[Kangkook Lee's Column] Is the future of AI and economic growth really rosy?

Dr. Kangkook Lee is Professor of College of Economics at Ritsumeikan University in Japan and Affiliated Research Fellow at CUKPE

As discussed in the previous article, even though evidence that artificial intelligence (AI) technology has boosted productivity is scarce to date, the future could be vastly different. Various studies have already presented scenarios forecasting the growth effects of AI. However, these studies also face a clash between optimism and pessimism. Will AI truly bring about a rosy future that significantly accelerates economic growth?

AI and Economic Growth: Optimism vs. Pessimism Scenarios

Immediately after the emergence of the generative AI boom, exemplified by ChatGPT, various institutions published studies forecasting how much AI technology would boost productivity and economic growth in the future. A Goldman Sachs report published in 2023 predicted that widespread adoption of generative AI would increase the annual labor productivity growth rate in the U.S. by 1.5 percentage points over a decade and boost global GDP by approximately 7%. Additionally, a McKinsey report projected that rapid AI adoption would increase U.S. labor productivity by an additional 0.7 percentage points annually by 2040. Forecasts from investment banks and consulting firms are generally optimistic, though this must, of course, be considered in light of their close ties to the AI ​​industry. There are also studies targeting Korea. The Bank of Korea's 2025 report , based on a model by International Monetary Fund (IMF) researchers , forecasts that the introduction of AI will increase total factor productivity by 1.1–3.2% and output by 4.2–12.6% over 10 years, depending on the extent of labor complementarity and productivity improvement effects.

What about the research of economists? Even top economists offer very different answers to the question of how much AI will boost productivity in the future. Daron Azzemoulou, a professor at the Massachusetts Institute of Technology (MIT) and the 2024 Nobel laureate in Economics, dampened optimism in his 2024 paper by stating that generative AI technology would bring about an annual increase in total factor productivity of about 0.06% over 10 years, suggesting that the effect would not be significant. On the other hand, Professor Aghion of France, the 2025 Nobel laureate in Economics, presented very different results using a similar model . According to his findings, AI brings about a significant increase in productivity of about 0.7% annually over 10 years, which is nearly 10 times higher than Azzemoulou's research.

The reason the results of the two studies differ so significantly is that they differ in their estimates regarding how profitablely AI will replace human labor. Their study calculates this using the formula: (10-year increase in total factor productivity due to AI = share of GDP accounted for by jobs affected by AI x average cost savings for those jobs). While Professor Ajemolou estimates the share of affected jobs to be merely 4.6%, Professor Aghion estimates this variable much higher at 30%. Although other studies report higher growth effects than Professor Ajemolou's, these results nonetheless clearly demonstrate how difficult it is to predict the future of AI. The key is how quickly and extensively AI will replace human labor, and the growth effect varies significantly depending on the extent of this replacement.

The outlook from an Antropic report released last September, which presented short-term scenarios regarding the economic effects of AI , is similar. This study projects that between 2026 and 2030, GDP will increase by 1.6% compared to the non-AI scenario under moderate change, 8.3% under significant change, and as much as 32% under extreme change. The difference depends on the spread of AI and job displacement by 2030; the assumptions are that moderate change affects only 4% of all jobs in the economy, significant change affects 20% of total cognitive labor—or 12% of all jobs—and extreme change automates about half of cognitive labor and 30% of jobs in the economy by AI. Interestingly, survey results among U.S. adults closely align with the 'significant change scenario.' Furthermore, the report predicts that as AI has a greater impact on economic growth, inequality will also deepen.

Of course, while these studies primarily examine the effects of generative AI, they are not without limitations in analyzing the effects of physical AI, which combines robots and AI. Moving forward, economic research regarding the advancement of AI and the automation of the entire economy is expected to develop rapidly.

The weak link blocking the growth effects of AI

 In fact, the field of economics has recently been rapidly developing theoretical models to address the growth effects of AI. For example, Professor Jones of Stanford University, a leading expert on economic growth theory, has developed a semi-endogenous growth theory in criticism of Romer's endogenous growth theory, which posits that R&D leads to growth by fostering new ideas and technological innovation. While endogenous growth theory predicted that an increase in R&D would proportionally boost economic productivity, in reality, the growth rate of total factor productivity did not rise but rather stagnated. Regarding this, Jones argued that this was due to diminishing returns, where the productivity of R&D itself declines. His recent paper reports empirical analysis results indicating that new ideas are becoming increasingly difficult to obtain across various industries.

Professor Jones's recent research presents a less optimistic but realistic perspective on the growth effects of AI. First, he reports that, as shown in the figure below, while innovative new technologies such as electricity, internal combustion engines, and computers have been introduced one after another over the past 150 years, the U.S. per capita GDP growth rate has remained surprisingly constant at an annual average of about 2%.

Why is that? It could be because finding new ideas has become more difficult over time, or because it has taken a long time for new technologies to spread throughout the economy and increase productivity. He predicts that in the case of AI, both an accelerating growth scenario and a business-as-usual scenario, which brings about growth of about 2% per year similar to the past, are possible, and that the reality of the future will likely be somewhere in between.

