Can AI systems safely replace human peer reviewers?

The position paper argues that "today's AI systems should not be used to produce paper reviews," and grounds this in two necessary conditions: C1, preservation of review diversity, and C2, resistance to gaming. On C1 it reports a hivemind effect: AI reviewers agree more within papers (IntraSim +8.7% to +9.8%) and across papers (InterSim +4.1% to +39.8%) than humans do, both in simulation and in real ICLR 2026 reviews. On C2 it introduces "paper laundering," a fully automated zero-shot LaTeX rewrite. Across 24 conditions (60 sampled ICLR 2026 papers, four prompts, two launderer models, three reviewer models), laundering raised AI review scores by +0.45 (p < 0.0001), at about $0.25 per paper. The paper takes its scale from a third party: Emi (2025) labels 15,899 of 75,800 ICLR 2026 reviews (21%) as AI-generated.

The paper's reason for holding AI to a higher bar is distributed versus centralized error. Human biases are spread across reviewers with different expertise and "partially cancel out" in aggregation. AI errors are correlated because models trained on similar data share biases, which the paper calls algorithmic monoculture. Gaming follows the same logic: gaming one human reviewer does not transfer to others, while "a single rewrite strategy can boost scores across models." The homogenization is measured, not only argued. Laundered papers become more similar to each other (pairwise similarity +6.5%, Cohen's d = 1.02), so the paper's worry is that a centralized system shapes "not only which papers are accepted but also how those papers are written."

The gaming result overlaps with How much does rhetorical style shift AI review scores?, which reaches a similar conclusion through a paired design, with the same paper in more and less favorable rhetorical versions. This excerpt adds the diversity condition and the cross-paper homogenization, which that note does not report. The paper also draws the opposite inference from Can human review keep pace with AI-accelerated research generation?: "the peer review crisis is real. An effective solution requires validated tools, not a simple replacement of human judgment." The ICML 2026 two-policy framework the excerpt cites is measured in Does banning LLM use in peer review change review outcomes?, but that study concerns reviewers' own LLM use, not AI systems writing reviews. The tested reviewers are general-purpose models and the agents of Bianchi et al. (2025b), not a purpose-built pipeline like Can inference scaling help reviewers catch errors humans miss?.

The excerpt does not establish several things. It gives no figures for the claim that AI-generated ratings are "less informative about final acceptance decisions" than human ratings, which appears only in the conclusion. The AI-review labels come from a third party, and the authors name a single reviewer prompt as a limitation, with the rest left to an appendix not included here. The claim that laundered scores reflect no "genuine improvement of scientific content" rests on the rewrites being purely textual; no human assessment of the laundered papers' substance appears in the excerpt. The narrower reading the evidence supports is that, for the models and prompt tested, AI reviewer scores move with textual rewrites, so a score cannot be read as merit without a robustness check. The broader case for accountability and democratic legitimacy is argued, not measured, and the paper's own answer is "cautious automation" until those risks are studied.