Table of Contents
The Comforting Slogan That Falls Apart Under Scrutiny
Every time a new technology threatens to eat a category of work, the same reassuring line gets wheeled out: artificial intelligence will create more jobs than it destroys. Politicians repeat it. Consultants monetize it. Executives use it to soften the edges of layoff announcements. And most people, exhausted by economic anxiety, nod along.
The problem is that the slogan hides a math problem, and the math is uncomfortable.
The claim rests on a historical pattern that may not apply this time, a definition of “job creation” that quietly excludes the people who matter most, and a timeline mismatch so severe it renders the whole comforting story functionally useless for anyone under 50. This essay is about the AI job creation myth, why intelligent people keep repeating it, and what the actual arithmetic looks like when you refuse to flinch.
The point here is not to predict doom. The point is to insist on honesty. Because if the numbers do not add up the way the slogan promises, then the policies built on top of that slogan will fail the people they claim to protect.
Where the Slogan Comes From
The optimistic case has a respectable pedigree. Economists point to the mechanization of agriculture, the rise of the automobile, the personal computer, and the internet. In each case, entire categories of work disappeared, and yet total employment kept rising. Blacksmiths became mechanics. Typists became data analysts. The economy absorbed the shock and expanded.
This is the pattern that gets invoked whenever someone worries about AI net jobs. The historical record, we are told, is on the side of the optimists.
John Stuart Mill, writing in the middle of the industrial revolution, was one of the first to notice something the modern slogan tends to skip over. In Principles of Political Economy, he acknowledged that machinery raised aggregate wealth over the long run, but he refused to pretend the transition was painless. He wrote that the suffering of displaced workers was real, that it lasted for a generation or more, and that it required active policy intervention to soften.
The introduction of machinery is often a great immediate evil to the labouring classes, though it may be for the general and permanent advantage of the whole community.
Mill was doing something the modern slogan refuses to do. He was distinguishing between the aggregate outcome and the individual outcome. The economy, in the long run, adapts. The individual worker, in the short run, often does not.
The Aggregation Trick
The phrase “creates more jobs than it destroys” is an aggregation. It sums across the entire labor market, across decades, across skill levels, across geographies. When you aggregate that broadly, almost any technology looks benign.
But no one lives at the aggregate level. A 47 year old paralegal in Cleveland does not experience the labor market as a summary statistic. She experiences it as a phone call from her manager. And when the slogan tells her that AI will create more jobs than it destroys, it is telling her something that is technically true and personally irrelevant.
This is the first crack in the math. The claim is defended at a level of abstraction that erases the actual human experience of technological change.
The Three Numbers Nobody Wants to Talk About
To understand the real AI economic impact, you have to look past the aggregate and examine three specific numbers: the displacement rate, the reallocation rate, and the wage differential. Each of them tells a story the slogan glosses over.
Number 1: The Displacement Rate
Displacement is how fast jobs disappear. Historically, technological displacement has been relatively slow. The tractor took roughly 50 years to fully displace agricultural labor in the United States. The personal computer took about 30 years to transform office work. The internet took around 20 years to gut retail, newspapers, and travel agencies.
Generative AI, by contrast, is displacing certain categories of white collar work in windows measured in months. Copywriting agencies have reported client contract losses of 40 percent from 2021 to 2025. Translation services, illustration, entry level legal research, and basic coding have seen similar compression.
The speed matters because human beings do not retrain instantly. A worker who loses her job at 45 does not become a prompt engineer at 46. She becomes unemployed, then underemployed, then discouraged, then statistically invisible.
Number 2: The Reallocation Rate
Reallocation is how fast displaced workers find comparable new work. Every previous technological wave produced new jobs, but those jobs did not automatically land in the laps of the people who lost the old ones.
When manufacturing collapsed in the American Midwest, the new jobs appeared in software hubs on the coasts. When retail collapsed, the new jobs appeared in logistics warehouses that paid less and demanded more physical strain. The aggregate number of jobs held up. The specific worker in Youngstown did not benefit.
