Our view
What we think is happening
AI is making machine work cheap, and what machines cannot make dear. Its first effects are a squeeze: on the pay and hiring of people entering office work, on pay's share of income, and on capital, power and sites. We think that squeeze, with money already dear, brings a downturn about the size of 2001's; that the office jobs cut in it are not all refilled; and that afterwards the public response decides how far the damage goes.
The view and its predictions were set on 13 Nov 2026. The same view, in the model's repository
Now: the squeeze 2026–27
Next: the downturn 2027–29
Later: cheap machine work 2029–32
The mark beside each point is its status in the current assessment, written by Claude, an AI model, on 15 Nov 2026 under written rules for what may move a point. It may depart from our own view.
What the model shows if this view is right. The run, by date
Sixteen predictions were set on 13 Nov 2026. All are still open. The next deadline falls in Jun 2027: Technology stocks. All the indicators
Each prediction below is drawn as a track. Its right end is the threshold our view predicts; its left end is the reading that would count against our view or, for a signal of a downturn, where the measure started. The line marks where it stood when the prediction was set, the dot its latest reading.
Now: the squeeze 2026–27
AI can already do much office work, and its price is falling
Under wayOn tests of whole tasks, the length and difficulty of the work AI systems finish unaided has kept rising, and the price of a given level of ability has kept falling. Machines have stayed dear partly because people build them: in the US national accounts for 2023, pay for work 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 could lower that share.
In the model's run. The model does not compute this point. It takes what AI can do as given, through the pace at which AI replaces jobs (point 2).
What would count against it. A year without progress on tests of whole tasks, or prices for a given task that stop falling.
No single measure: followed through tests of what AI can do, in the news.
Latest news on this point
- Anthropic mathematician Alpöge claims a Lean-formalised proof of the Generalized Riemann Hypothesis made with Claude (20 Nov)
- METR (20 Nov): GPT-6 Astra and Claude Opus 5.5 reach ~40-hour task horizons; doubling every ~3.6 months
- Architect Labs (18 Nov): AI agents designed the Frostbite CPU from ISA to layout; 4.59x performance per watt claimed
- Google's Gemini 4 Neon (18 Nov): 61 on the AA index, #1 on Vending-Bench 2 ($16,843), at $3/$15 per m tokens
- Anthropic (17 Nov): Claude Fable 5.2 agents solve two open black-hole physics problems in ~6 weeks, Lean-verified
Firms hire fewer people for that work; the loss of jobs is small and growing
Early signsFirms appear to use AI first to hire fewer people and to thin office layers, and people entering office work are the first affected. On the US jobs data from September 2024, when employment among 25–54-year-olds peaked, to August 2026, the net loss of jobs to AI is at most 0.2% of the workforce a year, and the trend itself cannot be told from zero. Job growth has come from health care and social assistance: from Aug 2024 to Aug 2026 they added 102% of the net growth in US jobs, while the rest of the economy shed jobs.
In the model's run. The run starts from that upper limit and doubles it every two years: 0.3% of the workforce a year by late 2027, 0.4% by late 2028 and 0.9% by late 2030, and never more than 1%. The doubling time is assumed, not measured, and the later years depend on it more than on anything else. Of the people AI displaces in the United States, 59% are taken to leave the workforce, the middle of three measurements (18–94%).
What would count against it. Office jobs growing in line with other jobs, graduates' unemployment falling back against that of others, and employment among 25–54-year-olds recovering.
Dated predictions
- Private jobs outside careJobs added over a year: 424 thousand in Oct 2026Against our view at 1,078 thousandThreshold zero or belowOpenby Oct 2027
- Office jobsTheir growth minus that of all jobs: −0.50 points in Oct 2026Against our view at 0.04 pointsThreshold −2.41 points or belowOpenby Oct 2028
- People leaving the workforceTwo-year change in the employment rate: −0.30 points in Oct 2026Against our view at 0.77 pointsThreshold −0.40 points or belowOpenby Oct 2028
Also watched
Latest news on this point
- P&G (investor day, 19 Nov) plans to cut its product-supply workforce 20-30% by 2030, within up to $10bn of savings over FY2027-30
- Unionen (18 Nov): employment in Sweden's six most AI-exposed office jobs fell 13% in 2023-25, to 98,500
- Bunnings to cut about 350 Australian tech roles, ~40% of its tech staff, moving work to TCS in India and the Philippines (17 Nov)
- Fortnox gives notice on 200 of ~700 jobs while investing SEK 5bn in AI over five years (16 Nov)
Income is moving from pay to owners
Early signsLabour'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, while output in the business sector grew 3.0% a year and hours worked 0.2%. So far the change has come through pay and hiring. In our view it reaches jobs later.
