The Simple Macroeconomics of AI

Abstract

This paper evaluates claims about the large macroeconomic implications of new advances in AI. It starts from a task-based model of AI's effects, working through automation and task complementarities. It establishes that, so long as AI's microeconomic effects are driven by cost savings/productivity improvements at the task level, its macroeconomic consequences will be given by a version of Hulten's theorem.

Using existing estimates on exposure to AI and productivity improvements at the task level, these macroeconomic effects appear nontrivial but modest—no more than a 0.71% increase in total factor productivity over 10 years. The paper argues that even these estimates could be exaggerated because early evidence is from easy-to-learn tasks, whereas some future effects will come from hard-to-learn tasks.

Predicted TFP gains over the next 10 years are even more modest and are predicted to be less than 0.55%. The paper also explores AI's wage and inequality effects, finding that AI advances are unlikely to increase inequality as much as previous automation technologies but there is no evidence that AI will reduce labor income inequality.

Introduction

Artificial intelligence (AI) has captured imaginations with promises of rapid productivity growth and new pathways for complementing humans. ChatGPT, released in November 2022, became the fastest spreading tech platform in history. While AI will have implications for the macroeconomy, productivity, wages and inequality, these are very hard to predict.

Some forecasts predict transformative effects, including artificial general intelligence (AGI) performing essentially all human tasks. Goldman Sachs (2023) predicts a 7% increase in global GDP and a 1.5% per annum increase in US productivity growth over 10 years. McKinsey Global Institute (2023) forecasts suggest generative AI could offer a boost as large as $17.1 to $25.6 trillion to the global economy.

This paper uses the framework from Acemoglu and Restrepo (2018, 2022) to provide insights for these debates, focusing on medium-term (about 10-year) macroeconomic effects of AI. The analysis builds a task-based model where AI-based productivity gains can come from several channels: automation, task complementarity, deepening of automation, and new tasks.

Conceptual Framework

The model builds on Acemoglu and Autor (2011) and Acemoglu and Restrepo (2018, 2022). The economy produces a unique final good by combining a set of tasks, with tasks allocated to either capital or labor based on comparative advantage.

AI could affect production through:

• Automation: AI enables further automation, increasing the set of tasks produced by capital.
• Task Complementarity: AI can generate new task complementarities, raising the productivity of labor in tasks it is performing.
• Deepening of Automation: AI could improve performance in previously capital-intensive tasks.
• New Tasks: AI can generate new labor-intensive products or tasks.

The effects of new AI tools will depend on the extent of each of these effects, with the framework providing a way to analyze their consequences.

Hulten's Theorem and AI

Hulten's theorem provides a simple formula for competitive economies with constant returns to scale, specifying how micro-level productivity improvements translate into macro changes. Since the economy here is competitive, this theorem applies and disciplines the productivity effects.

When AI's microeconomic effects are driven by cost savings at the task level—due to either automation or task complementarities—its macroeconomic consequences follow Hulten's theorem:

d ln TFP = ∫ χ(z) π_L(z) dz

where χ(z) is the GDP share of task z and π_L(z) is the productivity improvement or cost savings in task z driven by AI.

This implies that total factor productivity gains can be estimated by:

d ln TFP = π̄ × GDP share of tasks impacted by AI

where π̄ is the economy-wide average cost savings.

Easy vs. Hard Tasks

A crucial distinction is between "easy-to-learn" and "hard-to-learn" tasks. Easy tasks have a simple mapping between action and outcome with reliable, observable outcome metrics. Examples include writing standard subroutines, summarizing text, or verifying identities.

Hard tasks lack a simple mapping between action and desired outcome, depend strongly on contextual factors, and have no clear metric of success. Examples include diagnosing complex medical conditions, providing personalized advice, or creative problem-solving.

AI productivity gains observed so far are from easy tasks. Productivity gains in hard tasks will be more limited because:

• Lack of simple mapping makes training AI models difficult
• Human complementarity is harder to achieve
• Learning from humans won't lead to better-than-human performance

This distinction is important because existing experimental studies (Peng et al., 2023; Noy and Zhang, 2023; Brynjolfsson et al., 2023) focus on easy tasks, potentially overestimating broader productivity gains.

Quantitative Evaluation

Using data from Eloundou et al. (2023) on tasks exposed to AI and Svanberg et al. (2024) on feasible automation, along with experimental studies on cost savings, Acemoglu provides quantitative estimates:

• 19.9% of US labor tasks are exposed to AI
• 23% of exposed tasks can be profitably automated within 10 years
• Average labor cost savings: 27%
• Average overall cost savings: 15.4% (adjusting for labor share)

This implies TFP gains over 10 years of 0.71% (0.046 × 0.154). Accounting for the distinction between easy and hard tasks (74% of exposed tasks are easy), TFP gains drop to 0.55%.

GDP effects will be somewhat larger due to investment responses, estimated at 0.9-1.8% over 10 years depending on assumptions about capital stock increases.

Inequality Implications

AI's distributional consequences are complex. Even when AI improves the productivity of low-skill workers in certain tasks, this may increase rather than reduce inequality due to general equilibrium effects and ripple effects across the labor market.

Key findings on inequality:

• AI exposure is more equally distributed across demographic groups than previous automation technologies
• AI is unlikely to lead to substantial wage declines for affected groups
• No evidence that AI will reduce labor income inequality
• AI will likely widen the gap between capital and labor income
• Low-education women, especially white native-born women, may experience real wage declines

Unlike some optimistic forecasts, the analysis suggests AI will not significantly reduce inequality and may have small positive effects on overall between-group inequality.

Conclusion

While AI technologies have impressive achievements and potential for beneficial economic effects, the extent of their macroeconomic consequences appears more modest than many forecasts suggest. TFP gains from AI advances within the next 10 years are estimated at 0.55-0.71%, substantially lower than predictions from Goldman Sachs or McKinsey.

More favorable wage and inequality effects, as well as more sizable productivity benefits, will likely depend on the creation of new tasks for workers, especially middle- and low-pay workers. However, this does not seem to be the focus of current AI research and development.

There are potentially much bigger gains to be had from generative AI if reoriented toward providing reliable information to workers in various occupations, but this would require fundamental changes in AI development priorities.