How much will AI actually boost productivity over ten years?

Acemoglu's NBER working paper builds a task-based model in which AI's macroeconomic effects follow "a version of Hulten's theorem: GDP and aggregate productivity gains can be estimated by what fraction of tasks are impacted and average task-level cost savings." Applying "existing estimates on exposure to AI and productivity improvements at the task level," he finds the resulting effects "appear nontrivial but modest—no more than a 0.66% increase in total factor productivity (TFP) over 10 years." This is set against forecasts such as Goldman Sachs (2023), predicting "a 7% increase in global GDP, equivalent to $7 trillion, and a 1.5% per annum increase in US productivity growth over a 10-year period," and McKinsey Global Institute (2023), suggesting generative AI could add "$17.1 to $25.6 trillion to the global economy."

The task-based framework, drawn from Acemoglu and Restrepo, treats automation, task complementarity, deepening of automation, and new-task creation as the channels through which AI affects output; the paper focuses on the first two. It then argues the early productivity evidence — from Peng et al., Noy and Zhang, and Brynjolfsson et al. — comes from "easy-to-learn tasks," where AI already had sophisticated applications or businesses were already pushing adoption, and that this evidence cannot be "directly extrapolated to the rest of the economy" because future gains will increasingly come from "hard-to-learn tasks, where there are many context-dependent factors affecting decision-making and no objective outcome measures from which to learn successful performance." Revising for this, the paper predicts TFP gains "even more modest" and "less than 0.53%" over 10 years.

The same author's companion taxonomy, Why do firms build automating AI instead of pro-worker AI?, supplies the qualitative sorting of AI technologies that this paper's Hulten's-theorem calculation turns into an aggregate TFP number. The modest estimate sits in direct tension with When do AI feedback loops trigger explosive growth?, whose calibrated feedback-loop model reaches a growth singularity within six years — the two papers bound the plausible range of AI's macro effect from opposite ends. It also cuts against the more optimistic reading in Is AI productivity finally showing up in economic data?, since Acemoglu's framework predicts the aggregate productivity signal should stay small and slow to show up in the data.

The estimate is a projection built on 2023-24-vintage evidence on task exposure and productivity improvement, not a measurement of realized GDP or TFP growth, and it says nothing about possible future capability jumps that could change which tasks count as hard-to-learn. The excerpt also does not quantify how new, AI-enabled tasks with "negative social value (such as design of algorithms for online manipulation)" would net against the modest productivity gain, leaving the paper's headline TFP figure as a lower-bound-style estimate conditioned on current-generation AI capabilities.