How do detection tools shape LLM use enforcement at ICLR?

ICLR's 2026 program chairs call the core of their policy "twofold." An author or reviewer who uses an LLM "must disclose this" and "is ultimately responsible" for its output. Whether or not an LLM was used, nobody may "make false or misleading claims, fabricate or falsify data, or misrepresent results." The first rule is enforced through volume: papers that "make extensive usage of LLMs and do not disclose this usage will be desk rejected." Hallucinated content, "including hallucinated references," is treated as a code of ethics violation, and reviewers whose posts carry such problems face consequences that can include desk rejection of their own submissions.

The enforcement path is what sets the excerpt apart. The chairs say they will "leverage recent LLM detection tools" to flag papers with "a significant amount of LLM-generated content," but the flags go to area chairs (ACs) and senior area chairs (SACs), and the chairs "will only take action if an AC or SAC identifies concrete evidence." Their stated reason is "the possibility of false positives from detection tools." The same two-step sequence governs reviews: authors who received poor reviews, with "many hallucinated references or false claims," are asked to send a confidential message to the ACs and SACs with evidence. Near-duplicate submissions are a separate problem, one the chairs say LLMs can facilitate and may worsen, and the consequences for them are still being written. The chairs also credit public reviews and discussions with letting the community see issues that are "only visible at scale."

Set against the neighboring notes, the excerpt answers a problem the Graphite note leaves open. That note says a flat detector share cannot be separated from detector blind spots, and its false-positive rates go unmeasured. ICLR does not settle the detector question technically. It puts a human between the flag and the sanction, so a false positive costs an area chair's time rather than a paper. The policy also keys on content, false claims and invented references, not on style. That fits the argument in the rhetorical polish note that polish should not be read as merit: a rule aimed at fabricated citations does not depend on how fluent a paper reads. The ICML experiment in the banning versus limiting note tests a different lever. There, rules that banned or limited LLM use in reviewing barely moved scores or decisions, and a substantial share of reviewers broke the rules they were given. The excerpt does not say whether ICLR's wording would change behavior. If the ICML result carries over, the enforcement path is what would matter, and that is the part this excerpt spells out.

The excerpt is a policy announcement, not a measurement. It reports no count of flagged or rejected papers, names no detection tool or its accuracy, and gives no outcome. The chairs promise a later post on the desk rejections, and that post is not part of this text. It also cannot show that undisclosed use is detectable at the threshold the chairs apply, or that the evidence standard catches the cases the policy is written for. At the strength the evidence allows, the policy is a stated design. Whether it works stays open until the follow-up numbers appear.