Abstract

This paper introduces a novel computational theory of humor by formally equating jokes with cognitive bugs - mismatches or misfires within the predictive models of intelligent agents. We argue that humor arises from the sudden detection and resolution of epistemic errors, and that laughter serves as a public signal of successful model correction. By extending this theory to artificial intelligence, we propose that the ability to generate and comprehend jokes constitutes a form of self-debugging and may serve as a proxy indicator for general intelligence. Importantly, we contend that humor has direct implications for AI safety: systems that fail to recognize joke-like incongruities may also overlook critical misalignments between their behavior and human values. We develop a taxonomy of joke types mapped to software bug categories, explore the limits of formalizing funniness, and propose that humor may represent an AI-complete problem. In doing so, we offer both a theoretical lens and a diagnostic framework for assessing epistemic integrity and alignment robustness in advanced AI systems.