Calibrating AI Automation and its Labor Market Impacts

We are early in the latest AI technological revolution, and there are vastly different projections circulating about the impact of AI on the labor market; most of these projections assume labor automation happens on a fixed schedule. I revisit this question by calibrating a model where labor automation is an endogenous outcome driven by the accumulation of AI capital. The speed of accumulation of AI capital is calibrated based on hyperscalers’ and data-center investment. The model has an intelligence sector, composed of intelligence industries, where AI capital represents a (small) share of the sector’s capital; this sector automates labor as more AI capital becomes available. The physical sector, composed of the rest of the industries, uses ordinary labor and capital, and does not have the option to automate labor. As AI capital grows, the intelligence sector automates, and workers quit into unemployment. Workers eventually find a job in the physical sector, where physical capital starts to grow in order to respond to the increased labor supply that raises its marginal productivity. The model incorporates three frictions: unemployment, a cap on physical capital investment rate, and an adoption friction for AI. I calibrate the model to the 2024 US economy. If AI capital grows through a series of limited booms, unemployment increases by at most 1.4 percentage points over 20 years. Even if AI capital keeps doubling every year, as assumed in the aggressive scenario of the central case, there is no increase in unemployment in the first two years, consistent with limited effects observed on unemployment so far. Unemployment then increases by 8.4 percentage points after 7 years, while average worker labor income (including zero income for the unemployed) decreases by 5.8%, and the labor share decreases by 13 percentage points. After automation in the intelligence sector completes (year 7), leaving only strictly necessary workers there, unemployment decreases, wages increase sharply (+47.5% at 10 years), and the labor share overshoots its initial value. Without frictions, labor income instead rises monotonically. Across plausible parameter values, the depth of the income trough and the size of the ten-year wage gain depend most strongly on the growth rate of AI capital; most other parameters play a smaller role. This pins both automation and labor market effects to the growth rate of AI capital, which is therefore a key variable to watch.