Hands-on is not one thing

“Hands-on” sounds like one thing. It isn’t. A plumber wedged under a sink and a worker on a production line are both doing physical work, and a robot would find one of them far, far harder than the other.

The main score on this site is about generative AI. It is silent on robots. To stop that silence reading as good news, every job that the data reaches carries a second figure: how much of the work is physical.

That figure did a job, and then it stopped being enough. A production line and an aged-care ward both came out heavily physical. So did a plumber under a sink and a driver in a cab. “This work is about the body” is true of all four and useful about none of them, because the thing you actually want to know is which body work machines are bad at.

The measure we added

Data version 0.7 adds a third figure beside the other two: how fiddly and cramped the hands-on work is.

It is not our invention, and that is the point of it. The variables are the perception-and-manipulation bottleneck from Frey and Osborne (2017): finger dexterity, manual dexterity, and how often the work happens in cramped spaces or awkward positions. Their three, in their combination. We read them from O*NET 31.0 and carry them onto the ISCO occupation list through the published BLS crosswalk chain, which is what limits the coverage to 419 of the 427 occupations — eight have no close US match, and those eight say so rather than guessing.

Keeping the variables somebody else chose keeps the most tempting decision on this site out of our hands. If we picked which measures counted as “hard for a robot”, we would pick the ones that made our existing rankings look right.

Three kinds of hands-on work

With two physical measures instead of one, the hands-on jobs sort into kinds. The sort is not measured — it is a set of rules applied to measured quantities, and the rules are printed on the page beside the result:

Kind	The rule	What it means
Hands-on, helping and caring for others	at least half physical, and caring at or above O*NET’s own anchor 4	Whether a machine could do the movements is half the question. Whether anyone wants one to is the other half.
Hands-on, fiddly work in tight spaces	at least half physical, bottleneck at or above the median of the physical jobs	More fine handwork and more cramped spots than the typical physical job. Exactly what machines find hardest.
Hands-on, less fiddly work	at least half physical, bottleneck below that median	Clears one hurdle for machines. Does not mean machines are doing this work.
Less than half hands-on	under half physical	Robots still matter, for a smaller share of the work. Not sorted further.

Two details in there are deliberate and worth arguing with.

Caring is checked first. A robot that can physically lift a person is not thereby a thing anybody wants lifting their mother. For this work the social question is as live as the mechanical one, so it decides the sort before the dexterity numbers get a look in. O*NET’s own anchor for that level is “help a medical patient find in-home assistance or healthcare” — their wording, their scale point.

The fiddly line is a median, not a round number. Averaging three ratings across a crosswalk compresses everything toward the middle, so a cut at a tidy 50 kept 17 of 195 physical occupations and left the plumber sitting exactly on the line. A median claims only “more than the typical physical job”, which is all the data supports, and it moves with the data instead of being tuned until the answers look right.

What it changed, and what it didn’t

No physical figure changed. Not one. This release added a measure and a sort; it did not revise a number.

What it changed is what the pages can say. Registered nurses are the clearest case, and they show the seam. The ISCO nursing group pools seven ONET occupations and averages out at under half physical — a desk-ish result for work nobody would call desk work. The registered nurse profile on its own does not: it has its own ONET occupation, no crosswalk averaging, and it lands in the caring kind where it belongs. Both figures are correct about different things. The profile page uses the sharper one and says which it used.

Heavy truck drivers go the other way. Very physical, not especially fiddly, not caring work — the kind where the first hurdle for machines is already cleared, and where the main score’s near-silence on robotics is doing the most damage to anyone reading it as a forecast.

What it still cannot tell you

Everything about whether robots can do the work today.

This is three measured quantities about the shape of a job — how much of it is physical, how fine the handwork is, how cramped the spaces are, plus how much of it is caring for people. None of them is a capability figure. A job scoring high on fiddliness is telling you that the task is hard for a machine in a way that has held up for a decade; it is not telling you that no robot has managed it, and it certainly does not become a forecast when you squint.

Read all three together: the main score for what a language model reaches, the physical figure for what it cannot see, and the fiddly figure for how much of that blind spot is genuinely hard. Where you are on the rankings means something different depending on which of the three put you there.