Pith.

Cryptography and Security · arXiv:2611.06100 v1 · submitted 4 Nov 2026 · 59 references

Reconciling Privacy with Provenance for Language Model Watermarks

Elise L. Han, Tatsunori Hashimoto

The paper says

“Watermarking has great potential for regulating the use of language models (LMs), but its deployment has been marred by controversy, as many users argue that fine-grained tracking of LM usage violates their privacy.”

more of the abstract

This tension reduces the effectiveness of watermarks, as users engage in evasion strategies such as paraphrasing that can remove much of the watermark signal. We show that provenance and privacy can be reconciled if we shift the goal of watermark design from tracking individuals to estimating aggregate usage. We develop locally differentially private watermarks which provably limit the evidence that any individual text can contribute to detection, while still enabling accurate estimation of population frequencies. We introduce the Red-Green Sequence (RGS) watermark, a private and computationally efficient watermark that achieves accurate frequency estimation at high privacy levels across various models, sample sizes and generation entropies in settings that reflect realistic deployment scenarios. On C4 language modeling tasks, we attain 0.0019 MSE with 0.0038 mean absolute bias at a low ε of 1, while in more challenging settings such as an essay writing task with only 100 text samples, we still find errors on the order of 0.0113 MSE with 0.004 mean absolute bias at ε = 2.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 5 Nov 2026 · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

Holds up, with conditions. A clean, well-executed paper that makes a real contribution: reconciling privacy and provenance for LLM watermarks by switching the goal from individual detection to aggregate estimation. This line is wrong

1 major objection · 6 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 59 references say.

The paper says, at §5, Assumption 1 (Ideal coloring)

The review did not quote this passage.

A classifier read this sentence of the paper as acknowledging the concern (probability 0.64):

“In the formal PRF definition, these are only computationally indistinguishable (Goldreich et al., 1986), but we omit the negligible distinguishing advantage from our notation.”

Pith says: Objection 1 of 1 · major

The entire theoretical analysis (Lemma 1, Propositions 3–5) rests on Assumption 1, which treats the keyed thresholded hash as providing independent Bernoulli(π₀) colors for distinct sequences, with the practical realization acknowledged as only computationally indistinguishable. While the paper is transparent about this idealization and tests empirically with BLAKE2b, the theoretical gap between the assumption and the concrete instantiation is not quantified.

more of objection 1

The authors should either (a) provide a concrete bound on how the distinguishing advantage of the PRF affects the bias and variance of the estimator (e.g., by adding an indistinguishable-error term to the MSE bounds), or (b) explicitly state that the theoretical guarantees are modulo that negligible error and discuss the conditions under which the empirical success can be expected to generalize. Without this, a reader cannot assess risk of degradation under adversarial key choices or large-scale deployment.

6 minor objections

The paper says, at Algorithm 1

Pith says: minor objection 1 of 6

The notation 'G ← {j : hk(Yj) = G}' is overloaded: 'G' is used both as the set of green indices and as the color label 'G'. Rename the set to e.g., 'GreenIndices' to avoid confusion. Also, the deduplication step is described but the effect on the theoretical guarantees (especially the i.i.d. candidate assumption) is not discussed.

The paper says, at §5.1, Eq. (6) and (7)

Pith says: minor objection 2 of 6

The definitions of gi(k) and qi(k) should explicitly state that they condition on the key k. The text does this for qi(k) but for gi(k) it is implicit from Eq. (6); clarify by writing 'gi(k) = ...' throughout.

show 4 more

The paper says, at §6.1, C4 experiment

Pith says: minor objection 3 of 6

The paper reports average absolute biases of 0.0038, 0.0037, 0.001 for ε=1,2,3, and percentile half-widths of 0.072, 0.040, 0.031. It would be helpful to also report the standard deviation of the estimates across keys and mixtures to facilitate comparison with future work.

The paper says, at §6.4, sample efficiency

Pith says: minor objection 4 of 6

Fact 1 states the expected number of samples is at most 1/(Pr[green] + e^{-ε} Pr[red]). In the text, the empirical average is said to be 'fewer than 1.5' at ε=0.5,1. Clarify whether this is the average per completion over the entire experiment and whether the cap M=20 ever causes truncation that biases the count downward.

