acceptodds
Under review as a conference paper at ICLR 2027

WHY PERFECT AI-TEXT DETECTION IS IMPOSSIBLE: INFORMATION-THEORETIC LIMITS AND A CALI- BRATED REPORTING PROTOCOL

Abstract

We recast post-hoc AI-text detection from a moving engineering target into a quantifiable, distribution-dependent limit imposed by the entropy of natural lan- guage, and we deliver a deployment protocol that respects that limit. Our first result is a Bayes-error ceiling: any binary authorship classifier under equal pri- ors has error at least Emin = 1 2 (1 − ∥PH − PAI∥TV), so as large language models (LLMs) close the total-variation distance between their output distribu- tion PAI and the human distribution PH on the observable feature manifold, this floor rises toward 1/2 regardless of detector architecture, training data, or com- pute. Our second result is a logically independent rate–distortion floor on the paraphrase channel: measured mutual information I(H; AI) ≈ 2.71 bits per pas- sage places the channel at only ≈ 4% of its zero-information distortion floor, cer- tifying strictly positive preserved source information—a channel-fidelity quantity distinct from the marginal separability that governs detectability. We ground both bounds on a pre-registered paired corpus of 4,072 human/AI passages across six domains, generated by three open-weight models at three sampling temperatures, with a broader eight-model panel (open-weight and proprietary) for a domain- controlled robustness test. Under Benjamini–Hochberg FDR control only 8 of 12 detector/domain cells remain significant: a fine-tuned RoBERTa degrades from AUC= 0.978 globally to 0.785 on news and inverts under cross-domain trans- fer and adversarial perturbation, zero-shot logit detectors collapse to chance, and a simulated human baseline reaches only AUC= 0.542. On the two bounds we build a deployable conformal reporting protocol that issues calibrated probabili- ties with abstain regions rather than binary verdicts, with a distribution-free cov- erage guarantee. Finally, as an explicitly negative finding, we test a content/style complementarity hypothesis suggested by an uncertainty-principle analogy: the domain-pooled product K = ∆C · ∆S = 5.83×10−5 superficially matches the prediction, but an eight-model single-domain re-run reverses its sign (r = +0.41 to +0.72), exposing it as a domain-heterogeneity confound. We report this failure directly, as a methodological caution for stylometric trade-off claims.

Primary area:
foundation or frontier models, including LLMs
Keywords:
AI-generated text detection, authorship attri- bution, information-theoretic limits, Bayes error, total variation distance, rate–distortion, conformal prediction, selective classifi- cation, large language models, stylometry, media forensics

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_sUwPOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies the theoretical limits of AI-generated text detection using Bayes error and rate–distortion theory.

Critical strengths

The paper addresses an important problem with practical implications. The motivation for uncertainty-aware reporting is reasonable, and I appreciate the authors' willingness to report a negative experimental finding.

Critical weaknesses

W1. Limited novelty and questionable theoretical soundness. The main Bayes-error result is a standard result from statistical hypothesis testing, and the rate–distortion bound follows directly from its definition. And the proposed selective-classification guarantee appears incorrect: a bound on the Bayes-optimal error does not guarantee the error rate of an actual classifier.

W2. Unconvincing experiments and serious citation issues. I found several results are hard to understand, including three detectors achieving exactly 0.500 AUC and rule-based paraphrasing improving AUC from 0.472 to 1.000. More concerningly, I found multiple incorrect references. The arXiv IDs for Fraser et al. (2501.05970), Opara et al. (2502.16105), Przystalski et al. (2502.10218), and Zhang et al. (2502.09861) all correspond to unrelated papers.

Questions for authors

See W1-W2.

1 comment

We thank the reviewer, and we take W2's citation concern with the seriousness it deserves. We address all points candidly.

