AI, Work and the Economy

The argument

The argument

Why cheaper machine work squeezes pay and makes scarce things dear, what the data show so far, how the squeeze could end in a downturn, and what decides the outcome after it.

1. What pins the wage

In standard economics, when a firm can do a task with a person or with a machine, the wage for that task cannot for long stay above what the machine would cost. Båge and Wilson (2026) ask what sets the machine's cost.

A machine is built from other machines, labour, energy, land, minerals and permits. Followed back far enough, those costs end in two things: the work of people, and inputs no one can produce more of, such as land, sites, grid connections and mineral deposits. So the wage a machine competes with can fall in two ways. Machines can get better at the task. Or machines can take over the building of machines, so that less of a machine's cost is wages and more of it is rent on the scarce inputs. In the US national accounts for 2023, pay for work in the US is still 47% of what it costs to build equipment, software and buildings, counting the work all the way down the supply chain. AI that writes software, designs chips and runs data centres works on that share.

The same scarce inputs price the alternative to working. Life without a wage still requires somewhere to live, food, energy and warmth. When manufactured goods become cheaper and housing does not, what people can afford without work falls, and with it, perhaps, the wage below which they would leave work altogether.

What an hour's US pay buys, by what it is spent on, 1950 = 100

Average hourly earnings divided by the consumer price index of each category. Before 1964 the wage series is spliced. Source: Båge and Wilson (2026), from US Bureau of Labor Statistics data.

Between 1964 and 2024 an hour's US pay came to buy 277% more durable goods and 21% less shelter: the two measures of the same wage diverged by a factor of 4.8.

The paper's results are about relative prices and where they tend. It gives no dates. The rest of this page is our reading of how the first years might go, set out as seven points with dated predictions.

Båge and Wilson, Pinning the Wage to Scarcity and Technology (2026).

2. The first effects: a squeeze

If machine work is getting cheaper, the first signs should be in what it competes with and in what it needs. It competes with office work. It needs capital, power and sites. On our reading both are visible, and neither yet takes the form of large job losses.

Labour's share of US business output is the lowest since records began in 1947. It has fallen 3.2 points since the second quarter of 2023, about a point a year, with no recession. Over the same period output in the business sector grew 3.0% a year while hours worked grew 0.2%.

Labour's share of US business output, index 2017 = 100

Nonfarm business sector. Shaded: US recessions.

US business output and hours worked, index 2019 Q4 = 100

Nonfarm business sector.

So far this has come through pay and hiring more than through jobs. On their trend from September 2024, when employment among 25–54-year-olds peaked, to August 2026, the US jobs data allow a net loss of jobs to AI of at most 0.2% of the workforce a year, and the trend itself cannot be told from zero. That figure is an upper limit on the pace so far. It cannot show whether the pace will grow; our view assumes that it doubles every two years.

It is not yet clear whether people who lose work look for other jobs or leave the workforce. The share of Americans aged 25–54 in the workforce, working or looking for work, stayed between 83.0% and 84.0% from 2023 to May 2026, then fell 0.6 points in June. Measured from the peak in employment in Sep 2024, 94% of the fall in the share in work is people who have left the workforce, against 12% in 2007–10. That figure depends on June's fall: measured from other starting points it is 18–59%. The model uses the middle of the three, 59%. People who stop looking are not counted as unemployed.

Job growth has come from health care and social assistance. From Aug 2024 to Aug 2026 they added 1,343 thousand of the 1,311 thousand net new US jobs: 102%. Outside care, private employers added 60 thousand jobs and government shed 92 thousand, as federal cuts (330 thousand) outweighed hiring by state and local governments. Women took 72% of the new jobs.

+1,343
thousand jobs added by health care and social assistance
+60
thousand added by private employers outside care
−92
thousand shed by government, as federal cuts outweighed state and local hiring

3. Scarce capital, power and sites

Cheaper machine work raises the demand for what it runs on. Data centres need chips, power, grid connections, sites and permits, and none of these can be supplied quickly. The largest builders spend most of the cash their businesses generate on new capacity, so further growth has to be paid for with outside money.

Their borrowing comes while money is dear. The ten-year US Treasury yield reached 5.31% on 5 Oct 2026, its highest since May 2002; the yield on inflation-protected ten-year Treasuries, which leaves inflation out, reached 2.95% the same day, its highest since Nov 2008. On 19 Nov 2026 they stood at 5.04% and 2.88%.

Ten-year US Treasury yields, %

Daily, drawn every fifth day. The inflation-protected series begins in 2003. Shaded: US recessions. Source: Federal Reserve Board via FRED.

The boom is narrower than it looks. Since 2019, US output of semiconductors has risen 69%, while manufacturing as a whole has grown 0.1% and employs 1% fewer people. Much of the equipment is imported: imports of computers and parts have risen from 0.7% of GDP in 2022 to 1.9%. Net of them, US investment in computing, software and power is 0.4 points of GDP below its 2015–22 trend. A fall in AI investment might therefore hurt mainly the countries that make the equipment; in the United States it would arrive through share prices and credit.

