Introduction: Reframe the Question#
“Will AI cause unemployment?” is a question guaranteed to produce garbage answers, because it comes with only two canned scripts. Optimists recite history: from the power loom to the ATM, every technological panic ultimately created more jobs than it destroyed, and the Luddites have always been wrong. Pessimists declare an exception: this time is different because AI is replacing intelligence itself. Neither side is offering analysis. Both are making professions of faith—the former treats two centuries of induction as a law of nature; the latter treats a slogan as an argument.
There is only one way to do better than faith: first determine what AI is in economic terms, then trace how the shock propagates through the labor market, identify where the macroeconomic loop might break, and finally examine how different institutional structures might absorb it. Prophecy is for prophets. Analysts only get to place bets. This article follows those four steps, then closes with my wager and the indicators that would prove it wrong.
1. Diagnosis: Thinking Is Getting Cheap#
What makes AI distinctive in economic history is not how smart it is, but what it replaces. The steam engine replaced muscle. Electricity replaced energy tied to a particular location. The assembly line replaced one slice of craftsmanship. Every previous wave of automation consumed a subset of human capability. AI is the first technology to put a price on general cognitive ability itself: reading, synthesis, drafting, translation, coding, search, and preliminary analysis. Capabilities once available only by paying a monthly salary are now priced by the token. The unit price of comparable capability has fallen by one or two orders of magnitude in two or three years, and it is still falling.
This is a price revolution. The best analogy is not a “new tool,” but an input whose price is approaching zero: what printing did to the cost of copying, electricity to the cost of energy, and AI to the cost of average cognition. When the price of an input collapses, two things happen.
The first supports the optimists: Jevons paradox. The cheaper an input becomes, the more of it gets consumed. Improvements in coal efficiency caused total coal consumption to soar. Cognition will work the same way. When the marginal cost of a legal opinion, a code review, or a research report approaches the cost of the electricity used to produce it, humanity will consume a hundred times more cognition than it does today. That much is certain, and it is the real foundation beneath predictions of a productivity explosion.
The second supports the pessimists, and optimists often gloss over it: exploding demand for an input does not imply exploding demand for its former suppliers. Jevons paradox saved coal mines. It did not save horses. After engines became widespread, total consumption of transportation services grew exponentially, while horses—the old suppliers of transportation—peaked in population in 1915 and then collapsed toward irrelevance. The horse’s problem was not a lack of comparative advantage. Under Ricardo’s arithmetic, the horse would always be “comparatively least bad” at something. The problem was that its market-clearing wage fell below the cost of feed.
So the real question was never whether AI would “replace people.” It is this: after the price of average cognition collapses, can people who make a living selling cognition still command a market-clearing wage above a decent standard of living? There is no a priori answer. It depends on the transmission mechanism and the institutional response. That is what the next two sections examine.
2. Transmission: Three Layers of Labor and the Scarcity of Accountability#
Jobs are not atoms. They are bundles of tasks. AI does not consume jobs; it consumes tasks. The fate of a job depends on what happens after its task bundle is split apart: does the remaining human component become more valuable, or does it lose its reason to exist? By degree of exposure, human labor can be divided into three layers.
Layer one: routine cognition. Gathering material, organizing it, formatting it, writing first drafts, and doing basic analysis—the entire foundation of academia, media, law, consulting, and administration. This layer is being consumed now. That is not a prediction; it is present tense. Stanford research on payroll data has found double-digit declines in the relative employment of young workers in occupations most exposed to AI. Hiring for entry-level roles in technology, law, and consulting has contracted systematically since 2024. What matters is the reversal in direction: every earlier wave of automation hit blue-collar workers first. This one hits the credentialed class first, starting with a kick at the bottom rung of the ladder.
That conceals a badly underestimated second-order disaster: the apprenticeship crisis. Senior experts are not born. They are produced by spending ten years doing junior work. If all junior work goes to AI, where will senior judgment come from? Companies are eating their own seed corn. The shortage of partners fifteen years from now is already brewing in today’s entry-level hiring freezes.
Layer two: embodied presence. Nursing, repairs, skilled trades, and face-to-face services. Moravec’s paradox still holds: what is hard for AI is easy for humans, and what is easy for humans is hard for AI. A plumber has a much deeper moat than a junior lawyer. But the length of this layer’s protection depends on how fast the robot cost curve burns down, and the fuse is already lit. Embodiment is a buffer, not a fortress.
Layer three: judgment and accountability. A common argument says experts are safe because “AI cannot ask the real questions.” The conclusion is broadly right, but the argument is not sturdy enough—and if the argument is wrong, its shelf life may be only three years. Of course AI can generate questions. It can generate an infinite number of excellent-looking questions, diagnoses, and strategies. What it cannot do is put its name on one. Society needs more than answers. It needs an entity that can be sued, lose a license, forfeit a reputation, or go to prison when an answer is wrong. Accountability presupposes individuality: a continuous “who” with interests to lose and a self that can be punished. AI in its current form can be copied and rolled back and has no persistent state. It is structurally devoid of self, and therefore cannot serve as the endpoint of any chain of accountability.
