The standard reassurance about AI and employment goes like this: every previous wave of automation destroyed some jobs and created others, net employment is fine, and AI will be the same. This argument is not wrong. It is, however, being applied with a selective reading of the historical evidence that strips out the parts of the story that are most relevant to present circumstances.

Let’s do the history more carefully.

The mechanization of British textile production between roughly 1760 and 1850 is the ur-example of automation displacing workers at large scale. The handloom weavers of Lancashire — an occupation numbering around 250,000 in 1820 — experienced a 60-year decline in wages followed by near-complete occupational elimination. Net British employment was ultimately higher at the end of the process than the beginning, and living standards eventually rose. The “eventually” covered roughly 60-80 years of genuine hardship for the workers in the displaced occupation and their communities.

The agricultural mechanization of the early 20th century in the United States tells a similar story at larger scale. Farm employment as a share of the workforce fell from roughly 41% in 1900 to 3% by 1970, an absolute reduction of millions of jobs. The workers who left farming mostly found industrial employment that paid better, in cities that offered more economic opportunity. The transition was managed — imperfectly but substantially — by the industrial expansion that pulled workers from farms. The timing worked: the demand for industrial labor grew approximately as fast as agricultural mechanization displaced farm labor, which meant the transition, while painful for many individuals, didn’t produce sustained high unemployment at the macroeconomic level.

The office automation wave of the 1980s-1990s is the most directly analogous precedent to the current AI moment, because it affected similar workers (white-collar, educated, doing information processing and communication tasks) and involved similar technology claims (computers will transform knowledge work, productivity will surge). The displacement was real: typing pools, data entry departments, and certain categories of administrative work were eliminated. The creation was also real: the personal computer industry, the IT services industry, software development, and the expanding service economy of the 1990s absorbed most of the displaced workers, typically in jobs that were (eventually) better-paid and more cognitively engaging.

But here is what the optimistic retelling consistently omits: the office automation wave of the 1980s-1990s happened inside a specific economic context that was unusually favorable for absorption of displaced workers. The US was running significant current account deficits, which supported domestic consumption; the Cold War defense spending created substantial public sector employment; the IT sector itself was creating entirely new categories of jobs at a rate that had no historical precedent; and the expansion of global trade, while displacing manufacturing workers, also expanded the US service export sector. The net positive employment outcome was not driven purely by the internal logic of automation creating jobs to replace the ones it destroys. It was driven by a specific conjunction of expansionary economic conditions that are not obviously replicable.

The steel industry automation of the 1970s-1980s is the case where the optimistic story doesn’t work at all. Steelworkers in Pittsburgh, Youngstown, and the Mon Valley were displaced by a combination of automation and trade competition, and the jobs that replaced them — retail, service, healthcare — paid substantially less and offered substantially worse benefits. The communities built around steel employment didn’t recover their pre-displacement economic status. Youngstown, Ohio’s population in 2026 is roughly 60,000 — down from 170,000 in 1950. The aggregate job count in the US is higher now than in 1970; the specific communities that bore the cost of steel automation are not.

This is the pattern that makes the “automation creates jobs” reassurance incomplete when applied to specific workers in specific places: the jobs created by automation are typically not created in the same places, in the same industries, or for the same workers as the jobs destroyed. The agricultural worker who left Iowa for Detroit found industrial employment that worked for her. The steel worker who lost his Pittsburgh mill job in 1985 did not typically find equivalent work in the information technology sector. The displaced worker and the net economic beneficiary of automation have frequently not been the same person.

For the current AI moment, the question is not whether net employment will eventually be higher — historical precedent suggests it probably will, over a long enough horizon. The relevant questions are more specific: how long is the lag between job destruction and job creation? Which workers and which communities absorb the transition cost? Are the jobs created genuinely equivalent in compensation and quality to the jobs destroyed? And are the economic conditions of the current moment — growth rates, public investment, new sector expansion — favorable to the kind of labor absorption that made previous transitions manageable?

On the lag question, the first-year data suggests we’re in the displacement phase without yet being clearly in the creation phase. New AI-native roles are growing but not at the pace required to offset even the first-wave displacement, let alone the projected second wave. Economic historians have found that the lag between automation-driven job destruction and the new sector job creation typically runs 10-20 years in previous cycles. If that holds, the workers experiencing displacement now are being asked to absorb costs that will be offset by benefits realized in the 2035-2045 period — a timeline that is not useful to someone who needs to pay their mortgage in 2026.

On the geographic and demographic distribution question, the current AI disruption is following the same pattern as previous automation waves: the displacement costs are concentrated in specific places (mid-tier cities) and specific demographic groups (mid-career workers with some but not maximum educational credentials), while the productivity gains are distributed broadly through lower prices and higher corporate profits. This has been the standard distributional signature of technological disruption since the industrial revolution. Acknowledging it doesn’t require technological pessimism; it just requires honesty about who “we” is when we say “we’re going to be fine.”

The job creation promise of AI is likely to be real. The historical evidence is reasonably clear that technological automation, over sufficient time, creates more work than it eliminates by expanding the total economic pie. What the evidence is equally clear about is that the timeline of “sufficient time” is measured in decades, that the geographic and demographic distribution of costs and benefits is deeply unequal, and that the favorable outcomes associated with previous transitions were not automatic — they required active public investment, institutional adaptation, and economic conditions that don’t maintain themselves without effort.

Repeating the reassurance without the context is intellectually dishonest. The context is the more important part.

Get the best articles from Think Different

One email a week with practical Mac workflows, useful comparisons and honest tech recommendations.