The Dynamics of Trust and Reliance in Human-AI Interactions

As artificially intelligent (AI) systems increasingly support human decision-making, it becomes essential to understand how people build and sustain trust over time. While trust is widely recognized as a key factor for effective human-AI interactions, much of the existing research relies on momentary insights into trust beliefs, often measured in simplified or static testing environments. In reality, however, interactions with AI are ongoing and dynamic: they unfold through repeated encounters shaped by changing AI system performance, diverse task contexts, and people’s evolving beliefs and growing experiences. The goal of this dissertation was to explore how trust beliefs and reliance behaviors develop over time in repeated or long-term human-AI interactions with varying AI performance levels, system properties, decision-making contexts, or interaction durations. It was guided by three research questions. 1) How do trust and reliance develop over time in human-AI interactions? 2) How do prior trust perceptions or reliance behaviors influence subsequent beliefs and behaviors? 3) How do various types and timings of algorithmic errors affect the development and recovery of trust and reliance over time? These questions were explored through multiple controlled laboratory experiments and qualitative fieldwork. Two studies examined how AI accuracy and the types of explanations supporting AI advice influence trust and reliance in a repeated legal decision-making task. A third study investigated the effects of early and late AI errors within a repeated legal decision-making scenario. The fourth study involved a delivery planning task inspired by real-world logistics, analyzing how changes in AI performance, AI framing, and varying levels of task uncertainty impact trust and reliance across four consecutive sessions over two weeks. Finally, a qualitative case study explored long-term human-AI interactions based on interviews with logistics professionals who have worked with an AI planning tool for several years. By varying study designs in terms of methodology, task context, and experiment duration, this dissertation provides both empirical contributions and methodological advancements. Empirically, this dissertation enhances our understanding of how trust and reliance develop through repeated or long-term interactions. Although the two constructs (trust beliefs and reliance behavior) are related, they follow distinct trajectories over time and are influenced differently in human-AI settings. The findings reveal that people do not always adapt optimally to changes in system performance. For instance, certain initial experiences can anchor people’s behavior, hindering proper calibration of reliance and leading to temporary overreliance or underreliance. Over time, increased familiarity with an AI system made participants more tolerant of errors, indicating that trust in AI can recover, although this may not always be appropriate. Finally, while AI performance was identified as a key factor, the development of trust and reliance was also influenced and moderated by contextual factors such as task-related, social, and ethical considerations. Methodologically, the dissertations build on previous research on trust and reliance in human-AI interactions by applying and comparing multiple perspectives (e.g., self-report, behavior), exploring trust and reliance across different time frames (repeated tasks, multi-session, long-term reflection), and including prior experiences in the analysis. These quantitative approaches are complemented by in-depth insights from studying experts in their professional environments. Overall, the findings emphasize the need to move beyond one-time or static assessments of trust. They highlight the importance of understanding how trust and reliance develop over time, how they are influenced by prior experiences, and how they fit into larger decision-making contexts. These insights inform future research and the design of AI systems that promote balanced human-AI collaboration. This means systems should not only be technically reliable but also designed with awareness of how users interpret, accept, and ultimately adapt to algorithmic advice for their reasoning, including considerations of responsibilities and social environments. Even a technically trustworthy system can fall short if users cannot or are unwilling to understand its role or implications. Supporting trust calibration, therefore, requires attention not only to AI performance but also to the context in which systems are used, the meaning they hold for users, and how well they align with human needs.