AI task exposure decomposition
The data
| Quantity | Value |
|---|---|
| US labor tasks theoretically exposed to gen AI | ~19.9% |
| Of those, share profitably automatable in 10 years | ~23% |
| → All US tasks profitably automatable | ~4.6% |
| Average task-level cost saving on automatable tasks | ~15.4% |
| Resulting 10-year US TFP gain (upper bound) | 0.55–0.71% cumulative |
| 10-year US GDP gain (upper bound) | 0.9–1.6% cumulative |
The arithmetic: aggregate gain ≈ 4.6% × 15.4% ≈ 0.71% cumulative TFP over 10 years (Hulten’s theorem).
Source
Daron Acemoglu, The Simple Macroeconomics of AI, NBER Working Paper 32487, May 2024. Surfaced through acemoglu-simple-macroeconomics.
What it shows
The full intermediate decomposition behind the bear-case TFP ceiling. The 4.6% number — the share of US tasks profitably automatable in a decade — is the load-bearing figure. The 15.4% cost-saving estimate is conservative; the 4.6% exposure figure is what makes the aggregate small.
Sensitivity is real: deployment cost, verification overhead, and the easy-vs-hard-to-learn distinction can all move the numbers materially. But the order of magnitude survives most reasonable perturbations of the assumptions.
How to cite
Acemoglu (2024) decomposes the US task base as 19.9% exposed × 23% profitably automatable → 4.6% of all tasks automatable, with ~15.4% average task-level cost savings.
Connections
- task-based-framework — the concept page this table sits inside.
- task-based-tfp-ceiling — the aggregate ceiling derived from these inputs.
- capital-labor-divergence — same decomposition supplies the labor-displacement projection.
- scaling-wall — the architectural argument for why the 23% profitability figure may also be optimistic.
- acemoglu-simple-macroeconomics — primary source.
- eloundou-beta-exposure — adjacent task-exposure framework that feeds into observed-exposure-measure in the Anthropic source.