A mathematical framework developed by statistician Kuancheng Ye at the University of California, Irvine establishes rigorous finite-sample bounds on the prevalence of watermarked artificial intelligence text across large document collections, even when texts undergo unknown degrees of human or automated editing. Released on November 12, 2026, in an arXiv preprint, the study addresses a critical gap: single-document detectors identify whether an individual text carries a watermark, but cannot accurately quantify aggregate contamination across journal submissions or repositories without robust bounds.
Single-Document Detection vs. Corpus Prevalence Estimation
Model providers increasingly deploy watermarking AI output to embed statistical signatures during generation. For publishers and repository curators, the primary operational question is aggregate: what fraction of submitted papers originated from a watermarked model after author revisions?
Applying binary detection tools for AI-generated text to entire corpora introduces severe bias. Human manuscripts yield false positives, while watermarked papers that have been lightly edited or have low entropy yield false negatives. Standard corrections like Storey’s estimator adjust for human false-positive rates but ignore edit-induced false negatives. Because human document p-values are uniform under a calibrated key, unconstrained mixtures leave the identified prevalence set bounded above by 100%, offering no useful constraint.
The Three-Budget Power Envelope Framework
The framework restricts post-generation editing through three user-defined budgets:
- Retention floor (εmin): The minimum proportion of token positions retaining both token and context intact.
- Substitution cap (σmax): The maximum proportion of in-place token substitutions at intact contexts.
- Entropy budget (δ): A bound on next-token predictability, maxw P(w) ≤ 1 − δ.
Crucially, for Gumbel-max watermarks, substituting a token at an intact context inserts a token that lost the pseudorandom key draw. Consequently, substituted positions look less watermarked than genuine human text. Standard edit models treating edited positions as purely null miss this negative signal. For green-list (KGW) schemes, the author derives a closed-form envelope, while for Gumbel-max, a two-point least-favourable distribution yields a conservative envelope producing sharp identified sets.
Empirical Benchmarks Across Language Models and Perturbations
The framework was tested on text from OPT-1.3B and Qwen3-1.7B, alongside 12,000 human texts from the C4 validation corpus scored under three independent keys. The human p-values showed no significant deviation from uniformity (Kolmogorov–Smirnov p-values 0.03 to 0.80; none significant after Bonferroni correction).
In 2,000-document simulated corpora at 20% true prevalence (π = 0.20), the framework produced valid 95% confidence intervals across diverse perturbations:
| Model / Watermark Scheme | Perturbation Regime | Assumed Budget (εmin, δ) | Realized Retention (Mean εeff) | Exact-Null 95% Bound | Worst-Case Detection Upper Bound |
|---|---|---|---|---|---|
| OPT-1.3B / Gumbel-max | Random scattered edits | (0.3, 0.2) | 0.43 to 0.52 | [0.14, 0.38] | 0.56 |
| Qwen3-1.7B / Gumbel-max | Random scattered edits | (0.3, 0.2) | 0.43 to 0.52 | [0.14, 0.37] | 0.64 |
| OPT-1.3B / KGW Green-list | Random scattered edits | (0.3, 0.2) | 0.43 to 0.52 | [0.14, 0.41] | 0.72 |
| Qwen3-1.7B / KGW Green-list | Random scattered edits | (0.3, 0.2) | 0.43 to 0.52 | [0.15, 0.38] | 0.75 |
| OPT-1.3B / Gumbel-max | Plausible in-place substitutions | (0.3, 0.2) | 0.43 to 0.52 | [0.13, 0.38] | 0.57 |
| Qwen3-1.7B / KGW Green-list | LLM Paraphrase (Qwen2.5-3B) | (0.3, 0.2) | Variable rewrite | [0.02, 0.27] | 0.34 |
Across all 200 trials per cell, exact-null bounds achieved 100% empirical coverage. Single thresholds tuned to maximize the gap between envelope and null matched multi-threshold precision.
Calibration Sensitivity and the Large-Corpus Dilemma
A critical finding concerns calibration sensitivity. In moderate corpora (N = 2,000), sampling slack (scaling as N−1/2) absorbs mild envelope violations (like machine paraphrasing) and minor null discrepancies. In large corpora (N = 100,000), this slack vanishes while calibration errors persist.
Simulations showed that an uncorrected null error of 0.01—comparable to normal human variation—causes exact-null coverage to collapse at N = 100,000. To preserve validity in high-volume settings, the author establishes a reference-calibrated bound using an independent human sample.
Implications for Scholarly Publishing and Integrity Auditing
For journal editors and institutional research integrity officers, these bounds enable defensible macro-level monitoring without requiring disputed per-document accusations. Understanding these statistical limits clarifies overall AI detection accuracy and guides the deployment of AI detection tools in academic publishing across institutional collections.
Sources
- Ye, K. (2026). Prevalence Bounds for Edited Watermarked Text: Power Envelopes and Calibration Sensitivity. arXiv:2611.17035. https://arxiv.org/abs/2611.17035 (Checked 2026-11-15).







