acceptodds
Under review as a conference paper at ICLR 2027

Normative Rewriting on the Reading Axis: How to Evade Chinese AI-Text Detectors While Preserving Quality

Abstract

Chinese, the largest unsegmented-script language, poses a fundamental challenge for AI-text detection: word boundaries are invisible, making a detector's reading unit unknown until measured. We measure the reading axis of six real Chinese detectors and establish a predictable link to their training objective. Detection-tuned generative LMs consistently read at the word scale (), while BERT classifiers remain axis-balanced ()—directly refuting the prevailing assumption that Chinese detectors are character-readers. Controlled contrasts confirm that this axis is governed by the objective, not tokenizer granularity. We further show that evasion efficacy is governed by coverage of this axis, and that saturated detectors resist deletion-based probing yet remain vulnerable to substitution, exposing deletion saliency as the wrong diagnostic. Under a normative constraint that preserves semantics, human-exemplar-guided beam rewriting lowers AUROC from 0.968 to 0.606, delivering the best quality-adjusted evasion efficiency among all baselines.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
AI-text detection, reading axis, Chinese detectors, evasion attack, saturation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_3gicOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper investigates how Chinese AI-text detectors respond to word-level and character-level perturbations. It introduces a “reading axis,” measured through the differential evasion index ξ\xi, a saliency-coverage index τ\tau, and multiscale sensitivity measurements. The authors also derive a bound on deletion-based probability sensitivity under sigmoid saturation and propose human-exemplar-guided iterative rewriting with semantic and formatting constraints. Experiments use a 5,000-text corpus comprising outputs from four generators and human-written Wikipedia sentences, with six publicly available Chinese detectors and additional controls. In the main rewriting comparison, the proposed method reduces the target detector’s AUROC from 0.968 to 0.606 while achieving a mean semantic cosine similarity of 0.903. The paper interprets these results as evidence that evasion depends on covering the detector’s native reading axis.

Critical strengths
  1. The paper identifies a worthwhile diagnostic question and proposes measurable behavioral comparisons. Comparing Chinese detectors under word-level and character-level perturbations is a useful way to examine assumptions about attack granularity. The study considers multiple released detectors and reports uncertainty rather than relying only on selected successful attacks. This provides a potentially valuable starting point for a more carefully controlled investigation of granularity-dependent sensitivity.

  2. The rewriting results suggest a potentially useful trade-off between detector-score reduction and semantic similarity. The main comparison evaluates methods on the same AI texts and human reference set while reporting both AUROC reduction and semantic similarity. Within that protocol, the full method improves both reported quantities relative to the exemplar-only variant. This is a promising empirical result, although semantic cosine similarity remains a proxy for meaning preservation, and the implementation and comparison conditions need clarification.

Critical weaknesses
  1. The proposed reading axis and axis-coverage principle are not adequately separated from perturbation strength and attack budget. The paper argues that replacing the same proportion of words or characters cancels attack strength and isolates reading granularity. However, equal proportions of substituted units do not equate realized character-edit distance, semantic distortion, fluency degradation, or replacement distributions. A positive $\xi$ may therefore reflect differences between the two attack procedures rather than an intrinsic word-reading property of the detector. The reported sign-change control is interesting, but it does not resolve these unmatched perturbation characteristics. The evidence for the axis-coverage principle has a related confound. Figure 3 increases the attack budget while observing increases in both $\tau$ and AUROC reduction. Under Equation (4), if saliency weights are fixed, total saliency is nonzero, and every unit has marginal inclusion probability rr, then $\mathbb{E}[\tau\mid x]=r.$ Thus, coverage increases mechanically with budget even without saliency-aware targeting. The observed correlation does not establish that $\tau$ explains evasion independently of budget. Fixed-budget comparisons, randomized-saliency controls, or analyses controlling for realized perturbation strength are needed to substantiate the central mechanism claim.

