Believing vs. Achieving -- The Disconnect between Efficacy Beliefs and Collaborative Outcomes

As artificial intelligence (AI) becomes increasingly integrated into workflows, humans must decide when to rely on AI advice. These decisions depend on general efficacy beliefs, i.e., humans'confidence in their own abilities and their perceptions of AI competence. While prior work has examined factors influencing AI reliance, the role of efficacy beliefs in shaping collaboration remains underexplored. Through a controlled experiment (N=240) where participants made repeated delegation decisions, we investigate how efficacy beliefs translate into instance-wise efficacy judgments under varying contextual information. Our explorative findings reveal efficacy beliefs as persistent cognitive anchors, leading to systematic"AI optimism". Contextual information operates asymmetrically: while AI performance information selectively eliminates the AI optimism bias, data or AI information amplify how efficacy discrepancies influence delegation decisions. Although efficacy discrepancies influence delegation behavior, they show weaker effects on human-AI team performance. As these findings challenge transparency-focused approaches, we propose design guidelines for effective collaborative settings.