The paper says, at Appendix B.2.3, Lemma 3

Pith says: minor objection 5 of 6

The variance bound for B(k) uses the Efron-Stein inequality. The statement says Vark[B(k)] ≤ (e^{2ε} π₀(1-π₀)/m²) ∑_{i,j} ⟨p_i, p_j⟩. The derivation appears to assume that the map f is e^{ε}-Lipschitz, which is shown for both samplers. However, the final constant may be off by a factor of 2 from a direct application of Efron-Stein; please check the derivation or clarify that the inequality is valid as stated.

The paper says, at §D, RGS examples

Pith says: minor objection 6 of 6

The examples are a nice qualitative addition. Consider adding a sentence noting that the completions appear coherent and natural, consistent with the low-distortion claim.

What would settle it

Replicate the C4 experiment (Section 6.1) using a truly random oracle for color assignment in place of the keyed BLAKE2b; compare the resulting MSE and bias distributions. If they differ significantly from the PRF-based results, the PRF approximation affects estimator performance.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 5 Nov 2026

  1. Reader holds up, with conditions confidence high
  2. Skeptic agrees 1 min 1 s · 27,483 tokens read, 5,472 written
  3. Editor wrote the verdict above 15 s · 28,059 tokens read, 944 written
  4. Referee minor revision: 1 major objection, 6 minor 29 s · 27,592 tokens read, 2,775 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

This paper does something that sounds like it shouldn't work and makes it work. The core idea — shift watermarking from tracking individuals to estimating aggregate prevalence, then apply local differential privacy — is genuinely novel and well-motivated. The RGS watermark itself, a sequence-level red-green scheme with randomized response style noise, is a clever construction, and the theoretical analysis of its bias-variance tradeoff (especially the π0 tuning result, Proposition 5) is the kind of careful statistical thinking that separates a real contribution from a trick. The authors are transparent about limitations, including the idealization in Assumption 1, the sensitivity to low entropy and low diversity, and the lack of edit robustness. The empirical evaluation is thorough across multiple models, tasks, and parameter settings, and the results are convincing: MSE 0.0019 at ε=1 on C4 is genuinely low.

The soft spots are real but not hidden. Assumption 1 (ideal coloring) is the linchpin of the theoretical guarantees; the authors acknowledge it is only computationally indistinguishable. The experiments use a concrete PRF (BLAKE2b) and show good alignment, but the gap between the ideal coloring assumption and practice could be explored more systematically — for example, by testing with weaker hash functions or deliberately adversarially constructed prompts that might bias the coloring. The estimator degrades noticeably in low-entropy settings (e.g., the coding task with a single prompt), which is honestly reported. The paper also explicitly notes that RGS is not robust to editing (future work) and does not provide privacy over repeated queries from the same user. These are real limitations, but they are stated, not hidden. The paper's main claim — that aggregate estimation can work with local DP — holds up.

Who is this for? Researchers working on LLM watermark deployment, privacy-preserving AI governance, and anyone interested in the surprising connection between randomized response and watermarking. It deserves a serious referee. I recommend acceptance after the authors address the PRF approximation gap more directly — perhaps a brief discussion or a small experiment with a contrasting hash.

I would bring this to a reading group and would cite it in my own work. It is a serious thinker paper.

The strongest objection

The paper's theoretical framework relies on Assumption 1 (ideal coloring), which the authors transparently acknowledge as an idealization; the empirical evaluation uses a concrete PRF and demonstrates low MSE and bias across diverse settings. The reader's concern about this assumption is valid but already noted and indirectly tested. No additional load-bearing concern beyond what the authors and reader have already identified.

The paper is thorough: it clearly states the ideal coloring assumption, provides finite-sample bounds conditional on the key, and validates empirically across high- and low-entropy settings. The reader's conditional verdict appropriately requests a discussion of the PRF approximation, which is the paper's acknowledged weak point. I find no new load-bearing concern that would alter the verdict.

The referee's report

If the claims hold, this work provides a principled framework to resolve the tension between watermark provenance and user privacy, which has been a major obstacle to real-world deployment. The theoretical connection to randomized response and local differential privacy is clean and well-articulated. The empirical validation is thorough and covers challenging low-entropy, low-sample settings. The paper also identifies and begins to address the bias that arises when the underlying text distributions have limited entropy. This is a timely contribution that could influence both watermark design and regulatory standards.

