Jobs, not work: what nine past inventions did to employment, and what carries over to AI

Glasp Research

Glasp Research

Oct 31, 2026

16 min read

Jobs, not work: what nine past inventions did to employment, and what carries over to AI

A Glasp Research memo. Nine technology episodes from the tractor to e-commerce, 28 dated job-loss statements, and an early read of AI against that record.

Summary

Every wave of mechanization has come with a warning that human work is about to run out. We went back to nine past inventions (farm tractors, factory electrification, the car, dial telephone switching, office computing, ATMs, shipping containers, industrial robots, and the internet) and asked four questions: what happened to the jobs most exposed to each one, where new work came from, how long the change took, and who paid for it. We then compared AI, as of October 2026, with that record.

  • In 8 of the 9 cases the most exposed occupation shrank or stopped growing, sometimes by 70% to 95%. No case shows total employment falling for 10 years or more (UK containerization before 1971 was not checked).

  • The replacement work was often not created by the same technology, and most measured declines took one to five decades.

  • Where the cost could be traced, it fell on specific people: southern sharecroppers, middle-aged dockers, incumbent telephone operators, and young people who were never hired.

  • AI resembles this history in adoption speed and in an early, descriptive pattern of fewer young hires in exposed jobs. It differs in breadth, in exposing higher-wage work, and possibly in being able to do the new tasks it creates. History cannot settle that last point.

Two of our pre-registered hypotheses did not hold as written; we report that below.

How we did it

The nine cases were fixed before any research, and the hypotheses were registered before outcomes were read. For each case we looked for a dated prediction of job loss, employment in the most exposed occupation over time, new work, and who bore the cost. We used two timing measures. T-lag is the number of years from the technology's introduction to the exposed occupation's employment peak. T-decline is the number of years from that peak to the year employment fell below half of it.

All sources are public (US Census, Bureau of Labor Statistics (BLS), FRED, USDA, UK Hansard, and published papers). Occupation definitions change over time, so comparisons across classification breaks are approximate.

Nine cases in brief

Farm mechanization

US farm employment peaked at 13.63 million in 1916 and fell to 4.52 million in 1970 (USDA). The tractor spread slowly (3.6% of farms in 1920, 80.8% in 1969) and was one cause among several (Olmstead and Rhode 2001). Olmstead and Rhode estimate the tractor saved about 1.72 million workers by 1960, 27.3% of the decline since 1910. No new occupation tied to the tractor absorbed the people who left; farm-machinery operatives numbered 67,000 in 1950 (Historical Statistics 1975). The cost fell on southern sharecroppers and Black farm workers. Black net migration out of the South was 4.59 million between 1940 and 1970 (Holley 2003, citing Census series).

Factory electrification

The electric share of machine-drive capacity in US manufacturing rose from 4.8% in 1899 to 86.1% in 1939 (computed from Devine 1982). This case has no clear losing occupation. Electrification pulled people into factories, and new trades grew with it: electricians went from 51,000 in 1900 to 332,000 in 1950 (Historical Statistics 1975).

The car and the horse

Teamsters peaked at 441,000 in 1910 and fell to 31,000 in 1940 (Historical Statistics 1975), the fastest collapse in this study. The same technology created more jobs than it removed, and quickly: truck, taxi and bus drivers and chauffeurs rose from about 248,000 in 1910 to 1,515,000 in 1940, and auto mechanics numbered 448,000 in 1940 (same source, approximate across classification changes). In 1898 a horseshoe-nail supplier said "the horse is the thing again" (American Farriers Journal 2024).

Telephone operators and dial switching

Bell opened its first fully automated office in 1921 (Richmond Fed 2019). Operator numbers kept rising for decades anyway, to 408,560 in 1970 (Census 1984), before falling to 119,000 in 2002 and 26,000 telephone operators in 2025 (BLS CPS). That is a T-lag of about 49 years. Locally the picture was sharper. After a city switched to dial, the number of women aged 16 to 25 working as operators "immediately, permanently fell by 50 to 80%" (Feigenbaum and Gross). Incumbent operators bore the cost: about 10% ended up in a lower-paying job a decade later (same source; figures from our research notes, not re-read on the primary page in this study). In 1964 one analyst suggested the decline of operators "may be nearing an end" (quoted by Richmond Fed 2019). It was not.

