AI Reviewers Trained on AI Reviews Start Grading Everything the Same

Peer review — the process where experts read a scientist's paper and decide if it's worth publishing — is quietly being handed partly to AI. Large language models now write review comments, summarize papers, and even recommend accept or reject. That's convenient, because there are far more papers submitted than humans available to read them. But it creates a strange loop: AI-written reviews get posted publicly, get scraped into future training data, and then tomorrow's AI reviewers end up learning from yesterday's AI opinions instead of from real human judgment.

The researchers built that loop in miniature. They started with Meta's Llama 3.1 8B, a free, widely used AI model, and trained it on real reviews from ICLR, a major AI conference, from 2018 to 2023. Then they trained four follow-up versions using a mix of real and AI-generated reviews. As more AI-generated reviews entered the mix, the model's scores squeezed toward the middle and its opinions got less varied — both for any single paper and across the whole batch. In plain terms, it stopped having strong taste. Good papers and mediocre ones started looking alike to it.

So what? This is a small version of a risk that reaches well beyond science. Any system where AI judges things — hiring screens, loan decisions, content moderation, medical triage — can degrade if it's trained on its own past output. You don't get dramatic failure. You get slow sameness: the unusual, the risky, the genuinely new gets quietly rejected, and the safe and familiar wins.

To fight it, the team released TrustReviewer, an open-source reviewer system. It does two things. First, it trains on a hand-picked, higher-quality set of examples so it doesn't inherit sloppy or repetitive judgments. Second, it uses a technique called activation steering — think of it as gently nudging the AI's internal settings mid-answer — to push back against bland verdicts without retraining or hiring new expert annotators. The takeaway: the problem is real, measurable, and at least partly fixable.