AI and Economic Growth: What Models Predict
5 min read · updated August 3, 2026
Economists modelling AI reach conclusions ranging from a modest productivity bump to a change in the growth regime. They are not using different data. They are using different values for three or four parameters, and the parameters are where the argument should be.
Two families of model
Task-based automation models
Associated most closely with Daron Acemoglu and Pascual Restrepo, these treat production as a continuum of tasks, each performed by labour or by capital. Automation moves tasks from labour to capital, which raises productivity and displaces workers; new task creation moves the boundary back. Growth and distributional effects both fall out of the movement of that boundary.
The framework’s virtue is that it makes the aggregate effect an explicit function of quantities you can in principle measure: what share of tasks is exposed, how much cost is saved on each, and how fast new tasks appear. Acemoglu’s own applications of it to AI produce deliberately conservative aggregate numbers, and the reasoning is transparent — the effect is bounded by the exposed share times the saving on that share, so a large aggregate effect requires both terms to be large.
Idea-production models
The semi-endogenous growth tradition, associated with Charles Jones, models growth as driven by ideas, with ideas produced by researchers. Its central empirical observation is that ideas are getting harder to find: research effort has risen dramatically while growth has not, so productivity per researcher is falling. Aghion, Jones and Jones applied this framework to AI directly, and the key move is that AI enters not as a better tool but as a substitute for researchers themselves.
That changes the mathematics qualitatively rather than quantitatively. If the population of effective researchers can be expanded by producing more compute rather than by waiting for demographic growth, the constraint that keeps growth steady in these models is loosened, and under some parameter values the models produce accelerating growth. This is the formal home of the economic version of the intelligence explosion argument, and it is worth noting that the same authors are explicit about the parameter sensitivity.
Where the disagreement actually lives
| Parameter | Description |
|---|---|
| elasticity of substitution | How readily capital substitutes for labour across tasks. If tasks are poor substitutes, the ones that resist automation become the constraint and dominate cost. If they substitute freely, automation of most tasks translates into a large aggregate gain. This single parameter separates the modest forecasts from the explosive ones more than any other. |
| automatable share | What fraction of economically relevant tasks can actually be automated, and by when. Estimated from occupational task data and exposure indices, all of which measure exposure rather than realised automation. |
| cost saving per task | Automation that halves the cost of a task worth two percent of output produces a small aggregate effect regardless of how impressive it is. Aggregate effects are shares times savings, and both terms are usually smaller than intuition suggests. |
| returns to research | Whether adding effective researchers still produces proportional ideas. The ideas-getting-harder finding says no; how steeply returns decline determines whether more automated research compounds or merely offsets. |
| diffusion rate | How fast capability becomes deployed practice. Historically slow for general-purpose technologies, for reasons of complementary investment rather than availability. |
When two forecasts differ by an order of magnitude, the difference is almost always in the first and the last of these. That is a more useful thing to know than either forecast.
The bottleneck argument
The most important structural argument against explosive growth is older than AI and is usually attributed to William Baumol. If an economy has sectors that improve rapidly and sectors that do not, the slow sectors come to dominate spending, because their relative prices rise as everything else gets cheaper. Overall growth ends up governed by the slow sectors rather than the fast ones.
Applied here: even complete automation of all cognitive work leaves activities constrained by physical throughput, by regulation, by physical human presence, or by the fact that the good being purchased is a human relationship. Those become a growing share of output, and aggregate growth is bounded by their improvement rate.
The counter-argument, and it is a real one, is that the sectors currently classified as resistant are resistant to cognitive automation specifically. Advanced robotics, or biological engineering driven by automated research, would move some of them into the fast category, and the Baumol argument holds only for whatever remains outside. Which of the two applies is a question about physical technology, not about AI, and that is why forecasts in this area quietly depend on robotics assumptions their authors sometimes do not state.
What the empirical work can and cannot settle
There is a growing body of measurement — firm-level studies of deployment, randomised trials of assistance on specific tasks, adoption surveys, exposure indices built from occupational task descriptions. It is worth understanding what each design can support.
- Task-level trials measure a real effect cleanly and generalise poorly. The task was chosen, the users were in a study, and firm-level output depends on reorganisation the trial does not capture.
- Firm-level studies capture reorganisation and suffer selection: firms that adopt early differ from firms that do not, in ways that also affect their productivity.
- Exposure indices measure how much of an occupation’s described tasks a system could plausibly touch. They are frequently reported as automation forecasts and they are not: exposure is an upper bound on a mechanism, not a prediction of realised substitution.
- Aggregate statistics are the thing anyone actually wants and arrive with a lag of years, confounded by everything else happening in the economy.
The methodological guide to the labour-market half of this evidence is in what the employment data can support.
One further asymmetry is worth holding on to when reading this literature. Growth effects and distributional effects are separate outputs of the same models, and they can move independently: a scenario with a modest aggregate growth effect can carry a large reallocation of income between labour and capital, and a scenario with a large growth effect need not. The task-based framework is explicit about this — the boundary between automated and non-automated tasks determines both the productivity gain and who captures it — and public discussion routinely treats a claim about one as a claim about the other. When a paper is described as optimistic or pessimistic, it is worth checking which of the two outputs the description refers to.
Reading any forecast in this area
Three questions, in order. Which model family is it, task-based or idea-production? What elasticity and what diffusion rate does it assume, and would a plausible alternative value change the conclusion? And does it assume automation of physical work, or only of cognitive work — because that assumption is doing more of the work than anything else in the paper.
The honest state of the field is that the models are sensitive to parameters that are not well identified, that the professional disagreement is real and is between people using the same tools correctly, and that this is normal for macroeconomic forecasting of a technology in progress. What is not defensible is quoting one model’s output as the economic consensus. There is no consensus, and the papers themselves say so.