The Paradox of Automation

Imagine an island with 100 people.

The nine businesses produce nearly everything the island consumes: food, clothing, housing services, transportation, entertainment, repairs, financial services and other necessities.

The model is not intended to reproduce every detail of a modern economy. It only needs to contain two things we want to examine: production and purchasing power.

Now one owner discovers a remarkable technology. Artificial intelligence and machines can perform the work of eight of the company's nine employees.

The machines are faster. They do not take vacations or call in sick. They can work around the clock. They require electricity, computing resources, maintenance and replacement—but even after those costs, they are substantially cheaper than the workers.

What should the owner do?

From the perspective of that company, the answer seems obvious.

Costs fall. Margins rise. Prices might fall. The company becomes more competitive.

Now the owner of the second company sees what happened. If she continues employing nine people while her competitor operates with one, she may eventually be unable to compete. So she automates too.

Then the third. Then the fourth.

Eventually all nine businesses reach approximately the same conclusion.

Nobody needed to meet in secret. Nobody needed to be cruel. Nobody intended to damage the island.

Every owner simply made what looked like the rational decision for his or her own company.

And now the island has a problem.

Eighty-One Workers Were Also Eighty-One Customers

Before automation, the 81 private employees earned wages.

Those wages were costs to the businesses.

But they were something else at exactly the same time.

They were purchasing power.

Workers used their wages to buy food, housing, transportation, clothing, entertainment, repairs, services and products made by the other businesses.

This is the strange dual role of labor in a modern economy.

That tension is the beginning of the Paradox of Automation.

The Owners Got Richer. Didn't That Solve the Problem?

Perhaps the nine owners now earn much larger profits.

Wouldn't they simply spend more?

Somewhat.

But there are limits to personal consumption. A person who becomes ten times wealthier does not generally eat ten times as many dinners or buy ten times as many refrigerators.

They may buy a larger home, better vacations, more expensive services and more investments. But their personal consumption does not automatically replace the purchasing power previously distributed among dozens of wage-earning households.

Federal Reserve research finds that consumer spending is less sensitive to wealth held by high-income households, and that wealth gains concentrated among higher-income households have translated into less consumption than the same gains would have if distributed more broadly.¹

So transferring resources from many workers toward a smaller group of capital owners can change not only who has the income and wealth, but also how much of additional resources are spent.

Then the Owners Can Invest the Money

Yes.

And this is an important objection.

The owners might invest their additional profits rather than consume them. They might buy more AI systems, construct factories, build data centers, purchase equipment or fund new companies.

That investment creates economic activity and employment. For a time, it could offset part of the lost consumer spending.

But there is a constraint.

Why build another factory if there are not enough customers for its output?

Why open another restaurant if fewer people can afford to eat there?

Why expand a fleet if demand for transportation is declining?

Capital investment can support demand. But it cannot permanently escape the need for final customers.

Production Is Not the Same Thing as a Market

A factory capable of manufacturing one million refrigerators has enormous productive capacity.

But if only 100,000 people are willing and able to buy refrigerators, the remaining capacity does not create revenue merely because it exists.

Economics requires more than producing something.

Someone must also be able to buy it.

This is the idea behind effective demand: demand backed by enough purchasing power to complete a transaction.

Discussions about artificial intelligence frequently concentrate on the production side:

AI will produce more. AI will increase productivity. AI will make services cheaper. AI will allow companies to do more with fewer workers.

All of those things may be true.

But there is a second side of the transaction.

That question matters enormously in a country such as the United States, where personal consumption expenditures represented 67.9% of current-dollar GDP in 2024.²

This Is Not an Argument Against Productivity

Productivity is one of the primary reasons modern societies are wealthier than their ancestors.

We should want agriculture to produce more food with less effort. We should want factories to produce safer goods more efficiently. We should want artificial intelligence to help discover medicines, reduce waste, improve engineering, accelerate science and eliminate dangerous work.

The Paradox of Automation is not an argument that productivity is bad.

It asks a different question:

The technology can work perfectly.

And the economic system can still develop a problem.

But Technology Has Always Created New Jobs

It has.

That history deserves enormous weight.

Agricultural mechanization eliminated vast amounts of farm labor. Industrialization created factory work. Machines eliminated trades while new industries appeared. Computers destroyed some occupations and created others.

Previous technological transitions repeatedly pushed human labor toward tasks machines could not yet perform. Physical automation increased the relative value of human cognition, helping generations move from farms and factories toward professional, technical and knowledge work.

Artificial intelligence raises a different question:

The relevant threshold is not whether every doctor, engineer, attorney or accountant disappears. A large reduction in the number of humans required across many such occupations could be economically profound even if every profession continues to exist.

The International Monetary Fund estimates that almost 40% of global employment is exposed to AI, rising to about 60% in advanced economies. The IMF also emphasizes that exposure is not the same as replacement: many jobs may become more productive because of AI, while in others AI could reduce labor demand, wages or hiring.³

So the question is not whether there will still be jobs humans can do.

The Paradox Does Not Require 100% Unemployment

The island takes the mechanism to an extreme because extremes make mechanisms easier to see.

Reality is more likely to exist on a continuum.

Perhaps AI eventually reduces the amount of human labor required by 5%. The economy could probably absorb that relatively easily.

Perhaps it reduces labor demand by 10%. Still potentially manageable.

What about 20%?

30%?

40%?

50%?

At some point, ordinary labor-market adjustment may become something qualitatively different.

