From Plausible to Actionable: A Position on LLM Self-Explanations

Regarding the self-explanation generated by large language models (LLMs), this paper points out that it is highly malleable, with questionable fidelity and high operability. The author believes that traditional evaluation protocols have limitations and suggests expanding evaluation criteria to include the dimension of operability. The article aims to provide a practical guide for evaluating the malleability and fidelity of LLM self-explanation, emphasizing the use of LLMs’ rationalizing capabilities to support informed decision-making and appropriate actions by multiple stakeholders.