Citations (research integrity). The reviewer is correct. We verified the flagged arXiv identifiers and confirm that 2501.05970, 2502.10218, 2502.16105, and 2502.09861 point to unrelated papers. This is a serious error and we take full responsibility. We are conducting a complete audit of every reference, verifying each title, author list, venue, and identifier against the primary source, and replacing or removing any entry we cannot confirm. The claims these four references supported are being re-grounded in verified literature or removed if no correct source exists. We will also correct our AI-use disclosure to accurately reflect how the bibliography was assembled. We are grateful to the reviewer for catching this and are committed to a corrected, fully verifiable reference list. W1 soundness of the selective-risk guarantee. This is a valid and important correction. The reviewer is right that a bound on the Bayes-optimal error does not bound the error of an actual classifier. Our protocol has two distinct components, and we conflated them. (1) The conformal coverage guarantee Pr[y ∈ C(x)] ≥ 1−β is valid, distribution-free, and holds for any detector; it is a statement about the prediction set, not the Bayes floor. (2) The TV-based statement ½(1−TV) ≤ η bounds only the best achievable error on the retained region; it is an information-theoretic floor, not a guarantee on a deployed detector's realized error, and we incorrectly presented it as the latter. We have rewritten §5 to state (1) as the operative guarantee, to label (2) explicitly as a floor on achievable (not realized) error, and to back all deployment claims with measured realized selective accuracy. On a held-out split the protocol leaves 18.5% undecided with 62.6% decided-accuracy at β=0.30, rising to 80.7% undecided / 75.5% accuracy at β=0.05, with empirical coverage meeting 1−β throughout. These are realized numbers, not Bayes bounds. W1 novelty. We agree the Bayes-error result is standard and the rate–distortion inequality follows from its definition; we do not claim either as our contribution. Our intended contributions are the deployable calibrated protocol (now stated with the corrected guarantee) and the pre-registered negative result on the ΔC·ΔS trade-off. We have rewritten the contributions list to avoid overstating the theory.

W2 experiments. The reviewer is right that these were unconvincing. The three detectors at exactly 0.500 were placeholders for zero-shot logit methods requiring white-box generator access we lacked; we have removed them and state the constraint. The rule-based paraphrase result (0.472→1.000, CI [1.000,1.000]) is a degenerate artifact a systematic textual signature from the rule-based transform combined with a very small attack subset and is not a meaningful robustness measurement; we have removed it and retain the semantic paraphrase attacks (DIPPER 0.860; back-translation 0.472), which are the informative ones.

We thank the reviewer for a rigorous review that has materially improved the paper's correctness and integrity.

Reviewer_NAStOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the theoretical limitations of AI-generated text detection and their implications for practical deployment. The paper first discusses the fundamental limits of detection using the Bayes-error bound based on the total variation distance between human-written and AI-generated text distributions. They also apply rate-distortion theory to analyze how much information from human-written text is preserved through paraphrasing by AI. The paper then evaluates several detection methods across six domains, generators, and distribution shifts, including paraphrasing, adversarial attacks, and cross-domain settings. Finally, they propose a conformal prediction protocol that allows detectors to suspend making predictions when uncertainty is high.

Critical strengths
  • The paper provides the theoretical limitations of AI text detection, and then translates them into practical actions for real-world deployment. Specifically, it proposes using conformal prediction to refrain from making predictions when the decision is uncertain.
  • The paper provides a broad empirical evaluation, covering multiple domains, generators, and distribution shifts, including paraphrasing and other adversarial attacks.
Critical weaknesses
  • The theoretical result that the detection error approaches 0.5 as the distribution of AI-generated text approaches that of human-written text is already well established in prior work. While the paper extends this argument to distributions over observable features for detectors, I am not fully convinced that this extension provides substantial new theoretical insights.
  • Similarly, I found it somewhat difficult to understand what new insights the theoretical analysis of information loss through paraphrasing provides.
  • The observation that detection performance can change under paraphrasing, adversarial attacks, and domain shifts has also been widely discussed in previous studies.
  • I have a few concerns about the experimental setup:
    • It is unclear what data were used to fine-tune RoBERTa. Providing this information would be helpful for interpreting the performance of such a supervised detector.
    • The generators used in the experiments are relatively old compared to currently available models. Including more recent and widely used models, such as GPT-5.6 and Gemini, would help assess how well the findings generalize to realistic deployment scenarios.
  • I find the idea of using conformal prediction for decision pending interesting, but I have some concerns about its practical applicability. It relies on the assumption that the calibration and test data are exchangeable. This assumption may be difficult to satisfy in real-world detection, where the domain and the generator used to produce a given text are often unknown in advance.
Questions for authors

See the weakness part.

