-family models have plateaued in deceptiveness, then classifiers like SVM and Random Forest may keep identifying GPT-generated text without frequent retraining.
- If Gemini's improvement is architectural, then each new model family may require fresh detector development rather than a one-time fix.
- The Deception Rate gives a practical, per-family measure for tracking the failure mode that matters: AI text that reads as human.
- The resource gap—simple classifiers versus billion-parameter models with heavy energy costs—means the detector side of the race can remain competitive if it focuses on the right features.
Reading between the lines
- A natural extension is to run the same rewrite-and-detect protocol on news, product reviews, or academic abstracts; the paper's architecture explanation predicts the Gemini-versus-GPT gap will not be limited to detective fiction.
- Because the prompt was optimized for GPT and reused for Gemini, part of the measured gap could be prompt-model fit; testing each family with its own tuned prompt would show how much of the effect is intrinsic to the model.
- The paper's suggested self-hosted experiment—increasing parameter count while holding architecture fixed—is the decisive test of its claim that architecture, not scale, drives deceptiveness.
- Detector design might shift from generic AI-text detection to model-family fingerprints, since different families appear to leave different statistical traces in rewritten prose.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper asks whether later versions of large language models (LLMs) become better at generating text that evades statistical detectors, using the domain of classical detective fiction. The authors create datasets by rewriting Agatha Christie excerpts with GPT 3.5-turbo, GPT 4o-mini, GPT 4.1, Gemini 1.5-flash, and Gemini 2.0-flash, then balance the text lengths post hoc. They train four classical classifiers (Random Forest, SVM, MLP, Naive Bayes) on an 80/20 split and report accuracy and a 'Deception Rate' D = 1 - Recall, defined as the proportion of AI-generated texts misclassified as human. The main empirical finding is that Gemini 2.0 shows a drop of more than 10 percentage points in detector accuracy relative to Gemini 1.5 across all four classifiers, while GPT 4.1 does not show an improvement over GPT 3.5. The authors interpret this as evidence that GPT-like models may be plateauing in deceptive ability and that model architecture, not just scale, matters for evading detection.
Significance. If the central finding holds, it is a useful empirical contribution to the debate about whether LLM text generation will outpace detection systems. The paper makes a direct version-over-version comparison within two model families, and the Deception Rate metric, while being a simple transformation of recall, makes the core quantity of interest transparent. The authors are also honest about the limitations posed by undisclosed model parameters and the unavailability of older model versions. On the other hand, the paper provides no uncertainty quantification, no repeated splits, and no length-matched control for the main Gemini comparison; the cross-family conclusion is further complicated by the use of different human text subsets. The significance of the result is therefore conditional on these issues being resolved. I find no circularity in the metric: D = 1 - Recall is a direct definition, not a fitted parameter.
major comments (3)
- [Section 3, Table 1] The central Gemini-2.0 deception gain is confounded with residual text length. After the balancing procedure, the balanced Gemini-2.0 AI texts average 490.98 characters, the balanced Gemini-1.5 AI texts average 479.97 characters, and the balanced human texts average 509.19 characters; Gemini-2.0 is therefore roughly 11 characters closer to the human mean than Gemini-1.5. The accuracy drops reported in Table 3 (for example, Random Forest from 95.25% to 82.25%, and Naive Bayes from 96.25% to 79.75%) are exactly in the direction one would expect if the classifiers partially key on length-related features. The paper does not report a length-matched analysis, such as subsampling Gemini-2.0 texts to match the Gemini-1.5 length distribution, or adding length as a covariate. Until such an analysis is provided, the conclusion in Sections 5 and 6 that the improvement is due to architectural deceptiveness rather than a surface length artifact is not supported.
- [Section 4, Tables 3 and 4] The entire comparison rests on a single 80/20 split of one corpus, with no confidence intervals, significance tests, or repeated random splits. The Gemini accuracy drops are large and consistent across classifiers, but the GPT 3.5 versus GPT 4.1 differences are small (for example, Random Forest 89.48% versus 90.78%, and SVM 92.94% versus 93.95%) and could easily change sign under resampling. The word 'significant' in Section 6 ('no significant improvement') is not backed by any statistical test. The authors should report repeated-split distributions or bootstrap confidence intervals for the accuracy and Deception Rate values, and ideally state whether the GPT plateau, and the Gemini increase, survive when the test set is resampled.
- [Section 3, Table 1] The GPT and Gemini experiments use different human subsets, so cross-family comparisons are not controlled. The balanced human set for the GPT experiment has 1735 rows with mean length 579.70 characters, while the balanced human set for the Gemini experiment has 1000 rows with mean length 509.19 characters. Accuracy and Deception Rate are therefore not directly comparable between GPT and Gemini: the statement in Section 5 that 'the Gemini models proved more effective at generating human-like text' conflates detector performance on different human test distributions. The within-family version comparisons (GPT 3.5 versus 4.1, and Gemini 1.5 versus 2.0) remain valid in principle, but any architecture-level conclusion across families needs a common or carefully matched human test set.
minor comments (6)
- [Section 3] The paper says excerpts were 'approximately 100 word excerpts', but the reported mean lengths in Table 1 are in characters (574.31, 502.21, etc.). Please clarify the unit and how the chunking was performed.
- [Section 3] The prompt used for generation was optimized on OpenAI and kept the same for Gemini. This is a reasonable consistency choice, but it means the cross-family comparison in Section 5 may reflect prompt compatibility rather than intrinsic model capability. A prompt-sensitivity check, or at least an explicit caveat, would strengthen the architecture-level conclusion.
