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

Policies that allow peer reviewers to use large language models for polishing text are currently unenforceable, according to a new study. Researchers found that AI-text detectors cannot reliably distinguish between human-written reviews and those refined by AI, risking false accusations of misconduct. The study, submitted to arXiv on 20 March 2026 and revised on 23 June 2026, directly tests the enforceability of emerging journal policies [1]. Several scientific conferences and journals have recently enacted rules that prohibit LLM usage by peer reviewers, but carve out exceptions for polishing, paraphrasing, and grammar correction of human-written reviews [2]. The authors assembled a dataset of peer reviews simulating multiple levels of human-AI collaboration to test whether these exceptions can be policed [2]. They evaluated five state-of-the-art detectors, including two commercial systems [2]. The analysis showed that all detectors misclassified a non-trivial fraction of LLM-polished reviews as AI-generated [2]. This failure rate means reviewers who follow the rules by using AI only for language polishing could be falsely accused of academic misconduct [2]. The research further investigated whether peer-review-specific signals, such as access to the paper manuscript and the constrained domain of scientific writing, could improve detection [2]. While incorporating these signals yielded measurable gains in some settings, the authors identified limitations in each approach and found that none met the accuracy standards required for identifying AI use in peer reviews [2]. The findings carry implications for public estimates of AI use in scholarly evaluation. The study states that recent public estimates of AI use in peer reviews, derived from AI-text detectors, should be interpreted with caution, as current detectors misclassify mixed human-AI outputs as fully AI-generated, potentially overstating the extent of policy violations [2]. The work was led by Rounak Saha and posted on arXiv, an open-access repository for electronic preprints that has hosted over two million articles as of late 2021 and receives about 24,000 submissions per month [1][8]. The paper is available through arXivLabs, a framework launched in 2020 that allows community collaborators to develop and share experimental tools directly on the site, guided by values of openness, community, excellence, and user data privacy [7].