How a retreat from the build-out's financing would travel is set out in five steps, each with public measures: the money behind AI investment. Two of the four steps that public data can show are on: steps 2 and 3.

The model has no long-term interest rate and no market for power or sites, so it does not compute this part of the argument. It carries the squeeze in one way: core inflation starts at 3.0%, above the Federal Reserve's target, and fades at the pace it has faded since 2024. In the run of our view the Federal Reserve therefore raises its rate from 3.9% to 4.1% in early 2027. Without that inflation the model would have had it cutting from the start, to 3.4% by mid-2027.

Dear money with scarce supply recalls the 1970s, when inflation and unemployment rose together. We expect inflation to fade instead: pay is not running ahead of prices, and the downturn we expect comes from lost spending. That is an assumption. If inflation rose again by three points through 2028, the model has the Federal Reserve's rate reach 7.2% and 9.6% of the US workforce out of work by mid-2032, against 8.6%; if people came to expect the rise to last, 19% and 12.3%. If inflation does not fade

4. How the squeeze could end

The squeeze could end much as the dot-com boom did: lenders and investors retreat from the firms that depend on outside money, share prices fall, and the households that own them spend less. What follows would depend on how large the fall is and on how governments respond.

The model runs four economies together, the United States, the euro area, Sweden and China, month by month, with the euro area in two parts, France and the rest of the union, which share the European Central Bank's rate. It includes the channels that turned 1929 into the Great Depression: lost wages cutting spending, bank losses raising the cost of credit, bank runs, collapsing trade, and falling prices making debts heavier. One set of parameters was fitted on nine past downturns, from the United States in 1929–33 to the euro area in 2011–13, and checked on one it was not fitted to, the dot-com bust. For 1933 the model gives US output at 71% of its 1929 level against 74% in the data, and unemployment at 26% against 25%. For 2008–10 it gives peak unemployment of 9.3% against 10.0%, and for the dot-com bust 5.9% against 6.3%.

In the run of our view, share prices lose half as much as in the dot-com crash, on that crash's own path from their high of 6 Oct 2026: US households' shares are 22% down at the bottom. US unemployment rises from 4.2% to 5.2% by the end of 2027 and 6.2% in early 2029, a rise of 2.0 points against 2.5 after the dot-com crash. The Federal Reserve cuts its rate to 2.7% by the end of 2028. At the worst of the downturn the four economies' output is 2.3% below what they could produce.

The fall in shares, and the jobs lost to AI

Each run goes on to June 2032. Out of work counts the unemployed and the people who have left the workforce after losing a job to AI. Output is the four economies together against what they could produce, at its worst month. The outcome names the whole run by its worst point. Each run opens in the model.

US unemployment, end of 2028US out of work, mid-2032Four economies' output at its worstOutcome
No fall in shares, and no jobs lost to AI4.5%4.7%−0.9%A slowdown
Jobs lost to AI, no fall in shares4.9%6.1%−1.9%A recession
The fall in shares, no jobs lost to AI5.5%4.3%−2.5%A recession
Both: our view5.9%8.6%−4.1%A deep recession
A fall the size of the dot-com crash5.0%9.6%−5.5%A deep recession
A fall half as large again5.8%10.8%−7.1%A deep recession
US unemployment rate, %

The fall in shares brings the downturn; the jobs lost to AI decide what is left after it. With the fall in shares alone, 6.3% of the US workforce is out of work at the worst and 4.3% by mid-2032, about where it started. With the loss of jobs to AI and no fall in shares there is no downturn in 2027; the share out of work rises slowly instead. Unemployment is then two points up only by August 2031, and the stimulus waits until then.

5. Why the jobs may not come back

AI can already do much office, professional and technical work. Those jobs are 61% of employment in the United States, 56% in the euro area outside France, 59% in France, 65% in Sweden and 20% in China. Physical work would have to wait for robots.

Recessions hit physical work hardest. From 2007 to 2010, US physical jobs fell 11% while office jobs fell 3%, and 37% of a typical US recession's job losses are in work AI can do. In past recessions most of the jobs cut were filled again. In the model, each quarter firms automate 5% of the jobs cut in the downturn and do not refill them, in office work only until robots arrive.

The response comes late. In the model governments add stimulus once unemployment has risen two points, which in the run of our view is April 2029 in the United States. Unemployment then falls back to 4.7% by the third quarter of 2030. The share of the workforce out of work, which counts those who have left it, falls no lower than 5.3%, against 4.2% at the start.

US unemployment, end of 2028US out of work, mid-2032Four economies' output at its worstOutcome
Firms automate 5% of the jobs they cut each quarter: our view5.9%8.6%−4.1%A deep recession
Twice as many6.0%9.3%−4.5%A deep recession
Three times as many6.1%10.1%−4.9%A deep recession
Robots do physical work from late 20295.9%9.5%−3.9%A recession
Robots from late 20285.9%9.8%−3.6%A recession
Robots from the start6.1%10.6%−4.1%A deep recession

If firms automate twice as many of the jobs they cut, 9.3% of the US workforce is out of work by mid-2032; three times as many, 10.1%. Our view has no robots doing physical work at scale before mid-2032. If they arrive from late 2029, 9.5% are out of work by then; from the start, 10.6%.

