Abstract
<p>The rapid development of generative artificial intelligence (AI), particularly largelanguage models (LLMs), has created new opportunities for measuring complex andculturally embedded phenomena. However, existing applications largely focus on textbaseddata and predefined classification tasks, leaving multimodal and interactionalphenomena underexplored.This study proposes a framework for analyzing humor in video data by introducinghumor events as the unit of analysis, defined as temporally bounded sequences linkingan utterance, its multimodal delivery, and audience laughter. Using a large corpus ofSaturday Night Live (SNL) clips, we extract and classify 5,690 humor events with amultimodal generative AI model. To examine how AI can annotate complex culturalphenomena, we compare three prompt strategies: a definition-based prompt, aninteraction-oriented prompt, and a theory-informed prompt grounded in sociologicaltheories of humor. Results show that the theory-informed prompt substantially improvesperformance, particularly for categories requiring contextual and multimodalinterpretation. Beyond performance, the study demonstrates that prompt designsystematically shapes how cultural phenomena are measured. While aggregate categorydistributions remain stable across prompts, label assignments at the event level vary instructured ways, indicating that prompts function as measurement frames rather thanneutral instructions.The study makes three main contributions. First, it introduces an interactional andmultimodal unit of analysis for video-based research, providing a concrete approach tooperationalizing meaning in context. Second, it reconceptualizes generative AI notmerely as a classifier but as a theory interpreter, capable of operationalizing sociologicalconcepts when guided by theoretical prompts. Third, it shows that prompts themselvesconstitute measurement frameworks, suggesting that prompt design should be treated asan object of methodological scrutiny—potentially subject to evaluation and review in itsown right. More broadly, the findings highlight that cultural phenomena are bothstructurally patterned and interpretively open, and that AI-based measurementinevitably reflects the analytical frames embedded in prompt design. Althoughannotating complex cultural phenomena remains challenging, the results suggest thatgenerative AI offers a viable and scalable approach for video-based analysis whenguided by explicit category definitions and theoretically informed prompts, providing asignificant methodological advance over existing approaches</p>