In the model's run. The model does not compute income shares. This point is followed in the data.
What would count against it. The share recovering outside a recession.
Dated predictions
- Labour's share of business outputThe share: 93.73 in Q3 2026Against our view at 94.63Threshold 92.74 or belowOpenby Q3 2027
- Taxes on wagesYearly change as a share of GDP: 0.00 points of GDP in Q3 2026Against our view at 0.08 points of GDPThreshold −0.14 points of GDP or belowOpenby Q4 2028
Also watched
- Productivity and pay+1.1 ptsChart only
- Wages' share of household income49.2%Chart only
- Spending and income+1.8%Chart only
Latest news on this point
- Fed Governor Barr (12 Nov): AI may be 'capital-biased'; labour's share hit a record low 52.8% in Q2 2026
- US labour share of business output 53.0% in Q3 2026 (BLS, 5 Nov), the lowest since the series began in 1947
- US jobless claims 189,000 (week to 31 Oct), ties lowest since 1969; Q3 productivity +3.6%, unit labour costs −0.7%
- NBER paper (Oct 2026): same AI path gives 11.7% annual growth if saving responds, 2.2% at a fixed 23% saving rate
Capital, power and sites are scarce, so borrowing stays dear
Under wayThe build-out of AI computing costs more than its builders earn, and they are borrowing to pay for it while long-term interest rates are at their highest in more than fifteen years. The ten-year US Treasury yield reached 5.31% on 5 Oct 2026, its highest since May 2002, and the yield on inflation-protected ten-year Treasuries 2.95% the same day, its highest since Nov 2008. Grid connections, sites and permits limit what can be built, so the money raises the price of what is scarce.
In the model's run. The model has no long-term interest rate and no market for power or sites. 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. So the Federal Reserve raises its rate a little further, 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. If inflation rose again instead of fading, the rate would go much higher: 7.2% with a second wave of three points.
What would count against it. Long-term interest rates falling back while the build-out speeds up, or its weaker borrowers funding themselves freely again.
How the build-out is paid for
- 1The builders spend more than they earnFollowed in the news
- 2They borrow as money gets dearerOn
What we watch
- Ten-year yields in six markets5.13%Chart only
- The premium on US government debtContext · set 25 Sep 2026, already triggered0.99 ptsNormal
- Expected interest ratesContext · three rules set 24 Sep 2026Normal
- US mortgage ratesContext · set 25 Sep 2026, already triggered7.28%Close to triggering
- Oil pricesContext · set 25 Sep 2026, already triggered$93Normal
- Euro-area borrowing costsContext · set 8 Oct 2026, close to triggeringFrance, BelgiumTriggered
- France's borrowing costsContext · five rules set 24 Sep 2026; that day, one was already triggered and one close to triggering157 basis pointsTriggered
Latest news on this point
- Trump (20 Nov) threatens a 'tremendous tax' or a ban on US firms building data centres abroad; names Google's Finland plan
- Anthropic closes its IPO (20 Nov): $91.6bn to the company plus $26.0bn from AMD, Nvidia, TSMC and other suppliers
- US ten-year inflation-protected bond auction draws a 2.903% real yield (19 Nov), highest at auction since July 2002
- US Senate passes permitting bill 62-33 (19 Nov) that would make data centres pay for grid upgrades and power they use
Next: the downturn 2027–29
Dear money and a retreat from AI's weaker borrowers bring a downturn
Not yet seenWe think the squeeze ends as lenders and investors retreat from the firms that depend on outside money to build AI computing, and that share prices lose about half as much as in the dot-com crash, over two and a half years from their high of 6 Oct 2026. Households that own shares spend less, credit gets dearer, and unemployment rises about as far as it did after the dot-com crash.
In the model's run. 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; the Sahm rule, a rise of half a point in its three-month average, triggers in July 2027. The Federal Reserve cuts from the third quarter of 2027, to 2.7% by the end of 2028. The four economies' output is 2.3% below what they could produce at the worst of the downturn, in early 2029.