Pith's reading of the paper (a machine's summary, not the authors')

Most language model watermarks aim to detect whether a single piece of text was AI-generated, which creates a tension with user privacy and motivates adversarial evasion. This paper argues that many regulatory and monitoring goals only require knowing the proportion of watermarked texts in a population, not which texts are watermarked. To that end, the authors propose the Red-Green Sequence (RGS) watermark, a sequence-level red-green list watermark whose reweighting is calibrated to satisfy ε-local differential privacy. They show that with RGS, an analyst can estimate the population prevalence of watermarked text using a moment estimator analogous to randomized response, with provable bounds on bias and variance. On the C4 language modeling task, the estimator achieves mean squared error of 0.0019 and mean absolute bias of 0.0038 at ε=1; even on a challenging essay-writing task with only 100 samples, the errors are on the order of 0.0113 MSE with 0.004 bias at ε=2.

What carries the argument. The RGS watermark works by drawing M candidate sequences from the base language model, coloring each as green or red using a keyed hash, and then reweighting the selection probabilities so that the likelihood ratio between watermarked and base distributions is bounded by e^ε. This construction satisfies ε-local differential privacy (Proposition 1 and Proposition 2). For frequency estimation, the indicator of whether a text is green forms a noisy binary report analogous to randomized response; the estimator plugs the observed green fraction into a moment equation: ˆλ = proj_{[0,1]}((H̄ - π₀)/(π₁ - π₀)), where π₀ is the expected green probability under the base model and π₁ is the green probability after watermarking (equation 9). The paper characterizes how per-prompt entropy and cross-prompt diversity control bias and variance through the norms ∥p_i∥₂ and ⟨p_i, p_j⟩ (Lemma 1, Proposition 3).

Core claim. The central claim is that provenance and privacy can be reconciled if one shifts the design goal from tracking individuals to estimating aggregate usage. The paper establishes this by constructing a locally differentially private sequence-level watermark—the Red-Green Sequence (RGS) watermark—and deriving a moment estimator for watermark prevalence that mirrors randomized response. The key insight is that sequence-level watermarks, while unattractive in non-private settings, are highly efficient for private frequency estimation because they encode exactly one red-green bit per sequence, analogous to the optimal one-bit randomized response. The paper proves that RGS satisfies ε-local DP, derives the bias and variance of the prevalence estimator, and identifies a bias-variance tradeoff that can be tuned via the null green probability π₀.

The premise it rests on. Assumption 1 (ideal coloring) states that for any finite set of distinct sequences fixed independently of the random key k, the indicators 1{h_k(y) = G} are independent Bernoulli(π₀) random variables; the paper acknowledges this holds only up to computational indistinguishability, and the entire estimator's unbiasedness and variance formulas depend on it.

If the paper is right (Pith's inferences, not the authors')

Regulators and platform auditors can estimate the prevalence of AI-generated text without being able to identify any individual user as a generator, potentially enabling broader deployment of watermarking with less user backlash.

Users gain plausible deniability: any individual text's green/red status could plausibly arise from random chance, making adversarial paraphrasing less necessary.

The paper provides an explicit bias-variance tradeoff (Proposition 5) that allows practitioners to tune π₀ to reduce persistent bias at large sample sizes, yielding better accuracy than naïve RR-optimal settings.

If aggregate prevalence estimates suffice for many regulatory use cases (e.g., detecting training data contamination, monitoring AI use in peer review, or enforcing disclosure thresholds), then the privacy-provocation dynamic around watermarking may be substantially defused, changing the policy debate.

The paper's reliance on Assumption 1 (ideal coloring via a PRF) means that rigorous implementations must use a provably secure PRF; any deviation from independence could introduce unmeasured bias that the estimator does not capture.

