Automation and AI Exposure
There is no official measure of how exposed an occupation is to automation. The best-known research estimates disagree by a factor of five — 47 per cent of US employment at high risk, 19 per cent, or 9 per cent — and the gap is almost entirely about whether you score whole occupations or individual tasks. What the official projections do say is which occupations are expected to shrink, and that list is dominated by routine information handling rather than by physical work.
Automation is the part of the labor market where confident numbers are cheapest and most misleading. Before using any figure here, it is worth knowing why the published ones do not agree.
The three headline estimates
47 per cent. Frey and Osborne scored 702 whole occupations for computerizability. The figure attracted enormous attention and equally strong methodological objection.
9 per cent. Arntz, Gregory and Zierahn redid it at the task level, allowing for the fact that two people with the same job title do different work. High-risk estimates fell to single digits across OECD countries.
19 per cent. Eloundou and colleagues measured how many workers have more than half their tasks exposed to large language models, where exposure means a model could cut the time for a task by at least half without loss of quality.
Same question, three answers, differing mostly by unit of analysis. The field has largely converged on the view that task-level decomposition is necessary and occupation labels are too coarse.
Exposure is not displacement
Every one of those studies measures technical capability, not what employers actually do. A task a machine could perform is not a job that disappears — cost, regulation, liability, customer preference and organizational inertia all sit between the two, and historically most of them have mattered more than the technology.
What the official data does say
The published employment projections name the occupations expected to shrink: data entry keyers, telemarketers, order clerks, payroll clerks, file clerks. Routine information handling under rules. That is the pattern automation has actually followed, and it is a much narrower claim than the exposure indices make.
The reversal worth noticing
Earlier automation pressed hardest on routine manual and low-credential work. Exposure to large language models does the opposite — it rises weakly with the difficulty of job preparation, meaning better-educated information work is more exposed, not less. Anyone reasoning from the last wave will get this one backwards.
Articles in this section
- What Exposure to Automation Measures, and What It Does Not 2 min read
- The Tasks Being Automated First 3 min read
- Occupations Most Exposed to Generative AI 3 min read
- Jobs Automation Has Reshaped Rather Than Removed 3 min read
- Exposure and Displacement Are Not the Same Thing 3 min read
- What Happened to the Jobs Earlier Automation Was Meant to End 3 min read
- Skills That Have Held Their Value Through Automation 3 min read
- How to Assess Your Own Role’s Exposure 3 min read
- Automation and the Pay Spread Inside an Occupation 3 min read
- Retraining Programs and What the Evidence Says 3 min read
Common questions
How much employment is at risk from automation?
Published estimates range from 47 per cent to 9 per cent depending on whether whole occupations or individual tasks are scored. There is no official figure.
Why do the estimates differ so much?
Unit of analysis. Scoring whole occupations treats everyone with a job title identically; scoring tasks accounts for the fact that they do different work.
Does exposure mean my job will go?
No. Exposure measures what a technology could do. Cost, regulation, liability and customer preference decide what actually happens.
Which occupations does the official data expect to shrink?
Routine information handling u2014 data entry keyers, telemarketers, order clerks, payroll and timekeeping clerks, file clerks.
Is AI exposure the same as earlier automation?
No, and it partly reverses it. Exposure to language models rises with the difficulty of job preparation, so better-educated information work is more exposed.
Is there a government exposure score?
No. The employment projections are the authoritative US source on occupational change, and they do not publish an automation risk score.
Layoffs and Job Security
Exposure is a long-run question. Layoffs are what actually happens to people, and the data on those is much firmer.
Read it →