Work
Automation, AI and jobs
The most-quoted automation statistic measures occupational susceptibility rather than job losses, and the two best studies of AI's labour-market effect point opposite ways.
The statistic that 47% of US jobs will be automated was never a prediction of job losses and carried no time horizon. It estimated the share of employment in occupations technically susceptible to computerisation, and redoing the analysis at task level rather than whole occupations brings it down to about 9%. Neither number is a forecast. Where automation has actually been installed the measured effects are real but local: industrial robots cut employment rates in the commuting zones that received them. On generative AI, the two most serious empirical studies disagree — one finds a 16% relative fall in early-career employment in exposed occupations, the other finds precise nulls. For a practical commercial comparison with this kind of workplace measurement, see how to be more proactive at work.
The 47% figure measures susceptibility, not job loss
The Oxford Martin working paper applied a machine-learning classifier to occupational descriptions.
About 47% of total US employment was estimated to be in occupations at risk of computerisation.
Direction: Increase. Strength of evidence: Weak.
Caveat A share of employment in occupations classified as susceptible, with no time horizon and no prediction of job losses, from a classifier trained on subjective hand-labelling of 70 occupations by workshop participants.
Jobs are bundles of tasks, and automating some tasks in a job is not the same as automating the job. Rebuilding the estimate on that basis produces a different order of magnitude.
Re-doing the analysis at task level rather than occupation level cuts the estimate to about 9% of jobs at high risk of automation.
Direction: Decrease. Strength of evidence: Strong.
Caveat Still technical automatability rather than adoption, so neither 47% nor 9% is a forecast of employment change.
Both figures describe what machines could in principle do. Neither accounts for cost, regulation or whether a firm chooses to.
Where robots were installed, employment fell locally
The strongest evidence on actual deployment comes from industrial robots in US manufacturing regions.
One additional industrial robot per thousand workers reduced the employment-to-population ratio by 0.18 to 0.34 percentage points and wages by 0.25% to 0.5% in exposed commuting zones.
Direction: Decrease. Strength of evidence: Mixed.
Caveat Identified from local labour markets, so it captures relative rather than aggregate effects, and the "each robot replaces N workers" figure changed between working-paper and published versions.
The countervailing argument is that automation substitutes for labour in some tasks while complementing it in others and raising output, which is why two centuries of mechanisation have not produced mass unemployment — the bank teller who survived the cash machine being the standard illustration (Autor, Journal of Economic Perspectives, 2015). That is a synthesis essay, and it says little about who bears the transition costs.
The historical case for job creation now has a measurement base, built from the job titles the US Census added decade by decade. The majority of current US employment is in job specialties introduced since 1940, and new-work creation shifted from middle-paid production and clerical roles between 1940 and 1980 to high-paid professional roles since (Autor, Chin, Salomons & Seegmiller, Quarterly Journal of Economics, 2024). That measure is sensitive to how granularly the Census recorded titles in each decade, and a new job title is not a new kind of work.
Exposure indices measure capability overlap, not adoption
The main framework for generative AI and occupations rates how much of each job's task description overlaps with what language models can do. It is widely reported as a measure of jobs at risk. It is not.
About 80% of the US workforce could have at least 10% of their tasks affected by large language models and 19% could see at least 50% of tasks affected, rising to 47-56% of all tasks with LLM-based tooling.
Direction: Increase. Strength of evidence: Mixed.
Caveat Ratings were produced partly by the authors and partly by GPT-4 itself, three of the four authors were at OpenAI, and "affected" means a task could plausibly be done faster, not that a job will be lost.
The two studies of AI's effect on jobs disagree
The claim that AI is already destroying entry-level work rests on one analysis of proprietary payroll records.
Early-career workers aged 22-25 in the most AI-exposed occupations show a roughly 16% relative decline in employment since generative AI became widespread.
Direction: Decrease. Strength of evidence: Mixed.
Caveat A relative decline within a proprietary panel rather than a national statistic, not peer reviewed, and unable to separate AI from the post-2022 tech hiring contraction or shifting entry-level hiring norms.
The best-identified study, using national administrative records, finds something different and finds it precisely.
Earnings and recorded hours show precise null effects, ruling out effects larger than 2% two years after ChatGPT's launch.
Direction: No detectable effect. Strength of evidence: Strong.
Caveat Denmark has strong employment protection and compressed wages, so nulls may not transfer to more flexible labour markets, and two years may be too early.
These are not reconcilable by picking the more congenial one. They differ in country, data source, outcome and identification, and the labour-market effect of generative AI is unresolved. One calibration puts AI's cumulative total factor productivity gain at under about 0.53% to 0.66% over ten years (Acemoglu, NBER, 2024), implying modest macroeconomic effects, though it is highly sensitive to the assumed share of tasks affected. Better established is the effect on measured output within particular tasks, covered in why productivity resists measurement, and on how applications are screened, in what happens to a job application online.
The short version
- Frey and Osborne's 47% was a share of US employment in occupations classified as susceptible to computerisation, with no time horizon attached.
- Repeating the analysis at task level across 21 OECD countries gave about 9% of jobs at high risk, which is still automatability rather than a forecast.
- Each additional industrial robot per thousand workers cut local employment-to-population ratios by 0.18 to 0.34 percentage points in exposed US commuting zones.
- A majority of current US employment sits in job specialties that did not exist in 1940, with new work concentrated in high-paid professional roles since 1980.
- The evidence has a hard limit: a proprietary payroll study reports a 16% relative fall in early-career employment in AI-exposed occupations, Danish administrative data rule out effects above 2%, and neither settles it.