The sequence-level RGS watermark is not robust to edits—as the authors acknowledge. Combining RGS with locality-sensitive hashing or error-correcting codes could yield a private watermark that resists paraphrasing, an important open direction.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

Watermarking has great potential for regulating the use of language models (LMs), but its deployment has been marred by controversy, as many users argue that fine-grained tracking of LM usage violates their privacy. This tension reduces the effectiveness of watermarks, as users engage in evasion strategies such as paraphrasing that can remove much of the watermark signal. We show that provenance and privacy can be reconciled if we shift the goal of watermark design from tracking individuals to estimating aggregate usage. We develop locally differentially private watermarks which provably limit the evidence that any individual text can contribute to detection, while still enabling accurate estimation of population frequencies. We introduce the Red-Green Sequence (RGS) watermark, a private and computationally efficient watermark that achieves accurate frequency estimation at high privacy levels across various models, sample sizes and generation entropies in settings that reflect realistic deployment scenarios. On C4 language modeling tasks, we attain 0.0019 MSE with 0.0038 mean absolute bias at a low ε of 1, while in more challenging settings such as an essay writing task with only 100 text samples, we still find errors on the order of 0.0113 MSE with 0.004 mean absolute bias at ε = 2.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-11-05

The authors' main claim, as the classifier picked it out 1.00

“We show that provenance and privacy can be reconciled if we shift the goal of watermark design from tracking individuals to estimating aggregate usage.”

How the main claim is stated: asserted 0.82

What the abstract itself states, as probabilities: claims first 0.05 · states limitation 0.36 · human participants 0.09 · controlled comparison 0.07 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.11 · mentions code or data availability 0.03

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
Watermarking has great potential for regulating the use of language models (LMs), but its deployment has been marred by controversy, as many users argue that fine-grained tracking of LM usage violates their privacy. background 1.00 0.04 0.02 asserted 0.92
This tension reduces the effectiveness of watermarks, as users engage in evasion strategies such as paraphrasing that can remove much of the watermark signal. background 0.96 0.06 0.03 asserted 0.98
We show that provenance and privacy can be reconciled if we shift the goal of watermark design from tracking individuals to estimating aggregate usage. contribution 0.65 0.65 0.86 asserted 0.82
We develop locally differentially private watermarks which provably limit the evidence that any individual text can contribute to detection, while still enabling accurate estimation of population frequencies. contribution 0.56 0.93 0.26 asserted 1.00
We introduce the Red-Green Sequence (RGS) watermark, a private and computationally efficient watermark that achieves accurate frequency estimation at high privacy levels across various models, sample sizes and generation entropies in settings that reflect realistic deployment scenarios. contribution 0.82 0.93 0.20 asserted 1.00
On C4 language modeling tasks, we attain $0.0019$ MSE with $0.0038$ mean absolute bias at a low $\varepsilon$ of 1, while in more challenging settings such as an essay writing task with only 100 text samples, we still find errors on the order of $0.0113$ MSE with $0.004$ mean absolute bias at $\varepsilon=2$. result 1.00 0.97 0.07 asserted 1.00

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.55

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-11-05: 1 request, 5,237 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (11)
Figure 1 of the paper
Figure 1. RGS prevalence estimates on C4. Left to right:
Figure 2 of the paper
Figure 2. RGS prevalence estimates on OUTFOX (left) and LiveCodeBench (right), with
Figure 3 of the paper
Figure 3. MAPE versus mixture size n for different π0. Left to right: λ = 0.2, 0.3, 0.4. Essay writing We randomly select 20 prompts from the OUTFOX dataset consisting of argumen￾tative essay prompts (Koike et al., 2024). Each mixture we test consists of 100λ texts sampled from the RGS decoder, and 100(1 − λ) from the same base model, all independently sampled under the same prompt. Coding We randomly select 20 problems from the LiveCodeBench dataset (Jain et al., 2024) and construct the same one-prompt mixture as the essay setting. We enable the thinking mode for Qwen3-1.7B and Qwen3-4B and…
Figure 4 of the paper
Figure 4. Mean candidates examined by the rejection sampler versus
Figure 5 of the paper
Figure 5. plots MAPE as a function of ε at a fixed n = 2000. We observe that the errors decrease with increasing ε, consistent with the looser privacy requirements
Figure 6 of the paper
Figure 6. MAPE versus mixture size n, with the ideal RR benchmark dashed. Left to right: λ = 0.1, 0.5, 0.9
Figure 7 of the paper
Figure 7. MAPE versus temperature for Llama-3.1-8B-Instruct with prefix lengths
Figure 8 of the paper
Figure 8. RGS parallel-sampling prevalence estimates on C4. Left to right:
Figure 9 of the paper
Figure 9. MAPE versus candidate budget M for rejection sampling. Left to right: λ = 0.1, 0.5, 0.9
Figure 10 of the paper
Figure 10. MAPE versus candidate budget M for parallel sampling. Left to right: λ = 0.1, 0.5, 0.9
Figure 11 of the paper
Figure 11. MAPE versus prompt count for Qwen3-4B, with

Figures are the authors', reproduced from the paper.

