Employees who lean more on AI report greater performance pressure and anxiety, and these psychological effects are linked to rising, inefficient workplace competition (involution); the relationship appears to run through perceived evaluative pressure rather than a direct impact of AI reliance.

Main Finding

AI reliance in digitally intensive workplaces is linked to greater workplace involution — not directly, but indirectly via higher perceived performance expectations and increased employee anxiety. Structural equation modeling on survey data (N = 550) found a significant sequential mediation: AI reliance → performance expectations → anxiety → involution. The direct AI reliance → involution path was not significant, suggesting the effect operates through employees’ evaluative perceptions and emotional responses.

Key Points

• Conceptual framing: AI can raise visible baseline performance (and perceived expectations), producing a Jevons-like effect where efficiency gains fuel intensified, inefficient competition (involution) rather than reducing workload.
• Behavioral mechanism: Employees’ habitual reliance on AI increases perceived external evaluation pressure; when expectations exceed perceived coping resources this raises anxiety, which motivates defensive, imitative, and escalating effort (theater effect).
• Empirical signals:
  • Correlations: AI reliance correlated with performance expectations (r = 0.242), anxiety (r = 0.352), and involution (r = 0.368), all p < 0.001.
  • CFA fit: χ2/df = 2.91, CFI = 0.972, TLI = 0.964, RMSEA = 0.059, SRMR = 0.045.
  • Reliability: AI reliance α = 0.758 (final 4 items); performance expectations α = 0.68 (3 items); anxiety α = 0.948; involution α = 0.863.
• Controls included demographic variables plus job satisfaction, job control, and organizational commitment.
• Methodological caveats noted by the authors: cross-sectional self-report data from a China online panel; common-method checks were performed (Harman’s test and ULMC).

Data & Methods

• Sample: 550 valid responses collected via Credamo (late 2025); 62% female; 86% aged 21–40; 84.6% bachelor’s degree or higher.
• Measures: established scales adapted and back-translated; final indicators — AI reliance (4 items), performance expectations (3), anxiety (13), involution (9). Item parceling used for larger scales.
• Controls: gender, age, education, income, job satisfaction, job control, organizational commitment.
• Analysis: Confirmatory factor analysis and structural equation modeling (lavaan in R). Mediation tested via bootstrapping (5,000 resamples); mediation significance judged by 95% CIs excluding zero. Common-method bias was assessed and addressed.
• Key statistical outcome: sequential mediation (AI reliance → performance expectations → anxiety → involution) significant by bootstrapping; direct AI reliance → involution non-significant.

Implications for AI Economics

• Efficiency vs. intensity: The paper provides micro-level evidence that productivity-enhancing AI can increase work intensity and competitive arms races within firms, an instance of a Jevons-type effect at the workplace level. Economists should treat AI-driven efficiency gains as potentially increasing costly input (effort/time) rather than unambiguously reducing labor.
• Externalities and intangible capital: Erosion of workplace social capital (trust, cooperation) is an internal negative externality of AI adoption that may reduce long-run firm-level productivity and raise turnover or wellbeing-related costs; these should be incorporated into firm-level cost–benefit analyses of AI investments.
• Measurement and welfare: Standard productivity metrics may overstate welfare gains if they ignore increased anxiety, reduced cooperation, and inefficient overwork. Labor supply responses (hours, intensity) and utility costs matter for social welfare assessments of AI diffusion.
• Incentives and manager behavior: AI-enabled transparency and automated evaluation can recalibrate what is considered "adequate" performance; firms should design appraisal and incentive systems to avoid ratcheting expectations that produce involution.
• Policy relevance: Regulators and organizations may need interventions (information/expectation management, worker protections, limits on continuous monitoring) to mitigate psychological and social-capital externalities from AI deployment.
• Research directions for AI economics:
  • Quantify macro and firm-level welfare impacts of AI-induced involution (hours worked, burnout, turnover, long-run productivity).
  • Longitudinal and quasi-experimental studies to identify causal effects and dynamic adjustment (do short-term gains erode via social-capital loss?).
  • Cross-country and sectoral comparisons to assess institutional moderators (labor market protections, norms about work intensity).
  • Modeling of endogenous employer expectation-setting and worker effort responses in presence of AI tools (to predict equilibrium overinvestment).

Short takeaway: AI can raise observable baseline performance and managerial expectations, producing psychological pressures that fuel inefficient competitive escalation. Economists and policymakers should account for these behavioral and social-capital externalities when evaluating the net benefits of workplace AI adoption.