Why AI risk is measured in tasks, not job titles
Short answer
Because the unit you count decides the answer you get. In 2013 Frey and Osborne scored whole occupations and put 47% of United States employment in their high-risk band. In 2016 Arntz, Gregory and Zierahn used task-level survey data on what workers actually report doing, and put the United States figure at 9%. Same country, same underlying assessment of what machines can do, different unit of counting. The second team's explanation is the whole argument: even inside a "high risk" occupation, people spend much of the week on things that are hard to automate.
The 47% and the 9%
The 47% is the most quoted number in this field. It comes from an Oxford working paper that scored 702 occupations covering 138.44 million United States jobs, using 2010 employment as its baseline, and placed those above a model probability of 0.7 in a high-risk band.1
The authors, working with a group of machine-learning researchers, hand-labelled 70 of the 702 occupations as automatable or not, drawing on the Oxford Machines and Employment workshop and on the O*NET task and job descriptions, then trained a classifier on those labels to score the rest.1 A label of 1 was given only where they judged the whole occupation automatable: in their words, "we only assigned a 1 to fully automatable occupations, where we considered all tasks to be automatable".1
Three years later a team at ZEW Mannheim re-ran the question using individual-level survey data on what workers actually report doing, rather than treating an occupation as a single block. For the United States they found "only 9% of jobs rather than 47%" at high automatability.2 Their explanation: "even in occupations that FO expect to be at a high risk of automation, people often perform tasks which are hard to automate".2
Nedelkoska and Quintini later ran a larger version through the same working-paper series, covering 32 countries with a bigger individual-level sample, and put the high-risk share at 14% across those countries and about 10% for the United States.3 Their 14% sits above the earlier 9% partly by design, because they include workers who lack basic computer skills or work in jobs needing no computer, a group at higher risk.3
They also count a middle band the earlier work did not report: another 32% of jobs at an automation risk between 50% and 70%, which they read as jobs where a significant share of tasks, but not all, could be automated. Counting both bands, their own summary is that close to one in two jobs are likely to be significantly affected.3 That middle band is the tasks-not-titles argument stated in the automation literature's own words: some tasks go, the job changes, and it does not disappear.
Things both papers say that almost nobody quotes
Frey and Osborne did not predict that 47% of jobs would be automated. The paper states: "We make no attempt to estimate the number of jobs that will actually be automated."1 It gives no calendar date either, but it is not time-agnostic: the authors describe the high-risk jobs as ones they "expect could be automated relatively soon, perhaps over the next decade or two".1 So the familiar "47% within 10 to 20 years" headline is a hardening of the authors' own hedge into a deadline, rather than something invented by the people quoting them.
The 9% carries an equally firm warning. Its authors write that the estimated share of jobs at risk "must not be equated with actual or expected employment losses from technological advances", and they add that their own figure probably runs high, because like Frey and Osborne it measures what experts judge technically possible rather than what is actually adopted.2 It is also a working paper, not an institutional position: the correct attribution is to Arntz, Gregory and Zierahn, not to "the OECD".2
Why your job title hides your work
A job title is an administrative label. Two people who share one can spend their weeks very differently: a consultant who runs workshops and a consultant who builds slide decks face genuinely different overlap with AI, and no occupation-level score can tell them apart.
That is why every occupation page on this site shows the full task table, with each activity's share of the role, rather than stopping at the headline score, and why the score is built from work activities rather than job descriptions. It is also why the personalisation questions exist: they adjust for the things an occupation average cannot see, such as whether you are physically present, whether you sign off on outcomes, and how routine your week actually is.
If your title fits you badly, the task self-assessment skips the title entirely and works from how your week is actually divided.
What this means for you
When you meet a dramatic percentage in a headline, ask one question before anything else: a percentage of what? Whole occupations judged fully automatable, tasks within occupations, workers in occupations above a threshold, and conversations with a chatbot are four different objects, and their numbers cannot be compared or averaged. The five-fold gap between 47% and 9% is not a disagreement about AI. It is a disagreement about counting.
Next: what AI is measurably good and bad at, using the activity-level data directly.
- Carl Benedikt Frey and Michael A. Osborne, The Future of Employment: How Susceptible Are Jobs to Computerisation?, Oxford Martin School, University of Oxford, 17 September 2013; published in Technological Forecasting and Social Change vol. 114 (2017), pp. 254-280. Full text (PDF). Quoted: 47% of total United States employment in the high-risk band at probability 0.7 or above; 702 occupations covering 138.44 million jobs on a 2010 baseline; 70 occupations hand-labelled; the "fully automatable" labelling rule; the statement that the authors make no attempt to estimate jobs that will actually be automated; and the hedge that high-risk jobs are ones they expect could be automated relatively soon, perhaps over the next decade or two.
- Melanie Arntz, Terry Gregory and Ulrich Zierahn, The Risk of Automation for Jobs in OECD Countries: A Comparative Analysis, OECD Social, Employment and Migration Working Papers No. 189, 2016. oecd.org. Quoted: 9% for the United States against 47%; the explanation that workers in high-risk occupations still perform hard-to-automate tasks; the warning that the figure must not be equated with actual or expected employment losses; and the authors' statement that their approach reflects expert-assessed technological capability rather than actual utilisation, and may therefore overestimate. OECD working papers carry an explicit disclaimer that they do not represent official OECD views.
- Ljubica Nedelkoska and Glenda Quintini, Automation, Skills Use and Training, OECD Social, Employment and Migration Working Papers No. 202, 2018. oecd.org. A larger task-based estimate across 32 countries: about 14% of jobs at high risk of automation across the countries studied, and about 10% for the United States. The same summary places another 32% of jobs at a risk between 50% and 70%, described as significant change to how the job is done rather than its disappearance, and states that close to one in two jobs across the 32 countries are likely to be significantly affected by automation. The authors attribute part of the gap between their 14% and the earlier 9% to their coverage of workers who lack basic computer skills or work in jobs requiring no computer. They also caution that while the distribution of risk is informative about relative risk, estimates of absolute risk depend heavily on the estimation method, and that the figures reflect current technological possibilities rather than adoption. Like No. 189 above, this is a working paper and does not represent official OECD views.
Every figure above is quoted with the caveat its own source states. Last verified 2026-08-22. Scores and bands used on this site are documented on the methodology page.