Automation Labor Displacement: A Historical Timeline From Switchboards To AI

Between 1920 and 1940, AT&T mechanised more than half of the US telephone network. Switchboards did not empty out in a single year, and operators were not all dismissed at once. They were absorbed into lower-paid clerical work, pushed out of the labour force, or simply never hired in the first place.

That twenty-year slog, not a cliff event, is the most useful historical model anyone has for AI displacement today. Get the duration and the shape right, and the investable question stops being “who gets replaced” and starts being “what does the displaced cohort vote for”.

Key facts at a glance

• Mechanisation of US telephone exchanges ran 1920 to 1940; by 1940 roughly 60% of AT&T exchanges had been converted to dial.
• Incumbent operators a decade after their city’s conversion were “slightly more likely to be in lower-paying occupations” or out of the labour force entirely.
• Aggregate employment was restored through growth in middle-skill clerical and lower-skill service work, with no measurable long-run unemployment effect.
• A 2025 Stanford study, surfaced in a 2026 policy synthesis, found a 13% relative decline in early-career employment in AI-exposed jobs versus controls.
• The political bill landed in the 1930s with the New Deal, federal wage boards, and the slow-burn antitrust campaign against AT&T (T).

How automation displacement actually works

The mental model most people carry is the cliff: a technology arrives, jobs vanish in a quarter, and the labour market clears the rubble in a year. That has almost never been the case, and the cleanest counter-example sits in the early twentieth-century US telephone system.

The canonical study by James Feigenbaum and Daniel Gross in the Quarterly Journal of Economics traces the displacement of switchboard operators by automatic dial exchanges across two decades and finds something more useful than “they all lost their jobs”. Aggregate employment recovered. The cohort that was working when mechanisation hit did not.

The pattern has three layers. First, the headline number, total employment in the affected industry, looks fine after a transition window. Second, the firm-level numbers, who AT&T actually employed at which wages, are also broadly fine because new clerical and service roles absorb the workers who used to plug cables.

Third, the cohort-level numbers, the same individual operators tracked through subsequent censuses by the NBER working paper that preceded the QJE version, show a measurable downgrade in occupational rank and earnings that persists for at least a decade. The aggregate hides the scar.

That structural mismatch, a healthy aggregate sitting on top of an injured cohort, is what we mean by an automation labor displacement historical timeline rather than an automation labour shock. A timeline can be priced. A shock cannot.

A worked example: the dial-phone decade

The mechanisation Feigenbaum and Gross measure was not subtle. The Cato Institute summary of the same body of work notes that telephone operation was, by some margin, the largest single occupation for young women in the early 1920s. The Federal Reserve’s Econ Focus history adds that AT&T’s labour cost calculus made the conversion inevitable once technical feasibility was demonstrated.

The popular telling on History.com compresses the human side: tens of thousands of women whose first job was the switchboard, displaced into a labour market that was not equipped to re-hire them at the same rank.

What did not happen is also worth saying. Operators did not march on AT&T. The Broadstreet economic history note and the Conversable Economist summary both stress that the labour market quietly reabsorbed displaced operators into adjacent clerical work over years, not weeks.

The companion HBS working paper from Feigenbaum and Gross sharpens this for young entrants: cohorts that aged into the workforce just after the conversion were the ones who carried the scar, because they were never hired into the high-skill operator role to begin with.

Set against today’s early evidence, the shape rhymes. The Carnegie Endowment 2026 synthesis of the AI labour debate cites Brynjolfsson and colleagues at Stanford finding a 13% relative employment decline for early-career workers in AI-exposed jobs, while incumbents and senior cohorts are absorbed sideways rather than fired outright. That is the dial-phone playbook, not a labour cliff.

