Will AI Create More Jobs Than It Destroys? The Question Economists Keep Getting Wrong

Ask an economist whether AI will create jobs, and you will usually get a reassuring answer: yes, technology always creates more jobs than it destroys, so of course AI will create jobs too. It is a comforting line, repeated so often that it has hardened into common sense. It is also one of the most misleading true statements in economics. The claim is not exactly wrong — over centuries, in aggregate, employment has grown alongside every wave of automation. But the way that truth gets deployed in the AI debate papers over almost everything that actually matters to a working person: when the new jobs arrive, who gets them, and who pays the bill while everyone waits.

I want to take the optimistic case seriously before I complicate it, because it deserves better than a strawman. Then I want to explain why economists keep getting this question wrong — not because they are stupid, but because they are answering a different question than the one workers are actually asking.

The reassuring claim, at its strongest

Here is the optimistic argument in its best form. When a technology automates a task, it makes the output of that task cheaper. Cheaper output means more demand for it. Automation also frees up human labour and capital to do other things. The classic example is agriculture: two centuries ago most people farmed; today a tiny fraction do, and yet we are not drowning in mass unemployment. The farmhands did not vanish — their descendants became factory workers, then office workers, then software engineers and nurses and logistics coordinators in industries that did not exist when the tractor arrived.

Automation also creates demand in less obvious ways. When ATMs spread in the 1980s and 1990s, plenty of people assumed bank tellers were finished. Instead the cost of running a branch fell, banks opened more branches, and the number of tellers actually rose for a time — even as the job itself shifted from counting cash to selling services. Economists call the mistake the lump-of-labour fallacy: the belief that there is a fixed amount of work to go around, so every task a machine takes is a job permanently lost. There isn’t a fixed amount. Human wants are effectively bottomless, and as some work disappears, new work has reliably appeared to meet demand we could not previously afford to satisfy.

All of this is genuinely true, and anyone forecasting doom should sit with it. The historical record does not support the idea that machines cause permanent, economy-wide unemployment. In the long run, in aggregate, work reconstitutes itself. If your only question is “will there be jobs in fifty years,” the honest answer is almost certainly yes.

Why economists keep getting it wrong

So where is the error? It is not in the aggregate forecast. It is in treating the aggregate forecast as an answer to a human question. Economists keep getting AI wrong because they are trained to model totals and equilibria, and the two things they systematically underweight are the two things that decide whether a technology is good or bad for actual people: timing and distribution. Underneath both sits a third variable most models ignore entirely — power.

Start with timing. The phrase “in the long run” is doing enormous, quietly dishonest work in the optimistic story. The long run can be a very long time. The shift from farm to factory to office did not happen in a smooth glide; it unfolded over generations, punctuated by depressions, migrations, and a great deal of misery for the people caught in the transition. “Eventually the economy adjusts” is true in the way that “eventually the fever breaks” is true. It tells you nothing about whether you survive the night. A displaced forty-five-year-old whose skill has just been made worthless does not get to draw down the prosperity his grandchildren will enjoy.

“Eventually the economy adjusts” is true the way “eventually the fever breaks” is true. It tells you nothing about whether you survive the night.

Then there is distribution — and this is the sharper blade. Even when new jobs do appear, they rarely appear for the people who lost the old ones. The laid-off call-centre worker does not become a prompt engineer. The new jobs tend to demand different skills, exist in different places, and go to different, often younger, people. The gains and the losses land on entirely different human beings. Economists net them out in a single number — “employment rose” — and that number hides a transfer: from the displaced to the well-positioned, from labour to capital, from one generation to the next. In aggregate the ledger balances. In individual lives it does not net out at all. A statistic that is true for a country can be a catastrophe for a person.

This is the pattern I keep coming back to in almost everything I write about technology, and it is worth naming plainly: a new tool arrives, the productivity gains are real, and then there is a fight over where those gains go. The question is never simply “does this technology create value.” Value is almost always created. The question is who captures it. If you want the longer version of that argument, I have written about how technology gets captured — how the surplus a new tool generates gets steered, deliberately, toward whoever already holds the power to steer it.

The lesson the Luddites actually teach

We use “Luddite” as an insult now — a word for someone too dim to understand progress. That caricature is precisely how the winners of that fight wanted the story remembered, and it gets the history backwards. The Luddites were not confused about technology. They were skilled textile workers who understood the new machines perfectly well. Their objection was not to the loom; it was to the way the loom was being used — to break their bargaining power, cut their wages, and route the gains to the factory owners, all without so much as a conversation about terms.

