Perceived Privacy Invasion and Consumer Adoption Willingness of AI-Powered Personalized Recommendations: The Mediating Role of Psychological Resistance and Moderating Role of Perceived Algorithm Transparency

Artificial intelligence (AI)-powered personalized recommendation systems can enhance relevance and convenience, but their reliance on extensive personal data processing may also heighten users’ perceptions of privacy intrusion. This study examines the relationship between perceived privacy invasion and consumer adoption willingness, with psychological resistance as a statistical mediator and perceived algorithm transparency (PAT) as a moderating condition. Drawing on the privacy-cost perspective of privacy calculus theory and psychological reactance theory, this study focuses on the privacy-related inhibition pathway of AI recommendation adoption. Survey data were collected from 518 Chinese adult Internet users with recent experience of AI-powered personalized recommendation services and analyzed using partial least squares structural equation modeling (PLS-SEM). The results show that perceived privacy invasion was negatively associated with consumer adoption willingness (β = −0.357) and positively associated with psychological resistance (β = 0.279), whereas psychological resistance was negatively associated with adoption willingness (β = −0.286). Psychological resistance statistically partially mediated the association between perceived privacy invasion and adoption willingness (indirect effect = −0.080). The interaction between perceived privacy invasion and PAT was negative and statistically significant (β = −0.162), indicating that the positive privacy invasion–resistance association was less pronounced at higher levels of PAT. Supplementary conditional process analysis further showed that the negative indirect association through psychological resistance weakened as PAT increased, with a significant index of moderated mediation (index = 0.0464, 95% bootstrap CI [0.0191, 0.0816]). Additional analyses using common-method-bias diagnostics, PLSpredict and CVPAT, and a targeted Gaussian copula robustness check for PAT provided complementary evidence regarding measurement robustness, predictive relevance, and the robustness of PAT-related estimates. Age-based multigroup analysis revealed no statistically significant age-based heterogeneity in the structural path coefficients across the 18–29, 30–49, and 50+ age groups. This study contributes to the literature by identifying psychological resistance as a motivational pathway associated with privacy-related adoption responses and by conceptualizing PAT as a user-level perceptual boundary condition in AI-powered personalized recommendation contexts.