  2. The theoretical interpretation and several evaluation statements contain substantive inconsistencies. Equation (S6) is not a general identity for joint mutual information because it omits higher-order interactions. A direct counterexample is three independent uniform bits with $Y=U_1\oplus U_2\oplus U_3$. Every marginal information term and every pairwise correction appearing in the equation is zero, whereas $I(Y;U_1,U_2,U_3)=1$ bit. The equation therefore cannot support the claimed general information-theoretic explanation without additional assumptions or an explicit approximation argument. The evaluation also requires reconciliation. The manuscript reports TPR@1%FPR equal to 1 while reporting AUROCs substantially below 0.99. If these quantities come from the same examples, scores, and standard ROC definition, they cannot hold simultaneously: reaching TPR $=1$ at FPR $=0.01$ implies AUROC of at least 0.99. Different evaluation sets could explain the discrepancy, but that distinction must be explicit. Separately, Appendix S7 reports mean semantic similarities as low as 0.728 while describing the experiment as using the same semantic guard, whereas Algorithm 1 enforces similarity of at least 0.85 to the original text. These discrepancies affect confidence in the reported quality-constrained evaluation.

Questions for authors
  1. Which exact datasets and scoring functions produce the reported TPR@1%FPR and AUROC values? Please explain whether these values correspond to different experiments. Also, why do the Appendix S7 similarities fall below the hard constraint in Algorithm 1? Is the guard applied relative to the original text or the preceding rewrite?

  2. Does the word-versus-character conclusion remain after matching realized edit distance and semantic degradation? At a fixed attack budget, does $\tau$ predict evasion better than random coverage or coverage computed with shuffled saliency weights?

  3. What precisely implements Sample? Please specify the rewriting model, prompts, decoding settings, exemplar selection, separation between the exemplar library and evaluation references, and detector-query budgets for all methods. What happens when no candidate satisfies the semantic constraint?

  4. Is Equation (S6) intended to hold under additional distributional assumptions, or is it a truncated approximation? Which central conclusions remain independently supported after correcting this equation and restricting the saturation theorem to the perturbations covered by its assumptions?

Reviewer_z6NnOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies the robustness of Chinese AI-text detectors to rewriting at different linguistic granularities. It introduces a “reading-axis” framework that compares word- and character-level substitutions, measures attack coverage of detector saliency, and analyzes the limitations of deletion-based probing under probability saturation. Experiments on six Chinese detectors reveal differences in sensitivity to word- versus character-level attacks. The paper also proposes human-exemplar-guided iterative rewriting with detector feedback and semantic constraints.

Critical strengths
  1. The paper systematically compares character- and word-level perturbations across six Chinese detectors and identifies distinct sensitivity patterns. Its observation that near-zero deletion-induced probability changes can coexist with substantial vulnerability is particularly useful for interpreting robustness evaluations. These findings provide a worthwhile empirical contribution even if the proposed “reading-axis” interpretation requires further validation.

  2. On the main evaluation set, exemplar-guided rewriting reduces the target detector’s AUROC while maintaining an average semantic cosine similarity. The reported bootstrap interval supports a substantial effect, and the full method improves both evasion and similarity over exemplar guidance alone. These results support the practical potential of combining exemplar guidance, detector feedback, and semantic filtering.

Critical weaknesses
  1. Equal word- and character-substitution ratios do not ensure comparable perturbation severity, replacement quality, or semantic damage. Consequently, ξ measures the relative effectiveness of two attack procedures rather than uniquely identifying a detector’s intrinsic reading granularity. Likewise, ξ ≈ 0 does not establish balanced reliance on both granularities. This ambiguity undermines the paper’s central construct.

  2. The τ–evasion correlation mainly comes from increasing attack budgets, which can increase both quantities without demonstrating an independent coverage effect. Fixed-budget comparisons controlling for semantic damage are missing. The training-objective ablation also uses different learning rates despite claiming matched optimization settings, and does not report a direct statistical test of the difference in ξ. These limitations weaken the conclusions that coverage governs evasion and training objective determines the reading axis.

