Can AR-Embedded Visualizations Foster Appropriate Reliance on AI in Spatial Decision Making? A Comparative Study of AR See-Through vs. 2D Minimap

Xianhao Carton Liu, Difan Jia, Tongyu Nie, Evan Suma Rosenberg, Victoria Interrante, Chen Zhu-Tian

Abstract: In high-stakes, time-critical scenarios-such as emergency evacuation, first responder prioritization, and crisis management -- decision-makers must rapidly choose among spatial targets, such as exits, individuals to assist, or areas to secure. Advances in indoor sensing and artificial intelligence (AI) can support these decisions by visualizing real-time situational data and AI suggestions on 2D maps. However, mentally mapping this information onto real-world spaces imposes significant cognitive load. This load can impair users' ability to appropriately judge AI suggestions, leading to inappropriate reliance (e.g., accepting wrong AI suggestions or rejecting correct ones). Embedded visualizations in Augmented Reality (AR), by directly overlaying information onto physical environments, may reduce this load and foster more deliberate, appropriate reliance on AI. But is this true? In this work, we conducted an empirical study (N = 32) comparing AR see-through (embedded visualization) and 2D Minimap in time-critical, AI-assisted spatial target selection tasks. Contrary to our expectations, users exhibited greater inappropriate reliance on AI in the AR condition. Our analysis further reveals that this is primarily due to over-reliance, with factors specific to embedded visualizations, such as perceptual challenges, visual proximity illusions, and highly realistic visual representations. Nonetheless, embedded visualizations demonstrated notable benefits in spatial reasoning, such as spatial mapping and egocentric spatial imagery. We conclude by discussing the empirical insights, deriving design implications, and outlining important directions for future research on human-AI decision collaboration in AR.

External IDs:dblp:journals/corr/abs-2507-14316