The spread of generative AI is astonishing. Just a few years ago, tasks such as data search, organization, translation, coding, and PPT creation that required several people to spend days on were now handled by one person with the help of AI in a short time. In fact, empirical results showing that AI improves efficiency continue to emerge. A study by Eric Brynjolfsson of Stanford University and Daniel R. of MIT analyzing over 5,000 customer service representatives found that the number of cases processed per hour increased by an average of 14% after the introduction of generative AI. Another study from the MIT Department of Economics reported that document creation time for professionals using ChatGPT decreased by an average of 40%, while the quality of results improved by 18%.
However, these effects are not clearly visible in the productivity of the economy as a whole. This is the so-called AI productivity paradox. The U.S. Federal Reserve also noted last month that while AI performance and investment are expanding rapidly and corporate utilization is increasing, it is still difficult to see their effects widely reflected in overall economic productivity and the labor market.
This phenomenon is not new historically. When computers were rapidly adopted by businesses and households in the 1980s, the rate of productivity growth did not increase as expected. U.S. economist Robert Solow pointed out at the time that although computers spread quickly throughout society, their effects were hard to find in productivity statistics. So why does AI not show up well in productivity statistics? Three reasons can be considered.
First is the issue of illusion. One must not mistake improvements in efficiency for individual tasks as improvements in productivity for the entire company or economy. Even if AI cuts document creation time in half, if the review and decision-making processes remain unchanged, the organization's final output may not increase accordingly. The same applies when time and personnel saved by AI are not utilized for high-value-added work. Ultimately, one should not judge the productivity effects of AI based solely on efficiency improvements observed in specific tasks. Efficiency gains at the task level must lead to changes in overall workflow and resource allocation before they can translate into meaningful productivity improvements at the corporate level.
Second is the issue of time lag. It takes time for the effects of AI seen in individual tasks to lead to improved productivity across the entire company. This is because data must be accumulated to match new technologies, employees must be retrained, and business processes and organizational structures must be redesigned. Professor Eric Brynjolfsson emphasized that complementary intangible investments are necessary to fully harness the potential of general-purpose technologies like AI. Such investments incur costs first, and their results appear only after the organization adapts and restructures. Therefore, in the early stages of adopting new technologies, productivity may stagnate or even decline before rising later; this is known as the productivity J-curve.
Third is the issue of measurement. It is possible that AI has already increased productivity, but we are failing to measure it properly. Consider translating documents or conducting research that used to be entrusted to translators or research firms now being handled directly in minutes using AI. While users' time and costs decrease significantly, the services traded in the market may actually decline. Since benefits such as time savings, quality improvements, and self-production created by AI are not sufficiently reflected in GDP, total factor productivity may be underestimated even if expanded AI investment leads to actual productivity gains.
Therefore, one should not immediately conclude that AI has no effect just because productivity improvements are not clearly visible despite its adoption. It must be distinguished whether efficiency gains from individual tasks through AI have not yet translated into overall organizational performance, whether complementary investments such as business redesign and workforce retraining have not yet been sufficiently made, or whether productivity is already rising but existing statistics fail to capture it properly. This is because the necessary responses differ depending on the cause.
in OPINION
