First-hand information for everyone
This paper looks at how faster, more productive artificial intelligence (AI) would change wages and what policies could spread the benefits. The key twist is that the authors model AI as something that is itself made by workers — engineers, researchers and others. That assumption creates a distinct “transition” when AI becomes good enough to replace some human tasks, and it affects who gains and who loses as AI spreads.
The authors use a simple two-sector model of the economy. One sector makes AI. The other sector makes final goods and services. Workers come in three types depending on their role, not their skill level. “Substitutable” workers do tasks AI can replace (for example, routine coding, standardized data work, content generation, and customer service). “Complementary” workers are needed to build and improve AI (for example, machine learning engineers and system architects). “Final-goods-specific” workers do jobs that AI does not replace and that simply benefit from AI (for example, many doctors, teachers, and managers). The model tracks how labor and wages shift as AI productivity rises.
The model produces a clear pattern. As AI becomes more productive, it competes with substitutable workers. Their wages fall in both absolute terms and relative to other wages, even while total output rises. Complementary workers see wages grow faster than overall GDP because they both build AI and benefit from working with it. Final-goods-specific workers’ wages rise roughly in line with GDP. The economy goes through a transition once AI clears a productivity threshold; the transition ends only if the displaced workers can move into making AI or into jobs AI does not replace. How fast this happens depends on how much AI production itself needs complementary workers.
The paper also compares competitive and monopolistic AI production. If AI firms behave as monopolies, they restrict how much AI is sold, slow down the transition, and reduce total economic gains. That compression also narrows wage gaps compared with a competitive market but leaves less surplus to redistribute. If policymakers can force monopoly prices down to competitive levels or tax monopoly profits fully, the problem becomes similar to the competitive case. If they cannot, redistribution must rely on combinations of profit taxes and taxes on complementary-worker wages.
Because the model predicts overall gains during the transition, the authors show that tax and subsidy schemes can make every worker better off — a Pareto-improvement (a policy that makes at least one person better off without making anyone worse off). One natural policy is to tax the rising wages of complementary workers and use the revenue to subsidize substitutable workers who lose out. If the policy goal is simply to keep substitutable workers’ consumption at pre-AI levels, the required tax on complementary workers rises and then falls as AI gets more productive. If the goal is to split AI gains equally across worker types, the required tax must rise steadily with AI productivity. Under monopoly, early profit taxes alone are often not enough to pay for these transfers; taxes on complementary wages are needed until profits grow larger.
These results come from a deliberately simple model. It assumes AI can perfectly substitute for certain tasks and that AI is produced by labor, among other modeling choices. The authors note extensions and technical caveats in later sections, and they show some results hold beyond a standard functional form. Still, mapping these theoretical findings to the real world will depend on how closely actual AI production and labor markets match the model’s assumptions, and on the political feasibility of the taxes and regulations the paper discusses.