Contra Acemoglu on AI

🌈 Abstract

The article discusses a 2024 working paper by Daron Acemoglu, titled "The Simple Macroeconomics of AI," which models the economic growth effects of AI and predicts them to be relatively small, around a 0.06% increase in TFP growth annually. This contrasts with many other predictions that forecast much larger impacts on economic growth from AI.

🙋 Q&A

[01] Acemoglu's Channels for AI's Impact on Productivity

1. What are the four channels through which Acemoglu believes AI could affect productivity?

• AI enables further (extensive-margin) automation, such as generative AI tools taking over simple writing, translation, and classification tasks.
• AI can generate new task complementarities, raising the productivity of labor in tasks it is performing, such as by providing better information to workers or automating some subtasks.
• AI could induce deepening of automation, improving performance or reducing costs in some previously capital-intensive tasks.
• AI can generate new labor-intensive products or tasks.

2. Why does Acemoglu dismiss the latter two channels (deepening of automation and new tasks) without good arguments or evidence?

• For deepening of automation, Acemoglu argues that the tasks impacted by generative AI are quite different from those automated by previous digital technologies. However, the article points out that transformers, the underlying technology in generative AI, are already being used to increase the productivity of tasks already performed by machines.
• For new tasks, Acemoglu acknowledges their potential for larger wage and productivity impacts, but dismisses them based on the possibility of "new bad tasks" like misinformation and targeted ads, without providing sufficient evidence or arguments.

[02] Acemoglu's Estimation of AI's Productivity Impact

1. How does Acemoglu estimate the productivity effects from the "automation" channel?

• Acemoglu derives a simple equation: the change in TFP is the share of GDP from tasks affected by AI multiplied by the average cost savings in those tasks.
• He uses estimates that 20% of tasks are "exposed" to AI and 23% of those can be profitably automated, resulting in 4.6% of GDP being exposed.
• He then combines this with experimental results showing 30% productivity gains for labor, which is about 50% of costs, resulting in a 15% total cost savings.
• Multiplying these numbers gives his overall estimate of a 0.064% increase in TFP growth annually.

2. What are the issues with Acemoglu's estimation approach?

• The article argues that Acemoglu's approach of using current estimates of AI exposure and productivity gains may not hold up over the next 10 years, as the rapid pace of AI investment and development is likely to lead to significant changes in the types of tasks affected by AI.
• The article suggests that Acemoglu's paper would be better titled "Cost Savings From Extensive-Margin AI Automation" rather than claiming to cover the full macroeconomic impacts of AI.