Artificial intelligence has diffused rapidly in recent years, but its impact on aggregate productivity is not clear. This column uses data from original surveys to estimate the impact of AI on macroeconomic productivity in Japan. Using data on the use of AI in the workplace and its impact on work efficiency, it estimates a 0.5-0.6% boost to labour productivity at the macro level. The productivity impact is likely to grow over time, but with diminishing additional gains. Furthermore, higher adoption of AI among high-wage and highly educated workers may widen overall labour market inequality in the near term.
Distinguished Senior Fellow Research Institute of Economy, Trade and Industry (RIETI)
With the rapid diffusion of artificial intelligence (AI), its impacts on productivity and the labour market have attracted attention. Many studies have been conducted on the impacts of industrial robots on productivity (e.g. Graetz and Michaels 2018, Kromann et al. 2020, Cette et al. 2021, Dauth et al. 2021) due to the availability of International Federation of Robotics (IFR) data on robot utilisation by country and industry. However, the quantitative impact of AI on productivity is not yet well understood, mainly due to a lack of statistical data on the use of AI.
Recently, several studies have reported findings from randomised experiments on specific tasks in which AI has a large positive effect on productivity (e.g. Brynjolfsson et al. 2023, Kanazawa et al. 2022, Noy and Zhang 2023, Peng et al. 2023). These studies are valuable contributions that reveal the causal effect of AI on productivity, but it is impossible to infer macroeconomic impacts from these results because the studies only cover the very narrowly defined tasks of customer support, taxi driving, writing tasks, and software programming.
Acemoglu (2024) estimates the medium-term effect of AI on productivity in the US as the percentage of tasks affected by AI multiplied by task-level cost savings based on these existing task-level studies. According to his study, the macroeconomic impact of AI is non-negligible but small, with a cumulative total factor productivity (TFP) increase of less than 0.7%. However, he noted that there is huge uncertainty about which tasks will be automated, and what the cost savings will be. More recently, Filippucci et al. (2024) assess the aggregate productivity gains from AI assuming cost savings from AI as 30%, and state that AI could contribute 0.25–0.6% points to annual TFP growth in the US over the next decade. 1
Against this background, I provide an overview of the use of AI and estimate its impact on macroeconomic productivity in Japan, using data from original surveys (see Morikawa 2024a, 2024b for details).
Research design
I conducted surveys in September 2023 and October 2024 targeting Japanese workers aged 20 and older selected to be representative of the Japanese workforce. The number of respondents to the 2023 survey was 13,150. The follow-up survey in 2024 was sent to those who responded to the 2023 survey and 8,633 responded. Data from 8,269 of these respondents, excluding those not working as of the 2024 survey, are used in the analysis.
The main survey items are (1) the use of AI (including generative AI) at work, (2) the percentage of tasks performed using AI, and (3) the effect of AI use on work efficiency. The second and third questions are only posed to those who answered that they use AI at work. The survey also collects information on the respondents’ gender, age, educational background, industry, occupation, type of employment, weekly working hours, and annual earnings from work.
Based on the answers to these questions, the percentage of workers who use AI for their work (AI_User), the percentage of tasks using AI (AI_Taskshare), and the efficiency gains (AI_Efficiency) are tabulated. The worker-level productivity effect of AI (AI_Productivity) is calculated for AI users as AI_Taskshare*AI_Efficiency. For example, if a worker uses AI for 30% of his/her tasks and the efficiency effect of AI is 20%, the overall productivity of his/her work is calculated to be 6% higher than when he/she did not use AI. To calculate the impact of AI on macroeconomic productivity, AI_Productivity is aggregated by using annual earnings as weight which is divided by the total of annual earnings including those who do not use AI.
Although the approach adopted here is extremely simple and depends on the subjective evaluations of workers, where measurement errors are inevitable, this approach has the advantage of avoiding endogeneity concerns arising from selective use of AI because the survey asks AI users about efficiency gains in comparison with the situation in which AI is not used.
Productivity impacts of AI
The number of AI_User in the fall of 2024 was 8.3%. In the 2023 survey, the corresponding figure was 5.8% (5.3% when limiting the sample to panel respondents who also responded to the 2024 survey), indicating that the number of AI-using workers has increased by about 1.5 times in the past year. 2
The percentage of tasks that use AI (AI_Taskshare) among AI_User has a mean value of 15.1%. In other words, even when AI is used for work, the percentage of tasks that do not use AI is more than 80%, on average. The mean value of using AI on work efficiency (AI_Efficiency) is 25.9% and the mean of AI_Productivity is 5.6%, 3 meaning that the overall productivity of workers using AI for their jobs is 5.6% higher than without AI.
The macroeconomic productivity impact calculated by weighting AI_Productivity by annual earnings and using the total of annual earnings of all respondents as the denominator is +0.58% (see Figure 1). Therefore, at this point in time, our preferred estimate is a 0.5-0.6% boost to labour productivity at the macro-level compared to the case without AI. The effect on total factor productivity (TFP) is about 0.3% if the labour productivity is converted to TFP using the labour share (0.535) used by Acemoglu (2024).
Figure 1 Macroeconomic impact of AI on labour productivity
Possible impact in the future
The percentage of respondents who answered that “I do not currently use AI at work, but I think I will in the future,” is about 28%, suggesting that the use of AI for work will continue to increase and that the macroeconomic impact of AI is likely to increase in the future. Assuming that AI_Taskshare and AI_Efficiency are the same as that of current AI users, the macroeconomic impact on labour productivity would be about four times greater: about 2% higher than the case without AI. The effect on total factor productivity by taking into account the labour share is about 1.1%. 4
However, the additional productivity gain may diminish gradually. Since the 2024 survey is conducted for respondents to the 2023 survey, it is possible to disaggregate AI users into those who have newly started using AI during the past year from those who already used AI in 2023. Figure 2 summarises the comparison of these two categories of AI users. Both AI_Taskshare and AI_Efficiency are significantly lower for those who have newly started using AI than for those who have continuously used AI. As a result, there is a significant difference in the effect of AI use on overall work productivity: The means of AI_Productivity for continuous AI users and new AI users are 7.8% and 4.4%, respectively. This result suggests that the diffusion of AI started with jobs for which the effect of AI implementation is large and has gradually spread to jobs for which its effect is small. If these trends continue, the additional contribution of AI to macroeconomic productivity may diminish gradually as the number of AI users increases.
Figure 2 Comparison of new AI users and continuous AI users
Notes: The bars indicate 95% confidence intervals. The figures in the bars are the means. New AI users are those who started using AI at work in the past year.
Impacts of AI on labour market inequality
The calculation results, disaggregated by education and annual earnings, are reported in Table 1. Highly educated and high-wage workers tend to use AI at work. On the other hand, differences in AI_Taskshare and AI_Efficiency by worker characteristics are limited. In other words, although less-educated and low-wage workers are notably less likely to use AI in their jobs, the productivity effects are not much different (or somewhat higher) when they do use AI for work.
Figure 1 shown earlier also indicates the productivity impacts at the aggregate level by workers’ education and annual earnings categories. The productivity impacts are larger for highly educated and high-wage worker categories, mainly because of the higher AI adoption rate (extensive margin). The aforementioned recent studies on specific tasks have shown that productivity effects of AI are greater for relatively less skilled workers within the same task. However, the finding from my study suggests that diffusion of AI may widen overall labour market inequality, at least in the near future.
Table 1 Use of AI and its effect by education and annual earnings
Author’s note: The main research on which this column is based (Morikawa 2024b) first appeared as a Discussion Paper of the Research Institute of Economy, Trade and Industry (RIETI) of Japan.
Source: cepr.org