1 comment

We thank the reviewer for the careful read and for crediting the translation of theory into a deployable protocol and the breadth of the evaluation. We take the "little new" concern seriously and respond directly.

What we do and do not claim as novel. We agree the impossibility bound (error→½ as P_AI→P_H) is established, and that empirical fragility under paraphrase, adversarial attack, and domain shift has been reported. We do not claim these as our contribution. What we claim as new is the synthesis the reviewer identifies as a strength: (i) a deployable, calibrated reporting protocol that turns the impossibility result into an operational selective classifier with a distribution-free coverage guarantee and per-domain disclosure; and (ii) a pre-registered negative result the content/style product ΔC·ΔS appears to satisfy a Heisenberg-style trade-off under six-domain pooling but reverses sign (r=+0.41 to +0.72) once domain is held fixed, exposing a domain-heterogeneity confound. We are not aware of this trade-off's failure being reported, and we present it as a methodological caution for stylometric trade-off claims rather than as a law. The subspace sharpening is deliberately modest in scope: its value is that it makes the bound apply to deployed detectors, which read features rather than full strings, closing the gap between global-convergence theory and real systems.

Rate–distortion / paraphrase (Corollary 1). We agree its role was unclear and have demoted it from a "second bound on detection" to a channel-fidelity diagnostic. Its single insight: paraphrase provably preserves a quantifiable amount of source information (I(H;AI)≈2.71 bits, ≈4% of the zero-information floor) a fidelity quantity logically independent of detectability. It is a measurement, not a second impossibility claim, and we now frame it as such.

RoBERTa fine-tuning data (concern 4a). We have added this to the appendix: RoBERTa is roberta-base fine-tuned (2 epochs, stratified 80/20 split, seed 42) on the paired human/AI passages of our own corpus; the "all" detector pools domains, each per-domain detector trains within that domain. Its high in-distribution AUC (0.978) together with its cross-domain collapse (news 0.785; several off-diagonal transfer cells <0.5) is precisely the intended demonstration: a supervised detector exploits thin, domain-specific asymmetries that do not transfer. Older generators; GPT-5-class / Gemini (concern 4b). A fair point, and one that cuts in our favor theoretically: newer, stronger models push P_AI closer to P_H, driving the Bayes error toward ½. Our open-weight 7–70B panel is therefore a conservative (easier-to-detect) test; a stronger generator can only raise the floor, supporting the impossibility thesis rather than threatening it. Our eight-model panel already includes proprietary API models, and we commit to adding a current frontier model (GPT-5-class / Gemini) on a stratified subsample in the camera-ready.

Exchangeability of calibration/test (concern 4c). We agree this is the key assumption and have made it explicit as a limitation. Two mitigations are in the protocol: calibration is performed per domain, and the TV-based abstain rule is the safety valve when the estimated separability is low. When the domain/generator is unknown in advance, the principled extension is group-conditional (Mondrian) conformal prediction, which we now discuss; under distribution shift the coverage guarantee degrades gracefully and the abstain category absorbs the uncertainty rather than issuing an overconfident verdict. This is exactly the failure mode the protocol is designed to make legible.

We hope this clarifies the contribution and addresses the experimental concerns, and we would welcome further discussion.

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

The authors present a theorem on the detection error rate and characterize the relationship between information retention and paraphrase fidelity. They further provide a pipeline for deciding when not to make a prediction.

Critical strengths

The proposed pipeline is easy to understand and requires no additional training.

