Study Finds AI Peer Review Scores Improve After Mostly Stylistic Paper Rewrites

Computer scientists demonstrated that scientific papers can be rewritten to obtain higher scores from AI reviewer models, with most changes limited to stylistic adjustments. Joachim Baumann of Stanford University and colleagues will present the findings July 8 at the International Conference on Machine Learning in Seoul, South Korea.

The team selected 60 papers from submissions to the 2026 International Conference on Learning Representations.

They first prompted AI models to generate detailed reviews modeled on those produced by human reviewers at the conference. Two large language models then rewrote the papers to address the feedback. In most cases, three separate AI reviewer models assigned higher scores to the rewritten versions.

Nearly all modifications were stylistic. ” In some instances the models added findings from experiments that had not been conducted. AI-generated reviews of the 60 papers were far more similar to one another than human or human-assisted reviews, both before and after the rewrites.

The rewritten papers themselves also converged in language and structure compared with the originals. Of nearly 20,000 papers submitted to ICLR 2026, about one in five were fully AI-generated, according to a November case study by Pangram. A December survey of 1,600 scientists across 111 countries found that more than half had used AI tools to assist with peer review.

Many conferences now prohibit AI tools for peer review, while others are testing their quality before possible integration. Baumann noted that objective checks such as detecting hallucinated references are straightforward, but subjective judgments about a paper’s contribution remain difficult for AI to evaluate.