Robots and employment
'47% of jobs at risk': what Frey and Osborne actually measured
47% of US jobs are at high risk of automation, one 2013 study estimated. It measured how susceptible jobs are to computerisation, not how many will actually disappear, and that difference is where almost every citation goes wrong.
The estimate is Carl Benedikt Frey and Michael Osborne's, at Oxford. Using a model over 702 occupations, they scored each for its probability of computerisation and summed the employment in the occupations scoring above 0.7, over what they called the next decade or two.
This page traces the 47% to the paper itself, shows what the number is and is not, and sets it beside other credible estimates that are far lower.
Data covers Frey and Osborne, "The Future of Employment" (Oxford, 2013). Last reviewed by a human editor before publication.
The figures and where they come from
Each figure is rated for how safely you can cite it today. Ratings judge current usability, not whether a number was ever correct.
| Figure | What it is | Source | Citation Confidence | Notes |
|---|---|---|---|---|
| about 47% | US employment in the high-risk category | [A] | High | The headline number. It is the share of employment in occupations the model rates as potentially susceptible to computerisation, not a forecast of jobs that will be automated. |
| 702 occupations | Occupations analyzed | [A] | High | The model estimated a probability of computerisation for each of 702 detailed occupations. |
| probability above 0.7 | High-risk definition | [A] | Medium | 'High risk' means a modeled probability of computerisation above 0.7. It is a susceptibility score, not a schedule. |
| next decade or two | Time horizon | [A] | High | The paper frames the risk over 'the next decade or two,' an explicitly unspecified window, not a fixed deadline. |
| 23% of work hours | A different estimate (McKinsey) | [B] | Medium | The McKinsey Global Institute estimated 23% of US work hours could be automated by 2030. A different method gives a very different number. |
| 9% of workers | A lower estimate (Arntz et al.) | [B] | Medium | Researchers Arntz, Gregory, and Zierahn estimated 9% of US workers hold jobs at high risk, a fifth of the Frey-Osborne figure, by accounting for task variation within occupations. |
Why the numbers disagree
The 47% is misread because a probability gets turned into a prediction. Frey and Osborne estimated how susceptible occupations are to computerisation, expressed as a probability, and summed the employment in the occupations scoring above 0.7. That is a measure of exposure, not a forecast that 47% of jobs will be automated, and certainly not by a specific year.
The headline also drops the time frame and the conditionals. The paper speaks of the next decade or two, an explicitly loose horizon, and describes what is technically susceptible, not what will be automated once you factor in cost, regulation, and whether firms actually choose to automate.
Other credible estimates are much lower because they count differently. McKinsey put 23% of work hours as automatable by 2030, and Arntz, Gregory, and Zierahn found 9% of US workers at high risk by looking at tasks within occupations rather than whole occupations. The spread from 9% to 47% is a sign these are model outputs with different assumptions, not a single established fact.
How to cite these figures
Cite the 47% as Frey and Osborne's 2013 estimate of the share of US employment in occupations at high risk of computerisation over the next decade or two, and say it is a susceptibility estimate, not a prediction of jobs lost.
If the claim needs a number people can defend, present the range: 9% (Arntz and colleagues), 23% of work hours (McKinsey), and 47% (Frey and Osborne), and explain that they count different things.
Never attach a hard deadline. The original horizon is deliberately vague, and 'by 2033' or similar dates are added by others, not by the study.
Where people go wrong
Saying 47% of jobs will be automated, or will be gone by a given year. The study estimated susceptibility, over a loose horizon, not job losses on a schedule.
Citing 47% as the consensus. Credible estimates range from 9% to 47% depending on method; 47% is the high end of one approach.
Attributing the figure vaguely to 'a study' or 'researchers.' It is Frey and Osborne, 2013, Oxford, and naming it lets a reader check what it actually claims.
How we checked
We traced the figure to the original 2013 Frey and Osborne paper and confirmed the exact claims in its text: about 47% of total US employment in the high-risk category, a model over 702 occupations, the 0.7 probability threshold, and the 'next decade or two' horizon. We did not rely on any secondary restatement for the core number, because the entire point is what the primary actually said.
For the contrasting estimates, we used a US Government Accountability Office report that assembles the McKinsey and Arntz figures alongside Frey and Osborne, and confirmed each appears in its text. Citing a GAO compilation keeps the comparison in one auditable, government-published place.
The Citation Confidence ratings reflect how safely each figure can be quoted as what it is. The 47%, the 702 occupations, and the horizon are High because they are stated plainly in the source. The threshold and the contrasting estimates are Medium because they require the surrounding definitions to be quoted correctly.
Full source list
Primary sources, with live links. Every figure above traces to one of these.
- [A]University of Oxford2013
Carl Benedikt Frey and Michael A. Osborne, "The Future of Employment: How Susceptible Are Jobs to Computerisation?" (Oxford Martin School)
https://www.oxfordmartin.ox.ac.uk/downloads/academic/The_Future_of_Employment.pdf - [B]U.S. Government Accountability OfficeMarch 2019
U.S. Government Accountability Office, GAO-19-257, "Workforce Automation: Better Data Needed to Assess and Plan for Effects of Automation on Jobs"
https://www.gao.gov/assets/gao-19-257.pdf
Common questions
- Where does '47% of jobs at risk' come from?
- A 2013 study by Frey and Osborne at Oxford. They estimated that about 47% of US employment is in occupations at high risk of computerisation, defined as a modeled probability above 0.7, over the next decade or two.
- Does it predict 47% of jobs will be automated?
- No. It estimates susceptibility, expressed as a probability, not a forecast of jobs that will actually be automated, and it gives no fixed deadline. Reading it as a prediction is the most common error.
- Do other studies agree?
- No. Estimates range widely by method: about 9% of workers (Arntz and colleagues), 23% of work hours by 2030 (McKinsey), and 47% (Frey and Osborne). They count different things, which is why they disagree.
- Is there a deadline for the 47%?
- Not in the original. Frey and Osborne used 'the next decade or two,' an explicitly loose window. Specific years attached to the figure are added by others.
More data, traced to source
- 'One robot destroys 5.6 jobs': what the study really measured
The viral statistic that one robot destroys 5.6 jobs comes from a real, careful study. But the 5.6 figure is a local-labor-market estimate, not a claim that each robot nationwide eliminates 5.6 jobs. Here is what the researchers actually measured.
- How many US firms actually use robots? Far fewer than you would think
The story is that everyone is automating. The US Census Bureau's own firm survey found only about 2% of US firms use robotics. A newer Census release puts it far higher. The gap is the whole story.
- The 2.1 million missing manufacturing jobs: what that number actually is
The claim that U.S. manufacturing will have 2.1 million unfilled jobs by 2030 is everywhere. It comes from a 2021 survey of about 800 executives, not a government projection, and it is usually mis-attributed. Here is exactly what the number is.
- 9%, 23%, or 47% of jobs at risk? The government says nobody knows
Estimates of how many jobs automation threatens range from 9% to 47%. The US Government Accountability Office reviewed them and concluded the underlying data is not good enough to say. The disagreement is the finding.