Can AI Take a Joke—Or Make One? A Study of Humor Generation and Recognition in LLMs

Knowing when to joke-and when not to-is a subtle skill often missing in large language models (LLMs).This study examines how well LLMs generate and recognize humor in emotionally sensitive, support-oriented conversations.We introduce two targeted datasets: one for evaluating humor generation across distinct styles, and another for testing humor and speaker role recognition in human-written supportive statements.Using GPT-4o, LLaMA3, and Gemini 1.5, we assess humor style alignment, emotional appropriateness, and role sensitivity.While models produce fluent and stylistically varied humor, they often struggle with contextual nuance and role interpretation.GPT-4o consistently performs best in tone alignment and emotional fit, but subtle humor types remain challenging across models.These results highlight current limitations in LLMs' pragmatic and relational understanding, underscoring the importance of human oversight in humor-sensitive applications.