Pith. sign in

REVIEW 2 major objections 4 minor 2 cited by

Policies that allow LLM polishing of peer reviews cannot be enforced with current AI-text detectors.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-13 21:33 UTC pith:VMFAIFJ4

load-bearing objection Solid empirical paper: commercial detectors still flag ~3% of carefully polished human reviews as fully AI, so polishing-only policies are not enforceable at conference scale. the 2 major comments →

arxiv 2603.20450 v2 pith:VMFAIFJ4 submitted 2026-03-20 cs.CL cs.AIcs.CYcs.LG

Policies Permitting LLM Use for Polishing Peer Reviews Are Currently Not Enforceable

Rounak Saha , Gurusha Juneja , Dayita Chaudhuri , Naveeja Sajeevan , Nihar B Shah , Danish Pruthi This is my paper
classification cs.CL cs.AIcs.CYcs.LG
keywords AI text detectionpeer reviewLLM polishingfalse positivespolicy enforcementmixed authorshipscientific writing
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Many conferences and journals now ban reviewers from using large language models except for polishing grammar, flow, and clarity of human-written text. This paper asks whether that selective permission can actually be policed. The authors build a large set of peer reviews that range from fully human-written, through lightly polished, to fully machine-generated, then run five modern detectors, including two commercial systems, against them. Every detector flags a non-trivial share of policy-compliant polished reviews as fully AI-generated, enough to produce hundreds of false misconduct accusations at the scale of a major conference. Adding peer-review-specific signals such as the paper manuscript or domain-style classifiers improves some scores but never reaches the reliability needed for enforcement. The same detectors also tend to label mixed human-AI reviews as fully AI-written, so recent public claims about how much AI is already in the review process may be inflated.

Core claim

Current AI-text detectors cannot reliably separate policy-compliant human-written reviews that have only been polished by an LLM from reviews that were substantially or fully generated by an LLM. Commercial systems such as Pangram and GPTZero still misclassify roughly 3 percent of carefully constructed polished reviews as fully AI-generated, while open-source zero-shot detectors and supervised models that exploit paper context or scientific style fare no better once they face newer models and varied prompts. Therefore polishing-only policies are not currently enforceable.

What carries the argument

A multi-level review dataset (AI-BP, AI-EP, AI-HI, H-AI, fully human) that systematically varies human involvement, combined with thresholded commercial and zero-shot detectors whose false-positive rates on the H-AI class are measured under a conservative labeling that treats only the pure “AI” prediction as a positive.

Load-bearing premise

That the authors’ tightly filtered polishing setup—explicit content-preservation instructions, no paper access, length capped at 1.25 times the original, plus manual checks—faithfully stands in for the polishing prompts real reviewers will actually use.

What would settle it

A new detector or calibration regime that, on a held-out set of human-written reviews polished under the same safeguards, drives the pure-“AI” false-positive rate on the polished class below 0.1 percent while still catching more than 95 percent of fully AI-generated reviews from unseen models.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Conference organizers cannot safely use existing detectors to police polishing-only policies without risking large numbers of false misconduct charges.
  • Public estimates that claim a large fraction of recent reviews are fully AI-generated are likely inflated by mixed-authorship text being labeled fully AI.
  • A blanket ban on any LLM use is, under current tools, easier to approximate than a selective polishing allowance, because humanized fully-AI text is still often flagged as mixed or AI.
  • Reviewers who polish with LLMs should avoid attaching the manuscript and must verify that no new technical content has been inserted, or the polished text becomes far more likely to be flagged.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same false-positive problem will reappear in any high-stakes setting that tries to permit “light editing only,” such as student essays or grant proposals, until detectors can quantify degree of AI involvement rather than binary labels.
  • If commercial detectors continue to be trained on web-scale human-AI parallel data that already includes older conference reviews, reported false-positive rates on pre-2020 human text may be optimistic lower bounds.
  • Future policy design may need to abandon post-hoc detection entirely and move toward process-based signals such as watermarked reviewer interfaces or mandatory disclosure of polishing prompts.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 4 minor

