Automation and employment: what the evidence supports

The claim that automation destroys jobs and the claim that it creates them are both supported by selective reading. The literature is genuinely mixed, and the parts that are not mixed are worth separating out.

Tasks, not occupations

The framing most economists now use, and it clarifies a great deal.

Technologies automate tasks rather than occupations. Most occupations comprise many tasks, some automatable and some not.

Which means the typical effect is a change in the composition of a job rather than its elimination.

A radiologist whose image analysis is assisted spends more time on the tasks that were not automated. A warehouse worker who no longer walks spends more time picking.

Whether that is an improvement depends on which tasks remain, and this is where the interesting disagreement sits.

The displacement and productivity effects

The standard analytical decomposition.

Displacement. Automation substitutes capital for labour in specific tasks, reducing labour demand.

Productivity. Automation lowers costs, which raises output, which raises demand for labour in complementary tasks.

Reinstatement. New tasks are created that did not previously exist, some requiring labour.

The net effect depends on the relative size of these, which is an empirical question with different answers in different periods and sectors.

Historical work suggests that in the mid-twentieth century reinstatement broadly kept pace with displacement, and that in more recent decades it has done so less reliably. That finding is contested and it is the most policy-relevant claim in the literature.

What the cross-country evidence shows

Studies correlating robot adoption with employment have reached conflicting conclusions, and the conflict is instructive.

Work on United States local labour markets has found negative effects on employment and wages in exposed areas.

Work on European manufacturing has generally found smaller or absent aggregate employment effects, with composition shifts.

Several of the most robot-dense economies have maintained substantial manufacturing employment.

The likely explanation is that institutions matter: labour market structure, training systems, sectoral bargaining and how productivity gains are distributed all mediate the effect.

Which means the outcome is not determined by the technology.

The distributional finding

Clearer than the aggregate one and less discussed.

Automation has consistently affected middle-skill routine work most — clerical, administrative and routine production tasks — while affecting both high-skill non-routine and low-skill non-routine manual work less.

The result is polarisation: employment growth at both ends of the wage distribution and hollowing in the middle, documented across many developed economies.

That has consequences for wage inequality and for the availability of the mid-career jobs that once provided mobility.

What is different this time, if anything

The honest answer is that nobody knows, and two arguments are worth stating.

The case that it is different. Previous automation waves affected physical and routine cognitive tasks. Current systems affect non-routine cognitive work — writing, analysis, translation, coding — which was the destination for workers displaced from routine work.

If the escape route is also automated, the reinstatement effect is weaker.

The case that it is not. Every previous wave produced the same argument, and the predicted mass unemployment did not arrive. Task automation historically raised demand for complementary human work in ways nobody predicted in advance.

Both arguments have force and neither is settled by assertion.

What the near-term evidence shows

Early studies of generative systems in workplaces have generally found productivity gains concentrated among less experienced workers, which compresses performance differences within a role.

That is a genuinely interesting finding with an ambiguous implication: it could widen access to skilled work, or reduce the wage premium on experience.

Employment effects in specific occupations — translation, transcription, stock illustration, entry-level copywriting — are already measurable and negative.

Aggregate effects are not yet detectable in the data, which is expected this early and should not be read as evidence of absence.

What the policy questions actually are

Not whether to permit automation.

Whether productivity gains reach workers or accrue entirely to capital, which is a question about bargaining power and taxation.

Whether transition support exists for displaced workers, which historically it has largely not.

Whether training systems can retrain mid-career workers, which is a much harder problem than educating young people.

And whether the jobs that remain are better or worse — pace, autonomy, monitoring, safety — which is decided by how systems are implemented rather than by whether they are.

That last question receives the least attention and is the one workers actually experience.