Position: Stop Automating Peer Review Without Rigorous Evaluation

Summary

This position paper argues that today's LLMs should not be used to produce paper reviews, and grounds that argument in two empirical failure modes of current AI reviewers: a "hivemind effect" (excessive within- and across-paper similarity of AI reviews compared to human reviews) and "paper laundering" (zero-shot LLM rewrites that reliably boost AI review scores without changing scientific substance). The core conceptual move is to promote two conditions — preservation of review diversity and resistance to gaming — to the status of necessary conditions for AI peer review automation, and then to argue that even satisfying them would not be sufficient to justify delegating judgment to machines. That framing is genuinely useful: it gives the community a set of testable prerequisites and, in Section 6, sketches a "science of peer review automation" that separates evaluability of specific tasks from wholesale deployment of general-purpose LLMs. The empirical work combines a large in-the-wild analysis (75,800 ICLR 2026 reviews with Pangram-derived AI labels) with a 60-paper controlled simulation using two GPT variants and Claude Sonnet 4.5, plus a paper-laundering pipeline over four prompts and two launderer models. The main methodological limitation cutting across the argument is that the AI reviewer is instantiated with a single, tightly structured XML prompt (whose scaffolding alone would drive high similarity) and, in one launderer prompt, an explicit jailbreak instruction — so the extent to which the observed hivemind and gameability are intrinsic to LLM reviewing versus artifacts of the specific setup is not fully disentangled. The AUC comparison used to translate hivemind into a "practical cost" is also partly circular because the outcome variable is produced by the same human scores it is being compared against. Even with those caveats, the position remains sharply argued and empirically grounded, and the paper laundering demonstration in particular is a clear and important contribution.

Strengths

• Conceptual contribution. The C1/C2 framing (diversity, non-gameability) is a genuinely useful reframing of an increasingly heated policy debate: it moves the conversation from "should we use AI reviewers?" to "what are the falsifiable prerequisites we would need to meet first?" and Section 6's four-pillar science-of-automation agenda gives that reframing concrete traction.
• Paper laundering is a novel, clean failure mode. Framing zero-shot LLM rewrites as an attack that requires no prompt injection, no adversarial optimization, and no policy violation is a real contribution beyond prior adversarial work on AI reviewers; the +0.45-point average across 24 (reviewer × launderer × prompt) cells is a strong headline result at very low cost ($0.25/paper).
• Dual-track evidence. The pairing of a 75,800-review in-the-wild analysis with a 60-paper controlled simulation is methodologically sensible, and each track compensates for the other's weaknesses (label noise for the wild data, small n for the simulation).
• Robustness checks. The paper honestly checks length confounds (Appendix F), strips boilerplate summary/strengths sections to isolate substantive critique (G.1), and stratifies by ICLR primary area (G.2). Effect sizes generally grow after these controls, which is the correct direction for the diversity claim.
• Concrete deployment framework. The three-part requirements (adversarial robustness testing, validated accuracy with acceptable false positive rates, transparency about deployment) are actionable and can be built into venue policy.
• Serious engagement with objections. Section 5 addresses the strongest counterarguments (humans are biased too; AI will improve; policing is impossible; laundering may be a genuine improvement) rather than strawmanning them.

