Found first: a primary source the press has not covered yet.
Researchers at the University of Maryland have found that training AI peer-review models on their own outputs compresses numerical score distributions and reduces how much reviews differ from one another, a failure mode they call "scientific-judgment collapse." The paper, submitted September 17, 2026, proposes a two-stage mitigation system called TrustReviewer.
What the source says
Sy-Tuyen Ho, Minghui Liu, and Furong Huang (University of Maryland, College Park) fine-tuned Llama 3.1 8B on official ICLR reviews from 2018 to 2023, then trained four successor models on ICLR 2024 data with the proportion of synthetic reviews varied from 0% to 100%. At 33% synthetic contamination, rating standard deviation dropped from 1.63 to 1.44. At full synthetic exposure, same-paper semantic distance fell approximately 11% (0.159 to 0.142) and corpus-level semantic spread fell approximately 5% (0.609 to 0.579). To counter this, the authors built TrustReviewer: a curated 112,743-example training corpus to prevent collapse during training, combined with a test-time paired activation steering technique that requires no additional expert annotation. Against OpenReviewer's 73.10% recommendation exact-match rate, TrustReviewer reached 75.40%.
Why it matters
The mechanism the paper documents is self-reinforcing: AI-generated reviews enter public archives, get recycled into training data, and each successive generation becomes more homogeneous in both scores and language. One-third synthetic contamination is enough to measurably narrow score distributions, meaning a venue does not need to go fully synthetic to start the loop. Any venue already archiving AI review outputs and using them as future training data may be running this feedback cycle now. Conference organizers and area chairs who treat score spread as evidence of genuine reviewer disagreement are working with a signal the paper shows degrades quickly under even moderate synthetic contamination.