How accurate have AI job-loss predictions been?

The 47%, five years later

The 47%, Frey and Osborne's high-risk share of United States employment, fixed no calendar date, though its authors did expect the high-risk jobs could be automated "perhaps over the next decade or two", as the tasks-versus-titles page covers.¹ That hedge is loose, but it is what makes the scores testable at all.

The closest thing to a direct test comes from Coelli and Borland, who ran the Frey and Osborne scores against subsequent United States employment change. They estimate that employment in the occupations scored as high risk grew by around 7.4% between 2013 and 2018, an employment-weighted regression estimate rather than a raw count.² Using the full continuous probabilities rather than the high-risk cut, they find a statistically significant negative relationship with employment change (p = 0.006), but an R-squared of 0.024: the scores explain roughly 2% of the variation.² And once they control for the routine-task measures economists were already using, the Frey and Osborne relationship loses significance entirely (p = 0.704), meaning the scores added no forecasting power over the existing framework.²

A separate tally by the Information Technology and Innovation Foundation, a think tank rather than a peer-reviewed retrospective, checked individual occupations to the end of 2021. Insurance underwriters, which Frey and Osborne scored at a computerisation probability of 0.99,¹ grew 16.4%.³ Recreational therapists, whom they ranked least likely to be automated, declined 8.9%.³

Where the two retrospectives disagree

The ITIF tally reports the correlation between Frey and Osborne risk and actual job loss as negative and modest, at 0.26. In their data the occupations rated riskiest lost slightly fewer jobs, the opposite of the direction the scores predicted.³

An OECD analysis across 21 countries points the other way. Between 2012 and 2019, employment in the riskiest half of occupations grew 6.1% while the least risky half grew 17.8%.⁴ On that measure the automation literature had the direction right and only the magnitude and the vocabulary wrong: slower growth is not disappearance, and "at risk" was read by almost everyone as the latter.

The two are not strictly contradictory, because they use different risk measures, different countries and different windows. The retrospective evidence is not settled, and the two studies point in opposite directions.

The bank teller story, finished

The most repeated counter-example in this whole field is the automated teller machine. The usual telling: ATMs arrived, and the number of bank tellers went up, because cheaper branches meant banks opened more of them. That telling is accurate and sourced: full-time-equivalent teller employment grew about 2.0% a year from 2000,⁵ and tellers per urban branch fell from 20 to 13 between 1988 and 2004 while urban branches rose 43%.⁶

It also stops in 2013, which is where the data in Bessen's working paper end. Here is what happened next. Teller employment in the United States was 545,300 in 2012 and 364,100 by 2022, a fall of 33%, and 347,400 by 2024. Official projections for 2024 to 2034 have it falling a further 13%.⁷

Both halves are true, and the order matters. Employment held up for roughly two decades while branch numbers grew, then fell hard once that expansion stopped. Bessen set the same shape out as an inverted U in later work: employment rises while a cheaper service expands demand, and falls once demand stops responding to price. He develops it for manufacturing industries such as textiles rather than for tellers.⁸ Anyone citing the first half without the second is quoting a genuine finding to support a conclusion its own author did not reach.

The professionals missed it too. The Bureau of Labor Statistics projected in its 2012 to 2022 round that teller employment would grow 1%, to 551,000. The actual 2022 figure was 364,100, meaning the official projection was about 51% above the outcome.⁷ These are professional statisticians using established methods, and the direction they missed was the optimistic one.

How often is an occupation actually eliminated?

James Bessen tracked the 271 detailed occupations in the 1950 United States census forward to 2010 and concluded that in only one case, elevator operators, could an occupation's disappearance be attributed largely to automation.⁵ He also notes that of the 37 occupations Frey and Osborne's panel judged fully automatable with 2013 technology, none had been completely automated.⁵ In the nineteenth century, 98% of the labour needed to weave a yard of cloth was automated and the number of weaving jobs went up.⁵

Against that, a genuine elimination case. Between 1920 and 1940 AT&T mechanised switching across more than half the United States telephone network, and most telephone operator jobs went away.⁹ The researchers who documented it add the part that matters for anyone drawing lessons: incumbent operators were measurably hurt, but later cohorts of young women moved into clerical and service work, so overall employment among later cohorts of young women in those cities did not fall.⁹ Occupations can be eliminated. The people already in them bear the cost; the aggregate often absorbs it.

The forecasts nobody graded

The World Economic Forum has published employer-survey forecasts for a decade, and three of them have now passed their horizon.

In searches conducted in August 2026 we could not find a published retrospective of any of the three expired forecasts, from the Forum or from anyone else.¹⁰ Note also the reversal: in 2016 the outlook was a net loss of about 5 million by 2020; two years later the same exercise projected a net gain of 58 million by 2022.¹⁰

How these numbers are built matters. In the 2018 edition, the survey responses applied to a sample of over 15 million workers implied a decline of 0.98 million jobs and a gain of 1.74 million. Those were then extrapolated across large-firm, non-agricultural employment worldwide to produce a range of estimates, one of which is the headline 75 million displaced and 133 million created. The report says the estimates "should be treated with caution, not least because they represent a subset of employment globally".¹⁰ The 2020 report's 85 million figure covers 15 industries and 26 economies; the executive summary states that scope once, then quotes the figure without it, and the Forum's press release turned the report's hedged "may be displaced" into "will disrupt 85 million jobs globally".¹⁰

The Forum is explicit about the limits of the method, at least inside the reports. Of the survey subset behind its 2025 projections, covering 1.18 billion workers, it writes that the conclusions "should not be treated as comprehensive, but rather as providing insights on selected segments of the global workforce".¹⁰

Why the misses look the way they do

Forecasts measure technical capability and skip adoption, which is slower, costlier and more regulated than capability alone suggests. They count the jobs a technology removes and not the ones it creates or expands, which the weaving and early bank-teller cases show can dominate for a long time. And they are usually reported without their own hedges: the underlying papers say "potentially automatable" and "may be displaced", and the coverage says "will lose".

What this means for you

Treat every big round number about AI and jobs, including the ones on this site, as a description of overlap rather than a schedule. Ask what is being counted, over what period, and whether anyone has ever checked the last forecast from the same source. On the graded record the two retrospectives still point in opposite directions and neither is decisive: the ITIF tally finds the riskiest occupations lost slightly fewer jobs, the OECD comparison finds they grew markedly more slowly. What both rule out is the headline reading, occupations vanishing on schedule. Evidence on economy-wide displacement remains thin, and what the labour-market data currently shows covers it in detail.