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

TL;DR

This study explores how humans' beliefs in their own and AI's abilities influence delegation decisions in collaborative tasks, revealing biases and the complex role of contextual information in human-AI teamwork.

Contribution

It introduces experimental evidence on the impact of efficacy beliefs on AI reliance, highlighting biases and providing design guidelines for better collaboration.

Findings

Efficacy beliefs act as cognitive anchors leading to AI optimism.

Contextual information can reduce AI optimism bias.

Efficacy discrepancies influence delegation decisions more than team performance.

Abstract

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.