Paper projects AI could raise US productivity ~20% over a decade
Merali split the productivity gain roughly 56% compute scaling and 44% algorithmic progress, and found it far smaller for agentic tasks needing tool use.
- Ideas & essays
- Minor
Yale economist Ali Merali published a preregistered experiment measuring how large language models affect professional productivity, in which more than 500 consultants, data analysts and managers completed realistic work tasks with the assistance of one of 13 different LLMs. Merali found that each year of model progress reduced task-completion time by about 8% on average, and attributed roughly 56% of that gain to increased training compute and the remaining 44% to algorithmic progress.
The gains were not uniform: they were substantially larger for non-agentic analytical work than for agentic tasks requiring tool use, where models’ advantage over unassisted workers was smaller. Extrapolating the measured 8%-per-year trend forward, Merali projected that continued model scaling could raise US productivity by roughly 20% over the next decade — a projection that depends on the trend continuing at its current rate and on task-level productivity gains translating into economy-wide ones, neither of which the experiment itself could establish.