Gaps in what firms and workers believe about the economy drive wage and job volatility: optimistic, slow-learning workers command higher hiring wages in downturns but face greater separation risk later. Faster firm learning calms short-run swings but lengthens transitions, implying AI tools that improve forecasting may reduce volatility while prolonging adjustment.

Main Finding

Belief heterogeneity and information frictions about the aggregate state—especially persistent gaps between firm and worker beliefs—are key drivers of wage behavior, unemployment volatility, and the economy’s dynamic response to shocks. Workers who form sluggish, optimistic expectations are hired at higher wages in downturns and subsequently face higher separation risk; firm learning about the aggregate state reduces volatility but increases the persistence of responses to shocks.

Key Points

• Survey evidence: Households form dispersed, backward-looking expectations about macro conditions; more optimistic workers demand higher wages.
• Model: A search-and-matching framework with sticky wages and endogenous separations is used; both workers and firms face information frictions about the aggregate state.
• Belief gap mechanism: The difference between firm and worker beliefs is a central determinant of unemployment volatility and wage outcomes.
• Firm learning: When firms update beliefs over time, aggregate responses to shocks become more persistent but less volatile.
• Worker heterogeneity: Variation in workers’ learning rates explains observed heterogeneity in employment transitions—workers with more sluggish learning remain overly optimistic in recessions, receive higher hiring wages, and are more likely to separate later.
• Equilibrium implications: Endogenous separations and wage stickiness interact with belief heterogeneity to generate realistic labor market dynamics.

Data & Methods

• Survey data: Uses the University of Michigan Survey of Consumers for household expectations and the Survey of Professional Forecasters (SPF) for professional/firm-side expectations.
• Empirical targeting: The model is disciplined/calibrated to match key empirical moments drawn from these surveys and labor-market data (e.g., wage dispersion, hiring/separation flows, unemployment volatility and persistence).
• Model structure: Search-and-matching model with
  • sticky wages (bargaining or wage rigidity channel),
  • endogenous separations (quit/layoff decisions affected by realized shocks and mismatch),
  • information frictions (agents do not observe the aggregate state perfectly and learn over time),
  • heterogeneity in worker learning speeds.
• Solution/estimation approach: Calibration/method-of-moments style matching of model-generated moments to survey and macro labor-market moments (paper disciplines parameters with the Michigan Survey and SPF).

Implications for AI Economics

• Measuring beliefs matters: When evaluating AI-related labor shocks, incorporate agent beliefs (workers’ and firms’)—surveys or high-frequency belief measures can be crucial inputs for forecasting employment and wage dynamics.
• Information technology and AI that improves firm learning about aggregate conditions could reduce short-run volatility of employment but increase persistence of AI-driven transitions (e.g., slower recovery paths after an automation shock).
• AI that reduces worker information frictions (better signals about macro prospects or firm demand) could change bargaining outcomes and separation risk; making worker beliefs less optimistic in recessions could lower wage bills at hiring but also reduce subsequent separation rates.
• Heterogeneous learning and belief updating across worker groups (by skill, age, or sector) imply differentiated impacts of AI adoption: groups with sluggish updating may be paid more at hiring yet be more vulnerable to displacement—targeted retraining and information programs should account for belief heterogeneity.
• Policy and firm responses: Communication, transparency, and targeted information interventions (including AI-driven forecasting tools) can alter learning dynamics and thus macro labor-market outcomes; models of AI labor impacts should include information frictions and belief heterogeneity to capture transitional dynamics accurately.