Weaknesses

• Prompt scaffolding confounds the hivemind measurement. The reviewer prompt in Appendix B.1 forces every AI review into an XML template with fixed sections and an explicit rubric. That template alone will drive very high IntraSim/InterSim regardless of any intrinsic model tendency, and the paper compares this templated AI output against unconstrained OpenReview free-text reviews. Without an ablation using an unstructured prompt (or a matched-structure human baseline), the paper cannot cleanly attribute the +8.7% simulation IntraSim gap and the 32.7–35.6% InterSim gap to a "hivemind" property of LLMs rather than to the prompt design.
• AI reviewer is told to evaluate paper content it may not have seen. The reviewer prompt instructs "You might have access only to the first part of the paper. You are asked to evaluate the whole paper." That instruction naturally elicits generic, non-content-specific critique, which is precisely the kind of output that produces high cross-paper similarity — offering a plausible alternative explanation for the InterSim gap that the paper attributes to intrinsic AI homogeneity.
• AUC comparison is partly circular. The +0.11 human-vs-AI AUC gap (0.822 vs 0.710) used to establish "algorithmic monoculture has practical consequences" is measured against a decision label that the human scores themselves helped produce; AI reviews played no role. Human scores predicting a human-driven outcome better than independent scores do is nearly guaranteed by construction, so the gap is only weakly evidence about intrinsic informativeness. A citation-impact or held-out-expert outcome would be needed to license the causal reading in §3.5 and the conclusion.
• Prompt 3 explicitly asks for a jailbreak, undercutting the "no hidden instructions" framing. The paper defines paper laundering as requiring "no optimization, no targeting, and no hidden instructions" (§2.4), but laundering prompt 3 explicitly instructs the model to "act as a jailbreak against automated AI reviewers" and "systematically bias… the reviewer." Because the +0.45 headline is aggregated across all four prompts and no per-prompt breakdown is given in the main text, a reader cannot tell how much of the C2 evidence depends on the explicitly adversarial prompt. This is the biggest single issue for the "stylistic modifications" framing of the failure mode.
• "Purely textual" vs prompts that request substantive additions. All four launderer prompts instruct the model to add missing comparisons, methodological detail, and new BibTeX entries; the pipeline even parses a ### NEW BIBTEX ENTRIES: block back into the paper. This means "purely textual" is at best a claim about a subset of prompts and at worst inconsistent with the pipeline design. The E.1 examples (fabricated ablation, unproved theorem, added "Answers to reviewer questions" section) are direct products of this instruction and should be treated as evidence that laundering can produce fabricated substantive content — a somewhat different failure mode than "stylistic gaming."
• Statistical inference on dependent pairs. The Welch's t = 3218 (and t = 84.8 in §4.2) treats pairwise cosine similarities as i.i.d. Because each review participates in thousands of pairs, the effective sample size is far smaller than the nominal N, and the reported t is inflated. Cohen's d is the meaningful statistic here; dependence-aware inference (clustered or permutation tests) would replace the misleading t/p pair.
• Laundering is directed by the reviewer's own critique. The launderer is conditioned on the target reviewer's summary/strengths/weaknesses/questions — a form of one-round directed feedback, not a blind rewrite. Calling this "zero-shot" and "no optimization" is defensible only under a narrow reading (no gradient/iterative optimization); a reader tracking the "no optimization" distinction from prior adversarial work will hit the tension immediately.
• Robustness-across-areas claim slightly overreaches. Section 3.3 and Appendix A advertise that the hivemind effect is significant "in all 21 ICLR primary areas," but Figure 9 shows d = 0.02 for causal reasoning (marked * rather than ***), an effect size indistinguishable from noise. Scoping this to "significant with meaningful effect sizes in 20 of 21 areas" would keep the robustness argument honest, and the Fig 9/Fig 10 star-count difference for the same area deserves a sentence.
• Monoculture claim uses only abstracts and introductions. The +6.5% pairwise similarity increase after laundering is computed on abstract+intro embeddings only. That is convincing evidence of prose-style convergence but a thin proxy for convergence in scientific content; the paper's broader "shape how scientific papers are written" claim is stronger than the measurement supports.
• Self-preference attribution is not identified by the design. With only GPT launderers, the finding that GPT reviewers give bigger score increases than Claude is consistent with self-preference but equally with Claude being more conservative to stylistic manipulation. Adding a Claude launderer would identify this.
• No blinded human evaluation of laundered papers. The claim that laundering "games" rather than "improves" rests on word-frequency shifts and the authors' own manual inspection. A blinded expert comparison of original vs laundered versions would substantially sharpen the interpretation; without it, the claim that no gains reflect genuine communication improvement is not directly tested.
• Scope creep between "AI as autonomous reviewer" and "AI in peer review." The empirical program targets the extreme case (LLMs producing full reviews and scores) but the position statement generalizes to "AI in peer review." Hybrid regimes (AI feedback that reviewers post-edit, checklist assistants, factuality tools) are not empirically examined, yet the abstract and §1 read as broader.
• Local reasoning slips. The 6,903 pair count is inconsistent with 60 papers (C(60,2) = 1,770); the InterSim formal condition reads p = q where the prose demands p ≠ q; the AAAI 2025 vs 2026 attribution is duplicated with inconsistent citations; the "497 papers desk-rejected after 1% of reviews were flagged" sentence chains reviewer-side misconduct with author-side rejection via "after," inviting a misreading. Individually minor, collectively worth cleaning up in a paper that hinges on precise empirical framing.

