AI and the Economy: What Happens When Intelligence Gets Cheap?
"Will AI take my job?" is the question everyone asks, and it is the wrong one. It frames the economy as a fixed pile of work that machines and humans fight over. Economies don't behave that way, and the question produces answers that are either falsely reassuring or falsely apocalyptic.
Here is a better frame. AI is a steep fall in the price of a production input. The input is cognition: reading, drafting, summarizing, classifying, translating, checking, planning. For all of human history, the only way to buy a unit of that work was to hire a person. Now a meaningful and growing slice of it can be bought by the token, and the price per unit of capability has been falling by orders of magnitude every couple of years.
Economists know what happens when an input gets dramatically cheaper, because it has happened before. Electricity collapsed the price of mechanical power. Three things follow, reliably. People use vastly more of the input than they did when it was expensive. They substitute it for other inputs, including labor. And they build things that were never economic before, which is where most of the long-run action is.
That frame doesn't answer the job question. But it turns one unanswerable question into several sharper ones: which tasks get cheap, what happens to the wages attached to the remaining tasks, and who pockets the difference.
What history predicts, and where it lies
Three analogies do most of the work in this debate, and each one teaches something real before it breaks.
Electricity is the diffusion lesson. Edison's Pearl Street station opened in 1882; American factory productivity didn't visibly respond until the 1920s. The delay wasn't the technology. It was the complements: factories had to be rebuilt around small motors at each workstation instead of one central steam shaft, and that meant new buildings, new workflows, new managers. The prediction for AI is sobering and probably right — the gains arrive when organizations are redesigned around the input, not when they buy it.
The spreadsheet is the task lesson. VisiCalc and Lotus 1-2-3 automated exactly what bookkeeping clerks did all day: arithmetic and recalculation. Clerk employment fell. But accountants and financial analysts multiplied, because when recalculating a model became free, everyone wanted ten scenarios instead of one. Automating a task expanded demand for the judgment wrapped around it.
Containerization is the distribution lesson. The shipping container gutted longshore employment — and almost none of the enormous surplus went to ports or shipping lines. It flowed through to global supply chains and, ultimately, to consumers as cheaper goods. Surplus from a cost collapse can bypass the disrupted industry entirely.
Each analogy also breaks in a specific place.
| Analogy | What it predicts | Where it breaks for AI |
|---|---|---|
| Electricity | Diffusion takes decades; complements decide the timing | Software needs no rewiring — the technology spreads instantly, even if trust doesn't |
| Spreadsheet | Task automation can grow the surrounding profession | Spreadsheets stayed put; AI's frontier of "what it can do" moves every year |
| Containers | Surplus flows past the disrupted workers to consumers | Containers cut the cost of moving what workers made; AI cuts the cost of making it |
The last break is the deep one. Every prior general-purpose technology cheapened something adjacent to human thinking — muscle, transport, calculation. This one cheapens a portion of the thinking itself. History remains the best guide available. It has just never run this exact experiment.
Jobs are bundles, and bundles get reshuffled
The labor-market question sharpens once you drop the word "job" and use the word "task." An occupation is a bundle of twenty or thirty distinct tasks. Automation rarely deletes the bundle; it removes some tasks, cheapens others, and changes the value of what's left.
Radiology is the canonical cautionary tale. In 2016, Geoffrey Hinton suggested we should stop training radiologists because image recognition was about to surpass them. A decade later there is a radiologist shortage. Reading scans is one task in a bundle that includes choosing imaging protocols, consulting with surgeons, performing procedures, and delivering findings. The machines made one task faster; the bundle rebalanced around the rest.
Translation shows the harsher path. The price of raw translation has collapsed, and the job that remains — post-editing machine output — pays less than the craft it replaced, because the residual task is quality control on someone else's cheap draft. Same mechanism, opposite wage result.
Software development sits in between and is the case to watch. Code generation automates the typing, the boilerplate, the first draft. The bundle is tilting toward specification, review, and architecture — deciding what to build and verifying that it works. Whether that raises or lowers developer wages depends on how scarce those judgment tasks turn out to be once the drafting is free.
The question is not whether a machine can do part of your job. It is whether the part it cannot do becomes more valuable once the rest is nearly free — or merely leftover.
That is the whole wage question in one sentence. Complement, and your remaining hours are worth more. Residue, and they are worth less. Most occupations contain some of each, which is why honest forecasts refuse to average out cleanly.
Who captures the surplus
When an input's price collapses, someone banks the difference. There are four plausible claimants.
Model providers look, at first glance, like the obvious winners. But frontier models are close substitutes for one another, training costs are brutal, and competition has been shredding prices: the cost of a given level of capability has fallen roughly tenfold per year. That is a terrible setup for fat margins unless a provider finds durable lock-in.
The application layer — companies that wrap models in workflow, distribution, and domain trust — may do better than the layer beneath it. This rhymes with the PC era, where the money migrated from hardware to software and services. Knowing a customer's problem is harder to commoditize than the intelligence used to solve it.
Incumbent firms are the quiet claimants. A bank or insurer that cuts processing costs thirty percent doesn't issue a press release; it books the margin. Much of the surplus may show up not as AI revenue but as slightly fatter profits spread across the boring middle of the economy.
Consumers usually win in the end — that is the consistent verdict of every prior general-purpose technology. But "in the end" has historically meant decades, and the interim distribution is precisely what politics will be fighting about.
The wildcard is open weights. If freely available models stay within shouting distance of the frontier, they put a ceiling on what anyone can charge for raw intelligence, pushing the surplus down the stack toward users and toward whoever sells the compute. Whether that gap stays narrow is one of the most economically consequential open questions in the industry, and nobody knows the answer.
What to watch instead of predictions
Forecasts about AI and the economy are cheap. Prices and quantities are not. Three indicators beat any op-ed.
First, productivity statistics against median wages. If measured productivity climbs while the median paycheck doesn't, the surplus is real but concentrated — the distribution fight arrives on schedule. If neither moves for years, we are living through electricity's long lag, and patience matters more than panic.
Second, task prices in freelance markets. Platforms like Upwork are where cognition is bought closest to the spot market, and rates for writing, translation, and basic design began sliding within months of late-2022 model releases. Watch which categories fall next and which stay stubbornly firm. Falling prices show what AI substitutes for; firm prices show what it complements.
Third, adoption depth versus announcement volume. Corporate AI press releases are a sentiment index, not an economic one. The number that matters is how many firms have rewired an actual production process — and survey data still shows that share is modest. The electricity lesson again: the motor purchased is not the factory redesigned.
The honest ending
The range of serious forecasts is enormous. Careful economists like Daron Acemoglu put AI's productivity contribution at a fraction of a percentage point over a decade. People building frontier models talk about growth rates without industrial precedent. These camps are not divided by intelligence or information; they are divided by assumptions about diffusion, complements, and how far capabilities travel.
That spread is not noise to be averaged away. It is the signal. When credible people disagree by two orders of magnitude, the outcome is not yet determined by the technology — it depends on organizational choices, competition, and policy that haven't happened yet. Distrust anyone selling a point estimate. Watch the prices instead.