Provoking unstable and evasive ChatGPT behaviour with anti-chain-of-thought: Japanese rock song lyrics case, a year later

Considerably attention has been directed towards assessing the capabilities of ChatGPT and other Large Language Models (LLMs) in various aspects such as syntax accuracy, factual correctness, comprehensive world understanding, ethical alignment, common sense, and formal logical reasoning. However, the bulk of the research has primarily focused on analyzing "statically"generated texts, where only a single interaction between a human and LLMs was documented. More sophisticated methodologies involving open-ended discussions or debates between humans and LLMs yield far more intriguing outcomes. These reveal flawed rhetorical patterns, including circular arguments, self-contradictions, topic evasion, inconsistency, and a blend of passive aggression with attempts to please the human disputant. In this paper, we present an original observation of such behaviour occurring during a ChatGPT dialogue session centred around the translation of Japanese song lyrics.