6. What the response decides

After the downturn, in our view, AI replaces jobs faster. In the run 8.6% of the US workforce is out of work by mid-2032, 7.4% unemployed and 1.1% who have left the workforce, and the share is still rising. How far it goes depends on something that cannot yet be measured, how fast the pace grows, and on what governments do.

The same run, by how fast AI replaces jobs

US unemployment, end of 2028US out of work, mid-2032Four economies' output at its worstOutcome
No faster than so far5.7%5.8%−2.8%A recession
Doubling every two years: our view5.9%8.6%−4.1%A deep recession
Doubling every year6.1%10.1%−5.0%A deep recession

The same run, by how governments respond

US unemployment, end of 2028US out of work, mid-2032Four economies' output at its worstOutcome
Today's policies: our view5.9%8.6%−4.1%A deep recession
An income guarantee5.5%7.9%−3.3%A recession
More care jobs, paid for by borrowing4.0%7.5%−3.4%A recession
Spending cuts, tariffs and no bank rescues7.6%17.7%−17.4%A labour depression
The rules of the 1930s9.9%50% (the model's limit)−57.8%A depression
Americans out of work, by how governments respond, % of the workforce

With spending cuts, tariffs and no bank rescues, the same run puts 17.7% of the US workforce out of work by mid-2032, with output 17% below what the four economies could produce. Under the rules of the 1930s output falls 58% below it, deeper than in 1929–33. In at least one economy output reaches the model's floor, where the model finds no bottom. An income guarantee leaves 7.9% out of work, at a cost of up to 1.7% of US GDP a year.

Care may take in some of the people who lose other work, but it is unlikely to replace their pay. It pays about the average in the United States (0.98 times) and less in most of Europe (Sweden 0.84). It is mostly paid for by the state: Medicare and Medicaid cover 64% of US spending on health care. And the people AI displaces may not be the people care hires: men work in care at about a quarter of women's rate. No rich country employs more than 20.1% of its workers in care (Norway); the United States and Sweden are at 15%. Across 34 OECD countries, care's share of jobs follows public spending on long-term care closely (R² 0.66) and the share of people over 65 hardly at all (R² 0.04). The limit appears to be how care is financed. Norway's oil fund paid the equivalent of 7–13% of mainland GDP into its budget each year from 2019 to 2025.

In the model, funding new care jobs up to Norway's share leaves 7.5% of the US workforce out of work by mid-2032, against 8.6%. Paid for by a tax on wages it leaves 7.6%; by a levy on the income AI moves to owners, 7.5%. How an income guarantee is paid for matters more, because it costs more: borrowed, it leaves 7.9% out of work; paid for by a levy on that income, 7.9%; paid for by a tax on wages, 8.7%, more than with no guarantee, since those who still work then spend less.

The research this site rests on points to one more response: a tax on the rent of land and other scarce inputs, funding a payment made equally to everyone, in work or not. A tax on land does not reduce the amount of land, and a payment that continues when someone takes a job need not discourage work. Fully captured, measured US land rents would pay about a third of a basic living income per person (2025).

Run the model yourself

7. Limits and sources

  • The view is stated for the United States. The model has no exchange rates, and it reads inflation a little under target as a heavier burden of debt, so it overstates the damage in Sweden and France: in the run of our view, 15.6% of Sweden's workforce and 14.2% of France's are out of work by mid-2032, from 8.6% and 8.2%. No point of our view rests on those figures.
  • The pace at which AI replaces jobs starts from a measurement. How fast it grows is assumed, and by 2032 the run depends on it more than on anything else.
  • With no fall in shares and no jobs lost to AI, the model's US unemployment still drifts from 4.2% to 4.7% by mid-2032: a small early gap is never fully closed. Readings far out lean high for that reason.
  • The model has no long-term interest rate, no market for power or sites and no shares of income. Those parts of the argument are followed in the data, not computed.
  • It recovers from 2009 faster than the United States did, so it may be optimistic about how long a slump lasts.
  • How far falling prices deepen a slump is only loosely pinned down by the downturns the model is fitted on.
  • It counts no harm from inflation itself, only from what central banks do about it. It can show what fighting a second wave of inflation costs; the cost of letting one run is outside it.
  • It shows what follows from its settings. It is not a forecast.

What we think is happening now, and the evidence that would show whether it is, are set out in our view.

Båge, J. and Wilson, S. (2026). Pinning the Wage to Scarcity and Technology. SSRN 7226858. The model, its data and its checks are public at github.com/wilsoniumite/labor. Data: FRED (Federal Reserve Bank of St Louis), US Bureau of Labor Statistics, US Bureau of Economic Analysis, Federal Reserve Board, Sveriges Riksbank, US Treasury, Statistics Sweden, ILOSTAT, OECD, BIS, Statistics Norway and the World Bank.