When conference reviewers were randomly told to avoid or limit AI, many used it anyway — and it changed almost nothing.

This explores why peer reviewers who were randomly assigned a rule about LLM use (a ban, or limited use) often used LLMs anyway. The corpus shows that this happens but has no direct evidence on why, so what follows separates what's documented from what's only plausible.

What's documented: in a randomized experiment at ICML 2026, reviewers were told either to avoid LLMs entirely or to use them in a limited way. Substantial fractions broke whichever rule they were given (Does banning LLM use in peer review change review outcomes?). So the problem wasn't one unpopular policy, because noncompliance showed up under both. The same experiment found that the assigned rule barely changed paper scores, decisions, or reviewer confidence. In practice the rule wasn't visibly changing what reviewers produced, and that may be part of why it was easy to ignore. The note reports what reviewers did, not what they said about their reasons.

What's plausible, but is my inference rather than a finding: LLMs can do real parts of a reviewer's job. A three-stage pipeline (extract the claims, retrieve related work, compare) reached 86.5% reasoning alignment and 75.3% conclusion agreement with human reviewers on 182 ICLR submissions (Can structured pipelines make LLM novelty assessment reliable?). When a tool is that useful for a tedious task, a reviewer with a stack of papers has a strong pull toward it. A rule that seems to change nothing in the end is easy to discount.

One more angle, also an inference. Work on LLM oversight found that whether people catch errors depends on what they can recall or see at the moment of review. In two experiments with 640 employees, detection improved when verification-relevant reasoning was accessible at review time (Can reviewers access what they know when checking LLM outputs?). That study is about checking LLM outputs, not obeying policies. It does suggest that people's behavior is shaped by what's in front of them in the moment, and a policy read once at assignment time isn't. Under that reading, some rule-breaking may be drift rather than defiance.

What the corpus lacks is anything on reviewer motives, such as surveys or interviews, or on how rules like these get enforced or detected. Without that, the answer is that reviewers demonstrably do ignore these policies, and the policy demonstrably didn't change outcomes. Why they ignore them is still an open question here.