There is a particular genre of prediction I've grown tired of: AI will replace most white-collar work, therefore mass unemployment, therefore catastrophe. It gets repeated at conferences, in board decks, in op-eds, and almost never with any economics attached. It's an assertion about a labor market delivered by people who have not thought for thirty seconds about how a labor market actually clears.
I want to take the claim seriously — more seriously than its boosters do — and then do the thing they skip: follow the money around the loop. Because the interesting question isn't whether a model can write a brief or read a scan. It's what happens to the system if it does. Suppose, for the sake of argument, that the strongest version is true and AI performs the bulk of knowledge work. Does the economy collapse? Or does it do what economies have always done under a productivity shock — convulse, reprice, and reorganize?
The loop everyone forgets to draw
Start with the thing every intro-macro course draws on day one and every AI-doom take ignores. Households supply labor to firms. Firms pay wages. Households spend those wages on goods and services. That spending becomes firm revenue, which funds more wages and investment. Round and round. It's called the circular flow for a reason — income and expenditure are the same dollars seen from two sides.
The doom argument, stated charitably, is a claim about this loop: cut the "wages" arrow and the whole circuit dies. Labor income drops, so consumption drops, so revenue drops, so investment drops, so the system contracts into a spiral. And if you genuinely hold everything else constant — same prices, same ownership, same policy — then yes, that's the logical endpoint. Remove the income and you remove the demand it was going to become.
But "hold everything else constant" is doing enormous work in that sentence. It's the assumption that quietly smuggles the conclusion in. The whole point of a productivity shock is that not everything else stays constant. Prices move. Ownership adapts. Policy reacts. The reason the doom take feels airtight is that it freezes every variable except the one it wants to scare you with.
The micro move the macro take skips
Here's the part I think about most, and it's a micro-level observation that the macro headlines flatten.
Picture a worker — call her an analyst — whose job is to provide a service to another business. Her salary is somebody else's cost. Now automate her role. The standard story stops here: she lost her income, demand falls, bad. But look at what she was going to spend that income on. A chunk of it was going to other services — a tax preparer, a paralegal, a designer, a tutor — and those are exactly the services that, in this scenario, AI has also made dramatically cheaper.
So two things happen at once, and they are not independent. Her nominal income falls, and the price of the basket she consumes falls. The doom take only counts the first. But real living standards are income measured against prices, not income in a vacuum. If the cost of a whole category of services collapses toward the cost of compute, then the income required to maintain a given standard of living collapses with it. You cannot hold wages down while holding prices up — the same automation does both, to the same dollars, at the same time.
This is why the naive subtraction — "remove X million salaries, subtract X billion in demand" — is wrong on its own terms. It double-counts scarcity. The salary and the price were two ends of the same transaction. Automate the transaction and you don't just delete the income; you delete the cost it was paying for.
That doesn't make the transition painless. It means the failure mode is not a simple demand void. It's a distributional problem — who captures the surplus when the cost of cognition falls — which is a very different, and more tractable, thing than "the economy stops working."
History destroyed tasks, not demand
The Industrial Revolution is the obvious precedent, and it cuts against the doom take more than for it. Something like 70–80% of people worked in agriculture before mechanization; that demand for human muscle was annihilated. It was wrenching — generations of genuine hardship, displacement, unrest. But the economy did not contract to nothing. New sectors that nobody had drawn on a chart appeared: manufacturing at scale, then logistics, then an entire services economy.
The durable lesson is that technology destroyed tasks, not demand. Human wants turned out to be effectively unlimited; as old needs got cheap, attention and money moved to needs that didn't exist before. Nobody in 1900 was budgeting for software, air travel, or a personal trainer.
The honest objection — and I think it's the strongest one the skeptics have — is that this time the thing being automated is different in kind. Every prior wave replaced or augmented physical labor and left cognition to humans. AI aims squarely at cognition itself: writing, designing, diagnosing, coding, analyzing, negotiating. If the machine does the thinking, the comforting "humans will move up the value chain" story has to answer a harder question — up to what? That's a real question, and I'll come back to it. But "this wave targets cognition" is an argument about which jobs go, not proof that demand itself disappears. Those are different claims, and the doom take relies on blurring them.
