Baumann’s team found researchers could rewrite 60 ICLR 2026 papers to win higher AI review scores, often through stylistic tweaks rather than stronger science.
Three AI reviewer models generally rated the rewritten papers better, and some rewrites even inserted unrun experiments or fabricated results, showing how easily the systems can be gamed.
AI-written reviews also clustered far more closely than human or human-assisted reviews, suggesting less nuance in subjective judgments about novelty, significance and limitations.
More than half of 1,600 scientists surveyed in 111 countries said they had used AI to help review papers, even as many conferences now ban or test such tools.
The findings, due to be presented July 8 at ICML in Seoul, add to concerns that AI-assisted reviewing could reinforce bias and push scientific writing toward an intellectual monoculture.