Detecting Unintended Changes of AI Character Personality on Large Language Models

This study proposes a method for calculating personality changes that occur when updating interactive AI characters generated on large language models (LLMs). With the popularization of content distribution services using AI characters such as games, character backgrounds and content story are often getting updated. During such updates, the AI character's personality may change unintentionally or may not change as intended, so that the consistency of the scenario in the content may be lost. To address these issues, the proposed method not only conducts a personality diagnosis on the AI character, but also quantitatively measures the degree of personality change that has occurred for each personality facet when updating the AI character to detect unintended personality changes. In this study, we evaluate the feasibility of the proposed method through experiments using a large language model and a personality diagnostic test.