The top layer’s moat is therefore institutional, not cognitive. Licenses, signing authority, audit liability, and medical malpractice law are all machines for attaching responsibility to a natural person. That protection is far more durable than “AI hallucinates.” Hallucination rates fall every year; society’s need for a scapegoat is eternal. The last human job is taking the blame. But honesty requires one caveat: an institutional moat is guild politics by another name, and cost pressure will erode it inch by inch. Watch for the first industry to trade liability exemptions for efficiency. That is where the levee will spring its first leak.
Finally, consider the much-heralded supervisory layer in the middle. The most common structure today is to eliminate dozens of junior roles and add two or three reviewers. That fits the current evidence, but it is a mistake to treat it as a steady state. Verification is easier to automate than generation because it has a standard answer; the cost of AI reviewing AI is falling much faster than the cost of human review. Supervisory jobs are a tollbooth built on a river that is changing course. They may collect for a few years, but do not value them like a bridge.
3. Breakdown: Who Will Buy the Output of Cognition?#
Everything above is still only the labor market’s microeconomic picture. The real danger lies in the macroeconomic loop: wages are not only a cost; they are also demand. On the production side, cognitive output is about to explode. On the demand side, if labor’s share of income keeps falling, where will mass purchasing power come from? This is the reproduction problem Marx worked through in Volume II of Capital. It is Keynesian underconsumption. The labels do not matter. What matters is that the AI age’s most orthodox Marxist crisis may play out first in its most capitalist economy.
Look no further than the United States today. The top 10 percent of households already account for nearly half of consumer spending. A substantial share of GDP growth is being sustained by AI capital expenditure. In other words, the economy is maintaining growth by “building cognition factories,” even though those factories produce the very thing that will compress the labor income meant to buy their output. Building demand-destroying capacity with borrowed demand is called a bubble or a crisis, depending on when the loop breaks.
Another divide is widening in the price structure: deflation in bits, inflation in atoms. Everything AI can produce—text, code, images, and consulting work—falls toward zero in price. Everything that must be bundled with human presence—housing, health care, education, caregiving—keeps becoming relatively more expensive under Baumol’s cost disease. Ordinary people will have a strange experience: intelligence becomes free while life becomes more expensive. That scissor alone is political dynamite.
There are only three ways to repair the loop. First, a redistribution loop: tax compute, capital, and AI rents to fund transfers, universal basic income, or public services. Second, an ownership loop: broaden ownership of AI capital through sovereign-wealth-fund dividends or universal shareholding—a scaled-up Alaska model. Third, a new-scarcity loop: move the basis of employment toward things machines cannot provide—attention, presence, status, and care. The third is often presented as a complete answer, but it contains a circular dependency. Service industries can absorb labor only if the public has money to buy services, and the public having money is precisely the result of the first two loops. Technology determines the size of the cake; distribution determines whether anyone can afford a slice. Services are what grow after the distribution problem has been solved, not a substitute for solving it.
History offers only a grim control group. The last general-purpose technology revolution took decades to close the gap between productive capacity and purchasing power. A century, several depressions, and countless strikes separated the steam engine from the eight-hour workday. The gains from electrification became mass prosperity only after the Great Depression and the New Deal rewrote the rules of distribution. Institutional lag is the danger period. We are entering it now.
4. Mirror Image: What Ails China and the United States#
One common script says that the United States will make a smooth transition while China inevitably falls into crisis. I think the correct picture is a mirror image, not a one-sided story: the two powers are entering the same equation from opposite ends.
China’s disease is on the demand side, but the wound is often mislocated. “Migrant workers competing with robots” was the script ten years ago. The real problem in Chinese manufacturing today is that it cannot recruit young people. Aging is offsetting automation, while China itself accounts for half of all industrial robot installations worldwide. The people truly under the guillotine are college graduates. Two decades of university expansion batch-processed farmers’ children into white-collar workers who gather, organize, and format information—exactly the layer AI cuts first. Kong Yiji’s scholar’s gown—the symbol of an education that confers status but no livelihood—has come due. AI is collecting.
The cruelest symmetry is that the sectors best able to absorb these graduates—eldercare, health care, nursing, and education—face enormous real demand created by aging, yet have been starved for years by a production-first fiscal system. Money flows to chips, infrastructure, and industrial capacity, not to hospitals, pensions, and transfers. On one side are surplus workers. On the other is unmet demand. Between them stand the fiscal system and the household-registration and social-insurance regimes. The problem is not that China “doesn’t understand consumer economics.” Beijing’s economists understand it as well as anyone. The problem is that distribution is a redistribution of power. Moving a large share of national income from the government and corporate sectors to households changes not economic theory, but the structure of power. China’s hidden card lies in the same place: if embodied intelligence is the next wave, the manufacturing ecosystem of the world’s factory is home turf. Administratively, an authoritarian fiscal state could pivot toward welfare overnight. Politically, the shift may be immovable.