References (59)
  1. [1] Bengio, Yoshua and LeCun, Yann , booktitle =. Scaling Learning Algorithms Towards on Pith
  2. [2] Hinton, Geoffrey E. and Osindero, Simon and Teh, Yee Whye , journal =. A Fast Learning Algorithm for Deep Belief Nets , volume = on Pith
  3. [3] Deep learning , author=. 2016 , publisher= on Pith
  4. [4] Fan, Angela and Jernite, Yacine and Perez, Ethan and Grangier, David and Weston, Jason and Auli, Michael. ELI 5: Long Form Question Answering. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. 2019. doi:10.18653/v1/P19-1346 on Pithdoi:10.18653/v1/p19-1346
  5. [5] Mark Chen and Jerry Tworek and Heewoo Jun and Qiming Yuan and Henrique Pond. Evaluating Large Language Models Trained on Code , journal =. 2021 , url =. 2107.03374 , timestamp = on PithPith's reviewarXiv
  6. [6] Jacob Austin and Augustus Odena and Maxwell I. Nye and Maarten Bosma and Henryk Michalewski and David Dohan and Ellen Jiang and Carrie J. Cai and Michael Terry and Quoc V. Le and Charles Sutton , title =. CoRR , volume =. 2021 , url =. 2108.07732 , timestamp = on PithPith's reviewarXiv
  7. [7] Colin Raffel and Noam Shazeer and Adam Roberts and Katherine Lee and Sharan Narang and Michael Matena and Yanqi Zhou and Wei Li and Peter J. Liu , title =. CoRR , volume =. 2019 , url =. 1910.10683 , timestamp = on PithPith's reviewarXiv
  8. [8] Watermarks in the Sand: Impossibility of Strong Watermarking for Generative Models , author=. 2025 , eprint= on Pith
  9. [9] Optimizing Adaptive Attacks against Content Watermarks for Language Models , author=. The 1st Workshop on GenAI Watermarking , year= on Pith
  10. [10] Duchi, John C. and Jordan, Michael I. and Wainwright, Martin J. , booktitle=. Local Privacy and Statistical Minimax Rates , year= on Pith
  11. [11] Peter Kairouz and Sewoong Oh and Pramod Viswanath , title =. CoRR , volume =. 2014 , url =. 1407.1338 , timestamp = on PithPith's reviewarXiv
  12. [12] Roch, Sebastien , year=. Modern Discrete Probability: An Essential Toolkit , DOI= on Pith
  13. [13] Cynthia Dwork and Frank McSherry and Kobbi Nissim and Adam D. Smith , editor =. Calibrating Noise to Sensitivity in Private Data Analysis , booktitle =. 2006 , url =. doi:10.1007/11681878\_14 , timestamp = on Pithdoi:10.1007/11681878
  14. [14] Dwork, Cynthia and Roth, Aaron , title =. 2014 , issue_date =. doi:10.1561/0400000042 , journal = on Pithdoi:10.1561/0400000042
  15. [15] Randomized response: a survey technique for eliminating evasive answer bias. , author=. Journal of the American Statistical Association , year= on Pith
  16. [16] A Watermark for Large Language Models , author =. Proceedings of the 40th International Conference on Machine Learning , pages =. 2023 , editor = on Pith
  17. [17] Goldreich, Oded and Goldwasser, Shafi and Micali, Silvio , title =. J. ACM , month = aug, pages =. 1986 , issue_date =. doi:10.1145/6490.6503 , abstract = on Pithdoi:10.1145/6490.6503
  18. [18] OUTFOX: LLM-Generated Essay Detection Through In-Context Learning with Adversarially Generated Examples , author=. 2024 , eprint= on Pith
  19. [19] Pan, Leyi and Liu, Aiwei and He, Zhiwei and Gao, Zitian and Zhao, Xuandong and Lu, Yijian and Zhou, Binglin and Liu, Shuliang and Hu, Xuming and Wen, Lijie and King, Irwin and Yu, Philip S. M ark LLM : An Open-Source Toolkit for LLM Watermarking. Proceedings of the 2024 Confer... on Pithdoi:10.18653/v1/2024.emnlp-demo.7
