Humor as a Natural-Language Cipher: Why ChatGPT Misinterprets the "Mume" Episode (Analysis Based on the Situation up to December 2025)

This paper analyzes why ChatGPT systematically misinterprets the “Mume” episode from Makura no Sōshi, focusing on the structural limitations of large language models as of December 2025. The study argues that classical Japanese humor operates as a form of natural-language cipher based on semantic discontinuity, multi-stage reinterpretation, and intention inference—features that cannot be formalized into a single deterministic mapping. In contrast, ChatGPT is optimized for semantic continuity, pattern completion, and the production of unique, stable interpretations. This structural mismatch explains why the model defaults to symbolic or metaphorical readings instead of recognizing humor. The paper further distinguishes humor from quizzes, slang, and jokes, demonstrating that humor relies on contextual jumps and cultural knowledge that resist algorithmic extraction. The analysis provides a framework for understanding the inherent limits of LLMs in interpreting culturally embedded, discontinuous forms of meaning.