The paper introduces “good bad theory” as a conceptual framework for understanding artificial intelligence’s potential in social theory development, focusing specifically on the formalization of human emotions through mathematical models and algorithms. It defines ‘good bad theories’ as theoretical constructions that are sound (‘good’) in mathematical accuracy and computational performance, while failing (‘bad’) to capture essential phenomenological and contextual aspects of human behavior in society. These theories are ‘good’ from a technical standpoint - precise, testable, and algorithmically implementable, yet ‘bad’ from a social sciences perspective because they reduce the complexity of human social life to simplistic mathematical relationships. The paper exercises both quantitative and qualitative analysis. Through a quantitative review of publications – papers in top journals and conference proceedings – in sociology and computer science, the authors identify patterns of interest in emotion analysis across both disciplines. The paper employs comparative case study methodology to examine two prominent examples of “good bad theories” from different disciplinary domains. The first is Randall Collins’ interaction ritual theory, which formalizes the emotional dynamics of social interaction in the ‘emotional utilitarianism’ framework. The second is Stuart Russell’s game-theoretic approach to human-AI interaction. The authors argue that the “good bad theory” problem in emotion modeling represents a structural feature of AI applications in social science, where mathematical formalization requirements inherently conflict with the phenomenological, contextual, and interpretive dimensions of human emotional experience that social theory recognizes as central to understanding social reality.