AI as Area Chair: Model Ranks All 6617 ICML Papers, Diverges Sharply From Humans

Shayne Redford and Seungone Kim's team ran a pilot study called "AI as Area Chair": language models played the role of area chairs, anonymously ranking all 6617 accepted papers at ICML 2026 without seeing authors or human review scores, and the results were compared against human ACs' oral/spotlight decisions to probe how AI's "research taste" differs from that of human researchers.

Confirmed

Why it matters

As AI agents increasingly take part in the research process, models with consistent, systematic "tastes" influencing topic selection, reviewing, and writing could create homogenizing bias in academia. This study is the first to quantify such divergence at the scale of a full top-conference submission pool, providing a data foundation for discussing the risks of AI involvement in peer review

2026-10-07 ~ 2026-10-07 · 6 related posts

Primary sources