What Automation Actually Did to the Typing Pool

1Why clerical work is the right comparison

Discussion of automation reaches reflexively for the loom and the assembly line, and both are poor analogies for what is happening now. Those technologies replaced physical effort in manufacturing, a sector that was already concentrated, unionised, and geographically visible. The disruption was loud, and its loudness is why we remember it.

Clerical work is the better comparison because it was information work: reading, transcribing, filing, checking, routing, calculating. It was performed largely in offices, largely by women, and at its peak it accounted for a very substantial share of all employment in industrialised economies. It was then subjected, over roughly four decades, to a sequence of automating technologies: the photocopier, the electronic calculator, the word processor, the spreadsheet, the relational database, and finally the networked personal computer.

We have unusually good data on what happened, because censuses tracked occupational categories carefully during exactly this period. And what happened does not support either the comfortable story or the catastrophic one.

2The occupation that was eaten from inside

Take the typist and the stenographer. These were real, numerous, defined occupations with training pipelines, professional associations and career ladders. A competent stenographer had a skill that took years to acquire and that commanded a wage premium over general clerical work.

Word processing did not fire them. What it did was fold their task into somebody else's job. The manager who previously dictated now typed, because typing had become cheap enough in cognitive effort that the coordination cost of delegating exceeded the cost of doing it. The occupation did not shrink because people were dismissed from it in large numbers; it shrank because nobody entered it. Hiring stopped, existing staff aged out or moved sideways, and within two decades a category that had employed millions was a rounding error.

This matters because the political and statistical signatures are completely different. Mass dismissal produces protests, headlines and policy responses. A slow closure of entry produces nothing visible at all, and its cost falls almost entirely on people who never got the job in the first place and therefore never appear in any dataset of the displaced.

3Bank tellers, and the argument against panic

The standard counterexample deserves its status. Automated teller machines were introduced from the late 1960s and spread rapidly. The obvious prediction was that teller employment would collapse. It did not; teller numbers in the United States rose for roughly three decades after ATM deployment began.

The mechanism is well understood. ATMs reduced the cost of operating a branch, so banks opened more branches, so total teller employment rose even as tellers per branch fell. Meanwhile the job changed: less cash handling, more sales and customer service. The technology was labour-saving per unit of output and employment-increasing in aggregate, because demand expanded to absorb the saving.

This is a genuine and important result and I do not want to explain it away. But it is often quoted with a crucial condition dropped. It works when demand is elastic. If cheaper branch operation had not produced more branches, the arithmetic would have run the other way, and the same mechanism that saved teller employment would have destroyed it. The ATM story is not a law; it is a case where a particular elasticity happened to be high.

4Who actually pays

Aggregate employment is the wrong unit of analysis and it has been distorting this debate for thirty years. Total employment recovers because new work appears; that is true, well-evidenced, and almost irrelevant to the person whose specific occupation went away at fifty-two in a town with one employer.

The literature on trade shocks, which is methodologically much stronger than most automation research because the exposure is measurable at regional level, finds large and persistent local effects: depressed wages, reduced labour force participation, and worse health outcomes lasting a decade or more in affected areas, even while national employment recovers fully. There is no obvious reason automation shocks would behave differently, and some reason to think they diffuse more broadly and are therefore harder to target with policy.

The distributional question is also a compositional one. When a task is automated, the residual job is not a smaller version of the old job; it is the leftover parts. Sometimes those parts are the interesting ones and the job improves. Frequently they are the tedious ones that resisted automation precisely because they are irregular, and the job gets worse while the title stays the same.

5What is different this time, and what is not

Two things are genuinely different. The first is the breadth of exposure: previous waves hit one task family at a time, and the current one touches drafting, summarising, coding, translating, and first-pass analysis simultaneously. Breadth matters because the standard adjustment mechanism is workers moving to an adjacent occupation, and adjacency helps less when the adjacent occupation is exposed too.

The second is direction. Earlier computerisation hollowed out the middle: it automated routine codifiable tasks while leaving both high-end judgement work and low-end manual work intact. That produced a well-documented pattern of wage polarisation. The current wave points at parts of the high end, which is a different shape and one we have less historical evidence about.

What is not different is the timescale, and this is where I part company with most commentary in both directions. Clerical automation took four decades to work through, and the binding constraint was almost never the technology. It was organisational redesign, retraining, capital replacement cycles and, repeatedly, the simple fact that firms are slow. There is no evidence I find credible that these frictions have disappeared.

6The honest conclusion

I do not know how many jobs this will displace and neither does anyone else. The published estimates span an order of magnitude and their spread is driven almost entirely by assumptions the authors put in at the start, particularly about whether exposure at task level translates into displacement at job level. It usually has not.

What the historical record supports is narrower and still worth saying. Automation of a task family rarely eliminates the occupation outright; it changes the job, closes the entry pipeline, and shifts the wage. Aggregate employment tends to recover. Specific cohorts and specific regions frequently do not, and the recovery of the aggregate is no comfort to them whatsoever.

If that sounds like a refusal to make a forecast, it is. The forecasts on offer are not measurements and dressing them in percentages does not make them so. What we can do, and what the historical evidence actually licenses, is prepare for a distributional problem rather than a level one — which is a different policy question, with different answers, and it is being asked much less often than it should be.