Not all AI failures are equal: How failure type and user relationship norms shape retention in healthcare chatbots

As artificial intelligence health assistants become an integral component of daily life for millions of users, a critical question emerges: what recourse do individuals pursue when these text-based chatbots malfunction? This research reveals that not all service failures carry equivalent consequences. The study distinguishes between two distinct categories of chatbot service failures: functional failures and non-functional failures. By integrating attribution theory, expectation disconfirmation theory, and relationship norms, we construct a comprehensive framework that elucidates users' continuance intentions. Four scenario-based experiments demonstrate that functional failures exert substantially more detrimental effects on continuance intentions than their non-functional counterparts. The underlying psychological mechanism operates as follows: functional failures precipitate stronger internal attributions of responsibility and greater expectation disconfirmation, thereby engendering more pronounced negative responses. The principal innovation of this research lies in the moderating role of relationship norms: users' relationship norms fundamentally reshape their responses to identical failures. For exchange-oriented users, functional failures generate the most pronounced expectation disconfirmation; conversely, for communal-oriented users, non-functional failures inflict greater harm. This finding suggests that standardized service recovery strategies are inherently untenable. By deepening theoretical understanding of user behavior within human-computer interaction contexts, this research furnishes evidence-based recommendations for online health platforms: organizations should develop personalized recovery strategies tailored to both failure typology and user psychological profiles, rather than implementing generic apologetic responses.