Workers’ reliance on AI reduces innovation not by the tool itself but by eroding employees’ confidence, and ethical AI leadership can neutralize that harm. The study finds no direct drag from AI dependence on innovative behavior once self-efficacy and leadership are accounted for.

Main Finding

Using survey data from 421 full‑time employees in South Korea (three‑wave, time‑lagged design), the authors find that dependence on AI does not directly suppress employee innovative behavior. Instead, AI dependence reduces innovation indirectly by eroding employees’ general self‑efficacy (a full mediation). Importantly, AI ethical leadership (leaders who insist on AI transparency, accountability, fairness and reaffirm human judgment) neutralizes the negative effect of AI dependence on self‑efficacy, thereby preserving innovation.

Key Points

• Theoretical framing: Bandura’s Social Cognitive Theory (personal vs proxy agency). AI dependence ≈ proxy agency; innovative behavior ≈ high personal agency; self‑efficacy is the central personal‑agency mechanism.
• Hypotheses tested:
  • H1: AI dependence → (negative) innovative behavior.
  • H2: AI dependence → (negative) self‑efficacy.
  • H3: Self‑efficacy → (positive) innovative behavior.
  • H4: Self‑efficacy mediates the AI dependence → innovation link.
  • Moderation: AI ethical leadership buffers the AI dependence → self‑efficacy pathway.
• Empirical results:
  • No significant direct negative effect of AI dependence on innovative behavior.
  • AI dependence significantly reduces self‑efficacy.
  • Self‑efficacy positively predicts innovative behavior.
  • The indirect pathway (AI dependence → ↓self‑efficacy → ↓innovation) accounts for the suppression of innovation (full mediation).
  • High AI ethical leadership neutralizes the agency‑eroding effect of AI dependence on self‑efficacy; under high ethical leadership, the negative mediated pathway is attenuated or disappears.
• Conceptual implication: the problem is not AI per se but deprivation of mastery experiences (psychological de‑skilling); managerial / institutional context determines whether AI undermines or coexists with human agency.

Data & Methods

• Sample: 421 full‑time employees in South Korea.
• Design: three‑wave time‑lagged survey (to reduce common method bias and establish temporal ordering among variables).
• Constructs measured (survey scales): AI dependence (degree to which employees rely on AI for core tasks/decisions), general work self‑efficacy (personal agency), innovative behavior, AI ethical leadership (leader behaviors around AI transparency/accountability/fairness and safeguarding human judgment).
• Analytical approach: hypothesis tests of direct effects, mediation (self‑efficacy as mediator), and moderation (AI ethical leadership moderates AI dependence → self‑efficacy), integrated as a moderated‑mediation model. (The paper reports full mediation and a significant buffering/moderation effect of ethical leadership.)
• Theoretical validity: builds on Bandura’s sources of self‑efficacy (enactive mastery, vicarious experience, social persuasion) to explain how leadership can substitute for lost mastery experiences.

Implications for AI Economics

• Human capital dynamics: Organizational AI adoption can produce short‑term productivity gains while gradually eroding workers’ general problem‑solving confidence (a form of deskilling). Growth models that treat AI as a pure productivity multiplier may overestimate sustained innovation unless they account for human‑capital depreciation through psychological channels.
• Complementarity vs substitution: The study underscores that AI is not deterministically substitutionary — managerial/leadership practices are crucial complements. Investments in leadership, governance, and practices that preserve employees’ mastery experiences can convert AI from a potential human‑capital depressor into a lasting productivity complement.
• Firm investment decisions: Cost–benefit analyses of AI should internalize the downstream risk to organizational innovation capacity. Firms should budget for non‑technical complements (ethical leadership training, role redesign, rotation policies that preserve cognitive engagement) alongside AI system procurement.
• Policy and regulation: Public policy aimed at sustaining innovation (industrial policy, workforce development) should consider interventions beyond transparency/XAI technical fixes — e.g., leadership standards, corporate governance requirements for AI deployment, incentives for firms to maintain human oversight and training.
• Measurement and empirical research: Macroeconomic and firm‑level empirical models of AI impacts should include mediating psychological or organizational variables (e.g., measures of self‑efficacy, managerial AI governance) and allow for heterogeneous effects by leadership/organizational context. Cross‑sectional productivity regressions that omit these channels risk biased estimates of AI’s long‑run effect on innovation and growth.
• Labor market and inequality: If AI adoption without ethical leadership reduces workers’ innovative capacity, then occupations or firms lacking such leadership may experience longer‑term declines in upward mobility and wage growth for knowledge workers, amplifying heterogeneity across firms and sectors.
• Practical recommendation for firms/economists: treat “AI ethical leadership” and similar social governance mechanisms as essential complements when modelling or investing in AI — they preserve human agency, sustain innovation, and thereby protect the long‑run returns to AI adoption.

If you want, I can (a) extract the paper’s empirical tables/figures and translate effect sizes into economic magnitudes, or (b) suggest indicators and survey items to operationalize AI dependence and AI ethical leadership for macro/firm‑level empirical work. Which would be most helpful?