Capital That Clocks In: Why Economists Say AI Won’t Stay in the Capital Box

In the 1950s the economist Nicholas Kaldor listed a handful of regularities that growing economies seemed to obey. One of them became the quiet foundation of modern macroeconomics: the share of national income that goes to workers stays roughly constant over time. Machines get better, output rises, and labor keeps its slice.

That regularity is not a law of nature. It follows from an assumption about how capital and labor relate. If they are complements — if a machine needs a person to run it — then piling up machines makes people scarcer relative to the machines and therefore more valuable. The worker’s slice holds because the worker is the bottleneck.

The labor share stayed steady because the worker was the bottleneck. Remove the bottleneck and there is nothing holding the slice in place.

Everything downstream inherits the assumption. Payroll taxes assume income flows mostly through wages. Growth accounting assumes you can separate what the machines contributed from what the people did. Unemployment insurance assumes displacement is temporary because the displaced worker will be needed somewhere else. Pull out the complementarity assumption and all of it wobbles at once.

Anton Korinek, at the University of Virginia, has spent most of a decade formalizing that scene. His first serious paper on it, written with his doctoral adviser Joseph Stiglitz and published in the 2019 NBER volume The Economics of Artificial Intelligence, gave it a name: worker-replacing technological progress. The paper identifies what the authors call a singularity point — the moment it becomes cost-effective for machines to fully replace human labor rather than assist it.

The sharper version came in 2024. In Scenarios for the Transition to AGI, Korinek and Donghyun Suh model work as a range of tasks of rising complexity and let automation climb the range. The result, in their words, is that the effect on wages depends on “a race between automation and capital accumulation.” If automation is slow, there is always more for humans to do and wages can rise forever. If it is fast, and “full automation is reached, then wages collapse.” Worse, “declines may occur even before if large-scale automation outpaces capital accumulation and makes labor too abundant.”

That last clause is the synthetic-labor argument in one line. A later survey by Korinek and Philip Trammell spells out the mechanism: when automation outpaces capital accumulation, it “renders the two factors perfect substitutes on the margin.” Perfect substitutes are, for pricing purposes, the same thing. If an AI system can do the marginal task, the wage for that task cannot sit above what it costs to run the system. Human pay stops being set by what humans are worth and starts being set by the price of compute.

The Trammell–Korinek review, published this year in the Annual Review of Economics, is blunt about what that does to the textbook. Fully automating production, so that machines can build more machines, “would dramatically raise the growth rate and lower the labor share, breaking the Kaldor Facts.” Not bending them. Breaking them. And wages, which are the product of exploding output and a collapsing labor share, could go either way depending on things no one can yet measure well.

Stiglitz, who won the Nobel in 2001 for work on how markets fail when information is uneven, comes at the same machinery from the distribution side. His question is less “does the model break” than “who gets paid when it does.”

In a 2021 paper with Korinek on AI and developing economies, the two describe the gains from automation as largely a quasi-rent. When automation lowers the demand for labor, the return to capital rises correspondingly, and as they put it, “It wasn’t that capital owners did anything to receive these additional returns.” The money moves because the boxes moved, not because anyone earned it. That is a very Stiglitz observation, and it is the moral core of the synthetic-labor argument: if AI does labor’s work but pays capital’s return, the income that used to flow to people flows to whoever owns the systems.

Their most recent joint paper, Steering Technological Progress (NBER, March 2026), turns that into policy. It asks how to steer innovation toward jobs that pay better, and it identifies three things that make an innovation good for workers: its “technological complementarity to labor, the relative income of the affected workers, and the factor share of labor in producing the goods involved.” Then it concedes the limit. As technology devalues labor, the benefits of steering first grow, but beyond a critical threshold they decline, and “optimal policy shifts toward greater redistribution.” Put plainly: there is a point past which you cannot save wages by nudging research, and you have to share the output instead.

In public, Stiglitz has been more immediate. Speaking to Fortune in March, he tied the near-term risk to the investment boom rather than to the long-run theory: “Our economy is right now being supported by AI investment—the AI bubble,” he said, and “If I’m right, and there is this bubble, then the breaking of any bubble is really bad in the short term for the macroeconomy.” The long-run synthetic-labor problem and the short-run bubble problem land on the same people.

Here the headline version goes wrong. Daron Acemoglu, who shared the 2024 Nobel with Simon Johnson and James Robinson, is regularly named alongside Stiglitz and Korinek in this debate. But he is not arguing that AI breaks economic models. He is arguing close to the opposite.

