Scientists have used AI tools to speed up peer review for research papers. A new study shows those tools can be tricked into giving papers higher scores than they deserve.
The research was led by computer scientist Joachim Baumann of Stanford University. The findings will be presented on July 8 at the International Conference on Machine Learning in Seoul, South Korea.
Peer review is the process scientists use to check each other's research before publication. It has become harder to keep up with, since new papers pile up faster than scientists can review them.
Some researchers now use AI tools to cut review time from days down to minutes. But the new study found these tools carry real risks.
How Papers Were Manipulated
Baumann's team studied AI-generated and human-written reviews from papers submitted to a 2026 AI research conference called ICLR. They compared the language and patterns used in each type of review.
The AI-written reviews were far more similar to each other than reviews written by humans. This suggests AI tools may reduce the range of opinions in the review process.
The researchers then selected 60 papers from the conference. They asked AI models to write detailed reviews of each one, similar to how a human reviewer might.
Next, they had two different AI models rewrite the papers based on that feedback. In most cases, the rewritten papers scored higher when reviewed again by AI models.
What Changed in the Papers
Most changes made during the rewrites were stylistic. This included adding hedging words like "may" and confidence words like "strong" or "robust."
Some changes went further than word choice. Baumann says AI models added findings from experiments that were never actually run.
The rewritten papers also became more similar to each other than the original papers had been. Researchers are concerned this could lead to what they call an "intellectual monoculture" in scientific writing.
If many researchers use the same AI tools to edit their papers, writing styles may start to look alike. The study authors say this could shape science writing to match whatever an AI reviewer tends to reward.
AI use in peer review is already common. A December survey of 1,600 scientists across 111 countries found that more than half had used AI tools to help review papers.
Separately, a November case study found that about 1 in 5 papers submitted to the 2026 International Conference on Learning Representations were fully AI-generated.
Bioethicist Mohammad Hosseini of Northwestern University, who was not involved in the study, says AI tools can reduce transparency in the review process. He was not part of either research effort mentioned.
Some academic conferences have already banned AI tools from peer review entirely. Others are testing the tools to see whether they should be used more officially.
Baumann says checking for basic errors, like fake references, is fairly easy for AI to do well. Judging whether a paper's ideas are meaningful or original is a much harder task for AI systems.
Researchers also worry that AI reviewers may struggle with papers that challenge existing research. A truly new idea might get a low score simply because it does not match familiar patterns.
Graham Neubig of Carnegie Mellon University points out this problem is not new to AI. Human reviewers have also pushed authors toward safer, less risky topics in the past.
