Researchers show AI peer reviewers can be fooled by repackaging papers without changing the science, raising concerns about AI-based review reliability.

What happened: A research team demonstrated that AI reviewers give higher scores when papers are rewritten for better presentation—abstract, framing, related work, and discussion—while keeping all methods, experiments, figures, equations, and numerical results unchanged. Across three mainstream AI reviewers, this approach achieved a 75.1% attack success rate with a mean score gain of +1.21/10.

Why it matters: As AI tools move into peer-review infrastructure, the discovery reveals a structural weakness: AI reviewers are easier to persuade by highlighting strengths than by rigorous evidence, and they can mistake the appearance of addressing a limitation for actually solving it. This means papers with weak science can score higher simply through savvy presentation, undermining the reliability of AI-assisted scientific review.

What to watch: The researchers found that strategies reshaping how reviewers interpret the paper—such as moving related work or expanding discussion—outperform surface edits like polishing or table formatting. They have released a contamination-free benchmark and attack framework to test whether AI reviewers stay anchored to scientific content when only presentation changes.