A theoretical model suggests AI and automation can raise productivity while decoupling wages from labor’s share of output: wages respond to knowledge accumulation, but labor share is set chiefly by capital intensity, allowing policy to target them separately.

Main Finding

The paper develops a task-based growth model that endogenously links knowledge accumulation with the structure of production and explicitly models frictions (e.g., technological lock‑in, costs of knowledge generation/validation). Numerical simulation and machine‑learning analysis of the model show that wages and labor share of output are not tightly bound: labor share is highly sensitive to the capital–labor ratio (automation intensity), while wages respond positively to larger knowledge stocks. Thus, distinct policy levers can influence wages and labor shares independently.

Key Points

• Model innovation
  • Introduces a task-based production framework that endogenously couples knowledge accumulation (innovation, validation) with task assignments across capital and labor.
  • Incorporates frictions that limit unbounded automation/growth: physical capital frictions, knowledge accumulation costs, technological lock‑in, and validation burdens.
  • Explicitly models different modes of knowledge generation: autonomous and adaptive processes.
• Main behavioral results (from simulations)
  • Labor share depends sensitively on the capital–labor ratio: higher capital intensity (more automation) tends to lower labor share.
  • Wages correlate positively with the stock of knowledge: more knowledge/innovation raises worker wages even when labor share falls.
  • Wages and labor share can therefore be influenced separately by different structural parameters and policy levers.
• Role of simulation and data analysis
  • Uses forward numerical simulation of the analytical model to explore wide parameter spaces and policy regimes.
  • Applies machine‑learning techniques to the simulated data to identify patterns and quantify the effect of structural parameters on outcomes (output, wages, labor share).
• Policy insight
  • Simple interventions (e.g., affecting capital accumulation, taxation, R&D incentives, validation costs, knowledge diffusion) can steer outcomes toward desired wage and distributional objectives because the two outcomes respond to different mechanisms.
• Caveats
  • The results are derived from a stylized, simulated model—empirical calibration and country‑specific institutional details will affect quantitative predictions.
  • Machine‑learning analyses are applied to model outputs (not observational economic data), so findings indicate structural regularities within the model rather than direct empirical validation.

Data & Methods

• Analytical model components
  • Task-based production structure: occupations decomposed into tasks that can be performed by labor or capital (automation).
  • Endogenous knowledge accumulation: ideas/knowledge are non‑rival inputs that affect productivity; costs and frictions to generate and validate knowledge are modeled.
  • Frictions and refinements: physical capital frictions, knowledge accumulation costs, technological lock‑in, autonomous vs. adaptive knowledge generation modules.
• Quantitative strategy
  • Full model is solved and evaluated through forward numerical simulation across wide parameter ranges to explore dynamic responses.
  • Key outcome variables tracked: aggregate output, wages, labor share of output, capital–labor ratio, knowledge stock.
  • Machine‑learning methods are used to analyze simulated output and extract patterns relating structural parameters to outcomes (the paper refers to ML generally for trend identification and parameter effect quantification; specific algorithms are not detailed).
• Experiments and policy design
  • Parameter sweeps and scenario simulations examine how changes in capital accumulation, knowledge production costs, validation burdens, and other structural parameters affect wages, labor shares, and output.
  • Simple policy interventions are simulated to illustrate how targeted levers (e.g., R&D subsidies, capital taxes, measures to lower validation costs) translate into macroeconomic and distributional effects.

Implications for AI Economics

• Decoupling of wages and labor share
  • Policymakers should not assume a single policy will simultaneously secure worker wages and labor’s share of output. Policies that expand the knowledge stock (education, R&D, diffusion) can raise wages even if capital accumulation/automation reduces labor share.
• Targeted policy design
  • To protect labor share: consider interventions that influence the capital–labor ratio and the pace/choice of automation (e.g., capital taxation, regulation on adoption incentives, support for labor‑augmenting technologies).
  • To raise wages: invest in knowledge production and diffusion (public R&D, subsidies, validation infrastructure, credentialing and retraining).
  • To avoid adverse lock‑in: reduce validation burdens and lower barriers to entry for innovators to prevent dominant firms or technologies from locking the economy into suboptimal paths.
• Institutional importance
  • Strong institutions that manage technology diffusion, competition, and the returns to innovation (e.g., antitrust, IP design, standards for validation) are critical to steer the effects of AI and automation on distribution and growth.
• Research and measurement
  • The task‑based approach underscores the value of measuring task content, automation potential, and knowledge flows across industries and occupations to inform policy.
  • Model‑based simulation combined with empirical calibration should be used to tailor policy advice to specific countries and sectors.
• Long-run growth vs. short-run distribution
  • Policies that favor knowledge accumulation may yield long‑run wage gains, but short‑run distributional impacts depend on capital accumulation dynamics; transitional policies (e.g., profit sharing, redistribution, retraining programs) may be needed.

Limitations and next steps - Empirical calibration and validation using micro and macro data (task measures, firm‑level automation adoption, R&D/validation cost data) are needed to translate model insights into quantitative policy guidance. - More explicit specification and disclosure of the ML methods used on simulated outputs would clarify robustness and interpretability of the pattern extraction.