QuarryCan an economy get so good at producing things that it weakens the incomes needed to buy them?
AI's demand problem won't be output. It will be who gets paid for it.
Between 1983 and 2019, America's wealthiest 1% saved 53.6% of their disposable income. Everyone else saved 7.1%. If AI moves income from wages towards the owners of capital, that gap decides how much of the new output finds a buyer.
Jaidev Karthickeyan7 min readData: Mian, Straub and Sufi, NBER WP 26941
What happened
Between 1983 and 2019, America's richest 1% saved more than half of every dollar of disposable income they received. Everyone else saved about seven cents. Keep that gap in mind, because AI may be about to move a great deal of income from one group to the other.
Wages do two jobs at once. They pay people for their work, and they hand out the purchasing power that buys what the economy makes. Most of the debate about AI and jobs is about the first job. This note is about the second. If machines do more of the work, does the money that used to arrive in payslips still reach the people who do most of the spending?
The sharpest formal version of the worry comes from Luca Fornaro and Martin Wolf, writing in April. In their model, AI raises the share of tasks machines can do, which shifts income "from high-spending workers to low-spending capitalists". Without a policy response, consumer demand falls, firms cut investment, and output can shrink while productivity grows. They call it a paradox of productivity: sensible for every firm, self-defeating for all of them together.
Read on, and the same paper argues the other side. If the central bank aims for full employment, AI produces a boom, and in time workers' consumption rises because the bigger economy more than makes up for their smaller slice. There are two snags on the way: early automation behaves like a cost-push shock, giving "rapid productivity growth, weak labor market and high inflation", and later the natural interest rate can fall below zero. The authors say plainly they are "not predicting that AI will necessarily lead to a slump", and call their calibration, a 3-point rise in the capital share and a 10% gain in potential output, "highly speculative".
Everything hangs on one assumption: that the people who gain from automation spend less of their income than the people who lose. That is testable, and the chart tests it. Atif Mian, Ludwig Straub and Amir Sufi rebuilt US saving by wealth from 1963 to 2019. Before 1983, the top 1% saved 42.9% of their disposable income and the bottom 99% saved 13.8%. Afterwards, 53.6% and 7.1%. The assumption holds, and the gap has widened.
The arithmetic
United States, Mian, Straub and Sufi
Top 1% saving rate, 1963–82 → 1983–2019
42.9% → 53.6%
Bottom 99% saving rate, same periods
13.8% → 7.1%
Extra annual saving by the top 1% after 1982
+2.9pp of national income
In 2024 dollars
over $680bn a year
Change in annual saving by the bottom 99%
−6.7pp (−60%)
Bottom 99% spending financed by debt, 1982–2019
1.3% of national income a year
Share of that rise financed by the top 1%
57.2%
Federal primary deficit, same period, for scale
1.1% of GDP
Illustration: $100 moved from the bottom 99% to the top 1%
Spent at the bottom 99%'s post-1982 rate
$92.90
Spent at the top 1%'s post-1982 rate
$46.40
Spending someone else must borrow, invest or tax to replace
$46.50
An illustration, not a forecast: it uses averages rather than marginal rates, ignores how prices and interest rates would respond, and assumes the new owners of AI capital behave like today's top 1%, which they may not.
Now the scale. Fornaro and Wolf's calibration moves 3 points of income from labour to capital. Suppose, as an upper bound, all of it moved from the bottom 99% to the top 1% at those average rates. Direct spending would fall by about 1.4% of national income a year (3 × 0.465), roughly the debt-financed spending the bottom 99% added each year after 1982. The bound is loose, because capital income is spread wider than the top 1% and the two papers measure income differently. Treat it as an order of magnitude.
Going to ground
First order
A bigger pie, a smaller slice. That is Daron Acemoglu and Pascual Restrepo's description of automation: it "increases the size of the pie, but labor gets a smaller slice". What rescues workers is new tasks. In their US decomposition, new tasks added 0.47% a year to labour demand between 1947 and 1987 and only 0.35% between 1987 and 2017, while displacement rose from 0.48% to 0.7%. That evidence predates generative AI. It shows the balance tipped once; it doesn't show AI will tip it again. Whether AI mostly replaces tasks or mostly creates them is the biggest unknown in this whole argument.