Acemoglu’s framework, built with Pascual Restrepo over the last decade, already treats production as a long list of tasks, each done either by people or by machines. In that framework, AI doing a human task is not a paradox. It is just automation, the same thing a mechanical loom did. And his 2024 paper The Simple Macroeconomics of AI uses exactly that framework to put a number on it.

The number is small. On existing evidence, AI raises total factor productivity by “no more than” 0.66% over ten years, and Acemoglu argues even that is generous, because the early evidence comes from easy tasks; his adjusted estimate is “less than 0.53%.” GDP rises about 0.93% to 1.16% over the decade. That is real money, but it is not the end of the Kaldor Facts. It is a modest technology working through a model that handles it fine.

What Acemoglu does share with the other two is the distributional warning. The same paper finds no evidence that AI will reduce labor income inequality and says AI “is predicted to widen the gap between capital and labor income.” And in a September conversation on the Justified Posteriors podcast he gave the cleanest statement of the mechanism anyone has offered: “Whenever you do automation in a sector, that reduces the labor share in that sector. And whenever you introduce new tasks, that increases the labor share in that sector.”

That sentence is his entire position. AI is not synthetic labor that breaks the model. It is capital that can be pointed either at replacing what people already do or at creating things people could not do before. The labor share falls or rises depending on the ratio. In his interview with Fortune in June he called the second path, AI that creates “new tasks,” “the real pathway to true human complementarity, not just enabling you to do what you were doing before in a better way, or in a faster way.”

Where all three converge most concretely is tax, and it is the place where the two-box assumption does the most practical damage.

In 2020, Acemoglu, Andrea Manera and Restrepo concluded that “the US tax system is biased against labor and in favor of capital,” and that the bias encourages more automation than is economically optimal. A firm that replaces a worker with a system swaps an input taxed through payroll and income tax for one that is depreciated and written off. The tax code is subsidizing the substitution the other two economists are worried about.

Korinek and Lee Lockwood take the problem further in Public Finance in the Age of AI (NBER, February 2026). Transformative AI, they write, “may gradually erode the two main tax bases that underpin modern tax systems: labor income and human consumption.” Their paper shows formally that the most revenue a government can raise by taxing labor, as a share of output, shrinks toward zero as capital’s share of income approaches one. If AI really is labor on the capital side of the ledger, the state is taxing the side that is emptying.

This is the point where “AI breaks the models” stops being an academic phrase. Social Security, Medicare and unemployment insurance are funded out of the labor box. They do not care whether economists agree on the right growth model. They care whether wages keep flowing.

Having read the papers rather than the headlines, here is where I come down.

The models are not breaking because AI is secretly labor. They are breaking because the line between capital and labor was always a proxy. It stood for something simpler: the things only people can do. For two hundred years that set was large and stable enough that you could treat it as a fixed factor and build on it. AI moves the boundary of that set, and it moves it at the speed of software rather than the speed of steam.

Korinek is right that if the boundary keeps moving fast enough, the two-factor model has nothing left to say about wages. Acemoglu is right that we are not there, that the measured gains are modest, and that the direction of the technology is a choice rather than a fate. Stiglitz is right that in either case the income is moving to whoever owns the systems, and that our institutions were built for a world where it did not.

For anyone running a business, and that is most of the people I talk to in Charleston, the practical version is simpler. When you buy AI to do a job, you are buying labor on a capital budget. That is why it looks so cheap and why it feels so strange. The same thing that makes it a bargain for the firm is what makes it a problem for the payroll tax, and eventually for the town.

Already true

Labor’s share of US business output is at its lowest point since 1947. Three of the most-cited economists on AI, from very different starting points, all predict AI will widen the gap between capital and labor income. The US tax code, by Acemoglu’s own analysis, favors the substitution.

What has to happen

For the strong version of the synthetic-labor thesis to hold, AI has to automate hard, context-heavy tasks at scale, not only the easy ones, and it has to do so faster than new human tasks are invented. That is Korinek’s race, and it has not been run yet.

Where I am probably wrong

If Acemoglu’s numbers hold, the next decade looks like ordinary automation with a sharper distributional edge, not a regime change, and “AI breaks economics” will read as a phrase from a bubble. The record-low labor share also has older causes — globalization, market concentration, cheaper equipment — and it is entirely possible that AI’s contribution to it stays small.

The work Charleston AI does is the small, local version of this: finished work delivered by systems, for owners who need the work done rather than another tool to manage. If you want to see what that looks like in practice, it is at ai-chs.com.