Analysis · 23 September 2026 · 6 min read

The 47% study: what it actually said

The 2013 Oxford paper behind the 47 percent figure never estimated how many jobs would be automated. What it measured, and why others found 9 percent.

The claim that 47 percent of US jobs are at risk of automation comes from one paper: The Future of Employment: How Susceptible Are Jobs to Computerisation? by Carl Benedikt Frey and Michael Osborne of Oxford, dated 17 September 2013. It is probably the most quoted number in writing about AI and work. It is also routinely quoted as something it never claimed to be.

What the paper did

Frey and Osborne took 702 occupations from the US O*NET database. They hand labelled 70 of them as automatable or not, then trained a classifier on O*NET job characteristics to estimate a probability of computerisation for all 702. Occupations were then grouped into low, medium and high risk.

Weighting by employment, about 47 percent of US jobs fell into the high risk group.

What it did not claim

The paper is explicit about this, and the sentence rarely travels with the number. The authors write that they "make no attempt to estimate the number of jobs that will actually be automated", and that they focus on potential automatability "over some unspecified number of years". Elsewhere they describe the high risk jobs as ones that could be automated "perhaps over the next decade or two".

So 47 percent is a share of employment sitting in occupations a model rated as technically automatable. It is not a prediction that 47 percent of people would lose their jobs, and the authors did not present it as one.

Why other researchers got 9 percent

In 2016 Melanie Arntz, Terry Gregory and Ulrich Zierahn ran the question again for the OECD, at the level of tasks rather than whole occupations. Workers in the same occupation do different mixes of tasks, and most do some things that are hard to automate. Counting only jobs where nearly every task could be automated, they found about 9 percent of jobs at high risk across OECD countries, with the US close to that average.

Same underlying question, a different unit of analysis, a number five times smaller. Most research since has followed the task based approach.

What the 2013 paper expected, and what happened instead

Frey and Osborne named three bottlenecks they thought would hold for a decade or two: perception and manipulation, creative intelligence and social intelligence. Their model put most workers in transportation, logistics and office support in the high risk group.

Generative AI arrived through a different door. Its first measurable use in the workplace is in text, code and conversation. In Anthropic's 2026 measure of observed use, the most exposed occupations are computer programmers and, second, customer service representatives, a job built on talking to people. Truck driving, part of the transportation group the 2013 model put at high risk, sits in our lower risk band today: Heavy and Tractor-Trailer Truck Drivers (19).

None of this means the paper was careless. It was written a decade before today's language models, and it said clearly what it was measuring. The problem is the way its headline number kept circulating long after the technology it described had changed.

How current research differs

Our index does not use the 2013 paper. It aggregates 20 publications from 2023 to 2026 across 961 occupations, all of them published since generative AI arrived. Each occupation page lists the studies behind its score. How the scores are built explains the weighting, and why exposure studies disagree covers the split between language model exposure and broader automation potential.

If you want a figure rather than a history, how many jobs will AI replace sets the main estimates side by side. If you want your own, the assessment scores your role and your mix of tasks in about two minutes.

Figures come from the AI Job Risk Index and were current when this was published. Scores change as new research is ingested, so the index is always the live version. See how scoring works. Informational guidance based on published research, not professional career or financial advice.