AI tools raise technical empowerment but also heighten job insecurity, cancelling out performance gains; workers with moderate self‑efficacy suffer the largest negative indirect effects.

Main Finding

AI use produces a suppression effect on employee performance: its direct technological empowerment is counteracted by a psychological threat pathway via job insecurity. Job insecurity mediates a negative indirect effect of AI on performance (standardized indirect effect = −0.22). Self-efficacy moderates the AI → job insecurity link in an inverted U‑shape: employees with moderate self‑efficacy experience the highest job insecurity and the strongest negative indirect effect on performance.

Key Points

• Dual pathways: AI → (1) technological empowerment (positive) and (2) psychological threat via job insecurity (negative). The negative indirect path offsets the positive effects, yielding a suppression structure.
• Mediation results:
  • AI application → job insecurity: β = 0.416, p < .001.
  • Job insecurity → supervisor-rated job performance: β = −0.528, p < .001.
  • Direct AI → job performance: β = 0.072, p = .241 (non‑significant).
  • Indirect (AI → job insecurity → performance): standardized effect = −0.22 (95% CI [−0.347, −0.114]).
• Curvilinear moderation by self‑efficacy:
  • Self‑efficacy shows a significant inverted U‑shaped moderating pattern such that job insecurity is highest at moderate levels of self‑efficacy (supporting the authors’ Hypothesis 5).
  • The curvilinear conditional indirect effect (AI → insecurity → performance) is strongest at moderate self‑efficacy (Hypothesis 6).
• Construct reliability and validity reported as acceptable (Cronbach’s α > .70, CR > .80, AVE > .50).
• Model explanatory power: R2 ≈ 0.17 for job insecurity; R2 ≈ 0.31 for job performance.

Data & Methods

• Design: Cross‑sectional, multi‑source matched survey.
• Sample: 392 employee–supervisor pairs from 45 Chinese cross‑border e‑commerce firms (data collected July–August 2025). Sample features: 59.2% female; 85.8% bachelor’s degree or higher; majority ≤40 years old.
• Measures:
  • AI application: adapted Information Systems Use (ISU) scale for AI functions (12 items).
  • Job insecurity: Hellgren et al. (1999) 7‑item scale.
  • Self‑efficacy: General Self‑Efficacy Scale (10 items).
  • Job performance: supervisor rated 10‑item task + contextual performance scale.
  • Controls: gender, age, education, tenure.
• Analysis:
  • Reliability/validity checks (pilot + main sample).
  • Mediation and moderated mediation estimated using Hayes’ PROCESS macro with 5,000 bootstrap samples.
  • Curvilinear moderation tested by including linear and squared terms of self‑efficacy and interactions; conditional indirect effects computed at moderator levels (−1 SD, mean, +1 SD).

Implications for AI Economics

• Measurement and accounting of AI gains:
  • Productivity measures that ignore psychological costs (e.g., job insecurity) may overstate net benefits of AI. Economists modeling AI’s labor impact should incorporate worker‑level psychological channels as mediators of realized productivity.
• Heterogeneous labor effects:
  • The inverted U moderation implies non‑monotonic heterogeneity: workers with moderate self‑efficacy are most vulnerable to AI‑driven insecurity. Aggregate labor-market models should allow for nonlinear heterogeneity in worker responses (not just high vs low skill).
• Human capital and retention:
  • Job insecurity induced by AI can depress performance and raise turnover risk, potentially eroding human capital and offsetting automation gains. Firms and policymakers should factor in these dynamic human‑capital losses when evaluating AI investments.
• Policy and organizational interventions:
  • To capture technological gains while minimizing psychological losses, interventions should combine skill training, transparent workforce planning, and measures to reduce perceived occupational threat (re‑skilling subsidies, internal mobility programs, psychological safety initiatives).
  • Targeting support to moderately self‑efficacious workers may yield disproportionate returns because they appear most likely to experience heightened insecurity.
• Research and evaluation guidance:
  • Causal and dynamic assessment: cross‑sectional evidence points to suppression and nonlinear moderation but longitudinal and experimental designs are needed to estimate persistence, reallocation effects, and net welfare impacts.
  • Broader economic models should include: (a) psychological transmission channels (job insecurity, technostress), (b) heterogeneous worker beliefs/resources (self‑efficacy), and (c) potential feedbacks (turnover, learning, firm productivity).
• Policy caution:
  • Macroeconomic or sectoral forecasts of employment/wage effects from AI should not treat AI adoption as a simple productivity shock — psychological and behavioral responses can materially alter realized output and labor supply decisions.

Suggestions for economists: when evaluating AI’s labor-market consequences, integrate micro‑foundations that capture psychological mediators and allow for nonlinear heterogeneity in worker responses; design empirical strategies (panel data, employer experiments) that can separate technological augmentation from psychological displacement.