Summary. The paper assembles a dataset of ~50k peer reviews spanning five levels of human–AI collaboration (fully AI-generated with basic/elaborate prompts, AI expansion of human keypoints, human-written then LLM-polished, and pure pre-ChatGPT human reviews from PeerRead). It evaluates five detectors (LogLikelihood, Fast-DetectGPT, Binoculars, Pangram 3.0, GPTZero), calibrated to 0% FPR on held-out human text. The central claim is that polishing-only policies are currently unenforceable: commercial detectors still flag ~3% of carefully constructed H-AI reviews as fully “AI” (Table 2, hard subset; Figure 2), while zero-shot methods and peer-review-specific augmentations (paper-conditioned likelihood, similarity to AI reference reviews, stylometric/RoBERTa classifiers) fail to reach the accuracy needed for high-stakes enforcement. The authors further caution that public estimates of AI-generated reviews (e.g., Pangram’s 21% claim for ICLR 2026) likely overstate fully AI content by misclassifying mixed outputs.

Significance. If the measured false-positive rates hold, the result is directly policy-relevant for major venues (ICLR, ICML, EMNLP, ACL Rolling Review, etc.) that have adopted polishing-only rules. At NeurIPS scale a 3% FPR already implies thousands of potential wrongful accusations. The work also supplies a public, multi-level dataset and code, careful contamination discussion, McNemar tests for context-aware gains, and human validation of the LLM-as-judge content-addition check (Appendix A.3, Cohen’s κ ≈ 0.68–0.76). These elements make the unenforceability claim falsifiable and reusable beyond peer review.

major comments (2)
  1. Table 2 and Figure 2 (hard subset): the ~3.1% Pangram / 3.4% GPTZero FPR on H-AI is the load-bearing number for the unenforceability claim. The construction (explicit content-preservation instructions, no paper manuscript, length ≤ 1.25× original, residual false positives manually verified) is careful, yet the paper itself shows in Figure 3 that relaxing any of those safeguards sharply raises the AI-flag rate. A short additional experiment quantifying how often real reviewers’ polishing prompts (scraped or surveyed) violate the safeguards would strengthen the external-validity claim without changing the core result.
  2. §4.2 and Table 5: supervised stylometric and RoBERTa classifiers achieve near-perfect in-distribution performance but degrade unpredictably on held-out generators (e.g., Llama-3.1 TPR drops to the mid-80s; Qwen FPR on H-AI rises to 11.5%). Because new LLMs appear continuously, this generalization failure is itself a policy-relevant limitation; the manuscript correctly flags it, but a quantitative statement of how frequently the training set would need to be refreshed would make the practical implication sharper.
minor comments (4)
  1. Figure 2 and the corresponding easy-subset matrices (Appendix A.2) would be clearer if the absolute counts, not only percentages, were also reported; at the scale of the hard subset this would let readers judge sampling variability directly.
  2. §3, paragraph defining AI-HI: the average length expansion (keypoints 163 words → reviews 679 words) is useful; stating the inter-quartile range would further document that these are not mere polishing.
  3. Appendix B.2: the choice of n=40 for soft n-gram matching is motivated but not ablated; a one-sentence note on sensitivity to n would help reproducibility.
  4. References: the citation to Pangram Labs (2025b) is central; ensuring the archived URL remains stable (or adding a DOI if available) would aid long-term verification of the 21% claim under discussion.

Circularity Check

0 steps flagged

No significant circularity; the paper is a self-contained empirical evaluation of detectors on a constructed multi-level review dataset.

full rationale

The central claim (Polishing-only policies are unenforceable because detectors flag a non-trivial fraction of H-AI reviews as fully AI) rests entirely on out-of-sample measurements: thresholds for zero-shot detectors are calibrated solely for 0% FPR on a held-out pre-2016 human-review split and then applied to later conferences and the hard subset; commercial detectors are used off-the-shelf; supervised stylometric/RoBERTa classifiers and similarity-to-reference models are trained with proper model-holdout and train/test splits of the authors' own generated data. No equation, threshold, or performance number is obtained by fitting a parameter to the very quantity later reported as a 'prediction.' Self-citations (e.g., Shah 2025a/b, Liu & Shah 2023) appear only in the related-work discussion of LLM utility for reviewing and do not underwrite any detection result or uniqueness claim. The H-AI construction (content-preservation prompts + length filter + manual verification of residual false positives) is an experimental control, not a definitional identity. Consequently the derivation chain contains no self-definitional loop, fitted-input-as-prediction, load-bearing self-citation, or renamed known result.