Office computing

Between 1983 and 2002, years that share occupation codes, typists fell from 906,000 to 571,000 and secretaries from 3.89 million to 2.30 million, while accountants and auditors rose from 1.11 million to 1.70 million and computer systems analysts and scientists from 276,000 to 1.74 million (BLS Employment and Earnings 1984; BLS CPS 2002). By 2025 typists and word processors were down to 62,000 and software developers numbered 2.25 million (BLS CPS 2025). Women carried most of the clerical loss: in 1983 they were 95.6% of typists and 99.0% of secretaries (Employment and Earnings 1984).

ATMs and bank tellers

ATMs went from 19,000 in 1980 to 415,000 in 2007 (Federal Reserve Bulletin 2009). Tellers did not fall for decades: 480,000 in 1983 and 453,000 in 2010 on the CPS basis. Bessen explains why: tellers per urban branch fell from 20 to 13 between 1988 and 2004, but the number of urban branches rose 43% (Bessen, IMF Finance and Development 2015). The decline came after 2010, to 224,000 in 2025 on the CPS basis, and BLS ties it to online banking and branch closures (BLS Occupational Outlook Handbook). A widely quoted figure of about 600,000 tellers in 2010 was not found in a primary table and conflicts with the CPS count.

Containerization and dock work

Registered UK dock workers fell from 71,846 in 1959 to 45,542 in 1970 and 9,319 in May 1989 (Hansard, 16 May 1989). In 1987, 6,408 of the registered dockers listed by age were over 45, and only 693 were aged 25 to 35 (same source). On the US West Coast a 1960 union agreement traded work rules for a $29 million fund, earnings guarantees and early retirement. A local union president predicted that year that "within two months ... the work of 800 men will be eliminated". Coastwide longshore hours rose from 23.8 million in 1960 to 26.7 million in 1966, then fell to 14.8 million in 1971 (Bradley 2016). The loss came, but six to eleven years later.

Industrial robots

US manufacturing employment peaked at 19.43 million in 1979 and was 12.63 million in 2025 (BLS CES via FRED, computed). Most of that fall is not robots; the 2000 to 2010 drop coincides with the China trade shock. Acemoglu and Restrepo estimate robots cost about 400,000 US jobs over 1993 to 2007 (JPE 2020). In Germany robots had no effect on total local employment; manufacturing losses were offset by business services and came through fewer hires of young entrants, not layoffs (Dauth et al. 2018).

The internet and e-commerce

E-commerce rose from 0.6% of US retail sales in late 1999 to 17.1% in Q2 2026 (Census via FRED). Department store employment fell from 1.755 million in 2001 to 0.932 million in 2025, while total retail employment was about the same at both ends (15.26 million in 2000, 15.44 million in 2025), and warehousing rose from 517,000 to 1,866,000 (BLS CES via FRED, computed). In 1995 Jeremy Rifkin forecast "the steady and inevitable decline of jobs". US payrolls rose from 117.4 million in 1995 to 158.4 million in 2025 (FRED, computed).

What the cases add up to

The broad pattern

Over the twentieth century the six US occupation groups that shrank went from 76% of employment in 1910 to 25% in 2000 (Wyatt and Hecker 2006). Autor, Chin, Salomons and Seegmiller estimate that "roughly 60%" of 2018 employment is in job titles added since 1940 (QJE 2024). Lin found that 4.4% to 8.5% of workers in 1980, 1990 and 2000 were in kinds of work not catalogued about a decade earlier (Lin 2011).

That engine has slowed. Acemoglu and Restrepo find that, compared with 1947 to 1987, displacement by automation sped up and reinstatement through new tasks weakened in the three decades after 1987 (JEP 2019). The balance is an empirical regularity, not a law.

What we pre-registered, and what did not hold

Our first hypothesis was the four-part historical pattern: the exposed occupation shrinks, total employment does not fall persistently, identifiable new work tied to the technology appears, and losses are concentrated on identifiable groups. We set a threshold of 7 of 9 cases. Counting all four parts, the pattern is documented in 1 case on a strict reading and 5 on a lenient one. It is not met. The shortfall comes mostly from parts our sources did not document, not from cases that contradict the pattern. New work tied to the technology itself was clearly found in only 4 cases (electrification, the car, office computing, the internet); in the others the absorbing jobs were elsewhere in the economy or could not be sized. Concentrated losses could not be documented for electrification, the car or the internet.

