Directing AI toward labor-augmenting technologies can raise wages and social welfare when redistribution is limited, but as AI progressively devalues labor the returns to steering fall and policy should shift toward redistribution and improving non-monetary aspects of work.

Main Finding

Korinek & Stiglitz (NBER WP 34994, Mar 2026) develop a formal framework for "steering" technological progress (especially AI) toward innovations that raise labor demand and improve workers' welfare. They show that a planner who internalizes pecuniary externalities from innovation will prefer labor-complementary (or labor-augmenting) technologies and may tax or otherwise discourage labor‑substituting innovation. The welfare gains from steering are larger the less effective redistribution (or safety nets) are; however, as technology progressively devalues labor, the marginal returns to steering eventually fall and optimal policy shifts toward stronger redistribution and, increasingly, non‑labor dimensions of well‑being.

Key Points

• Conceptual taxonomy of innovations:
  • Labor-complementary (augmenting, creating new tasks) → raises labor demand and wages.
  • Labor-substituting (automation, direct replacement) → lowers labor demand and wages.
  • Impact depends on complementarity to labor, which workers are affected (their relative income), and the labor share in producing the relevant goods.
• Main mechanism: innovators do not internalize the pecuniary externality that lowering wages imposes on workers (and on aggregate welfare when redistribution is costly), so laissez-faire tends to overproduce labor‑displacing innovation.
• Optimal steering depends on redistribution capacity:
  • If redistribution is costless or full, production efficiency is optimal (choose the most efficient technology and redistribute).
  • If redistribution is costly/incomplete, the planner distorts technological choice toward technologies that raise earnings of poorer agents.
• Nonlinear response as labor is devalued:
  • Initially, as technological progress reduces labor value, steering becomes more valuable (protect workers).
  • Beyond a critical threshold of labor devaluation, steering effectiveness declines and policy emphasis moves to redistribution and non‑monetary well‑being.
• Applications analyzed:
  • Robot taxation: positive optimal robot taxes when planner weights worker welfare sufficiently; tax rises with concern for workers.
  • Factor-augmenting innovations: with capital and labor gross complements, capital-augmenting tech can raise wages and be preferred by a pro-worker planner.
  • Task automation: welfare-maximizing planner automates fewer tasks than a production-efficiency benchmark when workers’ welfare is heavily weighted.
  • Multiple goods: planner can steer technology to lower prices of goods disproportionately consumed by poorer agents (an additional redistribution channel).
  • Market power: firms may choose technologies that increase monopsony power or erode worker bargaining power; a planner would prefer technologies that preserve worker market power.
  • Non-monetary aspects: firms underweight non-monetary costs (job meaning, safety); steering should include these considerations and shift toward improving non-monetary quality of remaining jobs as labor’s economic role shrinks.
• Policy implications emphasized: encourage labor-augmenting R&D, consider robot/automation taxes or other technology‑specific instruments, reduce the relative tax burden on labor, use public R&D funding and procurement to steer directions, and redesign redistribution/public finance if labor tax base erodes.

Data & Methods

• Methodology: theoretical, analytical model building and comparative‑statics.
  • Core setup: economy with I heterogeneous agents (different factor endowments) and H factors of production; production F(ℓ; A) depends on factor inputs ℓ and technological parameters A (endogenous choice).
  • Planner types: (i) baseline planner who cannot make transfers (so uses technology as a distributional instrument), and (ii) extension with a nonlinear income tax but with costly redistribution.
  • Equilibrium benchmark: laissez‑faire where agents rent factors at factor prices and competitive firm chooses A and factors to maximize profits.
  • Social planner internalizes pecuniary externalities from innovation on wages and chooses A to maximize a weighted sum of utilities subject to resource constraints.
  • Uses Atkinson–Stiglitz style public economics logic to separate production and distribution when transfers are available, and to show when distortionary steering is optimal otherwise.
• Analytical results: derived conditions for when steering is optimal, sign and magnitude of optimal technology taxes, and comparative statics linking redistribution costs, labor devaluation, complementarity parameters, and welfare weights to optimal policy.
• Extensions: multi-good setting (price/consumption distribution effects), analysis of firm incentives under market power/monopsony, inclusion of non-monetary job utility.
• Empirical context: the paper cites and builds on empirical and theoretical literature (e.g., Eloundou et al. 2024; Brynjolfsson; Acemoglu & Restrepo; Dechezleprêtre et al. 2025; Azar et al. 2023) to motivate parameter plausibility and channels; the paper itself contains no new microdata or empirical estimation — its contribution is a formal policy‑oriented theory.

Implications for AI Economics

• Normative guidance for AI direction: policy and public R&D should favor AI that augments human labor and creates complementary tasks (assistive agents, tools that raise worker productivity without substituting workers).
• Technology‑specific taxation and regulation: robot/automation taxes, R&D subsidies targeted to labor‑augmenting lines, and procurement rules can be justified when redistribution is costly and market incentives favor automation.
• Tax policy design matters: current tax biases (heavier taxation of labor, cheap capital) can steer innovation toward labor‑saving technologies; correcting these distortions (e.g., reducing labor tax wedge or adjusting capital taxation) is a lever for steering.
• Distributional tradeoffs change over time: as AI reduces labor returns, the planner’s optimal mix of steering vs redistribution shifts — initial emphasis on steering, later emphasis on redistribution and non‑labor well‑being as labor’s role potentially diminishes.
• Market structure complementarities: antitrust and labor‑market policies (curbing monopsony, supporting bargaining power) can complement steering by reducing firms’ incentives to automate to erode worker power.
• Measurement and implementation needs: to operationalize steering, economists and policymakers need metrics of (i) labor‑complementarity of innovations, (ii) which worker groups consume which goods, and (iii) the costs of redistribution — empirical work to estimate these elasticities is crucial.
• Research agenda: quantify the elasticities (automation response to wages and taxes), evaluate real‑world welfare gains from steering interventions, and study dynamic/general equilibrium effects (education, task creation, capital accumulation) to refine policy prescriptions.

Limitations to keep in mind: the framework is theoretical and qualitative in parts; quantitative policy magnitudes depend on empirical parameters (elasticities of innovation to wages/taxes, redistribution costs, complementarity measures) that remain uncertain. The optimal policy also depends on normative welfare weights (political economy) and on long‑run structural responses (retraining, task creation) that require further study.