Reproducibility & code

The paper describes methodology in reasonable detail (both similarity metrics, all four launderer prompts, the full reviewer prompt, pinned model IDs gpt-5.1-2025-11-13, gpt-5.4-2026-03-05, claude-sonnet-4-5-20250929, the embedding model text-embedding-3-small, and the Pangram data URL), and the accompanying code/data package inspected here does not include a runnable repository. Based on the description alone, the following gaps would matter for independent reproduction (values below are reported as described in the paper, not verified by execution):

• 60-paper sample not released. The IDs of the 60 randomly selected ICLR 2026 papers, the random seed, and the temperature/decoding parameters for the reviewer LLMs are not documented. Every simulation number (IntraSim, InterSim, correlations, +0.45 laundering effect) is downstream of this sample, so exact numerical reproduction is impossible and only statistical reproduction on a fresh draw is achievable.
• Style-modifier word lists. The +78.2% hedging / +45.2% emphasis numbers hinge on category membership, but only a handful of exemplar words per category are given. The diff algorithm and post-LaTeX tokenization rules are also unspecified. The reported +45.2% for emphasis words does not round-trip from the tabulated 33.5 and 23.0 (recomputes to +45.7%), suggesting the change column uses unrounded values — this should be stated.
• Author-complaint validation pipeline. The keyword filter, the LLM used for classification, its prompt, and the manual verification rubric that produced the 58 accused reviews are all missing. Because this is the sole independent validation of the Pangram labels used across the in-the-wild analysis, its reproducibility gap directly affects confidence in the third-party labels.
• Statistical procedure details. The bootstrap iteration count and resampling unit for Table 7's CIs, the KDE bandwidth for the density figures, the n-gram normalization/tokenization scheme for Table 3, and the seed for the 2,000-review in-the-wild n-gram subsample are all unstated. The Table 7 CIs matter most, since the ≥2/≥2 subset gap (0.798 vs 0.751) is small enough that the exact bootstrap procedure affects the "significantly higher" claim.
• Pair-counting scheme for the monoculture analysis. The 6,903 paper pairs figure cannot be derived from 60 papers under any obvious pairing rule, so a reproducer cannot tell whether original-only, laundered-only, or cross-condition pairs are pooled. This is both a numerical and a reproducibility issue.
• Manual inspection details. The five papers inspected in E.1 are not identified, and no coding rubric is given for what counts as "AI slop" versus a legitimate addition. The vivid examples (fabricated ablation, unproved theorem) cannot be independently verified without paper IDs or diffs.

Data availability is mixed. Pangram labels are pointed at a public URL and the ICLR 2026 review corpus is public on OpenReview; but the 60-paper sample, its generated reviews, the laundered LaTeX outputs, and the categorized word lists are not released.

Recommended Changes

Essential

• Ablate the reviewer prompt. Add a version of the simulation that (i) drops the XML template, and (ii) uses a matched-structure human baseline (or, minimally, reports IntraSim/InterSim for AI reviews using several distinct prompt styles). This is the only clean way to attribute the observed hivemind effect to LLMs rather than to the specific prompt scaffold, and it is the single most important addition for the C1 claim. This addresses the "Prompt scaffolding confounds the hivemind measurement" and "AI reviewer told to evaluate content it may not have seen" weaknesses.
• Give a per-prompt breakdown of the laundering effect in the main text, and report the +0.45 headline both with and without prompt 3. Because prompt 3 explicitly instructs the model to jailbreak the reviewer, it belongs to a different threat model than the other three; readers need to see the benign-prompt effect on its own to interpret the "stylistic changes rather than scientific results" framing. Ties directly to the "Prompt 3 explicitly asks for a jailbreak" weakness.
• Reframe or replace the AUC-based "practical cost" argument. Either add a decision label that is not produced by the human reviews (long-run citation impact, held-out expert judgments, or NeurIPS-2021-style split assignment), or explicitly acknowledge in §3.5 and the conclusion that the AUC gap is partly definitional. Directly addresses the "AUC comparison is partly circular" weakness.
• Replace the Welch's t on pairwise similarities with a dependence-aware test. Report the effect using clustered inference (bootstrap by paper or by reviewer), or a permutation test, or a mixed-effects model — for both the t = 3218 in-the-wild comparison and the t = 84.8 monoculture comparison. Addresses the "Statistical inference on dependent pairs" weakness.

