The peer review process for top AI conferences has quietly become a minefield of desk rejections, and the rules are no longer about the paper itself. A recent analysis of NeurIPS, ICML, and ICLR policies reveals that since 2023, editors can reject submissions without reading them for reasons ranging from a co-author's negligent reviewing to an out-of-date OpenReview profile or even suspected LLM use in certain cases. This shift from content-based checks to behavioral and procedural policing is understandable, but it risks punishing researchers for infractions that have little to do with the quality of their science. As we Navigating AI-Generated Peer Reviews at ICLR and Why ICLR authors and reviewers are going silent on rebuttals have shown, the review ecosystem is already strained by ghost reviewers and silent rebuttals, adding more pre-review hurdles without addressing the underlying capacity crisis feels like treating a fever with a stricter dress code.
What this means for you as a researcher or practitioner is that submitting to a top venue now requires navigating a compliance checklist that extends far beyond your actual contribution. You can be desk rejected because a colleague on a different paper submitted a negligent review, or because your OpenReview profile hasn't been updated in a year. These rules are responses to real problems, collusion rings, fake conflicts, and reviewer gaming, but they place the burden of proof on honest authors. The ICML 2026 blog post on LLM review policies makes a reasonable case for transparency, but it also signals that conferences are shifting enforcement costs onto submitters rather than investing in better reviewer matching or training. We should ask whether a hard cap on submissions, as discussed in Exploring a hard cap on ML conference papers to sharpen research focus, might do more to restore sanity than adding another layer of pre-screening bureaucracy.
The core tension is between trust and scale. When conferences received a few thousand submissions, desk rejections for formatting errors were rare. Now that the big three see well over ten thousand papers each cycle, every line of policy is a friction point designed to reduce load. But the trend is troubling: rejection without review for behavioral infractions introduces a subjective element that contradicts the very purpose of blind peer review. A co-author's negligence is not your negligence, yet the policy treats the lab as a collective risk. This is not a call to abandon enforcement of ethics, but a warning that we are quietly normalizing a system where your work can be dismissed before anyone reads a word of it, based on metadata and reputation rather than ideas.
The specific detail to watch is how conferences handle the LLM-use rule. ICML 2026's policy allows desk rejection if an author fails to disclose AI assistance in writing or reviewing, but the enforcement mechanism is unclear, will there be automated scanning, or will it rely on whistleblowers? If the latter, we are entering an era where the most powerful tool in a reviewer's arsenal is not their expertise but their suspicion. That is not a sustainable foundation for scientific progress.