Workforce down, output per worker up — so far
Each country is indexed to 100 at the year its working-age population peaked. The question of the whole study is whether the dashed line can keep climbing fast enough as the solid one falls.
Lower birth rates have gone with faster per-worker growth
Every country with more than one million people, 1970 birth rate against annualized growth in GDP per working-age adult since. The same qualitative pattern Acemoglu, Autor, Beirne & Scott document with different data and definitions — here rebuilt entirely from public World Bank series.
What the next 35 years project
UN World Population Prospects 2024, medium variant; the shaded band on the second chart shows the high–low fertility variants. The decline ahead is not a forecasting artifact.
Technology measures are related questions, not one variable
Robot density on the factory floor, resident patent applications, and total factor productivity remain separate from the robot-import flow proxy. None of these series isolates AI — that absence is part of the finding.
Open-data replication and post-2015 holdout
Annual imports under HS 847950 are a trade-flow proxy, not installed robot stock. Imports can miss domestic production (the reason Japan is excluded from the regressions), and the “not elsewhere specified” commodity label can undercount robots captured by IFR. Japan remains in the descriptive chart.
Required versus achieved — the compensation bar
An identity, not a model: required per-worker growth = target GDP growth − projected workforce growth. The gap between required and achieved is what something — AI or otherwise — would have to add.
The denominator check: if only output per worker rose while total output shrank, "compensation" would be an arithmetic artifact of a falling denominator. It isn't — aggregate GDP has not fallen since any of these countries' workforce peaks. That is the anchor paper's aggregate-GDP result, visible here in East Asia.
How much would AI have to add? Move the levers.
The projection uses the UN medium-variant working-age path and holds the employment rate, participation (unless you move the lever), and hours constant — assumptions stated in the paper's methods. This is arithmetic on public numbers, not a forecast.