The forecast that would not die
For over a decade, a single number has anchored public fear of automation: 47%. In 2013, Oxford Martin School economists Carl Benedikt Frey and Michael Osborne estimated that share of US employment sat in occupations at high risk of computerisation, over an unspecified future window they described only as “perhaps a decade or two.” The paper became one of the most cited economics working papers of the century, and the 47% figure still circulates in headlines and boardroom slides as though it were a settled prediction rather than a probabilistic estimate built on 2013-era assumptions about which tasks software could plausibly learn.
By 2022, the Information Technology and Innovation Foundation’s Robert D. Atkinson had the natural experiment other forecasters lack: nine years of actual labor-market data. His verdict was blunt. The US had added roughly 16 million jobs since the paper’s release, unemployment sat at 3.7%, and when he checked Frey and Osborne’s occupation-level risk scores against what actually happened to those occupations, the correlation was weak — around 0.26. The 47% did not happen. That does not mean automation is harmless; it means a single point-in-time susceptibility score, however rigorously modeled, is a poor guide to a decade of technological, economic, and institutional change. Any claim about “safe” trades below should be read with that failure mode in mind.
Two paradoxes people keep mixing up
Discussions of manual work and AI regularly invoke two ideas as though they were interchangeable. They are not, and the difference matters for planning.
Moravec’s paradox is a computational observation, named for roboticist Hans Moravec’s 1988 book Mind Children and articulated around the same time by Rodney Brooks and Marvin Minsky. It says abstract reasoning is computationally cheap for a machine, while the sensorimotor and perceptual skill of a toddler — balance, grip, depth perception, object recognition in clutter — is extraordinarily expensive. It explains why a model can pass a bar exam before a robot can reliably fold laundry.
Polanyi’s paradox is the labor-economics term, and it is the one that actually predicts which jobs resist automation. MIT economist David Autor coined it in a September 2014 NBER working paper, building on philosopher Michael Polanyi’s 1966 observation that “we can know more than we can tell.” Skilled tradespeople routinely diagnose a failing circuit, a stuck valve, or a wheezing compressor using pattern recognition they cannot fully articulate as a rule set — which means it cannot be cleanly codified into training data or a decision tree either. Autor’s own conclusion still holds up: “challenges to substituting machines for workers in tasks requiring adaptability, common sense, and creativity remain immense.”
The distinction matters because Moravec’s paradox is about robot bodies, and Polanyi’s paradox is about tacit human judgment. A trade can be automation-resistant for either reason, both, or neither — and conflating them produces sloppy predictions.
What 200,000 real conversations show — and what they don’t
Frey and Osborne modeled susceptibility from the outside. In 2025, Microsoft Research took a different approach: Kiran Tomlinson and colleagues examined 200,000 anonymized Bing Copilot conversations against the O*NET database of occupational work activities, producing an “AI applicability score” for hundreds of occupations. The lowest-scoring occupations — phlebotomists, nursing assistants, massage therapists, equipment operators, and roofers — share an obvious trait: the work happens on a body, in a specific place, using hands.
The researchers were explicit that this is a measurement of task overlap with a chatbot, not a forecast of layoffs. In a follow-up blog post responding to breathless press coverage, they wrote: “our study does not draw any conclusions about jobs being eliminated; in the paper, we explicitly cautioned against using our findings to make that conclusion.” A low applicability score today describes where generative AI is currently unhelpful as a chat assistant. It says nothing about a five-year-old with better cameras and cheaper actuators — which is precisely the territory the site’s humanoid-robots chapter tracks.
The complication: “manual” is not one thing
Here is where the clean story breaks. The World Economic Forum’s Future of Jobs Report 2025 asked employers what skills they expect to matter more or less by 2030. Nearly every skill category was expected to grow in importance. Only two were expected to net decline — and one of them was “manual dexterity, endurance and precision,” at 24% net negative.
That is not a contradiction of the trades data above; it is a reminder that “manual work” spans wildly different jobs. Warehouse picking, assembly-line stations, and repetitive industrial handling are manual, physically demanding, and precisely the kind of structured, repeatable motion that industrial robotics has been automating for years — long before generative AI existed. A licensed electrician troubleshooting an unfamiliar panel in an unfamiliar house is also “manual,” but it bears almost no resemblance to a warehouse task in its unpredictability, liability, or need for judgment. Employers surveyed by the WEF appear to be picking up on the former even as demand holds for the latter. A trade is not protected by being manual; it is protected by combining tacit judgment with either a hard physical-presence requirement or structural replacement demand that a robot can’t yet meet cheaply.
Five trades, two different reasons to expect demand
The Bureau of Labor Statistics’ 2025–2035 Occupational Outlook Handbook projections illustrate both mechanisms at once:
| Trade | Projected growth | Annual openings | Median wage (May 2025) |
|---|---|---|---|
| Electricians | 9% | ~72,700 | $63,190 |
| Plumbers, pipefitters, steamfitters | 7% | ~42,000 | $63,800 |
| HVAC/refrigeration mechanics | 11% | ~40,600 | $61,010 |
| Elevator/escalator installers | 6% | ~1,900 | $109,910 |
| Home health/personal care aides | 18% | ~760,500 | $35,800 |
Electricians, plumbers, and HVAC technicians combine Polanyi’s-paradox judgment (diagnosing an unfamiliar system) with an aging incumbent workforce and a multi-year apprenticeship pipeline — roughly four to five years and about 2,000 paid hours annually for electricians — that cannot be compressed just because demand rises. Elevator installers carry the highest wage of the five, reflecting a narrow, licensed, safety-critical specialty with few entrants. Home health and personal care aides sit at the other end of the wage scale but post by far the largest raw opening count, driven by demographics rather than scarcity of training slots: an aging population needs more hands-on care than the current workforce provides, and that demand is comparatively insensitive to automation because it depends on physical presence, trust, and in-person judgment that current robotics cannot supply at scale or cost.
The throughline
None of this amounts to a permanent exemption. A trade looks durable when it needs tacit skill that resists codification, when the workforce providing it is shrinking faster than demand, or when the job requires being physically present in a way no robot yet performs cheaply — ideally more than one of these at once. It looks fragile when the “manual” component is actually repetitive and structured, which is exactly the kind of manual work industrial and warehouse robotics have targeted for a generation. Reading BLS growth and wage data is a reasonable first filter for a household or a young worker weighing options, but it is a filter, not a guarantee — the same discipline that should have kept people from treating 47% as gospel in 2013 applies to every number in this article.