Why do people resist AI even when it helps them? A narrative review of the psychology of AI resistance

People encounter artificial intelligence (AI) in workplaces, public services, and consumer products that demonstrably improve outcomes, yet a substantial fraction of users refuse, avoid, or actively resist these systems. This paper asks which psychological factors most strongly predict AI resistance, and how strong the evidence is for each: rational threat perception (job loss, status loss, loss of control) or general technophobia and low understanding. It is a narrative synthesis of more than 80 primary sources (peer-reviewed empirical studies, three large meta-analyses, large-scale public-attitude surveys, and recent field experiments), assembled through targeted rather than systematic search; effect sizes are not pooled and the review should not be read as a PRISMA-style systematic review. The headline verdict is that rational threat perception and, above all, loss of autonomy dominate the resistance picture, while the popular deficit model (people fear AI because they do not understand it) runs against the limited experimental evidence available (one adult RCT, N=94; one preregistered 27-country study): the only published adult AI-literacy randomized trial shows that brief literacy training increases fears of bias, privacy, and job displacement, and the strongest preregistered evidence (Tully, Longoni and Appel 2025; 10 studies across 27 countries) finds that lower objective AI literacy predicts greater AI receptivity, mediated by perceiving AI as magical. A non-negotiable principle-based floor also exists: Friis and Riley 2026 estimate that 12% of U.S. occupations categorically refuse AI even under perfect conditions, and the underlying capability-by-personalization framework of Qin et al. 2025 explains the wide heterogeneity. For practitioners, the intervention hierarchy implied by the evidence is: (1) preserve user agency; (2) frame the task, not the AI; (3) build trust through use, not communication; (4) involve workers in design; (5) match anthropomorphism to context; (6) provide calibrated explainability; (7) accept the moral-refusal floor.