What we think is happening

1. It may have begun two to four years ago

AI appears already to have taken some work outright, in translation, writing and some entry-level jobs, and some more quietly: existing staff do more, hiring slows, and people who leave are not replaced. Health care and social assistance seem to have taken up the slack: from Aug 2024 to Aug 2026 they added 107% of the net growth in US jobs, while the rest of the economy shed jobs. Measured on its trend since 2024, the loss of jobs to AI is still small: at most 15% of the pace the model assumes from here.

What would show it:

• Labour's share of income93.4Triggered
• People leaving the workforce−0.33 ptsTriggered
• Office jobs−0.0%Normal
• Jobs outside care−24kTriggered

What would count against it. Labour's share of income recovering, and employment among 25–54-year-olds holding up.

2. It could speed up as AI companies change how they sell

A few firms that can do most things may absorb the smaller ones, and begin to sell the replacement of work rather than help with it.

In the model. The loss of jobs to AI rises from the pace seen so far toward the model's full pace, half-way there by Q2 2028. This is still slower than the model's central case, which assumes the full pace from the start: on its own it lowers the share of Americans out of work in mid-2032 by 3 points.

What would show it:

• Office jobs−0.0%Normal
• Permanent job losses25%Normal
• US hiring3.3%Normal

What would count against it. Adoption stalling on cost or reliability, and firms hiring again as they grow.

3. Government debt could become the central worry

If jobs go, taxes on wages and spending fall while the cost of support rises, and interest costs climb as old debt is refinanced at higher rates. Markets may charge for that risk. For the US the danger is more likely to be inflation and pressure on the central bank than default; in the euro area, the gap between countries' borrowing costs, France's above all.

In the model. Investors charge a point more to lend to the US and half a point more in the euro area, reached over the first year and passed on to firms and households. By mid-2032 this adds 5 points to the share of Americans out of work. The model has no tax revenue and no default: it prices the worry, it does not compute the deficit.

What would show it:

• The premium on US government debt1.08Triggered
• France's borrowing costs128 basis pointsTriggered
• Sudden moves in bond yieldsClose to triggering
• Long-run inflation expectations2.35%Normal

What would count against it. US government bonds rallying in a slowdown, as they usually do.

4. A stock market bust, possibly larger and faster than the dot-com crash

The dot-com crash took about two and a half years; the S&P 500 fell by about half. Much of the money for the AI build-out is borrowed, a good part of it through private credit funds that borrow from banks, so a fall in AI investment could reach lenders quickly.

In the model. From Q4 2026, US share prices fall 60% and AI investment 90% within 2 quarters, and losses on lending to the build-out reach the banks. By the end of 2028 this is the largest single part of the reading: it adds 3 points to US out of work.

What would show it:

• Technology stocks+0%Normal
• Corporate credit3.0%Normal
• Financial stress−0.47Normal
• Bank lending to non-bank lenders$2.07tnNormal
• Korea's exports+71%Normal

What would count against it. AI investment plans raised and paid for out of profits rather than borrowing.

5. A recession, as spending falls with wages

Lost wages cut spending, falling share prices cut the spending of those who own them, and lower spending cuts more jobs. If firms replace the jobs they cut with AI instead of rehiring, the losses may not come back.

In the model. How many of those jobs firms replace is the assumption that matters most after 2028. If twice the usual share, it adds 15 points to US out of work by mid-2032.

What would show it:

• Market bust and job losses togetherjobs 1 · markets 0Normal
• US unemployment4.2%Normal
• US jobless claims200kNormal
• Permanent job losses25%Normal

What would count against it. Spending holding up on wealth and government support, and laid-off workers rehired.

6. Care may stop absorbing the jobs

Health care and social assistance have carried US job growth and are mostly paid for by the state. As budgets tighten, they could stop growing.

In the model. Care stops growing a year from the start in all four economies. This matters less than it might appear: at its normal pace care takes in about a point of US employment in six years, so capping it adds 0.5 points to US out of work by mid-2032. A funded expansion of care is what would make a difference.

What would show it:

• Jobs outside care−24kTriggered

What would count against it. Care jobs still growing through a downturn.

7. Tariffs in the US, spending cuts in Europe

We think it likely that the US raises tariffs and other countries answer, and that Europe and Sweden avoid tariffs of their own but cut spending as unemployment rises.

In the model. US tariffs rise 25 points and US exports meet the same; other economies face them on what they sell to the US. Tariffs add 4 points to US out of work by mid-2032. Spending cuts add 0.9 points to out of work in the euro area and 1.7 in Sweden. If Germany's spending offset the cuts elsewhere, the euro area would have 0.5 points fewer out of work.

What would count against it. Germany's spending and joint EU borrowing carrying Europe through the downturn.