Task Exposure Versus Job Replacement
Why the unit of analysis is the task rather than the occupation, what exposure measures and what it does not, and the two opposing forces that determine whether exposure becomes displacement.
Public discussion asks which jobs AI will replace. Labour economics has largely abandoned that framing, because a job is a bundle of tasks and technology acts on tasks. An occupation whose tasks are half automatable does not become half unemployed; it becomes an occupation that does the other half plus whatever new tasks the technology creates.
The task framework
Acemoglu and Restrepo's task-based framework is the standard structure. Output is produced by a set of tasks, each performed by labour or capital, and technology changes the allocation. Two forces run in opposite directions.
The displacement effect removes tasks from labour, reducing the demand for it at a given output.
The reinstatement effect creates new tasks in which labour has a comparative advantage, and productivity gains raise output and therefore demand for the remaining tasks.
Whether employment in an occupation rises or falls depends on which dominates, which depends on the technology, the elasticity of demand for the output, and how quickly complementary tasks appear. Historically, both effects have been large, and the net has varied by occupation and by period rather than following a rule.
What exposure studies measure
The most-cited estimate for language models found that around 80 percent of US workers are in occupations where at least 10 percent of tasks could have completion time reduced by half using a language model, and around 19 percent are in occupations where at least 50 percent of tasks meet that bar (Eloundou et al., 2023, GPTs are GPTs, arXiv:2303.10130).
What that measures is technical exposure: the possibility that a task could be affected. It is not a prediction about employment, and the paper is explicit on this. Exposure says nothing about whether adoption occurs, whether the quality is acceptable, what regulation permits, what the organisational change costs, or what the reinstatement effect adds.
The exposure distribution is also unusual relative to previous automation waves. It concentrates on higher-wage, higher-education occupations involving writing, analysis and coding, which is close to the inverse of the pattern for industrial robotics and earlier software.
When it breaks
Exposure is measured against today's models. Capability moves, so an exposure estimate is a snapshot whose usefulness decays. It is a description of a moment rather than a forecast.
Task decomposition is coarse. Occupational task inventories describe work at a granularity that misses the coordination, judgement and context that occupy much of a real job. A task rated highly exposed may be a small part of the work in practice.
Adoption lags capability by years. Organisational change, workflow redesign, procurement, training and trust all take time, and the historical pattern for general-purpose technologies is a long delay between availability and measurable productivity effect. This is the productivity paradox in its modern form.
The distributional question is separate from the aggregate one. Even where total employment is unaffected, the composition changes, and the people whose tasks are displaced are usually not the people who take the new ones. Aggregate stability is compatible with substantial individual disruption, and the aggregate number is the one most often reported.
12 flashcards for this concept
Click a card to reveal the answer.