Mechanization turned labor into a supervisory input, decoupling wages from measured productivity; the pay–productivity gap is therefore historical and likely to persist as automation (including AI) reduces labor demand without raising the supervisory role's productivity.

Main Finding

Beaudreau argues that the “pay–productivity gap” is a long-standing structural feature of industrial economies that arises from a mis-specification of production processes in standard economics. With the advent of mechanized power (steam) and, more recently, advanced automation and control (including AI), labor has increasingly become a supervisory/information input divorced from the physical (energy-based) drivers of output. Measured output per worker can therefore rise without a corresponding rise in wages. A better consilient production model frames output as W = η[K, L] E, where E (energy/available work) and η (second‑law efficiency / organization, a function of capital and supervisory labor) are the fundamental inputs.

Key Points

• Empirical pattern: Since roughly the mid-1970s (and historically in earlier industrial revolutions), productivity has risen faster than wages in the U.S. and other advanced economies (EPI 2026; ILO 2013). The paper documents similar episodes in the 19th and 20th centuries.
• Historical precedent: Nineteenth- and early-twentieth-century writers (Robert Owen, Sismondi, Malthus, Veblen, Henry Ford) observed and debated the same phenomenon: technological increases in productive power did not automatically translate into higher wage income or aggregate demand.
• Conceptual diagnosis: The conventional neoclassical production-function framing treats labor as a productive force in physical terms; Beaudreau contends this is erroneous for modern material processes. When machines supply the force/energy, labor’s role becomes supervisory/informational rather than the primary source of physical work.
• EO (Energy–Organization) model: Output W is produced by the available work/energy E multiplied by an organization/efficiency term η that depends on capital (tools) K and supervisory labor L:
  • W = η[K, L] E
  • E corresponds to available work/negentropy (not simply “energy consumed”).
  • η captures second-law efficiency and is increased by better tools and better information/control (quality of supervision).
• Mechanism for the gap: Because output is driven largely by E, increases in machine energy or machine efficiency can raise output per unit of labor even if L (supervisory labor) is not the causal factor—so wages tied to labor bargaining need not rise.
• Energy- and information-deepening: Productivity growth can come from applying more energy per unit time (energy deepening) or from better process/control information (information deepening). Both can raise measured labor productivity without increasing labor’s bargaining leverage.
• Policy history: Responses to past gaps (e.g., Hoover’s associative state, New Deal wage policy, Ford’s wage experiment) reflect recognition of the distributional problem, but the underlying physical/technical understanding was often missing.

Data & Methods

• Empirical evidence:
  • Time-series comparisons of labor productivity and hourly compensation (U.S., 1948–2025; EPI 2026) showing divergence since the mid-1970s.
  • Cross-country/G20 evidence (ILO 2013) of the pay–productivity divergence.
  • Historical wage vs productivity data for U.S. manufacturing in the 1920s (U.S. Department of Commerce, Historical Statistics, 1975).
• Historical textual analysis:
  • Extensive citations and quotations from classical and early-industrial economists and social commentators (Owen 1821; Sismondi 1827; Malthus 1820; Veblen; Ford).
  • Uses these sources to show the persistence of the problem and the recurring diagnoses/policy responses.
• Theoretical approach:
  • Development of the Energy–Organization (EO) framework drawing on mechanics and thermodynamics (second-law efficiency) and engineering concepts.
  • Treats capital (K) and labor (L) as inputs that raise η, while E is the direct physical source of output.
  • Conceptual distinctions: labor as supervisory/informational input vs. labor as a source of physical energy/force.
• Methods are primarily conceptual and historical, combining qualitative analysis, representative data series plotting, and a formal re-specification (W = ηE) to reframe distribution and productivity questions. The paper does not present micro-econometric causal identification of AI effects, but it provides a theoretical lens and historical evidence.

Implications for AI Economics

• Framing AI as a supervisory/control substitute: AI and advanced automation extend the historical shift from labor-as-force to labor-as-supervision by automating information-processing and control tasks. This deepens the decoupling between physical output drivers (machines/energy) and human supervisory labor.
• Measured productivity vs. labor bargaining: As AI increases η and/or allows greater E utilization (faster, continuous machine operation), measured output per worker can rise even while the marginal role of supervisory labor declines — weakening the ordinary wage-bidding channel that would raise wages.
• Rethinking production functions in models of AI impact:
  • Macro and micro models should explicitly include E (available work) and η (organization / control quality), and treat labor as an informational input that can be partly or wholly automated.
  • Task-based models should distinguish (a) energy/force tasks (machine-driven) and (b) supervision/coordination/information tasks (amenable to AI substitution).
• Distributional and policy implications:
  • Increases in machine-driven rents (energy rents, efficiency rents) are likely to accrue to owners of capital or energy access unless institutions (bargaining, labor market structures, taxation) reallocate gains.
  • Historical responses (wage policy, stronger collective bargaining, institutional arrangements to expand demand) remain relevant; modern policy may need to focus on capturing part of AI/automation-generated rents (corporate tax changes, robot/automation taxes, energy-rent taxation, UBI, stronger labor institutions).
  • Measurement and market considerations matter: if markets for exchanging the new abundance are missing (Owen/Malthus theme), aggregate demand shortfalls can persist.
• Empirical research agenda for AI economics:
  • Decompose productivity growth into E-driven vs η-driven components and quantify how much is due to automation/AI substituting supervision.
  • Task-level analyses using occupational and task databases (e.g., O*NET) to estimate substitutability between human supervision and AI control systems.
  • Firm-level studies of wage outcomes where AI reduces supervision tasks, to identify bargaining transmission mechanisms and wage dynamics.
  • Develop metrics for “organizational efficiency”/η (control quality, process automation intensity) and measure their correlation with wages, profits, and employment.
• Modeling and normative work:
  • Incorporate energy/available work and information/organization explicitly in growth and distribution models when projecting AI impacts.
  • Study institutional reforms (bargaining, property rights over AI/energy rents) that can alter the distributional consequences of productivity gains driven by AI.

In short, Beaudreau’s EO perspective suggests AI will amplify a long-standing mechanical cause of the pay–productivity gap by further transforming labor into a supervisory/informational role. Correctly modeling and measuring this transformation is essential for predicting distributional outcomes and designing policies to ensure productivity gains translate more broadly into wage and welfare improvements.