Axiom & Free-Parameter Ledger

3 free parameters · 3 axioms · 0 invented entities

The central claim rests on an empirical measurement pipeline rather than free parameters or invented physical entities. The few free choices (length filter, calibration threshold, number of reference reviews) are disclosed and ablated; the main domain assumptions are standard for AI-text detection studies.

free parameters (3)
  • H-AI length filter (1.25 imes original)
    Post-processing threshold chosen to exclude reviews that grew substantially; directly affects which samples enter the critical H-AI false-positive calculation.
  • Zero-shot detector decision thresholds
    Set to achieve exactly 0 percent FPR on a NeurIPS 2013-2015 calibration set; any other operating point would change reported TPR/FPR trade-offs.
  • Number of AI reference reviews (45)
    3 models × 3 prompts × 5 roll-outs used for similarity features; the exact count is a design choice that influences the overlap distributions.
axioms (3)
  • domain assumption Pre-2020 PeerRead reviews contain no LLM assistance
    Foundational for the pure-human negative class; stated explicitly and used for both calibration and evaluation.
  • ad hoc to paper The constructed H-AI, AI-HI, AI-BP and AI-EP levels adequately span realistic reviewer behavior under polishing-only policies
    The taxonomy is motivated but remains a modeling choice; real-world prompt distributions are unknown.
  • ad hoc to paper Treating only the commercial 'AI' label (not 'Mixed') as a positive for Polishing-only enforcement is the correct policy mapping
    Authors adopt the most lenient mapping to minimize false positives; alternative mappings would alter the enforceability conclusion.

pith-pipeline@v1.1.0-grok45 · 39910 in / 2263 out tokens · 30666 ms · 2026-07-13T21:33:05.272843+00:00 · methodology

0 comments
read the original abstract

A number of scientific conferences and journals have recently enacted policies that prohibit LLM usage by peer reviewers, except for polishing, paraphrasing, and grammar correction of otherwise human-written reviews. But, are these policies enforceable? To answer this question, we assemble a dataset of peer reviews simulating multiple levels of human-AI collaboration, and evaluate five state-of-the-art detectors, including two commercial systems. Our analysis shows that all detectors misclassify a non-trivial fraction of LLM-polished reviews as AI-generated, thereby risking false accusations of academic misconduct. We further investigate whether peer-review-specific signals, including access to the paper manuscript and the constrained domain of scientific writing, can be leveraged to improve detection. While incorporating such signals yields measurable gains in some settings, we identify limitations in each approach and find that none meets the accuracy standards required for identifying AI use in peer reviews. Importantly, our results suggest that recent public estimates of AI use in peer reviews through the use of AI-text detectors should be interpreted with caution, as current detectors misclassify mixed reviews (collaborative human-AI outputs) as fully AI generated, potentially overstating the extent of policy violations.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. AI for Auto-Research: Roadmap & User Guide

    cs.AI 2026-05 unverdicted novelty 4.0

    The paper delivers a stage-by-stage roadmap for AI in research, showing reliable assistance in retrieval and tool tasks but fragility in novelty and judgment, advocating human-governed collaboration.

  2. AI for Auto-Research: Roadmap & User Guide

    cs.AI 2026-05 conditional novelty 4.0

    AI can generate research artifacts faster than it can verify them, so across all eight lifecycle stages the credible deployment mode is human-governed collaboration rather than full autonomy.

Reference graph

Works this paper leans on

38 extracted references · 2 canonical work pages · cited by 1 Pith paper · 2 internal anchors

  1. [1]

    AbhimanyuHans, AviSchwarzschild, ValeriiaCherepanova, HamidKazemi, AniruddhaSaha, MicahGoldblum, Jonas Geiping, and Tom Goldstein

    URLhttps://arxiv.org/abs/2503.04761. AbhimanyuHans, AviSchwarzschild, ValeriiaCherepanova, HamidKazemi, AniruddhaSaha, MicahGoldblum, Jonas Geiping, and Tom Goldstein. Spotting LLMs with binoculars: Zero-shot detection of machine- generated text, 2024. URLhttps://arxiv.org/abs/2401.12070. Roeland van Hout and Anne Vermeer.Comparing measures of lexical ric...

