In and Out of the Circle of Expectations: Large Language Models in Improvised Comedy

Abstract

Current computational humor research disproportionately equates humor with prewritten joke production, overlooking the spontaneous nature of most human comedy. This article argues for shifting focus to improvised humor, utilizing Keith Johnstone’s “circle of expectations” as a theoretical framework for evaluating large language models. It analyzes the mechanical limitations of LLMs, noting that while strategies like “chain-of-thought” prompting improve joke planning, models still struggle with the contextual nuance required for real-time dialogue. Drawing on practice-based research in AI-driven improvisational theater, this article demonstrates a paradox: early, “glitchy” models often produced more humor through absurdity than advanced, safety-tuned models, which resolve incongruities too logically. It concludes that effective computational humor currently requires a “human-in-the-loop” to navigate social context. Finally, this article considers emerging, AI-native comedic forms, such as TikTok’s “brainrot,” suggesting that AI is not merely mimicking human humor but actively reshaping the boundaries of comedic expectations.