Policymakers can steer AI development to protect jobs and wages: a planner who cares about workers' welfare will favor labor-complementary and capital-augmenting innovations and may tax automation. Yet if AI sufficiently devalues labor, steering becomes less effective and policy should instead prioritize redistribution and improving non-monetary aspects of well-being.

Main Finding

Korinek and Stiglitz develop a theoretical framework showing that society can and should "steer" AI and related technological progress toward innovations that raise labor demand and workers' welfare — but the desirability and form of steering depend critically on (a) how well the economy can redistribute (or insure) losers from innovation, (b) the extent to which new technologies complement versus substitute for labor, and (c) the degree to which labor’s economic value is being devalued. When redistribution is costly or incomplete, it is optimal to distort the direction of technological change (e.g., favor augmentation over automation). However, as labor becomes sufficiently devalued, the marginal gains from steering decline and optimal policy shifts toward stronger redistribution and a greater focus on non‑monetary well‑being.

Key Points

• Conceptual core: innovation creates pecuniary externalities — e.g., a labor‑displacing technology lowers wages — which matter for welfare when markets are incomplete or redistribution is costly.
• Planner vs. laissez‑faire:
  • First best (costless redistribution): choose production‑efficient technology and then redistribute — production efficiency is preserved (Diamond–Mirrlees).
  • Second best (costly/incomplete redistribution): optimal policy distorts technological direction to raise incomes of agents with high marginal social utility (i.e., favor technologies that increase demand for factors owned by poorer agents).
• Properties that make an innovation desirable for workers:
  • Technological complementarity with labor (augments worker productivity).
  • Benefits accrue to workers/skill groups with low income or high social marginal utility.
  • Goods whose price reductions disproportionately benefit poorer agents (multi‑good channel).
• Nonlinear response to labor devaluation:
  • At first, as technology devalues labor, steering toward labor‑friendly innovation becomes more valuable.
  • Beyond a critical threshold of labor devaluation, steering yields diminishing returns and policy optimally shifts toward greater redistribution and non‑labor well‑being improvements.
• Applications studied:
  • Robot/automation taxes: positive optimal robot taxes when planner places enough weight on workers and redistribution is imperfect; tax rates increase with concern for workers.
  • Factor‑augmenting technology: if capital and labor are gross complements (empirically plausible), capital‑augmenting innovations can raise wages and be favored by a planner seeking to help workers.
  • Task automation: welfare‑maximizing planner will automate fewer tasks than a production‑efficiency benchmark when worker welfare is heavily weighted.
  • Market power: profit‑maximizing firms tend to adopt technologies that increase workers’ replaceability (reducing workers’ market power); a planner would steer technology to preserve worker market power.
  • Non‑monetary aspects: firms under‑account for non‑monetary harms/benefits (e.g., meaning of work, safety); steering should incorporate these considerations, and as labor’s economic role declines steering should focus more on direct well‑being improvements.
• Policy levers and complementarities:
  • Direct R&D funding, procurement, and incentives can steer direction.
  • Tax policy matters: current relative tax treatment of labor vs. capital biases innovation toward labor‑saving technologies; reducing labor taxation can incentivize labor‑augmenting innovation.
  • Unions, work councils, and entrepreneurs can play roles in influencing direction.

Data & Methods

• Approach: analytical, theoretical model building.
  • Builds on Atkinson–Stiglitz public economics framework adapted to endogenous technological parameters A in a production function F(ℓ; A).
  • Agents differ by factor endowments; planner chooses technology parameters (and, in extensions, a nonlinear income tax) to maximize weighted social welfare.
  • Baseline assumes competitive representative firm (constant returns) and no transfers (planner steers technology as second‑best). Extensions introduce costly redistribution, multiple goods, and market power (monopoly/monopsony).
  • Comparative statics characterize how optimal steering depends on redistribution costs, welfare weights, factor complementarity/substitutability, and devaluation of labor.
• Empirical grounding: the paper is primarily theoretical but cites empirical literature to motivate assumptions and parameter plausibility (e.g., job exposure measures for generative AI, elasticity estimates of automation to low‑skill wages, studies on AI as complement vs substitute). No new micro or macro dataset is introduced in the paper itself.
• Key modeling assumptions: constant returns to scale in factors, welfare weights normalized, planner may be constrained from lump‑sum transfers in baseline, competition vs. market power analyzed in extensions.

Implications for AI Economics

• For policy design:
  • Where redistribution is limited or costly, active steering of AI direction (via R&D subsidies, procurement, sectoral incentives, or technology‑specific taxes/subsidies) is welfare‑improving and can mitigate labor displacement harms.
  • Robot/automation taxes or targeted subsidies to augmentation technologies are justified when innovation causes pernicious distributional effects and redistribution instruments are imperfect.
  • Tax reform (reduce labor tax burden and/or adjust the capital tax bias) is a powerful indirect lever to alter private incentives for R&D direction.
  • Antitrust/monopsony policy matters: curbing employer market power reduces incentives for firms to choose labor‑eroding technologies.
• For research and evaluation:
  • Empirical priorities: measure complementarity of modern AI tools to different worker types; estimate how R&D direction responds to factor returns (elasticity of innovation to wages); quantify non‑monetary welfare impacts of job displacement and augmentation.
  • Policy evaluation should include distributional channels and pecuniary externalities, not only aggregate productivity gains.
• For firms, funders, and unions:
  • Firms and entrepreneurs seeking positive social impact should prioritize augmentation technologies that expand labor demand or improve goods consumed by poorer groups.
  • Unions and worker representatives should engage in shaping technology procurement and R&D priorities as part of industrial strategy.
• Longer‑run and normative considerations:
  • If AI progresses to the point of large‑scale labor devaluation, attention should shift from steering toward redesigning public finance and robust redistribution (universal income mechanisms, wealth taxes, public provision of goods) and toward technologies that enhance direct well‑being.
  • The tradeoff between maximizing aggregate efficiency and protecting worker welfare is central; policy choice depends on social weights and feasibility/costs of redistribution.

Limitations / caveats - The paper is a theoretical contribution; implementing its prescriptions requires empirical calibration (how complementary is a given AI innovation to labor, how costly is redistribution, where is the devaluation threshold). - Normative results depend on chosen welfare weights and modeling choices (e.g., market structure, frictions).