Modeling Trust Recalibration in AI Dialogue: Conversational Repair Strategies in ChatGPT

Abstract

As conversational AI systems, like ChatGPT, become embedded in everyday activities such as education, decision-making, and communication, trust between users and these systems is emerging as a critical factor in their sustained effectiveness and acceptance. However, trust is not static; it can degrade when users encounter ambiguous, incorrect, or misaligned responses. While prior research has extensively addressed trust formation and erosion, little is known about how trust can be actively repaired during dialogue. This paper addresses that gap by introducing the concept of trust recalibration – the process through which conversational agents recover user trust after a breakdown in interaction. We analyze four distinct repair strategies (explicit correction, clarification, apology with rephrasing, and meta-cognitive reflection) using simulated ChatGPT dialogues with annotated trust trajectories. Based on this analysis, we propose a lightweight rule-based model for predicting trust drift and recovery, supported by both quantitative metrics and qualitative dialogue evidence. Our results show that timely, context-appropriate repair strategies significantly enhance trust recovery, especially when matched to the domain of interaction. The paper concludes with a trust recalibration flow model and design recommendations for building more transparent, self-aware AI dialogue systems.

Ähnliche Arbeiten