Chung2026
| Chung2026 | |
|---|---|
| BibType | ARTICLE |
| Key | Chung2026 |
| Author(s) | Yonhee Chung, Yo-An Lee |
| Title | When AI suspends the floor: users' interactional management of guideline disruptions in ChatGPT |
| Editor(s) | |
| Tag(s) | EMCA, In press, Conversation analysis, Human-AI interaction (HAI), Trust, ChatGPT, Accountability, Voice-enabled AI, AI Reference List |
| Publisher | |
| Year | 2026 |
| Language | English |
| City | |
| Month | |
| Journal | AI & Society |
| Volume | |
| Number | |
| Pages | |
| URL | Link |
| DOI | 10.1007/s00146-026-03246-9 |
| ISBN | |
| Organization | |
| Institution | |
| School | |
| Type | |
| Edition | |
| Series | |
| Howpublished | |
| Book title | |
| Chapter | |
Abstract
Voice-enabled conversational AI mediates consequential interactions from job interviews to counseling, yet little is known about what happens when it interferes with ongoing talk. This study examines guideline-triggered interruptions in ChatGPT-4o—moments when the system cuts off its own turn mid-turn with a refusal and topic-change request. Applying conversation analysis to 100 voice-enabled sessions, we identified 72 events across 18 participants. Users respond along three user-initiated trajectories: pursuing repair, bypassing the intervention, and opting out, and a fourth, structural pattern, the cascade, in which interferences fire repeatedly and foreclose uptake. The analysis indicates two independent systems, a conversational engine and guideline module, that share the interface without coordination, producing output overlap. In a smaller subset of events, cross-linguistic template leaks interrupt second-language users’ target-language production. These interferences disrupt turn-by-turn alignment and shift the burden of coherence onto users. These interventions raise questions about the trust, accountability, and competence of generative AI. In reality, they are managed by the user within the unfolding interaction. The findings carry implications for the design, evaluation, and governance of voice-enabled AI: trust is not a static disposition but emerges turn by turn through users’ interactional labor. Tracing the sequence poses a question, namely, how to rely on systems that are fluent and coherent in interaction, yet cannot register when they have broken it.
Notes
