Redistributed

Papers: Anton Korinek & Megan Juelfs. "Preparing for the (Non-Existent?) Future of Work." Brookings, 2023. Anton Korinek. "Generative AI for Economic Research: Use Cases and Implications for Economists." Journal of Economic Literature, 2023. Anton Korinek. "Scenario Planning for an A(G)I Future." IMF Working Paper, 2024. Links: Brookings | NBER | IMF

Why This Paper Matters

Most economic analysis of AI fixates on a single number: GDP growth. Will AI add $7 trillion or $15 trillion to global output? The question is treated as if aggregate growth settles the matter. It doesn't. A tide that lifts all yachts while swamping rowboats is not a rising tide.

Anton Korinek, an economist at the University of Virginia and the Brookings Institution, has spent the past three years asking the question that GDP figures conceal: when AI creates value, who gets it? His body of work — spanning papers for Brookings, the NBER, and the IMF — constitutes one of the most systematic academic treatments of AI's distributional effects. Where other economists model the aggregate, Korinek models the split.

This matters because distribution isn't a secondary concern. It's the primary mechanism through which AI will reshape society. A technology that doubles national income but concentrates all gains among capital owners doesn't make a country richer in any meaningful sense — it makes a country more divided.

The Argument

Three Scenarios, Three Futures

Korinek's central contribution is a scenario framework that gives structure to an otherwise chaotic debate. Rather than arguing about whether AI will be "good" or "bad" for workers, he identifies three distinct paths:

Scenario 1: AI as complement. AI augments human workers, making them more productive. The doctor with diagnostic AI sees more patients, catches more cancers, earns more. The programmer with code assistants ships more software. In this scenario, AI raises the demand for human labor because it makes each hour of work more valuable. Wages rise broadly. This is the optimistic case, and it has historical precedent — spreadsheets made accountants more productive without making them obsolete.

Scenario 2: AI as substitute. AI replaces specific categories of workers in specific tasks. Translation, customer service, basic legal research, first-draft writing — tasks that can be decomposed and automated get automated. This scenario creates clear winners and losers. Workers whose tasks are automated see their wages fall or their jobs disappear. Workers whose tasks are complemented see gains. The economy grows in aggregate, but the gains are unevenly distributed. Korinek argues this is the most likely scenario in the medium term, and the one that demands the most urgent policy attention.

Scenario 3: AI as transformative. AI becomes capable enough to perform most economically valuable tasks. Human labor becomes optional for a wide range of production. This is the scenario that most existing institutions are least prepared for. If machines can do what most people do for a living, the link between labor and income — the foundation of modern market economies — breaks. Korinek doesn't predict this outcome, but he takes it seriously as a possibility that requires planning now, not after it arrives.

The framework's power is its clarity. It replaces "AI will take all our jobs" versus "AI will create new jobs" with a structured way to think about what's actually happening, sector by sector, task by task.

The Capital-Labor Split

Beneath the scenario framework lies a mechanism that operates in all three cases but with different intensity: AI shifts returns from labor to capital.

Here's why. AI is a capital good. It's owned by the companies that build it, the firms that deploy it, and the investors who finance both. When AI performs a task previously done by a worker, the income from that task flows to the AI's owner rather than the worker. This is not a market failure. It's the market working exactly as designed — rewarding the factor of production that's doing the producing.

The result is mechanical. As AI takes on more tasks, the share of national income going to labor falls and the share going to capital rises. Korinek shows this isn't speculative — it's an extension of a trend already visible in the data. The US labor share of income has been declining since the early 2000s. AI accelerates this.

The compounding problem is ownership concentration. Capital ownership is far more unequally distributed than labor income. The top 10% of households own roughly 90% of stocks. When returns shift from labor to capital, they shift from a broad base to a narrow one. Inequality increases not because anyone intended it, but because of who owns what.

The Skill Premium Puzzle

Traditional automation followed a clear pattern: machines replaced routine manual and cognitive tasks, increasing demand for high-skill workers who managed and directed those machines. This drove the skill premium — the wage gap between college-educated and non-college-educated workers — upward for decades.

Large language models break this pattern. LLMs are unusually good at tasks previously considered high-skill: writing, analysis, coding, legal reasoning, medical diagnosis. Early evidence from studies by Brynjolfsson, Li, and Raymond (2023) suggests that AI tools disproportionately boost the productivity of lower-skilled workers while providing less marginal benefit to experts. The customer service agent using AI improves dramatically; the senior consultant, less so.

Korinek flags this as genuinely novel. If AI compresses the skill premium rather than expanding it, the distributional consequences are different from every previous wave of automation. High-skill workers can no longer assume their education insulates them. The political economy shifts too — opposition to AI may emerge from professional classes, not just factory floors. The GPTs Are GPTs paper by Eloundou et al. provides the empirical backbone for this claim, finding that higher-income, higher-education workers face the greatest LLM exposure.

The International Divide

Korinek extends the analysis beyond national borders, and the picture darkens. Countries that develop AI — primarily the US, with China as a competitor — capture the gains from ownership and innovation. Countries that consume AI products pay for access without building domestic capacity.

This replicates a familiar pattern. As Korinek notes, the global economy has seen this dynamic before with oil, pharmaceuticals, and software — a parallel explored in Compute Is the New Oil. But AI may be more consequential because of its generality. A country dependent on imported AI for its legal system, healthcare, education, and government administration is dependent in a way that goes beyond any single industry.

