The Transparency Paradox: Decoupling Perceived and Behavioral Trust Through Trust Calibration in Human–AI Interaction

While explainable AI (XAI) aims to foster trust, its impact on behavioral reliance remains contested. Contrary to the common assumption that transparency linearly promotes reliance, our findings reveal a significant perception-behavior gap. Through a multi-round experiment, we demonstrate that while XAI improves perceived interpretability, it fails to facilitate immediate perceived trust recovery following a system error; instead causing a pronounced decline in behavioral reliance. These results suggest that transparency may function as a cognitive forcing function, triggering analytical System 2 processing that heightens user sensitivity to AI fallibility. This shift reflects a dynamic recalibration of reliance patterns, supported by improved decision accuracy in the XAI condition. Additionally, AI literacy moderates this process as a cognitive scaffold for lower-literacy users. Ultimately, transparency alone is insufficient for trust recovery, highlighting the need to design for trust calibration rather than trust maximization to ensure resilient human-AI collaboration.