With AI, the geography is even less forgiving. The new jobs cluster in a handful of metropolitan areas, they require credentials that take years to acquire, and many of them are themselves vulnerable to the next iteration of the technology that just displaced their occupant.
Number 3: The Wage Differential
Even when displaced workers find new work, the wage differential is brutal. Studies of workers displaced by trade shocks and automation consistently show that reemployed workers earn 15 to 40 percent less than they did in their previous role, and that this gap persists for a decade or more.
The slogan counts a $45,000 warehouse job as equivalent to a $75,000 middle management job it replaced. Arithmetically, that is 2 jobs created for 1 destroyed. Experientially, it is a family losing its house.
The optimism about AI net jobs depends on a definition of “job” so loose that a career and a shift are treated as the same unit.
Why This Time Might Actually Be Different
Every generation is warned against believing that “this time is different.” The warning is usually correct. Bubbles, panics, and predictions of technological doom have a long history of embarrassing the people who make them.
But the honest analyst has to acknowledge that some things about the current wave are structurally unlike previous waves. Ignoring those differences is not prudence. It is cowardice dressed up as caution.
The Cognitive Frontier
Every prior wave of automation targeted physical or repetitive labor. The loom replaced weaving. The tractor replaced plowing. The spreadsheet replaced manual bookkeeping. In each case, the technology attacked a narrow band of human activity, and the vast territory of cognitive, creative, and interpersonal work remained a safe harbor for displaced workers.
Generative AI attacks the safe harbor directly. It writes, it analyzes, it summarizes, it codes, it designs, it advises. The territory that absorbed displaced factory workers in the 1980s is the same territory that AI is now colonizing. There is no obvious next frontier to which cognitive workers can flee.
Some optimists point to the trades. Plumbers, electricians, and HVAC technicians are, for now, protected by physical embodiment. But the trades cannot absorb 40 million displaced knowledge workers, and the wage structure of the trades would collapse under the influx even if they could.
The Speed of Capability Gains
Previous technologies had capability curves measured in decades. The car in 1910 was recognizably the car of 1930, just faster and more reliable. The personal computer improved on a predictable Moore’s Law slope that gave workers, firms, and regulators time to adapt.
AI capability is doubling on timescales measured in months. A skill that was a defensible human moat in early 2023 was a commodity by late 2024. The rational response of a firm facing this curve is to freeze hiring, wait for the next model, and see how much more of the workforce becomes redundant. This is exactly what surveys of large employers now show.
The Capital Concentration Problem
Previous technologies were expensive but ultimately diffused. Anyone could buy a tractor, a computer, or a website. The productivity gains, over time, spread across the economy.
Frontier AI, in contrast, is being built by 5 or 6 firms with capital requirements measured in tens of billions of dollars. The productivity gains accrue disproportionately to those firms and their largest customers. There is no clear mechanism by which the gains diffuse to the displaced workers whose economic role has evaporated.
The Jobs That “Get Created” Are Not What You Think
When economists say AI will create new jobs, the imagination fills in something dignified: AI ethicists, prompt engineers, machine learning specialists. These jobs do exist. There are perhaps 200,000 of them in the United States. There will not be 20 million of them.
The jobs that actually get created in large numbers during technological transitions are less glamorous. They fall into three categories, and each has a problem.
Category 1: Data Labeling and Model Supervision
Someone has to train the models. Someone has to review their outputs, flag hallucinations, and provide reinforcement feedback. This work exists at scale, but it is largely offshored to lower cost labor markets, it pays poorly, and it is itself being automated as models become better at self supervision. It is a bridge job, not a destination.
Category 2: AI Assisted Versions of Existing Jobs
The optimistic scenario is that most jobs do not disappear but become AI assisted. The lawyer keeps her job but produces 3 times the output. The doctor keeps her practice but sees 2 times the patients. The programmer keeps his role but ships 5 times the code.
