AI, Work and the Economy

Latest/Assessment

The assessment

Where things appear to stand, what the coming months may bring, and where the longer run may lead, judged from every item flagged on this site.

Written by Claude, an AI model, on 15 Nov 2026, 19:14 Stockholm time, from the 331 items flagged since 24 Sep 2026, the recent ones weighted most. It may depart from our own view. The numbers for the longer run come from runs of the model.

What changed. Written afresh for the week: the calls stand. Power scarcity is better documented (a reported $3.4bn purchase of generator orders by an AI developer), and the week's tests are now dated: Nvidia's results on 17 November and Anthropic's share sale on 18–19 November.

Now

The squeeze appears under way through rising capability and dearer capital and power, while the labour market has not yet turned.

Measured capability rose sharply this month: Anthropic's newest model completes 41.3% of the Remote Labor Index against 17.9% for its predecessor, and comparable capability keeps getting cheaper. The build-out is increasingly financed by debt (the Federal Reserve counts more than $680bn of AI-related bond issuance by October) and is meeting limits on power: Norway's grid operator stopped booking data-centre connections, and an AI developer reportedly paid $3.4bn to take over another firm's generator orders. The ten-year Treasury yield is 5.15%. Shares in firms that depend on outside money have fallen hard, Oracle about half from its high. Yet US unemployment is 4.1% and jobless claims are 200,000: a loss of jobs, if it is coming, does not show yet.

Our view, point by point

The coming months

A downturn in 2027 is possible on this evidence but not yet signalled; dear money is the more likely course through the winter.

With October consumer prices up 0.4% and a December Fed rise priced, money is likely to stay dear for some months. A downturn would most plausibly travel through lenders: equity has retreated from the weakest AI borrowers, but credit spreads and bank lending standards remain calm (high-yield spreads 3.20%, banks tightening a net 2%). If Nvidia's outlook on 17 November or Anthropic's share sale disappoints and lenders then follow equity, a correction of the size of 2001's could take hold in 2027, with US unemployment rising toward 5% by year end. If revenue holds and lenders stay open, the squeeze could run longer with little loss of jobs. The payrolls of 4 December and the Fed's December decision come next.

The longer run

Toward 2030-32, a larger share of people out of work than today looks likely, with inflation that fades more slowly than the case assumes.

The capability measured this autumn is consistent with the pace of job loss doubling every two years or so; nothing yet shows it faster. Inflation has not faded as quickly as the case's run assumes (core consumption prices 3.1% in September, consumer prices +0.4% in October), so the closest run is one in which it stays where it is and central banks have less room to cut when jobs are lost. The severe outcome would take a bust of the full dot-com size meeting that inflation: rates would stay high while unemployment rose, and the share of people out of work could climb well beyond the case's run before any response arrived.

Most likely, on this assessment

A deep recession

The four economies' output falls 4% or more below what they could produce, about as far as in 2008–09.

The model's run of our view, except: Inflation: Stays where it is.

5.9%
US unemployment at the end of 2028
9.3%
of the US workforce out of work when the run ends, in June 2032, and still rising
−4.7%
the four economies' output at its worst, against what they could produce

Open this run in the model

A plausible worst case

A deep recession

The four economies' output falls 4% or more below what they could produce, about as far as in 2008–09.

The model's run of our view, except: The fall in AI shares and financing: Dot-com size; Inflation: Stays where it is.

5.0%
US unemployment at the end of 2028
9.9%
of the US workforce out of work when the run ends, in June 2032, and still rising
−5.6%
the four economies' output at its worst, against what they could produce

Open this run in the model

Americans out of work in the two runs, % of the workforce

Out of work counts the unemployed and those who have left the workforce after losing their job. Each run shows what follows in the model from its settings. Neither is a forecast.

What would count against it

No measure of AI's effect on jobs has yet crossed its threshold: office jobs trail all jobs by 0.50 points against a threshold of 2.41, private jobs outside care rose 424,000 in a year, and claims are near 50-year lows. A Fed vice chair finds AI's productivity effect still limited, a fifth of US workers replaced by AI say they were later asked back, and expected inflation five to ten years ahead has fallen to 2.18%.

What it weighed most

Each story's share of the weight of all flagged items. An item counts fully on the day it is flagged and half as much 14 days later.

How it has changed

  • (current)Written afresh for the week: the calls stand. Power scarcity is better documented (a reported $3.4bn purchase of generator orders by an AI developer), and the week's tests are now dated: Nvidia's results on 17 November and Anthropic's share sale on 18–19 November.Now. The squeeze appears under way through rising capability and dearer capital and power, while the labour market has not yet turned.
  • First assessment.Now. The squeeze appears under way through capability and the cost of capital, while the labour market has not yet turned.