“The Simple Macroeconomics of AI,” dated April 5, 2024 and circulated as NBER Working Paper 32487, is MIT economist Daron Acemoglu’s attempt to put a number on how much generative AI will actually move the macroeconomy over the next decade. Written for the journal Economic Policy, it is one of the most-cited skeptical counterweights to the bullish forecasts that followed ChatGPT.
Acemoglu builds a task-based model and applies a version of Hulten’s theorem: aggregate productivity gains can be approximated by the fraction of tasks AI affects multiplied by the average cost saving on those tasks. Plugging in existing estimates of task exposure and task-level productivity gains, he concludes the effects are “nontrivial but modest - no more than a 0.71% increase in total factor productivity over 10 years.” He then argues even that is likely too high, because early evidence comes from easy-to-learn tasks while many future gains would have to come from hard, context-dependent ones, putting predicted TFP gains over ten years at less than 0.55 percent.
The paper explicitly contrasts this with far larger claims it quotes, including Goldman Sachs (2023) predicting a 7 percent, roughly 7 trillion dollar increase in global GDP and a 1.5 percent per year rise in US productivity, and McKinsey Global Institute estimates of 17 to 25 trillion dollars. Acemoglu also warns AI may widen the gap between capital and labor income and is unlikely to reduce inequality, and that some new AI-enabled tasks - such as algorithms built for online manipulation - could carry negative social value. The work is a standing reminder that exposure estimates and headline forecasts are not the same as measured economic growth.