The international redistribution is compounded by data flows. AI systems trained on global data generate value that accrues to the companies (and countries) that control training and deployment. The user in Lagos who provides training data through interaction with an AI service does not share in the value that data creates.

What It Gets Right

Takes distribution seriously. The most important contribution is simply insistence. Korinek treats distributional effects as central rather than as footnotes to aggregate growth projections. This is rarer in economics than it should be.

Scenario planning over prediction. Rather than making confident claims about which future will arrive, Korinek provides a framework for thinking about multiple possibilities and preparing for each. This is methodologically humble and practically useful.

Mechanism clarity. The capital-labor split, the skill premium question, the international dimension — each is explained through clear mechanisms, not hand-waving. You can trace the logic from technological capability to distributional outcome.

Policy range. Korinek surveys a genuine range of policy responses — from modest interventions like wage insurance and education investment to more ambitious proposals like universal basic income, sovereign wealth funds, and profit-sharing mandates. He doesn't prescribe a single answer, which is appropriate given the uncertainty.

Critiques and Limitations

Theoretical abstraction. The framework is elegant but operates at a level of abstraction that can obscure messy realities. Real labor markets are shaped by unions, licensing regimes, immigration policy, cultural expectations, and institutional inertia. The model is clean; the world is not.

Political economy gap. Korinek identifies the policies that could address distributional effects — conditional transfers, robot taxes, sovereign wealth funds — but says relatively little about why these policies don't exist already and what it would take to enact them. The obstacle to better AI policy is not a shortage of good ideas. It's a shortage of political power among those who would benefit from them. As Virginia Eubanks has documented, the populations most affected by technological displacement are also those with the least political voice.

Speed of change. The scenario framework treats the three futures as somewhat distinct. In practice, all three may operate simultaneously and transition rapidly. A sector can move from complement to substitute to transformative in years, not decades. The framework may underestimate how little time policymakers have.

Missing labor agency. The analysis is primarily top-down — economists modeling, policymakers responding. It says less about workers as agents: organizing, bargaining, resisting, adapting. Ursula Huws's work on platform labor and the gig economy suggests that worker response is itself a major determinant of distributional outcomes, not just a consequence of them.

The Redistribution Question

This body of work is, in many ways, the academic backbone of the redistribution argument. Korinek demonstrates what should be obvious but frequently isn't: AI creates value, and the question of who captures that value is not settled by the technology itself. It's settled by ownership structures, tax policy, labor institutions, and political choices.

The default path is concentration. Without active intervention, AI increases the return to capital, capital is concentrated, and inequality widens. This isn't a prediction about malicious intent — it's a prediction about structural incentives. Firms maximize shareholder value. AI helps them do that. Shareholders are disproportionately wealthy. Therefore, AI's gains flow disproportionately to the wealthy. Q.E.D.

But Korinek also shows this is not inevitable. The policy tools exist: progressive taxation of AI-derived income, sovereign wealth funds that give citizens ownership stakes in AI infrastructure, profit-sharing mandates, wage insurance for displaced workers, public investment in AI as shared infrastructure rather than private property. The UBI and AI debate takes on new urgency under Korinek's transformative scenario. AI Cooperatives offer another structural alternative — distributing ownership rather than just redistributing income. The question is political will, not technical feasibility.

This is where Kate Crawford's observation in Atlas of AI becomes relevant: AI is an extractive industry. It extracts labor, data, and natural resources, and it concentrates the resulting value. Korinek provides the economic formalism for what Crawford describes empirically. Together, they make a case that redistribution isn't a side effect to manage — it's the central issue.

How to Use This Paper

For policy analysis: The three-scenario framework is the clearest available structure for thinking about AI's labor impact. Use it to categorize what's happening in specific industries and match policy responses to scenarios.

For arguing against GDP-only thinking: Korinek provides the academic ammunition to push back on "AI will grow the economy" as a sufficient answer. Growth for whom? Distributed how? These are the questions the framework forces.

For understanding the capital-labor dynamic: The mechanism — AI shifts returns from labor to capital, capital is concentrated, inequality rises — is simple enough to explain in a meeting and rigorous enough to cite in a paper.

For exploring policy options: The survey of interventions (wage insurance, UBI, sovereign wealth funds, robot taxes, profit-sharing) provides a menu for policymakers. Each has trade-offs Korinek outlines clearly.

For anticipating what comes next: The transformative scenario — where human labor becomes broadly optional — is treated with appropriate seriousness. Even if you think it's unlikely, having a framework for the possibility is better than not.

Sources

Korinek, Anton, and Megan Juelfs. "Preparing for the (Non-Existent?) Future of Work." Brookings, 2023. Brookings
Korinek, Anton. "Generative AI for Economic Research: Use Cases and Implications for Economists." Journal of Economic Literature, 2023. NBER
Korinek, Anton. "Scenario Planning for an A(G)I Future." IMF Working Paper, 2024. IMF
Brynjolfsson, Erik, Danielle Li, and Lindsey R. Raymond. "Generative AI at Work." NBER Working Paper 31161, 2023. NBER
Crawford, Kate. Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence. Yale University Press, 2021. Yale
Acemoglu, Daron. "Harms of AI." Cambridge Journal of Economics 45, no. 4 (2021): 819–830. Cambridge
Eubanks, Virginia. Automating Inequality: How High-Tech Tools Profile, Police, and Punish the Poor. St. Martin's Press, 2018. Publisher

General

Paper Review: The Distributional Effects of AI