When wages are lognormal, tiny skill or AI upgrades pay off exponentially at the top, prompting costly arms races in education and AI deployment; such overinvestment lifts GDP and widens inequality but can leave society worse off as private costs outweigh social gains.

Main Finding

Technological progress interacting with a lognormal wage (or return) distribution creates exponential returns to incremental skill. That exponential structure drives inefficient “arms races” in education for humans and in AI-capital (chatbot) training by firms: agents/firms have strong incentives to invest up to corner/binding limits, amplifying inequality and GDP while potentially reducing expected welfare (for plausible risk preferences). Network/scale-driven monopoly rents further magnify this over‑investment.

Key Points

• Core assumption: individual effective earnings follow e(y) = A · exp(c y) with y ∼ N(µ, σ). This yields lognormal incomes and exponential marginal returns to skill.
• Proposition 1: Expected earnings grow exponentially in mean skill µ, in c^2, and in σ^2; marginal gains to raising µ therefore grow exponentially.
• Over‑investment pressure (Proposition 2): Because marginal returns rise exponentially, private optimal investment in schooling can exceed median income (agents are ex‑ante pushed to spend more than what >50% will recoup ex‑post).
• Technology and welfare (Proposition 3): With CRRA utility parameter ϕ < 1, more tech always increases expected utility. If ϕ > 1, there is an optimal technology plateau c* = µ / ((ϕ − 1) σ^2). For ϕ above this threshold, further c growth can lower expected utility despite higher mean GDP.
• Monopoly/rents and “Bell curve competition” (Proposition 4): Let α ∈ [0,1] be the fraction of pay tied to pure productivity (1 − α is rent capture). If α ∈ (0,1), individual incentives to invest exceed the social optimum — competition for rents induces extra, inefficient schooling.
• Analogy to AI firms: If returns to chatbot/AI capability are similarly lognormal, firms face the same corner incentives and therefore invest up to borrowing/financial limits; winner‑take‑all/platform effects (low α) intensify the arms race.
• Welfare and distributional facts: Tech increases mean incomes and GDP but shifts mass to upper tail (higher σ effect). This can raise inequality and, for sufficiently risk‑averse agents, reduce expected utility.
• Opportunity cost of children: Using schooling/IQ as skill, the model implies very large life‑cycle earnings foregone by taking time off for childbearing, further depressing fertility incentives.

Data & Methods

• Analytical model:
  • Skill-to-income map: e(y) = A e^{c y}, y ∼ N(µ, σ).
  • Mean income E[e] = A exp(c µ + ½ c^2 σ^2); median M = A exp(c µ); mean/median gap ρ = exp(½ c^2 σ^2).
  • Agents choose schooling I that raises µ(I) to maximize E[e(y)] − I (some variants assume µ = ln I for tractability).
  • CRRA preferences used for welfare comparisons: U(C) = C^{1−ϕ}/(1−ϕ).
  • Monopoly rents captured by splitting exponent: α c y + (1 − α) c(y − µ) reflecting rent capture by high tail.
• Propositions derive from first‑order conditions and comparative statics; key corner/convexities arise because exponentials can make payoff convex in skill.
• Calibration:
  • Source: US mean and median household incomes (census Table H-11) for 1975 and 2024. IQ as skill proxy: y ∼ N(100, 15).
  • Values solved from mean and median:
    • 1975: median M = $58,000, mean E = $68,000 → c1975 ≈ 0.0376, A1975 ≈ 1,339.4
    • 2024: median M = $83,000, mean E = $121,000 → c2024 ≈ 0.0579, A2024 ≈ 244.7
  • Interpretation: c rose ~50% between 1975–2024 (steeper returns to skill); A fell (lower intercept), meaning more mass concentrated in upper tail.
  • Welfare comparison: agents with CRRA ϕ ≳ 2.5 would be indifferent (or prefer 1975 distribution) despite higher mean incomes in 2024.
  • Marginal value of an extra IQ point grows massively with skill level and over time (example: annual marginal wage from +1 IQ at IQ=115: ~$4.5k in 1975 → ~$16.6k in 2024; life‑cycle present values are correspondingly large).
• Limitations and assumptions to note:
  • Skill proxied by IQ (normality) and schooling raises mean skill µ via a chosen functional form (µ(I)), sometimes log form used for closed form.
  • Partial equilibrium, static or comparative‑static framework — limited dynamic/longitudinal treatment.
  • No explicit modeling of borrowing constraints in some sections (but corner solutions arise given borrowing limits).
  • Model treats education as productivity‑increasing (not pure signaling), which is an assumption that matters for interpretation.

Implications for AI Economics

• Direct analogy: Training improvements for chatbots/AI models that yield lognormally distributed returns (small fraction of models capture outsized performance/revenue) create exponential marginal returns to incremental model quality. That motivates firms to invest up to financial constraints — i.e., corner solutions.
• Corporate behavior and systemic pressure:
  • Winner‑take‑all effects and platform rents (low α) make firms overinvest in marginal model improvements simply to capture the disproportionate market share that a small performance edge provides.
  • The model rationalizes very large planned AI capex (author cites ~$700B planned 2026 capex by a few large firms). Firms may issue debt or strain balance sheets to chase tail outcomes.
• Market‑structure implications:
  • Strong network/scale effects that allow a single slightly superior model/platform to dominate increase the social cost of the arms race. Antitrust or measures that reduce monopoly rents (increase α) would mitigate overinvestment incentives and raise average productivity.
• Welfare and policy:
  • Aggregate output/GDP may rise while average welfare falls for risk‑averse agents; distributional policy (taxation of supernormal returns, redistribution, safety nets) becomes more important.
  • Public investment to lower private education/AI training costs or provide universal access (e.g., subsidized childcare, education, compute credits for smaller firms) could reduce socially wasteful arms‑race spending.
  • Financial regulation: align incentives around longer‑term value rather than short‑term race for tail capture (e.g., constraints on levered financing for purely marginal model improvements).
• Research & measurement recommendations:
  • Empirically test whether firm‑level returns to model improvements follow lognormal tails; estimate α (share of pay due to rent vs marginal productivity).
  • Measure how marginal improvements in model capability translate into market share/revenue (nonlinearities/winner‑take‑all margins).
  • Study dynamic public‑good aspects of compute/data and whether subsidizing access reduces socially wasteful duplication.
• Broader caution:
  • If the AI training ecosystem reproduces the same exponential tail returns as human skills, social inefficiencies can arise at scale — not only higher inequality but also systemic financial risk if many firms lever up to chase tail outcomes.

Overall, the paper provides a compact theoretical lens: once returns to capability are exponential (lognormal outcome distributions), both individuals and firms face strong, potentially socially inefficient incentives to escalate investment to secure tail outcomes. For AI economics, that implies the arms race in model training is structurally driven and may require policy, market‑design, or antitrust interventions to prevent welfare losses and excessive concentration.