America’s disease is on the distribution side, and it is catching China’s disease. The consumer engine is still turning, but it increasingly runs on a single cylinder: households at the top. Meanwhile, hundreds of billions of dollars in AI capital expenditure are production-first economics in its purest form. The United States is mobilizing on a national scale to expand cognitive capacity while treating the demand side as an afterthought. The picture is laughably familiar. America also holds a card: the global rents from frontier models. Whoever owns the best models can charge seigniorage on the world’s cognition. That flow of national income is large enough to cushion the entire transition—provided it is distributed. And that is the deadlock. This political system has not produced a redistributive project on the scale of Roosevelt’s in half a century, and it shows no sign of producing one now.
The complete mirror image is therefore this: China’s crisis enters through employment; America’s enters through distribution. Both are furiously expanding cognitive capacity, and neither has a demand-side answer. Some people bet on America. I will bet only this: whoever first solves the distribution problem politically will win—and at present, neither system has demonstrated that ability. The hotter the chip war becomes, the more clearly both sides reveal the same evasion: using supply-side diligence to avoid demand-side cowardice.
5. The Bet: Three Scenarios, Six Indicators#
Forget prediction. Place bets instead. Here are three possible worlds, ordered by my current weights.
The electricity scenario: baseline, about 50 percent. AI is a general-purpose technology and follows the script of electrification: slow diffusion, a J-shaped productivity curve, and twenty years of organizational restructuring and institutional adaptation. Routine cognition is compressed, the accountability layer holds, and the embodied layer slowly gives way. Through sustained pain, society develops new tools of distribution. The pain is real, but history has rhymed this way before.
The loom scenario: 20 to 30 percent. Capability hits a wall around the level of an “excellent intern.” Hallucinations and accountability problems keep humans in the loop for the long term, and a structure of “expert, AI, and reviewer” becomes the steady state. Entry-level employment reaches a new equilibrium after a generation of pressure, and the standard historical analogy holds in full. This is the most comfortable world—and effectively the only world governments are currently prepared for.
The horse scenario: 10 to 20 percent, and rising. Cognition and embodiment fall in succession within fifteen years. The market-clearing wage for a substantial share of human labor drops below a decent standard of living, and the entire problem reduces to distributional politics. In that world, every debate about employment today is merely rearranging the seating chart in the Titanic’s first-class cabin.
The rational posture is not to argue over which scenario is true. It is to live in the electricity scenario while buying insurance against the horse scenario. That insurance policy is a distribution regime, and the sooner we buy it, the better, because institutions take decades to build.
A bet needs falsifiable indicators. I am watching six. Labor’s share of national income: a sustained decline is the horse scenario’s ECG. Entry-level hiring in cognitive professions: the thermometer for the apprenticeship crisis. The duration of tasks AI agents can complete autonomously: it currently doubles every few months, and whether that curve bends will decide the loom scenario’s fate. The unit-cost curve for robots: the length of the embodied layer’s fuse. The first licensed profession to trade liability exemptions for efficiency: the first leak in the institutional levee. And legislative progress on taxes on compute or AI rents: the only hard indicator that construction of the redistribution loop has begun.
6. Epilogue: When “Useful” Is No Longer a Human Trait#
Many discussions skip the sharpest half of the question: “Will AI create a new social psychology?” That is precisely where the argument ends.
The status order of modern society runs on an unstated premise: cognition is scarce, so ranking cognition is roughly equivalent to ranking people. Schools sort people by cognition. Workplaces price them by cognition. A person’s “usefulness” is almost identical to the market value of their cognitive output. When cognition becomes cheap, the sorting machine loses its frame of reference. Unemployment is an economic problem. The feeling of uselessness is a political problem. Whenever history has produced “surplus people” at scale, they have eventually found a political outlet—and rarely a good one. The implicit bargain of education will break first. The collapsing return on years of study is already visible in graduate unemployment on both sides of the Pacific.
New status games will reorganize themselves around new scarcities. The list of things genuinely scarce in the AI age is surprisingly ancient: accountability, a punishable signature; presence, a body that cannot be copied; taste, the right to select from infinite supply; care, being cared about by a real person rather than a process; and ownership, equity in the cognition factories. They have only one common denominator: individuality. AI can generate almost anything, but it cannot be someone. It can produce every kind of output, but it cannot be present, bear responsibility, or suffer loss. In a world where thought is no longer scarce, “who you are” becomes more valuable than “what you can do.”
So here is this article’s final answer: the productivity explosion is certain. Musk got the first half right. But a productivity revolution never automatically becomes broad prosperity. Steam did not. Electricity did not. Cognition will not. Prosperity is technology’s promise; sharing it is a prize won through politics. The real battlefield of the next twenty years will not be inside the models’ parameters. It will be the question of who collects the rents from machine cognition. Both superpowers are building cognitive capacity at full speed, while history watches coldly from the sidelines: the last time humanity solved this problem, it took a hundred years. This time, we do not have a hundred years.
