생성형 AI 음성대화의 구조적 특성: 대화분석을 통한 ChatGPT-4o 연구

초록

This study examines how Conversation Analysis (CA) contributes to research on real-time voice interactions with generative AI (GenAI). While ChatGPT-4o demonstrates considerable capacity for coherent information exchange, it remains structurally limited in approximating human interaction. Drawing on transcribed data from 15 participants engaged in mock-interview, counseling, and discussion tasks, this study focuses on three foundational features of conversational organization: turn-taking, adjacency pairs, and repair. The findings fall into two categories. The first concerns ChatGPT-4o's interactional capacity and its limits: while it handles question-answer sequences and responds to user-initiated repair, it treats silence as a turn-transition signal and fails to suspend talk during overlap, which drops information and triggers repair. The second concerns users' practices for managing these troubles, such as recycling turn beginnings, reformulating queries, and switching codes to elicit appropriate responses. By tracing these phenomena turn by turn, CA's sequential analysis offers an empirical method for identifying the structural limitations of GenAI voice systems and for guiding the design of more interactionally fluent voice interfaces.

키워드