Executive SummaryIn January 2026, GPTZero reported finding 100+ hallucinated citations across 51-53 papersaccepted to NeurIPS 2025, one of the world's most prestigious machine learning conferences.This analysis examines the incident through the lens of Jevons Paradox—the economicphenomenon where technological efficiency gains lead to increased total consumption ratherthan conservation.Core Thesis: The NeurIPS incident represents systemic stress, not system failure. GenerativeAI has reduced the cost of paper production, triggering a submission surge (220% increase2020-2025) that outpaces verification capacity. This is a governance synchronizationproblem—verification protocols haven't AI-accelerated at the same rate as production.Key Finding: The fact that 51-53 papers were caught demonstrates the system retainserror-correction capacity. The "gift" of this incident is visibility of the production-verificationlag—a signal that institutional metabolism must evolve, not collapse.
Smith et al. (Fri,) studied this question.