Variable	Telephone operators (1920-1940)	AI-exposed entrants (2023 onward)
Span of displacement	~20 years	Early data, 2-3 years observed
Aggregate employment effect	Restored by 1940	Restored at occupation level so far
Cohort wage scar	Persistent for ~10 years	13% relative employment decline (early-career)
Mechanism of absorption	Clerical and service growth	Routine cognitive and relational roles
Political tail	New Deal, wage boards, antitrust	Open question

The fresh angle: political durability of margin

The lesson most articles draw from the operator analogy is incomplete. The interesting investable risk is not unemployment, because aggregate employment did recover. The interesting risk is that a wage-suppressed cohort is also a voting cohort.

The same generation that absorbed the dial-phone displacement was the political base for the New Deal, for the Wagner Act, and for the long-running antitrust campaign against AT&T itself. Margin captured at the top of a displacement curve is durable only if the political system lets it stay there.

That reframes the AI windfall trade. The current crop of AI-rentier firms, Microsoft (MSFT), Nvidia (NVDA), Alphabet (GOOGL), Meta (META) and Amazon (AMZN), are pricing a perpetual margin take from a productivity transition whose social cost lands on a specific cohort.

AT&T once played a structurally similar role, and it spent the back half of the twentieth century paying for the political reaction. A twenty-year political risk premium, not a one-quarter unemployment headline, is the right framing.

The steelman against this view is real and worth engaging. David Autor’s NBER working paper 32140 argues that AI used as a decision-support layer can push the middle wage up, by extending expert-tier work to non-credentialed workers, rather than down. The point lands.

If the next generation of AI tools genuinely raises the bottom of the middle, the operator cohort scar does not repeat at scale and the political tail is smaller.

A second steelman comes from the Dwarkesh Podcast conversation with Alex Imas and Phil Trammell, recorded June 4, 2026.

Their argument, summarised on the show’s YouTube release of “The better AI gets, the smaller its share”, is that AI’s share of GDP may shrink as capability rises, with value migrating toward a relational sector that absorbs displaced workers at rising wages because human-in-the-loop is the scarce input. If that holds, the operator analogy overstates the cohort scar and understates the absorption.

Common misconceptions

The “this time is different” framing, in either direction, misses the timing question. The operator displacement was neither a cliff nor a smooth absorption. It was a slog with a tail. So is the AI version, almost certainly. Confusing the shape of the curve for its direction produces the worst kind of forecast: confident, popular, and unhedged.

The “X percent of jobs by 2030” headlines that anchor most AI labour coverage are the second misconception. Feigenbaum and Gross do not produce a percent-of-jobs figure for telephone operation either; they produce a cohort and a decade. Investors pricing AI displacement off a single aggregate percentage are pricing the wrong variable.

What this does not tell you

The operator analogy is one historical case, and the canonical study covers a single industry inside a single national labour market with stronger geographic frictions than 2026 has. The pace of AI adoption could be faster, the absorption mechanisms weaker or stronger, and the political reaction shaped by very different institutions.

The 13% Stanford figure quoted via Carnegie is an early-cycle estimate and may not survive replication. The political-tail thesis is a frame for risk, not a trade; nothing here is a buy or sell call on any of the named companies.

FAQ

How long did the telephone-operator displacement actually take?

Between 1920 and 1940 as the Feigenbaum and Gross dataset is constructed, with full conversion of AT&T exchanges continuing into the 1980s. Most of the cohort effect they measure sits inside the 1920 to 1940 window.

Did total employment really recover?

At the aggregate level, yes. Middle-skill clerical work and lower-skill service work grew enough to absorb the displaced operators in headline terms. The cohort that was actually working as operators in 1920 ended up downgraded, but the labour market as a whole did not lose jobs.

Is the 13 percent AI employment number reliable?

It is an early estimate from a Stanford team led by Erik Brynjolfsson, surfaced through the Carnegie Endowment 2026 synthesis. Treat it as suggestive of direction and shape, not as a settled point estimate.

What is the strongest argument against the operator analogy?

David Autor’s case that AI is more of a decision-support layer than a substitute for cognitive labour, which would push the middle wage up rather than down. If that view holds, the cohort scar shrinks materially and the political tail with it.