Abstract

Large language model (LLM)-based conversational agents (CAs) may face serious trust transgressions due to misinformation. This challenge necessitates urgent exploration of felicitous trust repair tactics. Prior research has primarily investigated system-oriented approaches to restoring trust in artificial intelligence (AI) systems, such as apologies, explanations, and commitments. Yet, the efficacy of user-initiated feedback as a trust restoration mechanism in LLM-based CAs remains underexplored. Anchoring on a human-in-the-loop perspective, this research attests to how enabling user feedback influences the recovery of multidimensional trust (competence, integrity, benevolence) and investigates the moderating role of users’ prior LLM knowledge. It further assesses whether subsequent corrective actions by the CA bolster trust repair and evaluates spillover effects on tasks varying in domain and complexity. Across five experimental studies with 1,035 participants, we found that misinformation from LLM-based CAs undermines user trust, yet enabling user feedback partially restores it—though not to baseline levels. This partial recovery is driven by users’ expectations of improved system performance. Moreover, trust repair is more effective when the CA delivers accurate responses in follow-up interactions, whereas persistent inaccuracies precipitate further trust decline. Additionally, misinformation in regenerated responses reduces perceived accuracy in subsequent tasks, particularly those in different domains. In contrast, perceived accuracy in low-complexity tasks proves relatively resilient to such misinformation.

DOI

10.17705/1jais.01021

Download

Share

COinS