The Expectancy-Capability Gap in Generative AI: When Privacy protection Fails to Follow Knowledge

Abstract

Generative Artificial Intelligence (Gen-AI) systems are reshaping digital interaction, but they raise urgent concerns about data privacy due to their opaque design and dynamic learning processes. Despite growing interest in algorithmic accountability, existing privacy behaviour models fail to account for how users manage privacy in environments where system logic is non-transparent and control options are limited. To address this gap, this study develops an integrated model to examine how knowledge, motivation, and behavioural skills influence privacy protection in Gen-AI, drawing on the IMB framework, Self-Efficacy Theory, and Expectancy-Value Theory. We conducted a survey with 298 participants recruited from universities and professional networks. Structural Equation Modelling using Partial Least Squares (PLS-SEM) was employed to analyse both direct and conditional pathways, including moderated mediation effects. Findings reveal two novel patterns: the expectancy–capability gap, in which users hold high expectations for AI privacy protections but feel limited in their ability to act; and the privacy-confidence paradox, where greater AI experience reduces users perceived self-efficacy due to awareness of system limitations. The study also introduces a dual-knowledge framework separating technical use knowledge from privacy-specific knowledge, each of which independently predicts privacy self-efficacy. Challenging traditional privacy calculus, this research reveals perceived benefits don't drive Gen-AI privacy protection. Instead, user privacy settings preferences and confidence in tool use are pivotal. It offers a theory-driven model for Gen-AI privacy dynamics, with practical implications for transparent, skill-fostering, and trust-calibrated AI systems. This research challenges traditional privacy models by showing that user confidence and privacy preferences, rather than perceived benefits, drive protective behaviour in Gen-AI. It offers a theory-based model highlighting motivational, cognitive, and contextual factors, with practical implications for designing user-centred, transparent, and trust-calibrated AI systems.