  20. [20] Boucheron, St \'e phane and Lugosi, G \'a bor and Bousquet, Olivier. Concentration Inequalities. Advanced Lectures on Machine Learning: ML Summer Schools 2003, Canberra, Australia, February 2 - 14, 2003, T \"u bingen, Germany, August 4 - 16, 2003, Revised Lectures. 2004. doi:1... on Pithdoi:10.1007/978-3-540-28650-9_9
show 39 more
  1. [21] LiveCodeBench: Holistic and Contamination Free Evaluation of Large Language Models for Code , author=. 2024 , eprint= on Pith
  2. [22] Monitoring AI-Modified Content at Scale: A Case Study on the Impact of ChatGPT on AI Conference Peer Reviews , author=. 2026 , eprint= on Pith
  3. [23] Qi Pang and Shengyuan Hu and Wenting Zheng and Virginia Smith , booktitle=. No Free Lunch in. 2024 , url= on Pith
  4. [24] Tianhao Wang and Jeremiah Blocki and Ninghui Li and Somesh Jha , title =. 26th USENIX Security Symposium (USENIX Security 17) , year = on Pith
  5. [25] Bahri, Dara and Wieting, John , year=. A Watermark for Black-Box Language Models , url=. doi:10.48550/arXiv.2410.02099 , abstractNote= on Pithdoi:10.48550/arxiv.2410.02099
  6. [26] Hou, Abe and Zhang, Jingyu and He, Tianxing and Wang, Yichen and Chuang, Yung-Sung and Wang, Hongwei and Shen, Lingfeng and Van Durme, Benjamin and Khashabi, Daniel and Tsvetkov, Yulia. S em S tamp: A Semantic Watermark with Paraphrastic Robustness for Text Generation. Proceed... on Pithdoi:10.18653/v1/2024.naacl-long.226
  7. [27] Robust Distortion-free Watermarks for Language Models , author=. 2024 , eprint= on Pith
  8. [28] Christ, Miranda and Gunn, Sam , month = jun, year =. Pseudorandom. doi:10.48550/arXiv.2402.09370 , urldate = on PithPith's reviewarXivdoi:10.48550/arxiv.2402.09370
  9. [29] My ai safety lecture for UT effective altruism , url=. Shtetl-Optimized , author=. 2022 , month= on Pith
  10. [30] Are We in the AI-Generated Text World Already? Quantifying and Monitoring AIGT on Social Media , author=. 2025 , eprint= on Pith
  11. [31] Provable Robust Watermarking for AI-Generated Text , author=. 2023 , eprint= on Pith
  12. [32] How Claude’s text watermark works , author = on Pith
  13. [33] Spymarks, not Watermarks , author = on Pith
  14. [34] The Mark of the machine , url=. The New Yorker , author=. 2026 , month= on Pith
  15. [35] Krishna, Kalpesh and Song, Yixiao and Karpinska, Marzena and Wieting, John and Iyyer, Mohit , booktitle =. Paraphrasing evades detectors of AI-generated text, but retrieval is an effective defense , url =. doi:10.52202/075280-1195 , editor = on Pithdoi:10.52202/075280-1195
  16. [36] Saarinen, M-J. and Aumasson, J-P. , title =. 2015 , institution = on Pith
  17. [37] Tourangeau, Roger and Yan, Ting , year =. Sensitive Questions in Surveys , volume =. Psychological Bulletin , doi = on Pith
  18. [38] Lensvelt-Mulders, Gerty and Hox, Joop and Heijden, Peter G.M. and Maas, Cora , year =. Meta-Analysis of Randomized Response Research: Thirty-Five Years of Validation , volume =. Sociological Methods & Research , doi = on Pith
  19. [39] Li, Xiang and Ruan, Feng and Wang, Huiyuan and Long, Qi and Su, Weijie , year =. A statistical framework of watermarks for large language models: Pivot, detection efficiency and optimal rules , volume =. The Annals of Statistics , doi = on Pith
  20. [40] A Watermark for Black-Box Language Models , author=. 2026 , eprint= on Pith
