AI productivity gains lift pay for workers who build AI while lowering pay for those whose tasks are replaced, and monopoly provision of AI curbs deployment and alters optimal redistribution and regulation.

Main Finding

A general-equilibrium model in which AI is produced by labor implies an endogenous transition as AI productivity rises: (i) workers who build AI (complements) see wages grow faster than GDP; (ii) workers whose tasks AI substitutes see absolute and relative wage declines during the transition; (iii) workers confined to final-goods tasks see wages track GDP. Because AI production itself requires labor, adoption reallocates labor endogenously, producing characteristic wage dynamics, a sharp vs. gradual transition determined by AI-sector labor intensities, and clear policy levers that can generate Pareto-improvements. Monopoly provision of AI slows adoption and reduces the surplus available for redistribution; optimal redistribution differs under competition vs. monopoly.

Key Points

• Labor classification drives outcomes: three types (S = substitutable, C = complementary to AI production, F = final-goods–specific).
• AI is produced by labor: Ya = Aa (LS_a)^{αS_a} (LC_a)^{αC_a}; final output uses perfect substitution [LS_f + Xa] for the substitutable tasks: Yf = Af [LS_f + Xa]^{αS_f} (LC_f)^{αC_f} (LF_f)^{αF_f}.
• Two threshold productivities A ≤ A*:
  • Aa ≤ A*: no AI production.
  • Aa ∈ (A, A*): transition — AI is produced and gradually substitutes S in final production; labor reallocates across sectors and wages change nontrivially.
  • Aa ≥ A**: transition complete; substitutable tasks fully displaced in final goods.
• Role of complementary intensity αC_a:
  • If αC_a = 0, the transition is bang–bang (immediate): no gradual reallocation and relative wages track GDP.
  • If αC_a > 0, adoption is gradual; higher αC_a slows adoption because AI production competes for complementary labor used in final goods.
• Wage dynamics:
  • w_C (complements) rises faster than GDP during transition.
  • w_S (substitutable) falls in absolute and relative terms during transition; they are reallocated into non-substitutable tasks or AI production.
  • w_F tracks GDP one-for-one.
• Labor-supply invariance: the wage ratio w_C / w_S during the transition is pinned down by AI productivity and sector parameters (not by relative labor supplies), because equilibrium AI pricing enforces it.
• Redistribution:
  • Because aggregate output rises during the transition, redistribution can make all worker types better off.
  • Two natural Pareto-improvement targets:
  • Maintain substitutable workers’ absolute consumption at pre-AI levels ⇒ required tax on complements is hump-shaped in AI productivity (rises at first, then falls as AI productivity makes smaller transfers suffice).
  • Maintain substitutable workers’ share of GDP ⇒ required tax on complements increases with AI productivity (because market wage ratios widen).
• Monopoly AI producer:
  • Charges higher price, restricts deployment, hires fewer C workers, displaces fewer S workers, and reduces aggregate surplus compared to competition.
  • Profit taxation alone is insufficient early in adoption (monopoly profits grow more slowly than the S deficit); must be combined with a tax on complements until profits become large enough.
  • Restoring competition (price cap or full profit taxation) always leaves complements better off after taxes; gap widens as AI advances.

Data & Methods

• Method: Theoretical general-equilibrium model (closed economy), competitive and monopoly cases.
• Functional forms and assumptions:
  • Cobb–Douglas technologies in both sectors; constant returns to scale.
  • Final-good price normalized to 1; AI price pa determined in equilibrium.
  • Perfect substitution between AI and substitutable labor in final goods: LS_f + Xa.
  • Inelastic labor supplies L_S, L_C, L_F.
• Equilibrium objects: prices (pa, wS, wC, wF), outputs (Ya, Yf), AI use Xa, and labor allocations satisfying profit maximization and market clearing.
• Key analytic results derived:
  • Expressions for thresholds A and A* as functions of technology shares and labor endowments.
  • Closed-form wage and price relationships under Cobb–Douglas; characterization of Pareto-improving tax schedules as functions of Aa and sectoral shares.
• Extensions and robustness:
  • Authors argue Cobb–Douglas is for tractability; main qualitative results extend beyond it (discussed in paper).
  • They analyze a monopoly provider and an extension where AI performs substitutable tasks inside AI production (can create multiple equilibria) and show core insights remain.
  • One labor-role cell (substitutable labor that cannot work in AI) is omitted for clarity but is straightforward to include.

Implications for AI Economics

• Theory: Modeling AI as a labor-produced input materially changes general-equilibrium dynamics relative to treating AI as passive capital. Endogenous AI production creates a transition phase with reallocation-driven wage dynamics and supply-invariant wage ratios.
• Policy design:
  • Effective redistribution should target the rising rents of complementary workers and/or monopoly profits, not just taxes on “AI products” or general distortionary taxes.
  • The appropriate tax path depends on the redistribution objective and AI productivity: policies that keep absolute incomes for displaced workers call for a hump-shaped complementary tax, whereas equal-share goals require taxes increasing in AI productivity.
  • Under AI monopolization, relying exclusively on profit taxation is insufficient early on; a mixed approach (complement tax + profit tax or price caps) is needed.
  • Restoring competition (price regulation or full profit taxation) increases the room for redistribution and benefits complements relative to preserving monopoly rents.
• Empirics and measurement:
  • Empirical work should classify labor by role (S, C, F), estimate the sectoral labor intensities (α parameters), and measure the extent to which AI production uses complementary labor—these drive transition speed and welfare trade-offs.
  • Key measurable predictions: (i) excess wage growth for AI-complement roles during adoption; (ii) absolute declines in wages for substitutable roles during the transition; (iii) a predictable relationship between AI productivity growth and the C-to-S wage ratio.
• Research directions:
  • Empirically quantify αC_a and αS_a across industries to forecast transition dynamics and the size/timing of redistribution needs.
  • Study political-economy constraints on implementing the model’s Pareto-improving taxes, and welfare consequences under partial/limited tax instruments.
  • Extend analysis to open economies, heterogeneous agents, capital owners, and dynamic human-capital investment responses.
• Limitations:
  • Model is stylized (Cobb–Douglas, inelastic labor supply, representative final good); real-world frictions (search/matching, retraining costs, heterogeneous skills, general equilibrium of capital/wealth) could alter quantitative implications though not the core qualitative mechanism that AI produced by labor generates reallocation-driven wage dynamics.