Our second hypothesis was that job-loss forecasts are more often right about a single occupation than about the economy. We collected 28 dated statements from 1821 to 2016. Ten were not forecasts (observations, a slogan, and technical-potential estimates such as Frey and Osborne's 47%). Of the 18 forecasts, we could check 6 at the occupation level and 5 at the economy level:

  • Occupation level: 1 of 6 fully consistent with the outcome (a 1935 forecast that cotton sharecropping "will have to be abandoned"), 3 of 6 at least partly. Both forecasts that an occupation would not shrink were wrong.

  • Economy level: 2 of 5 consistent (1 of 4 under the pre-registered 10-year rule), both reassuring forecasts with open horizons. None of the 3 checked alarmist forecasts has been borne out, including Norbert Wiener's 1949 letter to Walter Reuther. Wiener wrote that "in the hands of the present industrial set-up, the unemployment produced by such plants can only be disastrous" and guessed that "a critical situation is bound to arise under any condition in some ten to twenty years". That horizon has passed; US unemployment was 3.5% in 1969 (FRED UNRATE). Three other alarmist forecasts, including the 1964 Triple Revolution memo, could not be checked against what they claimed.

  • Keynes's 1930 fifteen-hour work week is not due until 2030.

These counts do not separate the two levels. The coding was done after outcomes were known, by one coder, on statements found by search, so none of this is a base rate. The set shows something narrower: in every checked forecast that an occupation would shrink, it did eventually shrink, though timing and size could rarely be checked.

How long it took

T-lag ran from about 0 years (teamsters, farm labour) to about 49 years (telephone operators). Where an occupation did halve, it took from under 12 years (US newspaper newsrooms, 2008 to 2020; Pew Research Center 2021) to as much as 50 years (farmers and farm labourers on the census basis). Several groups had not halved 24 to 46 years after their peak: bookkeepers, department stores, and manufacturing. The crossing years come from sparse census or survey points, so each is a bracket, not a point.

Who paid

  • By region: the cotton South; London and Liverpool for docks; New York City's piers.

  • By sex: women in telephone and clerical work.

  • By age and career stage: incumbent operators lost status and pay; UK dockers were mostly over 45; German robot-era adjustment fell on young people who were never hired.

  • By skill: routine manual workers in the US robot case.

Two mechanisms recur: attrition, where new entrants are not hired and the cost shifts to people who never get the job, and buy-outs funded from productivity gains, such as the 1960 West Coast longshore fund.

AI against this record (through October 2026)

Exposure

Eloundou, Manning, Mishkin and Rock estimate that "around 80% of the U.S. workforce could have at least 10% of their work tasks affected" (2023 preprint). The IMF puts "almost 40 percent of global employment" as exposed (January 2024). The ILO's more conservative index puts one in four workers globally in an occupation with some exposure, and 3.3% in the highest category, with women at 4.7% against men at 2.4% (ILO 2025). Unlike most past cases, exposure rises with wages and education. BLS projects office and administrative support employment to fall 4.0%, or 752,100 jobs, over 2025 to 2035 (BLS, August 2026). Exposure measures overlap with tasks; it is not a forecast of job loss.

Adoption

Bick, Blandin and Deming find that nearly 40% of US adults aged 18 to 64 used generative AI by late 2024, with work adoption "as fast as the personal computer" (NBER w32966; figures from our research notes, not re-read on the primary page in this study). The share of employed adults using it for work in the past week rose from 28.2% in Q3 2024 to 39.2% in Q2 2026, but only 6.3% of work hours were AI-assisted (FRED Blog, August 2026).

Early labour-market outcomes

  • Customer support: generative AI raised productivity by 15% on average, most for less experienced workers (Brynjolfsson, Li and Raymond, QJE 2025).

  • Denmark: administrative data rule out effects on earnings or hours larger than 2% two years after ChatGPT, while "new AI-related tasks are widespread" (Humlum and Vestergaard 2025).