  21. [41] Undetectable Watermarks for Language Models , author =. Proceedings of Thirty Seventh Conference on Learning Theory , pages =. 2024 , editor = on Pith
  22. [42] A Resilient and Accessible Distribution-Preserving Watermark for Large Language Models , author=. 2024 , eprint= on Pith
  23. [43] Unbiased Watermark for Large Language Models , author=. The Twelfth International Conference on Learning Representations , year= on Pith
  24. [44] Lee, Taehyun and Hong, Seokhee and Ahn, Jaewoo and Hong, Ilgee and Lee, Hwaran and Yun, Sangdoo and Shin, Jamin and Kim, Gunhee. Who Wrote this Code? Watermarking for Code Generation. Proceedings of the 62nd Annual Meeting of the A ssociation for C omputational L inguistics (V... on Pithdoi:10.18653/v1/2024.acl-long.268
  25. [45] textGrain: Entropy-Calibrated Watermarking for Language Model Text , author =. 2026 , publisher = on Pith
  26. [46] Scalable watermarking for identifying large language model outputs , author=. Nature , year= on Pith
  27. [47] Duchi, John and Wainwright, Martin J and Jordan, Michael , booktitle =. Local Privacy and Minimax Bounds: Sharp Rates for Probability Estimation , url = on Pith
  28. [48] Bassily, Raef and Smith, Adam , title =. Proceedings of the Forty-Seventh Annual ACM Symposium on Theory of Computing , pages =. 2015 , isbn =. doi:10.1145/2746539.2746632 , abstract = on Pithdoi:10.1145/2746539.2746632
  29. [49] Adaptive Text Watermark for Large Language Models , author=. 2024 , eprint= on Pith
  30. [50] The Composition Theorem for Differential Privacy , author =. Proceedings of the 32nd International Conference on Machine Learning , pages =. 2015 , editor = on Pith
  31. [51] Erlingsson,. RAPPOR: Randomized Aggregatable Privacy-Preserving Ordinal Response , year =. Proceedings of the 2014 ACM SIGSAC Conference on Computer and Communications Security , pages =. doi:10.1145/2660267.2660348 , abstract = on Pithdoi:10.1145/2660267.2660348
  32. [52] Liu, Yepeng and Zhao, Xuandong and Song, Dawn and Wornell, Gregory W. and Bu, Yuheng. Position: LLM Watermarking Should Align Stakeholders' Incentives for Practical Adoption. Findings of the A ssociation for C omputational L inguistics: ACL 2026. 2026. doi:10.18653/v1/2026.fin... on Pithdoi:10.18653/v1/2026.findings-acl.1290
  33. [53] Spirlet, Thibault , year=. Claude users are canceling their subscriptions, citing Anthropic’s new ai watermark , url=. Business Insider , publisher= on Pith
  34. [54] Dmitry Kobak and Rita González-Márquez and Emőke-Ágnes Horvát and Jan Lause , title =. Science Advances , volume =. 2025 , doi = on Pith
  35. [55] Shumailov, Ilia and Shumaylov, Zakhar and Zhao, Yiren and Papernot, Nicolas and Anderson, Ross and Gal, Yarin , year =. AI models collapse when trained on recursively generated data , volume =. Nature , doi = on Pith
  36. [56] Optimal Estimation of Watermark Proportions in Hybrid AI-Human Texts , author=. 2026 , eprint= on Pith
  37. [57] SimpleMark: Fast Multi-Bit Text Watermarking under f -Divergence Constraints , author=. 2026 , eprint= on Pith
  38. [58] Does Claude’s watermark solve the AI transparency problem? , url=. TechTarget , author=. 2026 , month= on Pith
  39. [59] Anthropic’s new AI watermark sparks backlash from Claude Subscribers , author= on Pith
Cited by 1 Pith paper

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment