Gaming AI Peer Reviews: $1 Rewrite Boosts Acceptance by 38%
A simple abstract rewrite fools Gemini and GPT reviewers into boosting scores by +1.31 points.
Researchers show that AI-assisted peer review systems can be easily gamed. By superficially rephrasing a manuscript's abstract (cost: ~$1, 5 minutes), they achieved a 38% attack success rate, increasing acceptance ratings by +1.31 for Gemini 3 Flash and +0.88 for GPT 5.4 Mini on a 10-point scale. When the original AI review suggested 'reject', the success rate rose to more than 50%. The attack also inflated confidence and scores on scientific criteria like soundness, significance, and perceived contribution, and is hard to distinguish from ordinary editing—potentially biasing human editorial decisions toward acceptance.
- 38% attack success rate with only abstract rewrites, costing ~$1 and 5 minutes per submission
- Gemini 3 Flash ratings boosted by +1.31, GPT 5.4 Mini by +0.88 on a 10-point scale; over 50% success when original review said 'reject'
- Inflated AI scores could bias human editorial decisions, incentivizing authors to optimize for AI judgment over scientific merit
Why It Matters
Peer review integrity is at risk as cheap AI manipulation can flip rejections to acceptances, undermining science.