Since the release of the conversational artificial intelligence (AI) “ChatGPT” by the U.S.-based company OpenAI in November 2022, the use of generative AI has been expanding. According to a recent study by Harvard University professor David Deming and a team at Open AI, 10% of the global population uses ChatGPT at least once a week, centering on highly educated professionals.
The study team analyzed anonymized data and found that about 30% of the recent use is work-related. Of generative AI users, 80% use generative AI to gain practical advice, search for information, and for writing tasks. Most generative AI users use generative AI as a writing tool.
◆◆◆
Interest is growing in how generative AI, which is increasingly being used in this way, will affect employment, labor productivity, and wages. There are roughly three categories of research in this field (see the figure).
The first category uses occupational databases to compute generative AI exposure by occupation and analyze what percentage of all workers are likely to be affected by generative AI.
The second category tests a group of workers who use generative AI and a group of those who do not for a specific occupation to analyze how productivity, including quality, changes in each group.
The third category uses data from across the economy to analyze how generative AI use has affected employment, working hours, productivity, and wages.
The first category analyzes a combination of two occupational databases: one that lists tasks performed in various occupations, and one in which experts and generative AI determine which tasks can be replaced by generative AI.
The most famous research in this category to date is a study published in 2013 by Carl Frey and Michael Osborne of Oxford University, which concluded that 47% of jobs in the United States would be replaced by AI and other machines within 10 to 20 years. Although the findings shocked the public, there have been no massive technology job losses since then, leading to a growing recognition that the technical possibility of a job being replaced by generative AI technology does not mean that the job will disappear.
This category of research has been continuously updated, and studies that limit the target technology to generative AI have been published by the Organization for Economic Co-operation and Development (OECD) and the International Labor Organization (ILO).
According to a report released by the ILO in May 2025, a quarter of workers worldwide are exposed to generative AI. It revealed that office workers are the most exposed, high-income groups are more exposed than others, and women being more exposed than men.
The second category of research randomly assigns workers in a specific occupation as test subjects to a group of those who use generative AI and a group of those who do not and compares changes in productivity in the two groups after the experiment.
There are two particularly well-known studies in this category. One is a study by Shakked Noy and Whitney Zhang of the Massachusetts Institute of Technology, which examined how the use of generative AI changed the speed and quality of writers’ work. The results demonstrated that the use of generative AI reduced writing time and improved manuscript quality, resulting in productivity gains. As the effect was found to be greater for lower-skilled workers, the study concludes that generative AI has a leveling effect in terms of worker productivity.
Eric Brynjolfsson of Stanford University and his colleagues found that using generative AI assistance in customer centers in the Philippines reduced the time taken to resolve problems and improved customer satisfaction. Results also showed that the effect was concentrated among beginners, demonstrating that generative AI had the effect of leveling productivity among workers.
Previous to generative AI research, a research team from the University of Tokyo, including the author, through a non-experimental study, demonstrated that installing AI navigation systems in taxis that provided drivers with routes where passengers are more likely to be found improved productivity and that the improvement effect was concentrated among drivers with initially lower productivity, which revealed this trend ahead of other research on generative AI.
Given that AI learns from models and applies them to all workers, the effect of leveling productivity can be considered a natural result. These studies on workers in specific occupations show that the introduction of generative AI has the effect of eliminating productivity differences among workers and narrowing gaps among them.
The third category of research uses data from the whole economy, which includes various occupations, to examine how the use of generative AI affects employment, productivity, and wages. Although research in this category is still limited, I would like to introduce two relevant papers published in 2025. However, it should be noted that these papers have not undergone peer review or been evaluated objectively.
The first paper represents a study in Denmark by Anders Humlum of the University of Chicago and Emilie Vestergaard of the University of Copenhagen. They combined the results of a questionnaire survey asking whether generative AI is used for work with individual-level administrative panel data. They analyzed changes in income and working hours after the release of ChatGPT in November 2022, comparing a group using generative AI with a group that did not in 2023 and 2024. They have found that the use of generative AI did not affect income or working hours.
On the other hand, a U.S. study by Andrew Johnston of the University of Texas and Christos Makridis of Arizona State University found that employment and wages have increased since the advent of ChatGPT in industries and states where jobs with high generative AI exposure are concentrated.
Furthermore, the study separated substitutive exposure, in which generative AI reduces working hours without human assistance, from complementary exposure, where generative AI reduces working hours with human assistance and found that employment and wages have decreased in industries and states where substitutive exposure is higher, while increasing in those where complementary exposure is higher.
The impact of generative AI on wages differs from study to study, indicating that no definite findings have been established. However, the finding that productivity, jobs, and wages have increased in occupations that are complemented by generative AI seems plausible.
◆◆◆
Based on the above knowledge, I would like to forecast the impact of generative AI on the future labor market. The use of generative AI reduces productivity gaps between workers in the same occupation. However, given that highly specialized occupations tend to be those with high exposure to generative AI and that generative AI improves productivity, disparities between occupations are expected to widen.
In occupations to which generative AI is complementary, worker productivity and wages are likely to increase. The overall impact on the labor market may depend on whether the effect of narrowing disparities within each occupation is greater than the effect of widening them between occupations. Given that wage disparities between occupations explain a large part of general wage disparities, it seems likely that the latter effect will be larger. Due attention should also be paid to new inequalities that generative AI may create.
>> Original text in Japanese
* Translated by RIETI.
October 6, 2025 Nihon Keizai Shimbun
