ICML 2026: Assigning Rules Is Not Assigning AI Use
A review critiques a research report; a policy states allowed behavior. Kim and colleagues’ author-posted draft (arXiv v1, 2026-09-16) studied ICML’s January–April 2026 review process. The rules covered AI language-model tools: one banned their use during review; the other allowed help understanding reports and related work and polishing reviewer-written text—not evaluating reports or drafting reviews. Among reviewers accepting either rule and assigned only reports allowing limited language-model use, rules were assigned by chance; within 940 reports reviewed under both rules, permission slightly increased length. Differences in report scores, reviewer confidence and supervisors’ quality ratings were near zero. Separately, 165 of 732 ban-assigned and 155 of 425 permission-assigned survey volunteers reported disallowed use. These are not all reviewers. Rules—not actual use—were randomized. Near-zero differences do not prove identical effects; quality impressions do not establish accuracy.
0
1
Tags
Scientific Oversight of AI-Assisted Research
1Cademy