Automation can raise wages even as it shrinks labor’s share: in a competitive constant-returns model, wages are maximized at a labor share determined by the capital-to-labor ratio, and many rich countries—including the U.S.—have labor shares above that level, so further automation likely boosted wages and explained roughly 16% of U.S. real wage growth since 1954.

Main Finding

Autor & Kausik (2026) prove that in a competitive, constant-returns-to-scale economy the labor share that maximizes the wage — and the maximum wage itself — depends only on the capital-to-labor ratio (K/L). Consequently the relationship between labor share and wages is non‑monotonic: if the prevailing labor share exceeds this wage‑maximizing level, further automation (which lowers labor share) can raise average wages. Empirically, using CES-based estimates for 12 advanced economies, they find current labor shares lie above the wage‑maximizing level in all of them, implying further automation should, on average, increase wages. In the U.S., falling labor share explains roughly 15.7% of real wage growth from 1954–2019.

Key Points

• Theoretical result
  • Theorem: For CRS production Y(K,L,λ) (λ = labor share), the wage-maximizing labor share λ and the maximum wage w depend only on k = K/L.
  • Intuition: λ affects wages directly (higher share of output to labor) and indirectly (through how λ affects output). Those effects can oppose one another; when the indirect negative output effect dominates, reducing λ can raise wages.
  • If an interior solution exists and a second‑order condition holds, λ* is unique and in (0,1).
• Empirical findings
  • Model: CES production function (Y = A(αK^ρ + (1−α)L^ρ)^{1/ρ}), log‑expanded around σ≈1 to get a tractable approximation linking y = Y/L to k and λ.
  • Data: 12 advanced economies (US 1954–2019; others mostly 1995–2021) from FRED and EU KLEMS.
  • Key estimates: US −∂ln w / ∂ln λ ≈ 2.40 (SE = 0.040) — i.e., a 1% decline in labor share is associated with ~2.4% higher wages at current US k and λ. Estimated σ̂ ≈ 0.92 for the US, pooled σ̂ ≈ 0.72.
  • Decomposition: Wage growth = contributions from TFP, labor share changes, and capital deepening. For the US, falling labor share contributed ≈15.7% of cumulative real wage growth (1954–2019). For the other countries pooled, rising labor share slightly retarded wage growth over 1995–2021 (≈−6.1% contribution).
• Important caveats
  • The positive wage effect from falling labor share requires a competitive economy where changes in λ reflect technology/capital accumulation. If λ falls because of rent extraction or market power, the countervailing wage gains do not obtain.
  • Aggregate results hide distributional effects: average wages can rise while some workers or sectors lose, and overall inequality can increase because capital income is concentrated.

Data & Methods

• Theory
  • General CRS production family Y(K,L,λ), derive w = λ(Y/L), show ∂ln w/∂λ = 1/λ + ∂ln Y/∂λ and solve first‑order condition for λ* as function of k.
  • Uniqueness and interior maximum under a second‑order condition.
• Empirical specification
  • Use CES production (σ = 1/(1−ρ)), express y = Y/L as function of k and λ, then take logs and Taylor‑expand around σ≈1 to obtain an estimable approximation.
  • Econometric estimation: OLS on the expanded log form; US estimates from 66 years of data (1954–2019); pooled estimates for 11 other countries (largely 1995–2021). Standard errors clustered by country for pooled regressions.
• Data sources and coverage
  • US: FRED series for GDP, TFP, capital stock, total labor hours, labor share (1954–2019).
  • Other advanced economies: EU KLEMS (1995–2021, Japan to 2020), with currency conversions to common units.
• Key empirical numbers
  • US α1 ≈ 0.087 (implying σ̂ ≈ 0.92), pooled α1 ≈ 0.392 (σ̂ ≈ 0.72).
  • US derivative −∂ln w/∂ln λ ≈ 2.40 (SE 0.040).
  • R²: 0.228 (US), 0.153 (pooled).

Implications for AI Economics

• Automation and AI do not mechanically imply falling wages for workers. If AI reduces labor share in an economy where λ currently exceeds the wage‑maximizing level given K/L, aggregate wages can rise even while labor’s share of output falls.
• However, positive average wage effects from AI-driven automation coexist with substantial distributional risk:
  • Capital ownership concentration means gains accrue disproportionately to capital holders; aggregate wage gains need not improve broad welfare.
  • Sectoral and skill heterogeneity: some workers/sectors may face wage declines or job loss even while average wages rise.
• Policy relevance
  • Ownership and redistribution matter. To translate automation-driven productivity gains into broad-based welfare improvements, policies that broaden capital ownership (e.g., employee equity, sovereign wealth funds, public investment returns), progressive taxation, and redistribution can be crucial.
  • Antitrust and competition policy remain central: if falling labor share is driven by rent extraction or monopoly power, automation may harm wages and inequality without the offsetting productivity channel.
  • Labor market policies (retraining, unemployment insurance, wage subsidies) are important to manage transitional costs and sectoral dislocations.
• Directions for research
  • Micro evidence: task‑level and firm‑level measurement of how AI changes task allocations, marginal products, and rent generation versus pure productivity.
  • Ownership and distribution: quantify how capital income gains from automation are distributed across households and the implications for consumption and political economy.
  • Dynamics and transition: study adjustment paths, labor reallocation frictions, and heterogeneous worker outcomes rather than static comparative‑statics alone.
  • Robustness to alternative production technologies and to violations of competitiveness (market power, bargaining dynamics).

Short takeaway: under standard competitive technology assumptions, falling labor share induced by automation can raise average wages if the economy’s labor share is above the wage‑maximizing level given K/L. But whether AI is broadly welfare‑improving depends critically on who owns capital, how gains are distributed, and whether declines in labor share reflect productivity or rent extraction.