The numbers everyone repeats · 6.2

The automation percentages

The automation percentages. What is actually the case, and how it compares with what is repeated.

A practical software reference for this part of the discussion is Monitask's overview of workforce analytics.

Two well-conducted studies asked what share of jobs could be automated and produced answers of forty-seven per cent and nine per cent. Both were right. They were not asking the same question.

The first

A 2013 study assessed seventy occupations with expert judgement, identified features associated with automatability, and extrapolated across some seven hundred occupations. It reported that around forty-seven per cent of American employment was in occupations at high risk over perhaps a decade or two.

The unit of analysis was the occupation, treated as a whole. If the core of an occupation looked automatable, the occupation counted.

The second

A 2016 study took the same question and used the task as the unit. It found that within most occupations some tasks were automatable and others were not, and that occupations composed almost entirely of automatable tasks accounted for around nine per cent of employment.

The five-fold difference between the two is entirely a difference of method, and it is the most instructive pair of numbers in this subject.

Four things that are not the same

Technically possible. Economically worthwhile. Actually adopted. And resulting in net job loss.

Every step between them loses a large fraction. Something can be demonstrably possible and remain unadopted for decades because the labour it replaces is cheap, because the capital cost is high, because the process would need redesigning, or because somebody has to be accountable.

Headlines routinely move from the first to the fourth in a single sentence.

The current generation

Studies of exposure to large language models use the word exposure deliberately, meaning that some share of a job's tasks could be affected. The authors generally state explicitly that this is not a prediction of displacement.

The distinction survives about one news cycle.

What history records

Automation has repeatedly eliminated tasks, changed occupations, and displaced specific groups of workers severely while raising employment in aggregate. All of those are true simultaneously and arguments usually select one.

The severe local effects are real and have been well documented in studies of regional exposure to trade and to industrial robots. Aggregate reassurance is cold comfort to a specific labour market, and aggregate alarm is not supported.

What determines the outcome

Not the technology. Whether displaced workers can move to other work, which depends on their age, on the local labour market, on transferable skills, and on whether anybody pays for retraining.

Those are policy variables and they are the ones least discussed in coverage of automation forecasts, because they are duller than the forecast.

How to read the next figure

Ask whether the unit was the occupation or the task. Ask whether the claim is about technical feasibility or about expected adoption. Ask over what period. And ask what the authors themselves said the number meant, since it is frequently more careful than the coverage.

The honest summary

Task composition within jobs is changing and will keep changing. The share of jobs disappearing entirely in any decade is small. The people affected are concentrated rather than spread, which makes an aggregate figure a poor guide to whether anybody should worry.

That is less quotable than either forty-seven or nine, and it is what the two studies jointly support.

The word at risk

It does a great deal of unmarked work. At risk in the 2013 study meant having characteristics associated with automatability by expert judgement, over an unspecified period described as perhaps a decade or two.

Reported as a share of jobs that will disappear, it changed meaning entirely, and the authors have since noted the distinction publicly more than once.

What employers actually did

Adoption of automation technologies has been slower and more uneven than either forecast implied, concentrated in large firms, and driven as much by labour cost and availability as by the technology becoming possible.

Where labour is cheap and flexible, automation is deferred. That is a straightforward economic point and it means a forecast based on capability alone is missing the variable that decides the timing.

The occupations that did go

Switchboard operators, typesetters, film processing technicians, most typing pools, and a large share of bank tellers, though that last one happened more slowly than the machines arrived.

In each case the occupation contracted over decades rather than disappearing in a wave, and the workers affected mostly moved to adjacent work rather than out of employment. Local exceptions to that were severe and are the part worth taking seriously.

For broader context, consult OECD analysis of AI and work.