What would count against it. No rise in unemployment by the end of 2027, or a fall in shares and credit that reverses while the build-out is paid for out of its builders' own cash.
Dated predictions
- Technology stocksThe Nasdaq's fall from its high: −13.5% on 19 Nov 2026Started at −8.1%Threshold −20.0% or belowOpenby 30 Jun 2027
- US mortgage arrearsYearly change in the arrears rate: 0.08 points in Q2 2026Started at 0.08 pointsThreshold 0.40 points or aboveOpenby Q3 2027
- US unemploymentRise above its low of the past year: 0.00 points in Oct 2026Started at 0.00 pointsThreshold 0.50 points or aboveOpenby Sep 2027
- Financial stressThe index: −0.87 on 13 Nov 2026Started at −0.49Threshold 1.00 or aboveOpenby 24 Sep 2027
- Bank lending to non-bank lenders13-week change: 4.8% on 11 Nov 2026Started at 4.7%Threshold −3.0% or belowOpenby 29 Sep 2027
- Corporate creditThree-month change in the extra yield: −0.08 points on 19 Nov 2026Started at −0.13 pointsThreshold 0.56 points or aboveOpenby 30 Sep 2027
- Swedish unemploymentRise above its low of the past year: 0.10 points in Oct 2026Started at 0.27 pointsThreshold 0.60 points or aboveOpenby Dec 2027
- US jobless claimsRise above their low of the past year: 1.3% on 14 Nov 2026Started at 0.0%Threshold 25.0% or aboveOpenby 25 Dec 2027
- Korea's exportsYearly change: 84.7% in Sep 2026Started at 67.8%Threshold zero or belowOpenby Jun 2028
How a retreat would travel
- 3Investors pull back from firms that live on outside moneyOn
- 4Lenders follow and become selectiveNot yet
- 5Investment, wealth, spending and jobsNot yet
Two of the four steps that public data can show are on: steps 2 and 3. The five steps and their measures
Also watched
Latest news on this point
- S&P 500 −3.7% on 20 Nov to 6,910, over 10% off its 6 Oct record; Nasdaq −10% on the week, about 18% off its high
- Nasdaq falls about 4.6% on 20 Nov, about 17.5% off its high; Nvidia −12% on a reported SEC probe
- Nasdaq −2.25% to 23,879 on 19 Nov despite Anthropic's debut, 13.5% off its high; Kospi falls ~5% on 20 Nov
- No new high-yield AI bond deals in a month vs ~$60bn expected this autumn (Goldman, 17 Nov); CoreWeave bonds near 15%
The jobs cut in it do not all come back
Not yet seenIn past downturns most of the jobs cut were filled again when demand returned. We think that this time firms automate part of the office work they cut. The public response comes late, because it waits for unemployment to rise.
In the model's run. Each quarter, firms automate 5% of the jobs cut in the downturn and do not refill them. Governments add stimulus once unemployment has risen two points, which in the United States is April 2029. 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.
What would count against it. Office employment back at its level before the downturn within two years of its low.
Dated predictions
- US hiringYearly change in continuing claims: −9.5% on 7 Nov 2026Started at −12.8%Threshold 10.0% or aboveOpenby 25 Sep 2027
- Permanent job lossesYearly change in their share of the unemployed: −1.9 points in Oct 2026Started at −1.9 pointsThreshold 5.0 points or aboveOpenby Mar 2028
Also watched
Latest news on this point
Later: cheap machine work 2029–32
Afterwards AI replaces jobs faster, and the response decides the outcome
Early signsAfter the downturn we expect AI to replace jobs faster, in office work first. Robots doing physical work at scale would add to it; we do not expect them within this period. By then interest rates are low, and what governments do decides how far the damage goes: support incomes and tax the rents that scarcity creates, or cut spending and raise tariffs.
In the model's run. By mid-2032, 8.6% of the US workforce is out of work and the share is still rising: 7.4% unemployed and 1.1% who have left the workforce. The Federal Reserve's rate is 1.8%. If AI replaced jobs no faster than it has so far, 5.8% would be out of work; if the pace doubled every year, 10.1%. With spending cuts, tariffs and no bank rescues, 17.7%, with output 17% below what the economies could produce. With an income guarantee, 7.9%.
What would count against it. A pace of job losses to AI that is still at today's upper limit in 2029.
