We use analytics cookies to understand how readers use the site. They only run if you accept. Privacy policy.
In 1964 a presidential commission was convened to determine whether automation was destroying American jobs. Its members included Robert Solow, the economist who had just formalised how technology drives growth; Daniel Bell, who would go on to name post-industrial society; Walter Reuther of the United Auto Workers; and Thomas Watson Jr. of IBM. Two years later they reported that technological change had not been the cause of the unemployment they were asked about — weak aggregate demand had — and then recommended a guaranteed minimum income, government as employer of last resort, and two free years of post-secondary education anyway. Congress adopted none of it.
That is the shape of every automation panic since — and of the ones before. When Nottinghamshire textile workers began breaking stocking frames in 1811, the grievance was not the machine but the fact that it let unapprenticed labour undercut a trade whose wages were collapsing; Parliament made frame-breaking a capital offence the following year. The pattern repeats: a genuine technological shift, a forecast expressed as a percentage of all jobs, a policy response that does not arrive, and an economy that adds jobs regardless while particular regions and cohorts absorb the cost. The question now is whether generative AI breaks it — and the honest answer is that the first payroll-level evidence arrived in 2025, and it is narrow, real, and pointed squarely at people in their early twenties.
BLS Current Employment Statistics, annual averages. 1939 is the first year of the survey — there is no earlier national payroll count. 2026 is the January–June average. Output kept climbing after 1979; employment did not.
Manufacturing, construction and mining as a share of nonfarm payrolls, 1939 — the earliest year available — to June 2026. The line has been essentially flat since 2010: whatever automation is doing now, it is not still hollowing out goods production.
Manufacturing employment is the strongest evidence that automation displaces workers at scale — and the strongest evidence that the displacement does not show up as unemployment. Payroll manufacturing peaked at 19.55 million in June 1979 and stands at 12.60 million as of June 2026: a fall of 6.96 million jobs, or 35.6%. As a share of all nonfarm employment the decline is far steeper — from a wartime high of 38.8% in November 1943, to 21.7% at the 1979 peak, to 7.9% today.
How much of that was machines rather than trade is genuinely contested. A widely cited 2015 Ball State study attributed roughly 87% of 2000–2010 losses to productivity growth; Autor, Dorn and Hanson’s work on the China shock put up to 2.4 million jobs on import competition between 1999 and 2011, about a quarter of the contemporaneous decline; and Susan Houseman has argued that the measured productivity gain is itself largely an artefact of how computer and semiconductor output is deflated. What is not contested is the destination: the workers did not end up permanently unemployed nationally, and the communities that lost the plants largely did not recover.