While most research treats trust as a single, general feeling toward an entire AI system, our study looks closer. We empirically explore feature-specific trust—how users calibrate their trust toward distinct, individual capabilities within an AI system.
Focusing on partially automated vehicles as a prime example of physical AI, we used a multi-method empirical approach. By combining quantitative trust measurements over time with qualitative think-aloud protocols and interviews, we uncovered two major insights:
- Trust varies by feature: Users don’t just trust or distrust a car; they trust individual features differently.
- Trust evolves over time: How users trust these features shifts dynamically as they gain firsthand experience.
Our findings show why we must distinguish between individual AI features and account for temporal changes when studying trust. Ultimately, this work offers a new conceptual framework for trust calibration literature and provides practical insights for designing physical AI systems that better align with real user expectations.