What we watch
- Factory and warehouse jobsContext · set 24 Sep 2026+0.6%Normal
- The US tariff rate5.54%Chart only
- Long-run inflation expectationsContext · set 24 Sep 20262.12%Close to triggering
Latest news on this point
- Fed Chair Warsh (16 Nov) sets out three AI scenarios, including job losses above full employment and a bust
- Riksbank raises its policy rate 0.25 points to 2.00% (4 Nov), its first hike in three years; more likely in 2027
- Democrats win the US Senate (AP, 4 Nov) with at least 54 seats and 7 flips, taking both chambers of Congress
- US midterms (3 Nov): Democrats projected to win the House; Texas elects Hinojosa on a data-centre moratorium pledge
The model's run of this view
The model starts from the economies as measured to the end of September 2026 and runs month by month to June 2032. It shows what follows from the view's settings. It is not a forecast. Run the model yourself
- Early 2027
- The Federal Reserve's rate reaches its high, 4.1%.
- By 30 Jun 2027
- The Nasdaq is 20% or more below its high of the past year.
- July 2027
- US unemployment has risen half a point above its low of the past year (the Sahm rule).
- The end of 2027
- US unemployment is 5.2%, from 4.2% in September 2026.
- The end of 2028
- US unemployment is 5.9%; the Federal Reserve's rate is down to 2.7%.
- Early 2029
- US unemployment reaches 6.2%, its high in the downturn; the four economies' output is 2.3% below what they could produce.
- April 2029
- The US government's stimulus starts, once unemployment has risen two points.
- The third quarter of 2030
- US unemployment is back to 4.7%; 5.3% of the workforce is out of work, counting those who have left it.
- Mid-2032
- 8.6% of the US workforce is out of work, and the share is still rising; the Federal Reserve's rate is 1.8%.
Before the model starts: the unemployment rate as measured. People who have left the workforce are not counted in the unemployment rate.
Before the model starts: the federal funds rate, and core prices of consumer spending over a year.
The runs around it
Each changes one part of the view and keeps the rest. None is a forecast.
Out of work counts the unemployed and those who have left the workforce after losing a job to AI.
Out of work counts the unemployed and those who have left the workforce after losing a job to AI.
Out of work counts the unemployed and those who have left the workforce after losing a job to AI.
| The run | US unemployment, end of 2028 | US out of work, mid-2032 | Four economies' output at its worst | Outcome |
|---|---|---|---|---|
| Our view | 5.9% | 8.6%, rising | −4.0% mid-2032 | A deep recession |
| No fall in shares: the build-out's funding holds | 4.9% | 6.2%, rising | −1.8% the end of 2030 | A recession |
| A fall in shares the size of the dot-com crash | 5.0% | 9.6%, rising | −5.4% mid-2032 | A deep recession |
| A fall half as large again as the dot-com crash | 5.8% | 10.8% | −7.0% the end of 2031 | A deep recession |
| AI replaces no jobs outside the downturn | 5.5% | 4.3%, rising | −2.4% the third quarter of 2029 | A recession |
| AI replaces jobs no faster than it has so far | 5.7% | 5.8%, rising | −2.6% mid-2032 | A recession |
| The pace doubles every year | 6.1% | 10.1%, rising | −4.9% mid-2032 | A deep recession |
| Firms automate twice as many of the jobs they cut | 6.0% | 9.3%, rising | −4.4% mid-2032 | A deep recession |
| Robots do physical work from late 2029 | 5.9% | 9.5%, rising | −3.7% early 2032 | A recession |
| An income guarantee | 5.5% | 7.9%, rising | −3.3% mid-2032 | A recession |
| More care jobs, paid for by borrowing | 4.0% | 7.5%, rising | −3.3% mid-2032 | A recession |
| Spending cuts, tariffs and no bank rescues | 7.6% | 17.7%, rising | −16.9% mid-2032 | A labour depression |
| The better case: funding holds and the pace stays as it is | 4.7% | 5.7%, rising | −1.8% early 2032 | A slowdown |
| The worse case: a dot-com-sized fall, with spending cuts and tariffs | 8.7% | 21.4%, rising | −21.2% mid-2032 | A depression |
Output is measured against what the four economies could produce, a quarter's average. The outcome names the whole run, to mid-2032, by its worst point. Each run opens in the model.