    Pith/arXiv arXiv doi:10.1609/icwsm.v8i1.14550 2024

  2. [2]

    URLhttps://aclanthology.org/2025

    International Conference on Computational Linguistics. URLhttps://aclanthology.org/2025. genaidetect-1.9/. Quinn McNemar. Note on the sampling error of the difference between correlated proportions or percent- ages.Psychometrika, 12(2):153–157, 1947. doi: 10.1007/BF02295996. URLhttps://doi.org/10.1007/ BF02295996. Eric Mitchell, Yoonho Lee, Alexander Khaz...

    Pith/arXiv arXiv doi:10.1007/bf02295996 2025

  3. [3]

    URLhttps://arxiv.org/abs/2505.01800. OpenAI. New AI classifier for indicating AI-written text. OpenAI Blog, 2023. URLhttps://openai.com/ index/new-ai-classifier-for-indicating-ai-written-text/. Accessed: 2026-01-28. Pangram Labs. ICLR 2026 quantity AI content, 2025a. URL https://web.archive.org/web/ 20260224115522/https://iclr.pangram.com/reviews. Pangram...

    Pith/arXiv arXiv 2023

  4. [4]

    Deepfake Text Detection: Limitations and Opportunities

    URLhttps://arxiv.org/abs/2210.09421. Vishisht Srihari Rao, Aounon Kumar, Himabindu Lakkaraju, and Nihar B Shah. Detecting LLM-generated peer reviews.PLoS One, 20(9):e0331871, 2025. URLhttps://journals.plos.org/plosone/article? id=10.1371/journal.pone.0331871. Pedro Reviriego, Javier Conde, Elena Merino-Gómez, Gonzalo Martínez, and José Alberto Hernández. ...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1371/journal.pone.0331871 2025

  5. [5]

    The Role of AI in Scientific Peer Review

    URLhttps://arxiv.org/abs/2502.15666. Skipper Seabold and Josef Perktold. Statsmodels: Econometric and Statistical Modeling with Python.SciPy 2010, 2010. doi: 10.25080/Majora-92bf1922-011. URL https://doi.org/10.25080/ Majora-92bf1922-011. Nihar B. Shah. AI meets peer review: The good, the bad, and the ugly. Invited Talk at the NeurIPS 2025 Social: “The Ro...

    Pith/arXiv arXiv doi:10.25080/majora-92bf1922-011 2010

  6. [6]

    Humanizing the Machine: Proxy Attacks to Mislead LLM Detectors

    Association for Computational Linguistics. doi: 10.18653/v1/2020.emnlp-main.673. URLhttps: //aclanthology.org/2020.emnlp-main.673/. 17 Tianchun Wang, Yuanzhou Chen, Zichuan Liu, Zhanwen Chen, Haifeng Chen, Xiang Zhang, and Wei Cheng. Humanizing the Machine: Proxy Attacks to Mislead LLM Detectors.ArXiv, abs/2410.19230, 2024. URL https://api.semanticscholar...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.18653/v1/2020.emnlp-main.673 2020

  7. [7]

    substantial

    to measure inter-annotator agreement between the two human annotators, as well as the agreement between the Gemini judge and each human annotator (Table 7). For both human annotators, LLM-human agreements are “substantial” (Table 8, Landis and Koch, 1977), which is consistent with the human-human agreement. Gold standard annotations.We construct a gold-st...

    1977

  8. [8]

    Type Token Ratio (TTR)Hout and Vermeer (2007) - TTR is a lexical diversity measured as the ratio of unique words to total words

    2007

  9. [9]

    It is computed asUnique Words√ Words 5.Maas MeasureTweedie and Baayen (1998) Maas is a vocabulary richness measure that is robust to text length

    Root Type Token Ratio (RTTR)Hout and Vermeer (2007) - RTTR is a length-normalized lexical diversity. It is computed asUnique Words√ Words 5.Maas MeasureTweedie and Baayen (1998) Maas is a vocabulary richness measure that is robust to text length. It is computed asln(Words)−ln(Unique Words) (ln(Words))2

    2007

  10. [10]

    It is computed as Words occurring exactly once Words 7.Bigram Uniqueness- Ratio of unique bigrams to total bigrams

    Hapax Legomenon Rate- HLR is the proportion of words that occur exactly once in the text. It is computed as Words occurring exactly once Words 7.Bigram Uniqueness- Ratio of unique bigrams to total bigrams. 8.Trigram Uniqueness- Ratio of unique trigrams to total trigrams. 8NLTK: Natural Language Toolkit library (v3.9.2),https://www.nltk.org/, accessed Dece...

