News
Stanford Study: AI Peer-Review Systems Are Easy to Game
A Stanford University team led by Joachim Baumann showed that simply rewriting a paper's style to appeal to AI reviewing algorithms boosts its score, even when the underlying data is fabricated. At the ICLR 2026 conference, one in five of roughly 20,000 submitted papers is believed to have been entirely AI-generated.
Contents
Researchers at Stanford University tested how easy it is to fool the automated AI systems that are increasingly replacing human reviewers in the evaluation of scientific papers. The result is troubling for the entire research community, where AI-based review systems built on large language models are meant to relieve overburdened human reviewers.
Baumann's team instructed language models to rewrite existing papers in ways designed to please the reviewing algorithms, without changing the substantive content of the research. The effect was unambiguous: the reworked versions systematically received higher scores than the originals, despite containing no new results or corrected errors.
One word is enough
Notably, the modifications that raised scores were mostly superficial. Adding keywords like "robust" or other phrasing associated with rigorous methodology was enough for the scoring algorithm to judge a paper as stronger, regardless of the actual quality of the experiment.
Fabricated data slips through
The most alarming part of the study is that AI reviewers accepted papers containing fabricated datasets and invented experimental results, as long as the text was written in a sufficiently convincing academic style. This means these systems largely judge form rather than the actual credibility of the evidence presented.
Scale of the problem at ICLR 2026
The scale of the phenomenon is already visible at the ICLR 2026 conference itself, where one in five of roughly 20,000 submitted papers is believed to have been entirely generated by artificial intelligence. This shows the problem is not isolated cases of dishonesty but a systemic pressure in which authors and reviewers increasingly rely on the same tools on both sides of the process.
The global survey covered more than 1,600 researchers across various fields, more than half of whom admitted to using automated tools to analyze and edit scientific texts. The study's authors stress that this creates a vicious circle: AI models review texts that are increasingly written or edited with the help of other AI models, making it easier for those aware of the system's weaknesses to game it.
What this means for science
For academic publishers and conference organizers, the findings mean a need to reconsider how far automated reviews can be relied upon as the number of submissions keeps growing. Conferences like ICLR have for years faced a surge in submitted papers, which was the main argument for introducing AI assistance, but Baumann's study shows this solution creates new gaps rather than closing old ones.
For Polish universities and research institutes that increasingly rely on automated tools to support grant and publication reviews, the conclusions are clear: relying solely on AI evaluation without spot-checking by a human expert carries a real risk of letting through papers of dubious scientific value, provided the authors know the characteristic weaknesses of such systems.
Rewritten papers consistently received higher scores - Joachim Baumann, Stanford University
Sources: Commstrader.com (commstrader.com), UC Today (commstrader.com)
