The guide

1. The wage and the cost of what could replace it

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 things no one can produce more of: land, locations, grid connections, mineral deposits. While building machines still takes a great deal of labour, part of every machine's cost is wages, and machines stay relatively expensive. As machines take over the building of machines, that part shrinks, more of a machine's cost becomes rent on the scarce inputs, and the wage it competes with would fall with it. 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.

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.

Measured in the price of durable goods and in the price of shelter, the same US wage diverged by a factor of 4.8 between 1964 and 2024: an hour's pay now buys far more manufactured goods, relative to housing, than it did.

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

2. Output growing faster than hours worked

The change may have started to show in the US national accounts. Since Q2 2023, output in the business sector has grown 2.75% a year while hours worked grew 0.23%. Labour's share of business output is the lowest since records began in 1947, down 3.3 points in a year. Since 1990 only recessions and the recoveries after them have seen a fall that fast.

It is not yet clear whether people who lose work are looking for other jobs or leaving 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; the rest are looking for work. That figure depends on June's fall: measured from other starting points it is 18–59%, and on the trend since the peak people are not leaving the workforce, while unemployment among 25–54-year-olds is rising slowly, by 0.17 points a year. People who stop looking are not counted as unemployed.

So far the loss of work is small. On their trend since Sep 2024, the jobs data to Aug 2026 allow at most 22% of the pace at which the model assumes AI replaces jobs; measured from the peak to the latest months, which include June's fall, 39%. These figures are upper limits on the pace so far; they cannot show whether it will speed up.

The investment boom may be 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, investment in computing, software and power is 0.4 points of GDP below its 2015–22 trend: measured this way, the US has no AI investment above trend to lose, and the model's US bust has no investment leg. A collapse in AI investment might therefore hurt mainly the countries that make the equipment.

3. Where the income goes

The model follows one economy as machines take over tasks, at three speeds.

At the medium speed, machines go from doing 83% of tasks to 92–94% over fifteen years. Pay for work takes 55% of income at the start: 47% for making goods and services and 8% for building the machines. By year fifteen it takes 4–9%. Interest paid to the owners of machines also falls, from 26% to 6–7% of income, because machines get cheaper as the technology improves. The income goes to the land under homes, whose share rises from 13% to 81–86%.

The real wage falls from the start, even as goods get cheaper. Building a machine takes so much work that a lower wage makes machines cheaper, and cheaper machines pull the wage down again: at the start, each 1% gain in what machines can do lowers wages against land rents by about 2.4%. By year six the real wage is 83–89% of where it started, and by year fifteen 24–41%. At the height of the change, goods prices fall 20–25% a year while housing costs rise 27–29% a year, with the overall price index on its 2% target. The share of people in work falls from 60% to 54–56% by year six and 23–34% by year fifteen.

Where two numbers are given, they come from two fits of the model to the same data; the chart shows the first.

4. Public finances and a shrinking wage base

Most tax revenue is raised on wages. In the model, taxes on wages fall from 80% of revenue to 11–24% as pay shrinks, while more people need support. The model's government has debts but no assets, so it starts with Sweden's average deficit over the last five years counted without the interest the Swedish state pays or the income it earns on its assets: 2.06% of income (1.80% of GDP). Public debt rises from the start: from 40% of income to 115–131% after fifteen years at the medium speed and to 152–172% at the fast speed.

In the model, investment falls as a share of income from the start, from 18% to 4–5% by year fifteen at the medium speed, because machines become far cheaper to build. The neutral real interest rate is linked to three things, each by an amount taken from published estimates: how much is invested, how much of their income owners save, and how much governments borrow. Falling investment and income moving to owners, who save more of it, push the rate down; deficits push it up, by less. So the rate falls from the start at every speed: at the medium speed from 0.8% to −0.6% by year six and −3.6 to −3.5% by year fifteen, at the fast speed to −3.7 to −3.6% and at the slow speed to −2.1%. At the medium and fast speeds rates that low are below where central banks can take their policy rates: the policy rate stops at its floor of −0.5% by year ten at the medium speed and by year six at the fast speed. At the slow speed it is still above the floor after fifteen years, at 0.0%. Linked instead to how fast tasks move to machines, as the model assumed until 27 October 2026, the rate would rise by less than a tenth of a point.

The published estimates were measured on changes far smaller than the model's: income moving from wages to owners, and investment falling, on this scale lie well outside them, and several of the estimates of how much more owners save measure only the response in the year or two after a change. The evidence also leaves open how much of the fall in investment adds to saving: the model counts all of it, as the estimate it uses does; counting none of it, on the view that less is then set aside for wear and tear, the rate would end at −1.3 to −1.2% by year fifteen at the medium speed, and it would still not rise at any speed. The model's figures for debt and interest rates are its reading of that evidence, not forecasts. Started in surplus instead (with the tax on wages the model used until 27 October 2026), debt would end lower: 108% after fifteen years at the medium speed (in the first fit; the second was not run this way).

One response would be 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).

5. Why markets may see it late

Markets would learn of this mainly from data on labour's share of income, which are noisy and revised. In the model, a market that starts with 5% odds that the change is under way moves from 10% to 90% between years 1.2 and 2.1, when about 3% of the eventual fall in labour's share has happened.

In the model, government bond yields move only once the market starts to learn: across that window the ten-year yield falls by 141 basis points. How far long-term rates move depends most on how much of the new income its owners save, which pushes rates down, and also on how much governments borrow, which pushes them up. Across 4,000 variations of the model, the ten-year yield ten years on ranges from −318 to +167 basis points around its start (5th to 95th percentile); it falls in 8 variations in ten. The variations also differ in how much of the fall in investment they count as saving, from none of it to all: the yield falls in about 7 in ten of those counting less than a third, and in about 9 in ten of those counting two-thirds or more.