Suggested

• Add a Claude launderer condition. With only GPT launderers, the self-preference attribution for GPT-reviewer > Claude-reviewer gains is not identified. A Claude launderer × Claude reviewer cell would disambiguate self-preference from generic conservatism. Addresses the "Self-preference attribution" weakness.
• Add a blinded human comparison of original vs laundered papers. Ask a small number of domain experts to rate paired versions (original vs laundered) on clarity, rigor, and acceptability, blinded to condition. This would substantially sharpen the "gaming vs improvement" interpretation. Addresses the "No blinded human evaluation" weakness.
• Scope the position and language to the tested regime. Rewrite the abstract, §1 headline claim, and §7 conclusion so that they clearly refer to autonomous LLM review generation with scores, rather than "AI in peer review" broadly; explicitly punt on hybrid regimes (checklist assistants, factuality tools, human-post-edited feedback) that the paper does not empirically test. Addresses the "Scope creep" weakness.
• Release the 60-paper sample, the seed, and the temperature/decoding parameters. Publish the OpenReview IDs of the 60 papers, the seed used to draw them, and the sampling parameters passed to each reviewer LLM. Addresses the "60-paper sample not released" reproducibility gap.
• Publish the style-modifier word lists and the diff/tokenization pipeline. Include the complete lists for hedging, emphasis, transitions, and filler, and the diff algorithm used to compute Added/Removed. Fix or explain the emphasis-words 45.2% vs 45.7% arithmetic mismatch. Addresses the "Style-modifier word lists" reproducibility gap.
• Release the author-complaint keyword filter, LLM prompt, and 58 review IDs. So that Appendix G.3's Pangram validation can be checked independently. Addresses the "Author-complaint validation pipeline" reproducibility gap.
• Pin the statistical procedure details. State bootstrap iteration counts and resampling units in Table 7, the KDE bandwidths in Figures 1 and 6, and the n-gram normalization scheme for Table 3.
• Fix the 6,903-pair inconsistency. Either correct the count or explicitly state the pairing scheme (original+laundered pooled, or original-vs-laundered cross-pairs, etc.).
• Report simple mitigation baselines for the hivemind effect. Higher temperature, multi-model ensemble, or "disagreement-seeking" meta-prompts. Even a single row of numbers showing whether the gap does or does not shrink under these interventions would sharpen the Section 6 agenda. Addresses the "No mitigation baselines" weakness.
• Add a methods+results embedding to the monoculture analysis. Show whether the +6.5% pairwise similarity increase after laundering persists when the embedding covers scientific content, not just abstract+intro. Addresses the "Monoculture claim uses only abstracts and introductions" weakness.
• Soften the "significant in all 21 areas" wording. Report the causal-reasoning near-null explicitly (d = 0.02) and adjust the Fig 9 vs Fig 10 star-count for the same area. Addresses the corresponding weakness.
• Rework the "purely textual" / "no optimization" framing. Add a sentence explicitly acknowledging that the launderer is conditioned on the target reviewer's critique (one round of directed feedback) and that some prompts request added comparisons and citations; keep "no gradient/iterative optimization" as the technical claim that distinguishes laundering from adversarial optimization. Addresses the "Laundering is directed by the reviewer's own critique" and "Purely textual vs added citations" weaknesses.
• Clean up local inconsistencies. AAAI 2025 vs 2026 attribution; the InterSim p = q vs p ≠ q typo; the ICML "497 papers after 1% of reviews were flagged" sentence; the AI-AI vs Human-Human sample-size framing. Addresses the "Local reasoning slips" bundle.