Three places this can actually settle
If you take the strong scenario seriously — AI does 70–90% of knowledge work, productivity explodes, the marginal cost of many services falls toward zero — there aren't infinite outcomes. There are basically three stable configurations, and which one we land in is a choice, not a law of physics.
Concentrated capitalism. The AI capital stack is owned by a handful of firms, labor is displaced, profits concentrate, and most people lack purchasing power. This is the configuration the doom take implicitly assumes — and it's genuinely unstable. An economy that produces enormously but distributes to almost no one has nobody to sell to. It contracts not because production failed but because demand was strangled by distribution. It also tends to resolve politically before it resolves economically, and not gently.
Redistribution. Output rises massively; some of the surplus is taxed or shared back — UBI, a sovereign dividend, public ownership stakes in the AI capital stack — and households consume without traditional wage labor. Demand survives because purchasing power was deliberately reconnected to the output. Work becomes, to a degree, optional. This is less exotic than it sounds: modern economies already redistribute heavily and already run partly on capital income.
Repositioning. Humans move into what stays scarce when cognition gets cheap — physical-world services, in-person craft, status and experience goods, care and relational work, governance and oversight, the creative frontier. AI compresses the cost of the old basket; new human-centric sectors expand into the freed-up spending. This is the historical pattern, and it's the most optimistic — but it's also the slowest, because new sectors take a generation to absorb displaced workers, not a quarter.
Reality is almost certainly a blend, and it varies by country depending on ownership structure and political will. The point is that none of the three is "collapse." Collapse is a specific and avoidable failure: concentration and failed redistribution and political paralysis, all at once. That's a governance failure, not a technological inevitability. The technology sets the possibilities; politics picks among them.
Why would anyone produce if they don't have to?
This is the question underneath the question, and it's where the economics shades into something older.
The premise "people produce because they need to consume" is true at the level of survival. But it's been visibly false at the top of the income distribution for a long time. The wealthy don't stop working when their needs are met; they work for status, power, meaning, curiosity, competition, recognition, the sheer wanting-to-build. Those motives don't switch off — they're what's left when subsistence is solved.
So if AI drives the cost of the survival basket — food, housing, basic services — down far enough, the economy doesn't stop. It tilts from survival production toward meaning and status production. That's not a utopian guess; it's a description of how rich societies already allocate their marginal effort. The scarce, expensive things in a wealthy economy are already not calories. They're attention, authenticity, access, trust, beautiful design, a great doctor who actually listens.
If intelligence becomes cheap, scarcity doesn't disappear. It relocates. And the economy reorganizes around wherever it lands.
The real risk is the transition, not the destination
So I don't think the collapse story holds up. Three things keep stopping it: cheaper goods mean a lower income is needed to live; ownership of AI capital diffuses through pensions, index funds, and sovereign wealth more than the "two companies own everything" image suggests; and no modern government will sit and watch aggregate demand crater — they are demand-managed systems and they intervene, hard, because the alternative is their own collapse.
But "no permanent collapse" is not the same as "fine." The dangerous part is the transition itself — a window, plausibly ten to twenty years, of skill mismatch, identity rupture, and political reaction. Economic collapse requires demand to fall permanently below supply. What automation actually produces is a frantic repricing, where the question of who owns the surplus and how fast people can reposition gets fought out in real time. That fight can be ugly without ever becoming collapse.
Which is the more useful way to frame the whole debate. The economy was never fundamentally about jobs. Jobs are the current mechanism for distributing claims on output. If AI breaks labor as that mechanism, the system doesn't end — it has to invent a new mechanism. That's a hard political and institutional problem. It is not an economic impossibility.
The honest answer to "will AI collapse the economy" is: only if we choose the configuration where it does. The structural pressure points toward turbulence and redistribution, not toward the void the headlines promise.
In part two, I take up the question this one defers: if cognition really does get automated, how does a society retrain itself — institutionally, legally, and culturally — and what does the shift actually look like, profession by profession?