Forecasts of jobs destroyed by automation

Carl Benedikt Frey and Michael Osborne published an estimate assigning a probability of automation to each American occupation, from a ranking of a sample of jobs by specialists and then a statistical extrapolation.

The aggregate figure, close to half of jobs, was disseminated worldwide on a considerable scale, often without the clarification that the study concerned occupations and not jobs actually eliminated. Nor did the authors set a precise horizon, which later citations often turned into a deadline.

A team at the Organisation for Economic Co-operation and Development, led by Melanie Arntz, Terry Gregory and Ulrich Zierahn, redid the exercise reasoning about tasks rather than occupations. Their work belongs to the task-based approach proposed by David Autor, Frank Levy and Richard Murnane in 2003.

Their result is markedly lower, around one job in ten for the countries studied, because an occupation contains automatable tasks and others that are not, and the mix varies greatly between holders of the same post.

The gap between the two estimates therefore stems not from a disagreement about the technology but from the unit of analysis. The same expert judgement produces very different orders of magnitude depending on whether it is applied to the whole occupation or to each of its components.

They estimate technical feasibility, not adoption. Adoption depends on relative cost, organisation, regulation and acceptability, variables absent from these estimates.

The history of earlier waves shows a substantial gap between feasibility and diffusion, and job creation in activities that did not exist at the time of the forecast. It also shows transformations within occupations, where automating one task shifts activity towards others rather than eliminating the post.

The fear is not new: John Maynard Keynes was already using the phrase technological unemployment in 1930. Forecasts of jobs destroyed published during earlier automation waves proved systematically pessimistic about total employment, and often inaccurate about the sectoral distribution.

That finding does not establish that current forecasts will be wrong, but it provides a benchmark that recent publications rarely cite. The opposite argument, that current technologies reach cognitive tasks hitherto spared, deserves to be set out with the same seriousness, and the available data do not settle it.

Empirical work on the actual introduction of industrial robots, notably that of Daron Acemoglu and Pascual Restrepo on American local labour markets, published in 2020, measures local effects on employment and wages, negative in exposed areas.

That work concerns one precise technology and an observed period, which makes it more solid than prospective estimates, and narrower in reach.