Generic Joke Generation with Moral Constraints

In this paper, we explore the frontier of AI-generated humor, aiming to create jokes that are not only entertaining but also ethically and emotionally intelligent. We present a novel approach that combines the principles of the Incongruity Theory with the psycholinguistic dimensions of Valence and Arousal, and intertwines them with the foundational elements of morality. To evaluate the humor generated by our models, we employ the Best-Worst Scaling (BWS) method, known for its efficiency in sentiment and preference analysis. Our experimental results, derived from comparing various methodological approaches and state-of-the-art language models, indicate that jokes considering incongruity, emotional, and moral dimensions are rated as funnier than those generated without these considerations. Among the language models assessed, ChatGPT4 exhibited superior performance in humor generation, both in terms of mean humor quality and consistency across evaluations. The findings underscore the importance of a multidimensional approach to computational humor and highlight the potential of AI to create humor that is sensitive to human emotions and ethical standards. Through this work, we contribute to the emerging discourse on AI and humor, offering a framework that can be expanded upon by future research to encompass personalized humor generation, intercultural humor translation, and dynamic humor adaptation. Our research paves the way for AI systems that not only understand and generate humor but do so with an awareness of its moral and emotional impact on diverse global audiences.