Training AI reviewers on AI reviews can distort judgment
AI peer review has a feedback-loop problem. A new paper warns that model-written reviews can enter training data and bend the judgment of later AI reviewers. The proposed defense starts with tracking where reviews came from and checking results against people.
AI Agency
What the source reports
An arXiv paper asks what may happen when AI-written peer reviews train the next AI reviewer. The digest says old views can pass into the next model and warp its calls. The paper points to two checks: mark where each review came from, and test AI calls against human calls. This brief is based on the abstract. It does not give the model setup, data size, or size of the effect.
Original source
- Title
- When AI Reviews Train AI Reviewers: Scientific-Judgment Collapse and Mitigation
- Publication
- arXiv
- Date
- Monday, September 21, 2026