PHILADELPHIA — The paper is an academic exercise, a mathematical model of an economy where machines do the work and people do the buying, and its conclusion is a paradox that has haunted economists since the first loom replaced the first weaver. Researchers at the Wharton School and Boston University built a model of what happens when firms substitute artificial intelligence for labor at scale, without changing the institutions around it, and the answer is an economy with rising productivity and collapsing demand, producing everything and selling nothing.
The mechanism is simple and brutal. Every firm that replaces workers with AI makes a rational decision: labor costs fall, output per dollar rises, and the firm’s own position improves. But the workers who leave the payroll do not simply disappear; they stop earning, and stopped earnings mean stopped spending. What looks like efficiency at the level of the firm looks like a shrinking customer base at the level of the economy. The model shows the two forces compounding until production outruns the ability to buy what is produced.
The paper’s authors do not claim this outcome is inevitable, and they are careful to distinguish the model from a forecast. The model describes what happens if current incentives continue without change: if firms keep automating, if the gains keep flowing to capital, and if no mechanism replaces the purchasing power that wages once provided. Change any of those conditions, the authors argue, and the trajectory bends. The point of the model is to show how much depends on the institutions around the technology, not the technology itself.
The intellectual lineage runs deep. The same anxiety appeared with the steam engine, the assembly line, and the computer, and each time the economy found new jobs to replace the old ones, new products to absorb the new capacity, and new forms of demand to fill the gap. The authors’ contribution is to ask whether the current wave is different, and their model identifies a specific reason it might be: AI substitutes for cognitive labor, the kind of work that previously absorbed displaced workers. When the machines can do the thinking, the exit from the labor force becomes harder to reverse.
The policy proposal is the part that has drawn attention. The paper floats what it calls a Pigouvian automation tax, named after the economist who argued that prices should reflect social costs. Under the proposal, a firm would pay a levy each time it replaces a task with AI, with the revenue funding the demand that the automation destroyed. The tax is designed not to stop automation but to price it honestly, forcing firms to internalize the cost that displaced workers impose on the economy rather than offloading it to society.
The proposal will be controversial on every side. Technology companies will argue that taxing automation punishes progress and that the economic history of technological displacement shows the market eventually reabsorbs workers. Labor advocates will argue that the tax is too small, too slow, and too easily avoided, and that it does nothing for workers displaced before it takes effect. The authors’ response is that the tax is a starting point for a conversation that the economy has never actually had: what the price of automation should be.
The timing gives the paper a wider audience than academic work usually commands. AI adoption has moved from pilot projects to production, and the public conversation has moved from whether AI will take jobs to when, and which ones, and what happens to the people who lose them. Corporate announcements about AI-driven efficiency have become routine, and each one raises the same question the paper formalizes: if everyone becomes more efficient, who is left to buy the results?
The model’s starkest implication is that the problem is not individual firms but the system they compose. A single company that automates gains a competitive advantage. A whole economy that automates, without new demand, runs into the wall the model describes. The distinction matters because it explains why no single decision-maker can fix the problem: the incentives that drive each firm are rational, and the collective outcome is irrational, which is exactly the situation for which Pigouvian solutions were invented.
Critics will note that the model’s assumptions are assumptions, and that models have a poor record of predicting the path of technology and work. The paper concedes the point in its own pages, describing itself as an exercise in what could happen rather than a prediction of what will. But the questions it raises are already being asked in boardrooms and policy offices, and the answers are being improvised without a framework. The paper’s contribution is to offer one, with a tax attached.
The debate that follows will define the next phase of the AI era. The technology’s boosters promise abundance, its critics warn of collapse, and the paper suggests both are right, in sequence, unless something changes between them. The authors’ answer is institutional: a mechanism that keeps demand attached to production as the machines take over. Whether the answer is adopted, adapted, or ignored, the question it addresses will not go away, because the economy it describes is already being built.
