Nomura Research Institute's "49% of Japan's Workforce Replaceable" — Where Japan's AI-and-Jobs Debate Began

The number that launched Japan's version of the "AI will take our jobs" debate is 49% of the Japanese workforce. On 2 December 2015 Nomura Research Institute (NRI), in joint research with Associate Professor Michael A. Osborne and Dr Carl Benedikt Frey of the University of Oxford, released an estimate of the probability that each of 601 Japanese occupations could be replaced by artificial intelligence or robots. Its conclusion: in 10 to 20 years, about 49% of Japan's workforce would be in occupations that could technically be replaced.

The 49% is not a forecast that half the workforce will lose their jobs. NRI itself noted that the figure describes technical replaceability by computers only, and that it had not considered the social and economic factors that decide whether replacement actually happens. Knowing what the number counts is the starting point for reading it.

What was estimated, and how

The method applied Frey and Osborne's 2013 US analysis to Japan. The method itself, and the 47% figure it first produced, are covered on the Frey and Osborne page. Here the focus is on what differed in the Japanese version.

JILPT's data came from surveys that quantified each occupation's interests, values, work environment, skills and knowledge; the 601 occupations were those for which at least 30 responses had been collected. Where the US study used the US Department of Labor's O*NET, the Japanese version swapped in Japanese occupational data.

In its note on method, NRI stated two limits:

• Whether replacement actually happens is expected to depend heavily on social and economic factors, including labour supply and demand, but the estimate did not consider them.
• It did not examine the probability that a computer would take over only part of one person's work.

Most and least replaceable occupations

NRI appended two reference lists of 100 occupations each, with high and low replaceability, sorted in Japanese syllabary order rather than by probability.

NRI summarised the pattern this way: occupations that require knowledge for organising and creating abstract concepts, as in art, history, archaeology, philosophy and theology, or that require cooperation with others, understanding, persuasion, negotiation and a service orientation, tend to be hard to replace; occupations that need no particular knowledge or skill, or that involve data analysis and orderly, systematic operations, tend to be replaceable.

Background to the estimate

The estimate was one strand of a research programme at NRI's Mirai Sohatsu (future creation) centre built around the theme of thinking about Japan from 2030 and preparing for it now. NRI described the aim as studying the social impact of using AI and robots to supplement the workforce in a Japan whose labour force was expected to shrink as the population declined. The starting question, in other words, was less about jobs being taken than about how to fill labour shortages.

The release came with lists of named occupations, and announced a research lecture on 12 January 2016 with Osborne and Associate Professor Yutaka Matsuo of the University of Tokyo. The 49% thus entered public debate in a form that could be read both as a threat to employment and as a measure of how much labour technology might supply.

How to read the number

Later work added two cautions to estimates of this kind.

First, treating whole occupations as the unit of replacement makes the effect look large. A 2016 OECD working paper by Arntz, Gregory and Zierahn found that, allowing for differences in tasks among people in the same occupation, only 9% of jobs were automatable on average across 21 countries. NRI itself had noted that it did not consider partial replacement.

Second, "substitution" and "complementarity" are not cleanly separable. The Cabinet Office's July 2024 report on the world economy (Sekai Keizai no Choryu 2024 I) said one must distinguish technology that replaces human work outright from technology that assists workers and raises productivity, but added that even in complemented occupations some tasks become automated and need fewer people, so the difference is smaller than the words suggest. The same report summarised Eloundou et al. (2023), whose authors included OpenAI researchers, as estimating that about 80% of US workers could have at least 10% of their tasks, and about 19% at least half, exposed to automation by large language models.

Japan's labour market ten years on

The estimate's horizon runs to roughly 2025–2035. For 2025, the Ministry of Internal Affairs and Communications' Labour Force Survey reported an average unemployment rate of 2.4%, unchanged from the year before, and employment up by 600,000 to 67.24 million. The aggregate figures, at least, show no sign of mass replacement.

That alone does not prove the estimate wrong. NRI's figure was never an unemployment forecast, and changes in what people do inside their jobs do not necessarily appear in unemployment or employment totals. Estimates made after generative AI arrived are on the Goldman Sachs page, and research using real employment data on young workers is on the "canaries in the coal mine" page.

The wider sequence of events is on the AI and society timeline.