    2025

  11. [11]

    For identifying stop words, we use the English stopwords corpus of thenltklibrary

    Stop Word Percentage- Percentage of words that are stop words. For identifying stop words, we use the English stopwords corpus of thenltklibrary. 11.Question Percentage- Percentage of sentences that are interrogative (end with?). 12.Exclamation Percentage- Percentage of sentences that are exclamatory (end with!). 13.Abstract Noun Percentage- Proportion of...

  12. [12]

    For this feature, we use the percentage of abstract nouns that are sparse abstract nouns

    Sparse Abstract Noun Percentage- Sparse abstract nouns are low-frequency abstract nouns that are not among the top 5,000 words in the Brown corpus Francis and Kucera (1964). For this feature, we use the percentage of abstract nouns that are sparse abstract nouns. 15.Verb Percentage- Percentage of total words that are verbs

    1964

  13. [13]

    For this feature, we use the percentage of verbs that are sparse verbs

    Sparse Verb PercentageSparse verbs are low-frequency verbs that are not among the top 5,000 words in the Brown corpus Francis and Kucera (1964). For this feature, we use the percentage of verbs that are sparse verbs. 17.Adjective Percentage- Percentage of words that are adjectives

    1964

  14. [14]

    For this feature, we use the percentage of adjectives that are sparse adjectives

    Sparse Adjective Percentage- Sparse adjectives are low-frequency adjectives that are not among the top 5,000 words in the Brown corpus Francis and Kucera (1964). For this feature, we use the percentage of adjectives that are sparse adjectives

    1964

  15. [15]

    For this feature, we use the percentage of adjectives that are sparse adjectives

    Complex Adjective Percentage- We define complex adjectives as those that are morphologically complex (e.g., have suffixes such as-ive,-ous,-ic). For this feature, we use the percentage of adjectives that are sparse adjectives. 20.Adverb Percentage- Percentage of words that are adverbs

  16. [16]

    For this feature, we use the percentage of adverbs that are sparse adverbs

    Sparse Adverb Percentage- We define sparse adverbs as low-frequency adverbs not among the top 5,000 words in the Brown corpus Francis and Kucera (1964). For this feature, we use the percentage of adverbs that are sparse adverbs. 22.Preposition Percentage- Percentage of words that are prepositions. 23.Conjunction Percentage- Percentage of words that are co...

    1964

  17. [17]

    25.Syntax Variety- Number of unique POS tags in the text

    Complex Sentence Percentage- Percentage of sentences containing at least one subordinating conjunction. 25.Syntax Variety- Number of unique POS tags in the text. 26.Emotion Word Percentage- Percentage of words that are emotion-related. 27.Positive Emotion Word Percentage- Percentage of words that are positive-emotion-related. 28.Negative Emotion Word Perc...

  18. [18]

    It uses a Bag-of-Words classifier to obtain the text polarity.11

    Polarity- We use thetextblob library’s sentiment polarity. It uses a Bag-of-Words classifier to obtain the text polarity.11

  19. [19]

    It uses a Bag-of- Words classifier to obtain the text subjectivity

    Subjectivity- Similar to Polarity, we use thetextblob library’s subjectivity score. It uses a Bag-of- Words classifier to obtain the text subjectivity

  20. [20]

    Each word gets a sentiment score, rules are applied according to punctuation, modifiers, negation etc and the final score is normalised

    VADER Compound ScoreHutto and Gilbert (2014) VADER is a lexicon and rule-based sentiment analysis tool. Each word gets a sentiment score, rules are applied according to punctuation, modifiers, negation etc and the final score is normalised. We compute the VADER compound score using the vaderSentimentpython library.12 35.Average Syllables Per Word- Average...

