Frey and Osborne's "The Future of Employment" — The Paper That Put 47% of US Jobs at High Risk

The most quoted number in the debate over whether AI will replace jobs is 47% of US employment. It comes from a working paper dated 17 September 2013 by the Oxford economist Carl Benedikt Frey and the machine-learning researcher Michael A. Osborne, "The Future of Employment: How Susceptible Are Jobs to Computerisation?" The paper estimated, for 702 US occupations, the probability that the work could be done by computers, and placed about 47% of employment in a high-risk category with a probability of 0.7 or more.

The number is not a forecast that 47% of people will lose their jobs. The authors wrote that they made no attempt to estimate how many jobs would actually be automated. The 47% is the share of workers in occupations that are technically automatable, and the time frame was given as "some unspecified number of years, perhaps a decade or two". Almost every later argument about the paper turns on that distinction.

What the paper measured

The question was which jobs are how susceptible to computerisation. Economists had mostly held that routine tasks were the ones machines replace. Frey and Osborne argued that advances in machine learning and mobile robotics, once pattern recognition became possible, extended automation to non-routine tasks as well, and they described the limits as three "engineering bottlenecks".

The method had two steps. First, drawing on a workshop at Oxford's Department of Engineering Science, the authors and a group of machine-learning researchers hand-labelled 70 occupations as automatable (1) or not (0), choosing only the occupations about which they were most confident. Second, they took the nine variables above from O*NET, the US Department of Labor's occupational database (2010 version), trained a Gaussian process classifier on the relationship between those variables and the hand labels, and estimated a probability for all 702 occupations. Weighted by 2010 Bureau of Labor Statistics employment, occupations at 0.7 or more were classed as high risk and those below 0.3 as low risk.

The 47% and its caveats

About 47% of employment landed in the high-risk category. At the extremes of the paper's ranking, telemarketers, tax preparers and library technicians scored 0.99, while recreational therapists were lowest at 0.0028. In law, paralegals and legal assistants were high risk and lawyers low risk. The paper also reported a strong negative relationship between an occupation's probability of computerisation and its wages and educational attainment: the lower-paid and less-educated were the more exposed.

The paper spelled out its own limits:

• It did not estimate how many jobs would actually be automated. Whether labour-saving technology is adopted depends on factors left out of the model, such as access to cheap labour and the price of capital.
• It did not examine how productivity gains would raise or lower employment.
• It made no attempt to forecast changes in the occupational composition of the labour market.
• Because the hand labels were subjective, the model treated them as noisy measurements.

So 47% is closer to a ceiling on technical feasibility than to a projection of employment statistics. The paper was later published in Technological Forecasting and Social Change, volume 114 (2017).

The 9% rebuttal

The best-known rebuttal came in a 2016 OECD working paper by Melanie Arntz, Terry Gregory and Ulrich Zierahn. Their point was that treating whole occupations as the unit of automation overstates the risk. People with the same job title do different tasks, and many workers in "high-risk" occupations spend much of their time on tasks that are hard to automate.

Allowing for that variation within occupations, they re-measured 21 countries and found that 9% of jobs were automatable on average, from 6% in Korea to 12% in Austria. They attributed the differences to how workplaces are organised, past investment in automation and workers' education.

Because the metrics differ, the three numbers cannot simply be compared. Japan's 49% is covered on the Nomura Research Institute page.

After generative AI

The 2013 paper was mainly about robots and pattern recognition. When ChatGPT arrived in 2022 and machines began writing, summarising and coding, the 2013 categories were questioned again.

In September 2023 Frey and Osborne published a reappraisal for the generative-AI era. According to the Oxford Internet Institute's summary, they accepted that generative AI widens the range of automatable tasks, but argued that firms would keep humans in the loop while AI hallucinations persist. Frey said that "if a task can be done remotely, it can also be potentially automated", while work that depends on long-standing in-person relationships would stay with humans, and Frey said that, in the absence of major breakthroughs, they expected the bottlenecks outlined in 2013 to keep constraining automation for the foreseeable future. They did not revise the 47% figure.

From 2023 a wave of estimates focused specifically on generative AI. The best known is Goldman Sachs's "300 million jobs", which measures how much of each occupation's workload is exposed. Research that uses actual payroll data to look at young workers is collected on the "canaries in the coal mine" page.

What remained

The paper's social impact lay less in the accuracy of its number than in the frame it set. The question "what share of jobs will automation affect?" became standard, and bodies such as the OECD and Japan's Nomura Research Institute produced their own estimates. The OECD rebuttal also showed that the framing itself, whether whole occupations disappear or tasks within them shift, moves the estimate a long way.

The finding that the lower-paid and less-educated face the highest probabilities points to a longer-running question: who carries the cost of technological change. For the wider sequence of events, see the AI and society timeline.