[D] ICML rejects papers of reviewers who used LLMs despite agreeing not to

The International Conference on Machine Learning (ICML), a premier academic venue, has made a landmark enforcement decision by rejecting papers based on reviewer misconduct. According to widespread reports on social media platform X (formerly Twitter), the conference identified reviewers who used Large Language Models (LLMs) like GPT-4 or Claude to write or assist their reviews, despite those reviewers having voluntarily selected a 'no LLM use' track during the assignment process. In a consequential move, ICML rejected all papers those reviewers were assigned to evaluate, effectively penalizing the submitting authors for the reviewers' breach of protocol. This represents the first known instance of a top-tier conference taking such direct and punitive action against AI-generated peer review, setting a stark precedent for academic integrity enforcement.

This decision has ignited a fierce debate within the research community. Critics argue the punishment is disproportionately harsh, primarily because it harms innocent authors whose work's fate was tied to a rogue reviewer. The core of the controversy lies in the imperfect science of AI text detection. Current tools for spotting LLM-generated content are known to have limited precision, occasionally flagging human-written text as AI-generated (false positives) and missing sophisticated AI outputs (false negatives). By acting on these detections to reject papers, ICML is wagering on the reliability of systems that the broader community often views with skepticism. The conference's stance forces a critical examination of how academic institutions can practically and fairly govern LLM use, balancing the need for human expert review with the evolving capabilities and temptations of AI assistance.