His paper presents a growth model based on various jobs that are difficult to replace one another, in which the overall production of the economy is determined by the weak link or bottleneck—the sector most difficult to automate. For example, although current computers are about 100 million times faster than they were in the early 1970s, the productivity of economists using computers has not increased dramatically because the people using them have not become as productive. If the economy's production function has such characteristics, the speed at which the weak link, or bottleneck, can be automated will determine the overall productivity of the economy.

Using this production function, his study shows that per capita income increases as the proportion of spending on automated jobs in the overall economy rises. However, assuming a low elasticity of substitution between jobs of 0.2, even if 50% of total GDP is automated, per capita GDP will increase by 19%, and 94% of GDP must be automated to double per capita income. He explains that income growth becomes very steep and can increase exponentially only after more than 90% of the entire economy is automated. This is because jobs that are still necessary for humans and difficult to automate exist in the economy, acting as an obstacle to raising overall income.

Professor Jones's model assumes low substitutability between jobs in the production function, resulting in weak links that are difficult to automate significantly constraining economic growth. However, the aforementioned report by Entropic estimates this figure higher at 0.5, reporting a much greater growth effect from AI than this study. In any case, if his assumption is correct, this implies that AI is merely an "ordinary technology" that will struggle to significantly boost productivity and growth for at least 10 to 20 years. Of course, rapid GDP growth would be possible in the long term if AI were to automate almost all jobs, but this might not be possible until the 2100s. Indeed, he poses the question of how long it would take for robots to safely replace teachers in kindergartens.

Nevertheless, can AI lead to explosive economic growth?

Contrary to these theoretical possibilities, other reasons are presented for why the productivity-boosting effects of innovative new technologies do not easily manifest in reality. A representative explanation is Professor Brynjolfsson of Stanford University's productivity J-curve theory. According to this theory, productivity can only increase after general-purpose technologies have spread throughout industry and society, and after continuous tangible and intangible complementary investments related to them have been made. Because this process takes a considerable amount of time, productivity growth initially stagnates upon the introduction of new technologies, but later rises rapidly to exhibit a J-shaped curve.

It is pointed out that , in fact, even general-purpose technologies of the past, such as electricity , took 40 to 50 years from their invention to the realization of increased productivity. This is because it took time to change the layout of factories and production methods that use electricity. He emphasizes that in the case of AI as well, complementary intangible capital investment—so-called organizational capital investment—is required to reorganize a company's organizational structure, management methods, and business models. However, as an optimist, he expects that, unlike past technologies, the productivity-enhancing effects of AI will become clearly apparent within the next five years.

Meanwhile, there is another theoretical possibility that AI could explosively boost productivity and economic growth. This involves AI automating research and development to significantly increase productivity and overcome the diminishing returns associated with discovering new ideas. Professor Korinek of the University of Virginia, who advises Entropic on economic research, suggests this possibility. As a leading researcher in recent AI economics, he co-authored the previous Entropic report with Professor Jones. His paper , published in April of this year , presents an economic model in which the so-called "singularity"—where economic growth explodes as R&D is automated by AI—emerges. The title of the paper is also "When Will Automated AI Research Create Explosive Growth?"

The authors demonstrate that explosive economic growth is possible through technical feedback, where the automation of one research sector enhances the productivity of others, and economic feedback, where increased output further promotes R&D and automation. Their model consists of three research sectors—software, such as AI algorithms; hardware, such as chips; and general research that generates new ideas—and an output sector that produces goods. According to simulation results, explosive economic growth is possible if all four sectors are 13% automated. Particularly surprising is that if software research is fully automated at 100%, explosive growth is possible within six years even if the other sectors are only 5% automated. Furthermore, this study points out that the bottleneck problem presented by Professor Jones can be overcome if the pace of automation is sufficiently fast. The figure below illustrates the degree of automation for each sector and the timing of achieving explosive growth in this paper.

In fact, if AI undergoes repeated 'recursive self-improvement' in the future to produce an artificial general intelligence that is significantly smarter than humans, humanity might achieve dream-like scientific and technological advancements. AI is already predicting protein structures, discovering new substances to contribute to new drug development, and solving difficult mathematical problems. This reality demonstrates that Professor Korinek's theoretical paper is not entirely without grounds. Of course, it remains to be seen how long this will take. I am not particularly optimistic; I suspect it will take decades more for such science fiction-like stories to become reality.

Greater Concern Than Economic Growth

Readers may have found it difficult to follow these complex and challenging economic studies. To summarize, there is considerable uncertainty regarding how much AI will boost productivity and growth in the future. Several factors are critical: first, how quickly and extensively AI will replace labor; and second, how much weak links in the overall economy will constrain production. Finally, the key lies in how quickly AI evolves to automate research and development and generate new ideas.

However, these theoretical studies in economics are not without limitations. While elegant economic models automatically derive economic growth forecasts, reality involves numerous factors to consider, such as workers' resistance to AI adoption and automation, as well as government policies. Furthermore, all of these growth models are supply-side models and do not account for issues of insufficient demand that may manifest as unemployment.

Above all, the faster human labor is replaced and the greater the increase in productivity, the more severe the problems of unemployment and inequality may become. Perhaps people are more concerned about these issues than economic growth. In the following article, I intend to examine the potential for mass unemployment and inequality caused by AI, as well as the recently raised risks of uncontrollable AI.

(This article was originally published as a column in Hankyoreh in Korean and translated into English with the help of Google Translate. The views expressed in this article are those of the author(s) and do not necessarily represent the official stance of the center.)