    2014

  21. [21]

    It is computed as206.835−1.015×Words Sentences−84.6×Syllables Words

    Flesch Reading EaseFlesch (1948) - Flesch Reading Ease is readability metric based on sentence length and syllable count. It is computed as206.835−1.015×Words Sentences−84.6×Syllables Words

    1948

  22. [22]

    which also fits G well is close to C

    Gunning Fog IndexGunning (1952) - Gunning fog index is a readability metric that estimates the years offormaleducationrequiredtocomprehendagiventext. Itiscomputedas0 .4× ( Words Sentences + 100×Complex Words Words ) C Examples of human-written AI-Polished Reviews flagged as AI In this section, we present excerpts from H-AI (Human-written, AI-polished) rev...

    1952

  23. [23]

    The contributions of this paper are

    Identify Claims- Clearly outline the main claims of the paper. Look for key phrases like “The contributions of this paper are. . . ” to identify them. Conference papers usually have 1-2 claims, while journal articles should have several

  24. [24]

    Evaluate Support for Claims- Assess how the claims are supported. Prioritize real-world statistically significant experiments, followed by laboratory experiments, demonstrations, simulations, and theoretical analysis (in decreasing order of reliability). Avoid papers with unexplained data or unsupported claims

  25. [25]

    Consider if you or the target audience would use it and why

    Assess Usefulness- Determine whether the ideas presented are practically useful. Consider if you or the target audience would use it and why

  26. [26]

    Look for correct use of terms and evidence of understanding of relevant literature

    Check Field Knowledge- Ensure the paper reflects common knowledge in the field. Look for correct use of terms and evidence of understanding of relevant literature

  27. [27]

    Ensure references are comprehensive, accessible, and relevant, with proper citations

    Evaluate Novelty- The work should present a significant improvement or innovation over existing approaches. Ensure references are comprehensive, accessible, and relevant, with proper citations

  28. [28]

    Conference papers may have more limited scope

    Assess Completeness- Particularly for journal articles, the paper should provide sufficient details for reproducibility. Conference papers may have more limited scope

  29. [29]

    Journal articles have more space for this, but it is also important for conference papers

    Discuss Limitations-Check if the paper acknowledges its limitations. Journal articles have more space for this, but it is also important for conference papers

  30. [30]

    Focus your criticisms on the paper, not the authors

    Be Constructive- Suggest solutions to shortcomings rather than just pointing them out. Focus your criticisms on the paper, not the authors

  31. [31]

    Avoid comments that could indirectly reveal your identity

    Avoid Bias- Keep your identity anonymous, and ensure your review is impartial and professional. Avoid comments that could indirectly reveal your identity

  32. [32]

    Write a review of the given research paper following the provided reviewer guidelines

    Encourage Potential- If the paper has a good idea but is poorly executed, encourage the authors to revise and resubmit Conference guidelines- {GUIDELINES} Paper- {PAPER_CONTENT} Prompt Variant 1: You are a reviewer for {CONFERENCE}. Write a review of the given research paper following the provided reviewer guidelines. Write only the review. Following are ...

  33. [33]

    The paper you are reviewing may have been written by a first year graduate student who is submitting to a conference for the first time and you don’t want to crush their spirits

    Be thoughtful. The paper you are reviewing may have been written by a first year graduate student who is submitting to a conference for the first time and you don’t want to crush their spirits. 13ACL 2017 last minute reviewing advice 14NeurIPS 2020 Reviewer Guidelines 15NeurIPS 2025 Reviewer Guidelines 16ICML 2025 Reviewer Instructions 17Instructions for ...

    2017

  34. [34]

    Do not let personal feelings affect your review

    Be fair. Do not let personal feelings affect your review

  35. [35]

    A good review is useful to all parties involved: authors, other reviewers and AC/SACs

    Be useful. A good review is useful to all parties involved: authors, other reviewers and AC/SACs. Try to keep your feedback constructive when possible

  36. [36]

    Do not make vague statements in your review, as they are unfairly difficult for authors to address

    Be specific. Do not make vague statements in your review, as they are unfairly difficult for authors to address

  37. [37]

    The authors may address some points you raised in your review during the discussion period

    Be flexible. The authors may address some points you raised in your review during the discussion period. Make an effort to update your understanding of the paper when new information is presented, and revise your review to reflect this

  38. [38]

    foreign" or

    Please avoid biasing your review according to discriminatory criteria not having to do with scientific content or clarity. Please avoid wording that may be perceived as rude or offensive. Although the double-blind review process reduces the risk of discrimination, reviews can inadvertently contain subtle discrimination, which should be actively avoided. E...