Second order
The savings didn't build factories. They funded mortgages. If the labour share falls, extra income goes to households that save most of it. Mian, Straub and Sufi followed the post-1982 savings and found "no appreciable increase in net investment done on behalf of the top 1%". The money was lent to the bottom 90% and the government instead, "largely against" households between the 51st and 90th percentiles of wealth: the American middle class, whose wages pay for their spending. The top 1% built up net claims worth an extra 20% of national income. The bottom 99% financed spending worth 1.3% of national income a year with debt, more than the average federal primary deficit.
Their companion theory, Indebted Demand, explains why that bites later. Borrowing pulls tomorrow's spending into today. Push it far enough and the economy lands in what they call a debt trap, where low interest rates stop reviving demand. The mechanism runs on differences in saving out of permanent income, not windfalls.
That is why windfall surveys mislead here: they mostly measure who has bills to pay. An AI-driven shift in income would be lasting, and the chart's long-run averages are closer to that.
Third order
Ownership, not employment, decides who shares. From here on, this is my deduction, not the papers'. If demand comes to lean on household borrowing, asset prices and deficits, three pressures follow. Taxes built on wages and payrolls collect less just as more is asked of them. Lower interest rates leave central banks less room in a downturn. And it gets harder to defend an economy whose output rises while most households feel no better off.
The thread running through all of it is ownership. If more income is paid to capital, the spread of capital decides who can buy what the machines make. The papers offer specific remedies: employment subsidies or labour-tax cuts paid for by a levy on capital (Fornaro and Wolf), redistribution and wealth taxes (Mian, Straub and Sufi), and ending the tax bias towards equipment over labour (Acemoglu and Restrepo). Wider ownership of the capital itself, through employee equity or public funds, is my extension. None of the authors proposes it.
Ears up: what to watch
The US labour share. Follow the BLS series for the nonfarm business sector. A sustained fall that tracks AI adoption is the first link in the chain; stability breaks it.
Household debt service relative to income, published quarterly by the Federal Reserve. If spending holds up only as debt rises, demand is being borrowed, not earned.
The real neutral rate. An AI building boom should push it up first. If it falls during heavy investment, weak consumer demand is winning.
Where AI profits go. Reinvestment absorbs saving; buybacks and dividends paid to high savers add to the glut once the build-out slows.
Bone of contention
The best case against my argument is that the mechanism has never run unopposed. Acemoglu and Restrepo's framework lets new tasks rebuild labour demand, and from 1947 to 1987 they did. If AI makes things cheaper, real wages can rise even as labour's share falls; purchasing power is a quantity, not a share. Fornaro and Wolf show that the right monetary policy turns the slump into a boom. Today's capital-hungry AI build-out is absorbing saving in a way Mian, Straub and Sufi didn't find after 1982. And AI's owners are not only the very rich: pension funds, index funds and retail investors hold much of it, and where those pay out to retirees who spend, the chart overstates the drag.
The evidence has limits too. The chart is American, sorted by wealth rather than income, made of averages rather than marginal rates, and ends before generative AI. The rise in top saving had other causes, from taxes to globalisation to financial deregulation. The authors call their work measurement, not a causal story. The models are stylised and, like the paper behind the chart, not yet peer-reviewed.
The thesis can be proved wrong, and here is how: a labour share that holds steady or rises through broad adoption; new kinds of work growing as fast as old tasks vanish; spending that tracks output without rising household debt or deficits; a real neutral rate that rises over the adoption period; and AI profits that are reinvested rather than paid out. If most of those hold over the next five years, the consumerless economy stays a model, not a forecast.
The question underneath is old; AI just sharpens it. For a century, paying people for work has been the main way purchasing power reached the people who spend it. If more of the income goes to machines and their owners, the economy's ability to produce is not the constraint. The constraint is whether employment can stay the main channel for sharing the returns of intelligence, and if it can't, what replaces it.