AI-Assisted Peer Reviews Vulnerable to Simple Manipulation, Study Finds

A new study posted on arXiv shows that AI systems used to assist with scientific peer review are alarmingly easy to manipulate. The researchers found that a simple, low-cost attack — superficially rephrasing a manuscript's abstract — can mislead these AI tools, even without knowing which AI model is being used and without making any change to the underlying scientific content or claims.

Many peer-review systems now rely on AI to help screen manuscripts, assist reviewers, and triage submissions to editors. Proponents argue these tools reduce reviewer burden and speed up publication. But the study reveals that this efficiency comes with a hidden cost: the AI's assessment can be gamed by attackers looking to get a more favorable review or suppress a competitor's paper.

The vulnerability poses a significant threat to scientific integrity. Since these AI systems influence editorial triage and reviewer decisions, a bad actor could exploit the weakness to unfairly sway a paper's evaluation process. The researchers emphasize the need for robust safeguards to prevent such manipulation before AI-assisted review becomes more widespread.

The full paper is available on arXiv: https://arxiv.org/abs/2606.10159