This is the scenario the industry loves to tell, and it contains a hidden problem. If each remaining worker produces 3 to 5 times the output, then the same market demand can be served by one third to one fifth as many workers. The individuals who keep their jobs feel more productive. The ones who lose their jobs feel the aggregate demand curve.
Category 3: Personal Services and Care Work
The last refuge is work that requires physical presence and human warmth: elder care, childcare, hospitality, personal training, therapy. These jobs will grow. They already are.
The problem is that these jobs pay poorly, offer few advancement paths, and are structurally difficult to unionize. Transitioning a workforce from cognitive knowledge work into personal service work is not economic progress in any meaningful sense. It is downward mobility with extra steps.
A society that answers technological displacement by pushing its middle class into elder care and gig delivery has not solved the problem. It has renamed the problem.
What Honest Policy Would Look Like
If the slogan is misleading, what should replace it? The answer is not doom, and it is not luddism. The answer is honesty about the transition, combined with policy that treats the transition as the actual problem, not an abstraction to be waved away.
Mill argued that the state had an obligation to soften the transition when technology displaced labor. He did not argue against the technology. He argued against pretending the displacement was cost free. That distinction remains the most useful frame available.
Stop Measuring Aggregate Jobs
The first policy move is to stop treating aggregate employment as the relevant metric. What matters is labor market transitions: how many workers move from displaced roles into comparable or better roles within a defined window, at what wage differential, in what geography.
Countries that measure transition well, such as Denmark with its flexicurity model, produce meaningfully better outcomes for displaced workers than countries that measure only aggregate employment. The measurement is not cosmetic. It changes what governments are politically obligated to fix.
Fund Real Retraining, Not Theater
Retraining programs in the United States have a long history of failure. They are often short, poorly matched to actual labor demand, and administered by contractors incentivized to enroll rather than to place. A 6 week coding bootcamp does not turn a displaced accountant into a machine learning engineer, and everyone in the room knows it.
Serious retraining looks like 2 year subsidized programs tied to specific employer commitments, paid at close to prior wages, with childcare and relocation support. It is expensive. It is also cheaper than the social costs of a generation of underemployment.
Tax the Gains Where They Accrue
If the productivity gains from AI concentrate in a handful of firms, then the fiscal capacity to fund transition support has to come from those firms. The alternative is to fund the transition by taxing the workers who still have jobs, which is politically unsustainable and morally incoherent.
This is not a call for punitive taxation. It is a recognition that when the returns from a technology are concentrated and the costs are diffused, some redistribution is not ideology but arithmetic.
The Question the Slogan Is Designed to Prevent
The reason the slogan persists is that it forecloses a question that powerful people do not want asked. If AI creates more jobs than it destroys, then there is nothing to discuss, no policy to enact, no distributional fight to have. The market will handle it.
If the slogan is wrong, or even partially wrong, then a much harder conversation opens. Who owns the productivity gains? Who bears the transition costs? What do we owe to the workers whose careers were built on skills that a $20 monthly subscription can now replicate?
These questions do not have easy answers. But they are the actual questions, and pretending they have been resolved by a comforting historical analogy is not analysis. It is anesthetic.
Mill understood, 175 years ago, that the aggregate case for technology and the individual case for displaced workers were separate arguments that had to be made separately. The modern slogan collapses them into one and hopes no one notices. Some of us are noticing.
The AI economic impact will almost certainly be enormous. It may well raise aggregate productivity, aggregate wealth, and, over a long enough horizon, aggregate employment. None of that tells you anything useful about what happens to the specific 40 something knowledge worker whose skills became obsolete in 2024. She lives in the transition, and the transition is what the slogan was designed to make you stop thinking about.
The honest answer is that we do not yet know how the math works out. The dishonest answer is to pretend we do, and to use that pretense to avoid the policy work that either scenario would demand. Whichever side of the debate turns out to be correct, the people repeating the slogan are not being helpful. They are being comfortable, which is a different thing entirely.