Critical weaknesses
  1. The theoretical results appear trivial and disconnected. Theorem 1 is trivial, since overlap between the distributions of AI-generated and human-written text necessarily leads to unavoidable classification errors. Corollary 1 seems irrelevant to the detection problem and, as the authors acknowledge, is logically independent of the detectability result. The authors should therefore clarify the theoretical novelty of this paper.
  2. The authors claim that one of their contributions is the calibrated reporting protocol. Section 5 provides a theoretical lower bound on the probability that the true label is included in the output set, i.e., the coverage probability. However, the protocol allows the output set to contain both AI and human, which means that the system does not make a definite prediction. Consequently, high coverage can be achieved simply by leaving most or even all samples undecided. The authors should report the proportion of texts left undecided and the detection accuracy on the remaining texts.
  3. The experimental evaluation appears incomplete. In Table 2, only RoBERTa has complete results across all domains. The first three baselines all have an AUC of exactly 0.500, with no domain-specific results reported. This gives the impression that the baseline experiments have not been fully completed.
  4. Table 3 does not specify which detector was used to obtain the reported AUC values.
Questions for authors

My main concerns are the weaknesses mentioned above.

1 comment

We thank the reviewer for a detailed, technically engaged review. We have corrected the presentation and added the requested analysis.

  1. Novelty of Theorem 1 / Corollary 1. We agree the form of Theorem 1 (Bayes error as a function of TV distance) is classical, and do not claim it as new we cite Sadasivan et al. (2023) and Chakraborty et al. (2023). Our contribution is narrower and we sharpened the text: prior impossibility results require global convergence P_AI=P_H over the whole string space, whereas we show the floor E_min→½ is driven by convergence of the projections onto the observable feature manifold alone, even when the distributions differ on rare/unobservable directions. This explains why detectors fail despite P_AI≠P_H globally; Reviewer 2qJ7 independently flagged this relaxation as the paper's main value. We also convert the ceiling into per-domain TV estimates with bootstrap CIs (Table 4), quantifying how far each domain sits from the floor. On Corollary 1: you are right it is logically independent of detectability. We state this deliberately, but agree the original text over-sold it. It is a channel-fidelity diagnostic (how much source information a paraphrase provably retains), not a second detectability bound. We retitled it accordingly and removed it from the contributions that bear on detection.

  2. The protocol could game coverage by abstaining. A correct and important point; coverage alone is meaningless without reporting how often the system decides. We added a selective-accuracy (risk–coverage) analysis reporting both the undecided fraction and accuracy on the decided set, on a held-out split 25/25):

| β | undecided | decided | acc. on decided | emp. coverage | |------|-----------|---------|-----------------|---------------| | 0.30 | 18.5% | 81.5% | 62.6% | 0.695 | | 0.20 | 41.9% | 58.1% | 64.9% | 0.796 | | 0.10 | 67.3% | 32.7% | 71.3% | 0.906 | | 0.05 | 80.7% | 19.3% | 75.5% | 0.953 |
Empirical coverage meets the 1−β guarantee at every level, and decided-set accuracy rises as the target tightens (62.6%→75.5%). The risk–coverage curve increases monotonically from 60.6% at full coverage to at 50% coverage. So the protocol does not obtain coverage by refusing to predict: at β=0.30, trading an explicit, reported abstention rate for a genuine accuracy gain. Per-domainabstention is disclosed (at β=0.10, 48% undecided on medical vs. 86% on legal). Added as Figure N / Table M. (Note: RoBERTa's per-item scores were not retained, so this study uses a reproducible stylometric detector; since the protocol is detector-agnostic, the mechanics and guarantee transfer, onects this weaker detector.) 3. Incomplete baselines (three detectors at 0.500, no per-domain cells). You are right the The three zero-shot logit detectors (Binoculars, Fast-DetectGPT, GLTR) require white-boxaccess to the token log-probabilities of the generating model, which was not uniformly available across our open-weight/proprietary panel; a flat 0.500 implied a measured result where there was none. We removed these rows and now state the access constraint explicitly, retaining only detectors with co

  1. Table 3 detector unspecified. Corrected: Table 3 reports the RoBERTa detector (the only ribution signal), now stated in the caption.

