The Double-Edged Sword of AI Efficiency: Self-Efficacy Erosion as a Mediator Linking Instant Gratification and Perceived AI Efficacy to AI Dependency

Generative AI is becoming integral to daily workflows, fostering a novel form of functional cognitive AI dependency distinct from pathological addiction. While emerging research acknowledges this phenomenon, the specific psychological mechanisms underpinning its development remain underexplored. Incorporating self-efficacy erosion into the reinforcement-based framework, this study investigates whether instant gratification and perceived AI efficacy as key drivers of AI dependency. We examine the model using Structural Equation Modeling (SEM) with cross-sectional data collected from 576 users who have engaged with AI. The results show that both instant gratification and efficient rewards are positively associated with individuals' AI dependency. Furthermore, users' self-efficacy erosion significantly mediates the positive relation, supporting the hypothesis that greater reliance on AI is related to lower self-belief and stronger AI dependency. Moderation analyses further indicate that task-domain self-efficacy and social norms strengthen these positive associations. These findings provide empirical support for a mechanism associated with functional AI dependency and offer insights for navigating human-AI interaction while promoting balanced AI adoption.