- [Section 4, Table 4] The row for GPT 4.1 is labelled 'GPT 4.1 Balanced' while the other rows are not; please make the labels consistent. Also, 'GPT 4o mini' should be written as 'GPT 4o-mini' consistently.
- [Section 4] The paper states that the false results were mostly AI-generated texts misclassified as human, but it does not provide confusion matrices or per-class error rates. Reporting false-positive and false-negative rates would help the reader assess the direction of the errors, especially because the Deception Rate metric only captures one side.
- [Methodology (global)] No API call parameters (temperature, top-p, max tokens), random seeds, or data/code availability statement are provided. For a study with only one split and no repeated trials, these details are important for reproducibility.
- [Abstract and Section 6] The phrase 'over a 0.5 version increase' is imprecise for the GPT family, which goes from 3.5 to 4o-mini to 4.1 rather than through a single incremental version step. Please state more precisely which models are being compared and how the version increment is defined.
Circularity Check
No significant circularity: the accuracy comparison and Deception Rate are direct measurements, with prior self-citation only as background.
full rationale
No circular derivation is present. The paper's central comparison is an empirical measurement: classifier accuracy on an 80/20 split of balanced human and LLM-generated text. The Deception Rate D = 1 - Recall is a direct algebraic transformation of the measured recall, not a fitted parameter, and Table 4 simply restates Table 3 in a different form. The GPT/Gemini version comparison is independent of any parameter fitted to the same data. The only self-referential element is citation [5], the authors' earlier study, which is used for background and for noting the difficulty of controlling generated-text length; it does not supply the accuracy values or the version comparison. The plateau/architecture conclusion is an interpretation of the measured drop, not an input to the calculation. The reviewer concern about text-length imbalance (Gemini 2.0 balanced texts are roughly 11 characters closer to human length than Gemini 1.5) is a possible confound and a correctness/robustness risk, but it is not circularity because the paper does not define deceptive ability as length closeness or derive the accuracy drop from length. Therefore the circularity score is 0.
Assumptions & free parameters
assumptions (4)
- domain assumption The six Agatha Christie novels are a representative sample of creative writing for drawing conclusions about fake-text detection generally.
- domain assumption A single 80/20 split of each dataset yields stable accuracy and Deception Rate estimates for comparing model versions.
- domain assumption The four statistical classifiers with unspecified hyperparameters provide a consistent detector benchmark across all models.
- ad hoc to paper The prompt optimized for OpenAI models is a neutral stimulus when applied to Gemini models.
Cite this review
Pith. "Pith review of Cat and Mouse -- Can Fake Text Generation Outpace Detector Systems?." pith.science (2026). https://pith.science/paper/TLIBPUCP
@misc{pith2026250621274,
author = {Pith},
title = {Pith review of: Cat and Mouse -- Can Fake Text Generation Outpace Detector Systems?},
year = {2026},
howpublished = {\url{https://pith.science/paper/TLIBPUCP}},
note = {Machine review of arXiv:2506.21274}
}
read the original abstract
Large language models can produce convincing "fake text" in domains such as academic writing, product reviews, and political news. Many approaches have been investigated for the detection of artificially generated text. While this may seem to presage an endless "arms race", we note that newer LLMs use ever more parameters, training data, and energy, while relatively simple classifiers demonstrate a good level of detection accuracy with modest resources. To approach the question of whether the models' ability to beat the detectors may therefore reach a plateau, we examine the ability of statistical classifiers to identify "fake text" in the style of classical detective fiction. Over a 0.5 version increase, we found that Gemini showed an increased ability to generate deceptive text, while GPT did not. This suggests that reliable detection of fake text may remain feasible even for ever-larger models, though new model architectures may improve their deceptiveness
Reference graph
Works this paper leans on
-
[1]
Creamer, E. (2025), ‘Meta has stolen books’: authors to protest in London against AI trained using ‘shadow library’. https://www.theguardian.com/books/2025/ apr/03/meta-has-stolen-books-authors-to-protest-in-london-against-ai-trained- using-shadow-library
work page 2025
-
[2]
N., Japkowicz, N., & Viktor, H
Crothers, E. N., Japkowicz, N., & Viktor, H. L. (2023). Machine-generated text: A comprehensive survey of threat models and detection methods. IEEE Access, 11, 70977-71002
work page 2023
-
[3]
Jegham, N., Abdelatti, M., Elmoubarki, L., & Hendawi, A. (2025). How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference. arXiv preprint arXiv:2505.09598
arXiv 2025
-
[4]
Koplin, J. J. (2023). Dual-use implications of AI text generation. Ethics and In- formation Technology, 25(2), 32. https://doi.org/10.1007/s10676-023-09703-z
-
[5]
McGlinchey, A. C. & Barclay, P. J. (2024). Using Machine Learning to Distinguish Human-written from Machine-generated Creative Fiction. arXiv.Org. https://doi. org/10.48550/arXiv.2412.15253
work page Pith review arXiv doi:10.48550/arxiv.2412.15253 2024
-
[6]
Rahman, A., Mahir, S. H., Tashrif, M. T. A., Aishi, A. A., Karim, M. A., Kundu, D., ... & Eidmum, M. D. (2025). Comparative analysis based on deepseek, chat- gpt, and google gemini: Features, techniques, performance, future prospects. arXiv preprint arXiv:2503.04783
arXiv 2025
- [7]
-
[8]
L., Virgolin, M., & Suvanto, M
Wahde, M., Della Vedova, M. L., Virgolin, M., & Suvanto, M. (2024). An inter- pretable method for automated classification of spoken transcripts and written text. Evolutionary Intelligence, 17(1), 609-621
work page 2024
Reviewed August 6, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.