And here is the part the sneering version leaves out: in the narrow sense, they were right. Their trade was destroyed. Their wages collapsed. The broad prosperity that industrialisation eventually produced took the better part of a century to arrive and went overwhelmingly to people other than them. The optimists’ timeline was accurate and the optimists’ comfort was useless, because “your great-grandchildren will be richer” is not a wage. The Luddites lost not because they misunderstood the technology but because they had no power over how it was deployed. That is the real lesson: the outcome of a technological shift is not decided by the technology. It is decided by the distribution of power around it — who owns the machine, who sets the rules, who can afford to walk away from the table.

What “eventually” costs a real person

It is easy to keep this abstract, so let me make it concrete. If AI can do a meaningful share of what a junior analyst, a paralegal, a copywriter, or a support agent does today, the firm does not slowly retrain those people into higher-value roles out of kindness. It cuts the roles and books the savings. The value is real; it flows to shareholders and, maybe, to customers as lower prices. The displaced worker is told the economy will generate new opportunities. It probably will. But not this quarter, not in her town, and quite possibly not for her at all.

The distributional problem is especially cruel at the bottom of the ladder, because the ladder is how people used to climb. The traditional path into a profession ran through the grunt work — the document review, the first-draft copy, the basic support tickets — exactly the tasks AI is best at absorbing. When you automate the bottom rung, you do not just displace the people standing on it; you remove the on-ramp for everyone who was going to start there. I have written separately about the vanishing entry-level job, because I think it is one of the most underrated risks of this whole transition: an economy can keep producing plenty of jobs in the aggregate while quietly pulling up the first few rungs of the ladder, leaving a generation of young workers with no obvious way to begin.

If you are trying to think this through for your own situation rather than the economy’s, I have taken a more personal run at it in will AI take my job. The short version is that “the economy will be fine” and “you will be fine” are two completely different claims, and the optimists routinely smuggle the second in under cover of the first. Do not let them. The aggregate is not your landlord. The aggregate does not pay your bills.

The one variable that actually decides it

So will AI create more jobs than it destroys? In the long run, in aggregate, probably yes — and I would not bet against the historical pattern. But that is the wrong question, and answering it confidently is how economists keep missing the point. The question that decides whether AI is good or bad for workers is not how many jobs but who owns the technology and who sets the rules around it. Every other variable — the pace of adoption, the retraining, the wage effects — bends around that one.

Consider the two extremes. In one world, AI capability is concentrated in a handful of firms that own the models, the data, and the compute; they capture the productivity gains as profit, workers have little leverage because their skills have been commoditised, and “the economy is booming” coexists with a shrinking, anxious middle. In the other, the gains are widely distributed — through ownership, through strong labour institutions, through rules that give workers a genuine seat at the table — and the same productivity shows up as shorter hours, higher wages, and more security rather than a wider gap. The technology is identical in both worlds. The outcome is opposite. What differs is only the distribution of power.

The technology is identical in both worlds. The outcome is opposite. What differs is only who owns it and who writes the rules.

This is why I find the “jobs created versus jobs destroyed” debate almost beside the point. It treats the outcome as a property of the machine, something we can forecast by studying the technology closely enough. It isn’t. It is a property of the arrangements we build around the machine — and those are chosen, not discovered. Automation has no opinion about who benefits. People decide that, and history is fairly unambiguous that the people already holding power tend to write the rules to suit themselves, unless they are made to do otherwise.

What would tilt it toward workers

If the outcome is chosen rather than fated, then the useful question is: chosen how? A few things would genuinely move the balance toward workers rather than away from them.

• Ownership that is spread, not hoarded. The single biggest lever is who owns the productive asset. If AI’s gains flow only to the owners of the models, the distribution is settled before the debate even starts. Broader ownership — through profit-sharing, employee stakes, cooperative structures, or public options — changes who the productivity belongs to in the first place.
• Bargaining power that survives the transition. The Luddites lost because they were atomised against concentrated capital. Workers with real collective leverage can bargain over how AI is introduced — as a tool that augments them or a replacement that discards them — instead of simply being told after the fact.
• Rules that make the timeline humane. “Eventually” is only acceptable if the gap is bridged. Portable benefits, serious transition support, and income floors are the difference between a hard adjustment and a ruined life. In aggregate the economy can afford it; the open question is whether it is made to.
• Protecting the on-ramp. If entry-level work is what AI eats first, then someone has to consciously rebuild the ladder — apprenticeships, training that is real rather than rhetorical, roles designed so that newcomers can still start and climb.

None of this is a prediction. It is a set of choices, contested precisely because the stakes are the trillions of dollars of value AI will create. That value is coming. The only open question — the one economists keep answering wrong by forecasting totals instead of asking who takes them — is which direction it flows. Will AI create jobs? Sure, eventually, in aggregate. But “eventually,” “in aggregate,” and “for whom” are not footnotes to that answer. They are the answer. And unlike the technology itself, they are still up for grabs.