We thank you again the selective-accuracy analysis and corrected table materially improvey, and we hope they address the core concerns.

Reviewer_2qJ7OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper provides theoretical results regarding the impossibility of separating between the AI-generated text and human-written text. They also propose a calibrated authorship-reporting protocol that can make a detector abstain when the total variation between AI and human text is low.

Critical strengths
  • The paper's topic is one that generated a lot of discussion within the community with researchers publishing papers and results that suggest both the possibility that AI-generated text is separable from human text as well as its impossibility. The current paper provides theoretical bounds supporting impossibility on more relaxed assumptions than prior work. For example, they allow the P_AI and P_H to differ on rare or unobservable directions which I believe it's valuable for the community.

  • The authors document a negative result as well.

Critical weaknesses
  • The paper is hard to read sometimes. For example, long sentences such as "Our second result is a logically independent rate–distortion floor on the paraphrase channel [...] distinct from the marginal separability that governs detectability"

  • The empirical results do not consider state-of-the-art detectors (e.g., Pangram).

Questions for authors

Did the authors consider black-box detectors such as GPT Zero or Pangram?

1 comment

We thank the reviewer for the careful reading, for recognizing that our bound relaxes prior assumptions by allowing PAIP_{\text{AI}} and PHP_{\text{H}} to differ on rare or unobservable directions, and for valuing our decision to report a negative result. We are also glad that the soundness of the formal results was not in question. We address the two weaknesses below.

  1. Readability (overlong sentences). We agree, and we thank the reviewer for the specific example. The abstract and introduction packed too many claims into single sentences. In the revision we have rewritten the abstract so that each result occupies its own sentence, and we have split the sentence the reviewer cites into three: (i) the rate–distortion floor itself, (ii) the numerical finding (I(H;AI)≈2.71I(\text{H};\text{AI}) \approx 2.71 bits, ≈4\approx 4% of the zero-information floor), and (iii) the conceptual point that channel fidelity is distinct from the marginal separability that governs detectability. We have made a pass over the whole paper to cap sentence length and to move multi-clause definitions into displayed equations.

  2. State-of-the-art / black-box detectors (Pangram, GPTZero). This is a fair point, and we want to address it directly rather than sidestep it.

The key observation is that our central result is detector-agnostic by construction. Theorem 1 bounds the Bayes error of any classifier under equal priors — including proprietary black-box systems such as Pangram and GPTZero. No such detector, regardless of its training data or architecture, can achieve error below Emin⁡=12(1−∣PH−PAI∣TV)E_{\min} = \tfrac12\big(1 - |P_{\text{H}} - P_{\text{AI}}|_{\text{TV}}\big), because that floor is a property of the human and AI distributions on the observable feature manifold, not a property of any particular detector. Our empirical benchmark is therefore not intended as a leaderboard to identify the best-performing detector; it is a measurement of how close deployed detectors sit to this information-theoretic floor. Including a stronger commercial detector would not alter the thesis: it would be bounded by the same ceiling, and its score would be one more point on the "distance-to-floor" axis rather than a counterexample to the theory.

We also note that the detectors we do evaluate were chosen to span the main deployed families a fine-tuned transformer discriminator (RoBERTa), zero-shot log-likelihood detectors, and a stylometric/meta-classifier baseline precisely so that the measurement is not tied to one architecture. Published evaluations of commercial systems report the same qualitative failure modes we characterize theoretically: degradation under paraphrase, cross-domain shift, and adversarial sampling. These systems are consistent with, not exceptions to, our ceiling.

That said, we agree a direct measurement of a commercial black-box detector would make the empirical section more complete and more convincing to practitioners. We were constrained here by access and reproducibility (closed APIs with changing models complicate a pre-registered, reproducible benchmark), but we commit to adding an evaluation of at least one commercial black-box detector (e.g., GPTZero) on a stratified subsample of our paired corpus in the camera-ready version, reported under the same cross-domain and paraphrase conditions as our other detectors.

We hope these revisions and clarifications address the reviewer's concerns, and we thank them again for a constructive and fair review.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.