  • Young US workers: the August 2026 update of "Canaries in the Coal Mine" finds "no evidence of widespread, economy-wide job displacement", but employment of 22 to 25 year-olds in AI-exposed occupations is 19% below where it would have been, mainly through reduced hiring. The authors call this descriptive, not causal (Brynjolfsson, Chandar and Chen 2026).

  • Anthropic's March 2026 study finds no systematic rise in unemployment among highly exposed workers, and a "just barely statistically significant" 14% lower job-finding rate for 22 to 25 year-olds entering exposed occupations.

  • Federal Reserve Governor Lisa Cook said on 28 September 2026 that there is "limited evidence that AI is yielding significant changes to the structure of the labor market".

This early pattern for young workers is consistent with the attrition channel seen with telephone operators and German robots, but it is descriptive, not a demonstrated mechanism.

Does the pattern carry over?

We pre-registered a four-part test of whether AI falls outside the range of the nine cases.

  • Diffusion speed: inside. Faster than the PC, comparable to the internet.

  • Breadth of exposure: outside. 80% of workers with at least 10% of tasks exposed, against farm work at 33% to 41% of employment for one occupation group.

  • Who is exposed: outside. Exposure concentrates in higher-wage, higher-education work.

  • New work: inside, visible but small. Whether AI will also perform the new tasks cannot be tested yet.

With two of four outside, our pre-registered rule gives "the past pattern carries over, with caveats". The verdict is fragile: benchmark diffusion against the PC instead of the internet and three dimensions are outside, which turns it into "a weak guide". Most measured historical declines took decades, and AI has had under four years, so 2026 data cannot establish that the pattern fails.

Where history is a weak guide

  1. The new tasks may be automated too. The technology-linked new work in these cases (driving, repairing, programming, logistics) was work the displacing machine could not do. A tractor cannot fix a tractor. A general-purpose model may do part of the new work it creates, and no historical case calibrates this.

  2. Exposure runs up the wage scale, into the professional and office work that absorbed displaced workers in the past.

  3. A system that can do any new task at or below human cost is outside this evidence base by construction.

What history suggests will remain

These are inferences from the cases, not measurements.

  • Work where a person is accountable, or the relationship is the product: tellers moved to sales and relationships, and accountants grew while bookkeepers shrank.

  • Physical work in unstructured settings: maintenance, construction, care.

  • Work whose demand rises when technology makes a service cheaper (branches, computing, warehouses), but only while demand keeps growing.

  • At risk first: entry-level, codified work learned on the job. The cost shows up as jobs never offered, which is easy to miss in unemployment data and hard to compensate.

Limits

  • Selection. The nine cases are well-known, mostly US episodes chosen before research but not sampled from any population of technologies. The predictions were found by search, so they cannot give base rates of forecast error.

  • Causation. Occupation counts mix the technology's effect with other forces: the China trade shock for manufacturing, out-migration for farm labour, online banking for tellers after 2010.

  • Definitions. Census and CPS occupation classifications break in 1950, 1983 and 2000 to 2010. Peaks and end years come from sparse series, and other choices would move T-lag and T-decline. No harmonized microdata (such as IPUMS) were used.

  • Coding. All prediction coding was done by one coder (an AI system), not blind to outcomes, with no second coder.

  • Total employment is an easy bar given population growth; employment rates and displaced workers' wages were not assessed for most cases.

  • Unverified figures. UK aggregate employment for 1959 to 1970 was not checked. The telephone-operator incumbent figures (Feigenbaum and Gross), the Acemoglu and Restrepo yearly rates, and the Bick, Blandin and Deming adoption figures were not re-read on the primary page in this study. AI evidence through 2026 is early and descriptive.

Data

The 17 case series, the 28 coded statements, the pre-registration and a script that recomputes every count are available on request.

Disclosure

The text was drafted with AI assistance (Anthropic Claude) and checked against the cited sources and the data files. The analysis was reviewed in three rounds by OpenAI Codex, plus a confirming round on the final fixes by a second AI reviewer (Anthropic Claude). No Glasp user data was used. No external funding. No known conflicts of interest.

References

Sources accessed 30 October to 1 November 2026.

Comments

Add a comment
    Glasp Research

    Written by Glasp Research