Developments in artificial intelligence (AI) have captured our imagination, filling us with expectations and challenges. It is not clear what limit they may have, if it exists. From an economic perspective, it is immediate to ask how it can be the impact on macroeconomic variables that we usually consider central. Projections are always difficult, especially in disruptive scenarios. How much could production increase? What will happen to productivity? Who will benefit from these changes? Addressing these questions in a disciplined way requires relying on economic models that reflect the functioning of specific parts of the economy and adjusting their assumptions. But how to carry out this analysis? Should we ask ChatGPT?
A few weeks ago, an article by Daron Acemoglu titled “The Simple Macroeconomics of AI” was released. Acemoglu, a Turkish economist of Armenian origin, earned his PhD from the London School of Economics and has been a professor at the Massachusetts Institute of Technology (MIT) since 1993. In 2005, he received the John Bates Clark Medal. This award is given by the American Economic Association in recognition of the outstanding contribution of an economist under the age of 40. Currently, Acemoglu is one of the most cited economists in the world (surely in the top 5).
Acemoglu’s model takes his previous work on automation as a starting point. In this adaptation he builds a model in which, for the production of a final good, the completion of a series of tasks is required and these can be accomplished using both labor and capital. The marginal productivity of capital and labor are different for different tasks. In the competitive equilibrium of the market, the assignment of tasks to labor and capital is achieved through cost minimization.
AI can generate productivity gains through various channels. First, through automation AI models can reduce the cost of certain tasks. For example, mid-level administrative functions, also text synthesis or data classification.
Second, AI can complement tasks that, although not fully automated, could increase work productivity. Various activities, such as those required to write this note, could be better executed through efficient access to accurate information. Another way to understand complementarity is that AI could automate some parts of tasks, allowing workers to specialize and increase their productivity in other dimensions of their work.
Third, AI could increase capital productivity in previously automated tasks. Finally, it is possible that new tasks will be generated that impact the productivity of the entire production process.
Acemoglu’s model allows us to do something that we economists understand is always key. It allows you to put a number on these speculations (especially the first two channels), and thus quantify how much the macroeconomy can be affected by AI.
Focusing on improving productivity, its aggregate effect will depend on the fraction of tasks affected and the average cost savings they imply. According to Acemoglu’s first estimate, an increase in total factor productivity of 0.66% will be achieved in 10 years. Although significant, the expected increase in productivity is modest.
The author warns that this estimate could be too optimistic regarding the possible effects of AI. Because? Because the first tasks that AI has tackled are naturally the easiest to automate. A greater challenge will be the more complex (more difficult to learn) tasks to come. Likewise, AI can generate tasks with negative social value such as designing algorithms for manipulation on-line (deepfake, misleading digital ads, AI-powered hacks). Incorporating this into macro estimates requires even more speculative guesswork.
It should not be concluded from this study that AI does not offer important benefits. Firstly, such an increase in total factor productivity as indicated is modest, but it is not irrelevant. Secondly and more importantly, Acemoglu points out that there may be other ways in which AI could generate more significant benefits from an economic perspective. If AI is used to create new tasks for workers, its effects could be broader both in terms of productivity, wages and inequality. To achieve these benefits, Acemoglu believes a change in the orientation of the industry is necessary. This includes, possibly, a review in the architecture of the most common generative AI models in such a way that the generation of reliable and useful information for different types of workers is prioritized, instead of the development of general conversation tools similar to human.