The other economies in the run, and why we do not rest on them
| Out of work, end of 2027 | End of 2028 | Mid-2032 | Output at its worst | |
|---|---|---|---|---|
| United States | 5.3% | 6.1% | 8.6% | −3.8% |
| Euro area outside France | 6.6% | 7.6% | 9.4% | −3.7% |
| France | 9.0% | 10.3% | 14.2% | −5.0% |
| Sweden | 9.5% | 10.5% | 15.7% | −7.7% |
| China | 5.3% | 5.5% | 6.4% | −4.2% |
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, where inflation starts below target. A falling krona cushioned Sweden in 1931 and in 2008–09.
If inflation does not fade
In the 1970s unemployment and inflation rose together. Our view has inflation fading as unemployment rises. This is the main way the squeeze could go further than we think.
We expect inflation to fade for three reasons. Pay is not running ahead of prices: labour's share of income is falling. Over the year to Q3 2026, output per hour rose 2.1% and pay per hour after inflation rose 1.0%. The downturn we expect comes from lost spending, which lowers inflation. And scarce power, sites and capital raise some prices against others, which need not raise prices in general. Inflation expected five to ten years ahead was 2.1% on 20 Nov 2026, inside the range of 1.5–3% it has kept to.
It is still an assumption. Core inflation was 3.1% over the year to Sep 2026. The run of our view takes it to fade at the pace it faded in 2024 and 2025. If it stays where it is, the same run leaves 9.1% of the US workforce out of work by mid-2032, against 8.6%. If it rises again by three points through 2028, the Federal Reserve's rate reaches 7.2% and 9.6% are out of work. If people come to expect the rise to last, as they did in the 1970s, the rate reaches 19% and 12.3% are out of work.
What would show it. Core inflation still near 3% in mid-2027 while unemployment rises. Inflation expected five to ten years ahead leaving the range of 1.5–3%. Pay per hour, after inflation, rising faster than output per hour, which would mean labour's share recovering.
The model counts no harm from inflation itself, only from what central banks do about it. These runs show what fighting a second wave could cost. The cost of letting one run is outside the model.
| The run | The Federal Reserve's rate at its highest | US core inflation, end of 2028 | US out of work, mid-2032 | Outcome |
|---|---|---|---|---|
| Our view: inflation fades | 4.1% early 2027 | 2.1% | 8.6%, rising | A deep recession |
| Inflation stays where it is | 4.2% mid-2027 | 2.6% | 9.1%, rising | A deep recession |
| Inflation rises again: a second wave | 7.2% early 2029 | 5.6% | 9.6%, rising | A deep recession |
| A second wave, and people half expect it to last | 10.3% mid-2029 | 7.1% | 10.2%, rising | A deep recession |
| A second wave, and people expect it to last | 19.4% the third quarter of 2029 | 11.0% | 12.3%, rising | A deep recession |
The second wave is assumed: three percentage points more inflation by the start of 2028, in every economy, held through 2028 and gone by the start of 2030. Each run opens in the model, where inflation and what people expect of it can be set.
Market context
Market prices move for many reasons. These are shown for context and never count as evidence for our view.
- Government borrowing costs in a US downturnContext · two rules set 8 Oct 2026; that day, one was close to triggering+16 basis pointsNormal
- Sudden moves in bond yieldsContext · twelve rules set 24 Sep 2026; that day, two were already triggered and three close to triggeringTriggered
- US house pricesContext · set 25 Sep 2026+2.1%Normal
- The yenContext · set 25 Sep 2026¥160Normal
- Japan's bond yields2.78%Chart only
What the view does not claim
A date for the long run. The research behind this site shows where wages and rents tend as machines take over work; it puts no years on it, and the view stops in mid-2032.
Numbers for Europe. The view is stated for the United States, and no point rests on the model's figures for other economies.
A call on which way US government bonds move in the downturn. In most downturns their yields fall; with large deficits and a build-out competing for savings they might not. We show both and count neither.
How sure we are
This is our view, and it could be wrong. The loss of jobs to AI measured so far is small and could stay small: at the pace seen so far, the same run has 5.8% of the US workforce out of work by mid-2032, and with no fall in shares as well, 5.7%, a slowdown without a recession. How fast the pace grows is assumed, not measured. The model, fitted to past downturns, may not fit this one; with no fall in shares and no jobs lost to AI its unemployment rate still drifts from 4.2% to 4.7% by mid-2032, so its later years lean high. The settings and the model are public: the code for this run.
Behind the headline numbers: twelve measures fixed in advance