An earlier warning, if there is one, is more likely to come from the real economy: wages, labour's share of income, tax receipts and the jobs data. A market move on headlines alone, with no new data behind it, may reverse if the data do not follow.

6. How a bust could become a depression

An end to the AI investment boom might begin much as the dot-com bust did: share prices fall, investment stops and lenders take losses. What follows would depend above all 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: the strain on the union's borrowing is French, so France carries a spread of its own over Germany's. 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%.

The same bust, different policies

A bust the size of the dot-com crash, AI replacing jobs at the central pace and robots arriving in two years. Each run goes on to the end of June 2032. Out of work counts the unemployed and the people who have left the workforce after losing their job, at its highest within the run; "rising" marks a run in which it is still rising at the end. Output is the four economies together against their capacity.

In the model, today's policies hold the fall in output well short of a depression. They do much less for employment: by June 2032, 18% of the US workforce is out of work and still rising at the end of the run, while output across the four economies falls 8.6% below capacity at its worst. The unemployment rate shows only part of it: it reaches 11%, because the model assumes most people whose jobs AI replaces leave the workforce, and 7% of the workforce has left it by then. An income guarantee lowers the share out of work only to 14% (from 18%), at a cost of up to 4.0% of US GDP a year. With spending cuts, failing banks and tariffs, the same bust becomes a depression: output falls 28% below capacity. Under the rules of the 1930s it falls 60% below capacity, far deeper than in 1929–33: at least one economy reaches the model's floor, where the model finds no bottom.

These runs take AI's pace from the research: each year it replaces jobs equal to 1% of the US workforce. On their trend since Sep 2024, the jobs data allow at most 22% of that pace; measured to the latest months, 39%. At those paces the same bust under today's policies leaves 10% and 11% of the US workforce out of work at its peak quarter, against 17% read the same way (18% at the peak month); with spending cuts and tariffs, 26% and 28%. In these runs the bust and the response to it matter more than the pace of job losses seen so far.

Today's policies, different shocks

The more of the jobs cut in a downturn that firms automate and do not refill, the more of the loss the unemployment rate misses. When firms automate three times as many, 29% of the workforce is out of work, but the unemployment rate shows 14%: 15% have left the workforce.

In the model, governments act on dates rather than on the data: each response starts in the month unemployment had risen by a set amount in the central case (for today's policies two percentage points, in the United States February 2028), and starts then whatever the shock, so a run with no bust gets the same stimulus.

In the model the bust costs the European economies the most jobs: against the same run with no bust, it puts 12.7 points more of Sweden's workforce out of work, 8.0 of France's, 7.5 of the rest of the euro area's and 5.2 of the United States'. The link between lost output and lost jobs is fitted on past downturns, and in the euro area's of 2011–13 and Sweden's of 1990–93 each point of output lost cost more jobs than in the United States. Sweden's prices fall furthest and its slump is the deepest; the model has no exchange rates, so it overstates Sweden's exposure.

Run the model yourself

7. Office work first, physical work later

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%. Until robots arrive, most jobs lost in a recession are filled again afterwards: only 37% of a typical US recession's job losses are in work AI can do. In the model, once robots arrive, those jobs are no longer refilled.

In the model, a later arrival of robots postpones the loss of jobs without preventing it.

8. Care and US job growth

From Aug 2024 to Aug 2026, health care and social assistance added 1,349 thousand of the 1,258 thousand net new US jobs: 107%. Outside care, private employers added 1 thousand jobs and government shed 92 thousand, as federal cuts (328 thousand) outweighed hiring by state and local governments. Women took 75% of the new jobs.

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.

Care is unlikely to grow without limit. 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 lowers the share of the US workforce out of work by June 2032 from 18% to 12%. How the jobs are paid for matters more as the crisis deepens: taxes on wages raise less as the wage base shrinks, while a levy on the income AI moves to owners raises more. In the model, borrowing costs do not limit an expansion of this size.

9. What to watch

A direct-evidence signal is a measure the argument predicts should move; it can count as evidence for our view, but only if no US recession signal was on during the period it measures and the prices US businesses charge were not rising faster than in 95% of quarters from 1990 to 2019. A downturn signal, when it triggers, is a sign that a market fall or a recession has begun; it can count as evidence for two points of our view, a stock market bust and a recession, judged against how often it was on from 1990 to 2019. A context signal never counts as evidence for our view.

• Jobs and pay: office jobs against all jobs, and labour's share of business output.
• Households: people leaving the workforce and private jobs outside care.
• AI investment and markets: technology stocks, corporate credit, bank lending to non-bank lenders and Korea's exports.
• The labour market: US unemployment, jobless claims, permanent job losses and hiring; then a market fall and job losses together, and office jobs not regained after a downturn.
• Public finances and policy: taxes on wages, government borrowing costs in a US downturn, and the borrowing costs of euro-area countries, France's above all.
• Market context: ten-year yields in six markets, oil, mortgage rates, the yen and others.

Policy may decide the outcome, so watch budgets, benefits, bank rescues and tariffs too.

Limits and sources

• The model is a simplified description of four large economies. Many of its settings are approximations; each is stated in its code.
• It has no exchange rates, so it overstates Sweden's exposure. A falling krona cushioned Sweden in 1931 and in 2008–09.
• It recovers from 2009 faster than the United States did, so it is optimistic about how long a slump lasts.
• How owners spend their income and how governments borrow are set by stated assumptions; the model does not estimate them.
• It shows what can happen and why. It is not a forecast.

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, Sveriges Riksbank, US Treasury, Statistics Sweden, ILOSTAT, OECD, BIS, Statistics Norway and the World Bank.