We do not know where that point lies. It may not even be a single number.

The result will depend on wages, prices, new occupations, government policy, debt, savings, demographics, ownership and many other variables.

But we do not need to know the exact threshold to recognize that order of magnitude matters.

What If AI Makes Everything Cheaper?

This may be one of the strongest objections to the Paradox.

If automation dramatically lowers production costs, prices may fall.

And if prices fall enough, people need less income.

The relevant variable is not nominal wages. It is real purchasing power.

If someone earns $100,000 and it costs roughly $100,000 to maintain a given standard of living, then in a simplified world where both income and every relevant cost fall proportionally to $10,000, the nominal numbers are dramatically smaller but real purchasing power is essentially unchanged.

That is why lower wages do not automatically mean lower living standards.

But now change the magnitudes.

And if income falls to zero because a job disappears and another does not replace it, cheaper prices do not solve the problem.

A service that once cost $2,000 might now cost only $500.

Someone with no income may still be unable to buy it.

And Not Everything Reprices at the Same Speed

Households also enter long-term financial commitments.

Mortgages. Leases. Loans. Insurance obligations. Tuition. Contracts.

Suppose technological productivity eventually pushes the market value of a home down substantially.

That may make housing cheaper for a future buyer.

But a homeowner who borrowed $500,000 yesterday does not automatically owe less because the market value of the house fell.

Debt is nominal. Asset prices can adjust. The two do not necessarily move together.

And the overall price level may not move cleanly in one direction. AI could make some digital and professional services dramatically cheaper while housing, healthcare, insurance, energy, taxes or debt service remain expensive—or even rise.

Government monetary and fiscal responses can add another layer. Policymakers may respond to unemployment or weak demand with lower interest rates, fiscal support or other stimulus, while supply constraints and existing debt push other prices in the opposite direction.

So the transition is unlikely to resemble a neat spreadsheet in which income and every expense decline by the same percentage.

The Real Question Is Distribution

Imagine AI doubles the island's productive capacity.

That sounds wonderful.

And it could be.

The island might become capable of producing twice as much food, clothing, transportation and services with a fraction of the labor.

The physical problem of scarcity has been reduced.

But another question appears:

If ownership of the productive technology is concentrated among nine owners while 81 former workers have little or no income, the island may possess extraordinary productive abundance alongside extreme inequality in purchasing power.

That is not a technological failure.

It is an allocation problem.

The Paradox Is Created by Individually Rational Decisions

This brings us back to the owners.

Each owner saves money by automating. Each becomes more competitive.

If one refuses to automate because of concern for workers, competitors may eventually force the issue.

The system therefore does not require malevolent executives. It may actually punish the executive who refuses to participate.

This resembles the Paradox of Thrift.

Saving money is generally prudent for an individual household. But if every household suddenly cuts spending at the same time, businesses lose revenue, employment can fall and total income may decline.

What is rational individually can produce an undesirable outcome collectively.

That is the Paradox of Automation.

Why Doesn't Competition Stop It?

One might expect businesses to recognize that eliminating customers is self-defeating.

But individual businesses generally cannot solve system-wide coordination problems.

The owner of Business A might understand perfectly well that broad unemployment would be bad for the island.

But refusing to automate does not preserve the incomes of workers at Businesses B through I.

It merely leaves Business A with higher costs.

Each business races toward the benefit.

The collective consequence emerges later.

The same logic can occur at the national level. A country may want a more measured transition but fear losing investment, technology leadership or strategic capability if other nations race ahead. One country can establish an example, but a genuinely systemic AI transition may ultimately require forms of international coordination.

AI Could Make the Race Faster

Many technologies diffuse slowly.

Factories take years to build. Oil must be extracted, transported and stored. Enterprise systems require integration across customers and suppliers.

AI will encounter plenty of friction too: legacy systems, regulation, bad implementations, security concerns, training, integration, human resistance and capital constraints.

Those are reasons not to assume an overnight transformation.

But AI also has characteristics that could accelerate diffusion.

Software can scale quickly. Its cost can fall rapidly. It has attracted extraordinary capital. Executives face pressure to demonstrate AI strategies. Nations increasingly treat AI capability as strategically important.

Unlike many invisible back-office technologies, AI is also culturally and financially exciting.

There Is Another Possible Future

The Paradox is not destiny.

Imagine a different outcome.

AI substantially raises productivity. Workers become more valuable because AI complements rather than replaces them. Workweeks become shorter. Wages remain strong. Prices fall. New industries absorb displaced workers. Ownership of AI-generated wealth becomes broad. Productivity gains flow into retirement accounts, pensions, wages, dividends or new forms of income.

In that world, automation could create extraordinary prosperity without destroying aggregate demand.

That outcome is possible.

And it should probably be our objective.

The question is whether existing institutions automatically produce it.

There is no obvious reason to assume they will.

The Question

The fundamental question is therefore not:

It almost certainly will not.

Nor is it:

It almost certainly will.

The question is one of order of magnitude:

If yes, history may repeat itself. The economy adapts. Productivity rises. People move into new forms of work. Living standards improve.

But if the answer is no—if productive capacity increasingly separates from human labor faster than purchasing power is reconstructed—then we face something genuinely different.

Machines may become capable of producing more than humanity has ever imagined.

Companies may become extraordinarily efficient.

The technology may succeed beyond expectations.

And yet the economy could discover that it optimized away part of the mechanism through which its customers obtained the money to buy what it produces.