When AI Reviews Science: Can We Trust the Referee?

The researchers propose a safety- and reliability-centered analysis of AI peer review, mapping attacks along the review lifecycle to five stages: training and data retrieval, initial review, in-depth review, rebuttal, and system level. On stratified samples of ICLR 2025 submissions, they conduct four groups of treatment-control probe experiments with two LLM-based AI reviewers, isolating the causal effects of prestige packaging, assertion tone strength, rebuttal sycophancy, and context poisoning on review scores. The paper notes that existing incidents show hidden prompt injections in manuscripts can steer LLM reviewers toward inappropriate positive judgments, and that such systems are vulnerable to adversarial phrasing, authority and length preferences, and hallucinated claims; this taxonomy and experimental audit provide a baseline for evaluating and tracking the reliability of AI review, and identify concrete failure points for targeted mitigation.