  3. The deletion-saturation bound does not establish that substitution escapes saturation; substitutions inducing bounded logit changes can exhibit the same compression. These issues compromise the theoretical support for both the coverage interpretation and the claimed diagnostic superiority of substitution.

  4. Algorithm 1 does not explicitly use ξ, τ, or measured saliency to guide candidate generation or selection. Thus, the experiments do not establish that measuring the reading axis contributes to the rewriting gains. Moreover, the algorithm retains only one candidate per round, leaving the “beam search” characterization unclear. This limits the demonstrated methodological contribution beyond exemplar-guided sampling and detector-score selection.

Questions for authors

see weaknesses

Reviewer_fg84OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies the robustness of Chinese AI-text detectors through their sensitivity to different perturbation units. It introduces a differential evasion index comparing word- and character-level substitutions, a saliency-coverage index, multiscale sensitivity measurements, and a bound explaining why probability-scale deletion saliency can become small under sigmoid saturation. Based on these observations, it proposes human-exemplar-guided rewriting that selects low-detector-score candidates subject to a semantic-similarity constraint. The experiments include six Chinese detectors and additional controls. On the main evaluation of 250 AI-generated texts and 200 human references, the proposed method reduces the target detector’s AUROC from 0.968 to 0.606, with semantic cosine similarity of 0.903 and reported quality-adjusted evasion efficiency of 3.73. The paper compares rewriting and substitution methods and includes objective and rewriting-component ablations. Its main claims concern the interpretation of a detector “reading axis,” the relationship between saliency coverage and evasion, and the quality-adjusted effectiveness of exemplar-guided rewriting.

Critical strengths
  1. The paper addresses a relevant robustness problem and reports useful empirical evidence. The study examines several Chinese detectors rather than a single target, and the main rewriting comparison uses common evaluation texts. The reported AUROC reduction of 0.362 and embedding cosine similarity of 0.903, is a promising empirical result. These observations motivate further investigation of detector robustness under content-preserving rewriting, even though stronger controls are needed to support the proposed mechanism.
  2. The saturation analysis identifies a useful diagnostic limitation. Under the stated bounded-logit-change assumption, the bound in Appendix S1 correctly shows that sigmoid saturation can make probability-scale leave-one-unit-out saliency small. This provides a useful warning that a small change in a saturated probability score need not imply insensitivity in the underlying model. The observation is consequential for evaluating and interpreting highly confident detectors.
Critical weaknesses
  1. The reading-axis and saliency-coverage interpretations are not sufficiently established. Equation (2) defines an operational contrast between two substitution procedures. Matching their nominal replacement ratios does not establish comparable realized edits, output quality, or search effort; ξ can therefore reflect properties of the attacks as well as the detector. The budget sweep in Appendix S10 does not resolve this confounding. Similarly, Figure 3 varies the perturbation budget while both saliency coverage and AUROC degradation increase. This association does not establish that coverage explains evasion beyond the amount of perturbation.
  2. The claimed quality-preserving advantage needs independent validation and clearer comparison controls. The same embedding cosine is used to constrain rewriting and to assess content preservation, and QAE derives its quality term from this measure (Sections 3.4 and 4.5). This does not independently establish preservation of the source text’s factual claims and meaning, or the fluency of the output. Thus, the QAE ranking in Table 3 supports a narrower conclusion than superior content-preserving evasion.
Questions for authors
  1. Does the qualitative word/character contrast in Table 1 persist after controlling for realized edit amount and output quality, under comparable search budgets?
  2. How does the reading-axis analysis contribute to the design or effectiveness of the proposed rewriting algorithm?
  3. Equation (S6) fails for three independent fair bits whose parity is Y: the joint information about Y is one bit, while all individual and pairwise terms on the right-hand side are zero. What corrected statement or additional assumptions are intended, and which claims connecting τ to evasion remain supported after this correction?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.