Every occupation, every dimension, and seven years of BLS employment —
sliced however you need it. Nothing here is computed on a server: the
whole dataset loads once and filters instantly, and you can download
exactly what you're looking at.
Employment for this selection
Total US employment across the filtered occupations, by year. Counted
by the BLS — the one series here that isn't our judgement.
Prefer a ready-made cut? Six rankings — safest,
most exposed, best paid and safe, largest, growing, and already
shrinking.
The safest big jobs are not the good jobs
The largest occupation in the United States is Home Health and Personal Care Aides — 4,305,810 people, scored 70, one of the most resistant on the
register. It pays $35,800. Cashiers, scored 23 and
COOKED, pay $32,880. Whatever protects the first job from automation is not
paying for it.
COOKED
EXPOSED
SAFE
· circle size is employment · 825 occupations
0.35correlation (r)score vs log wage
12%of wage variation explainedr² = 0.123
0.248r, weighted by employmentweaker where the people are
5.8Mworkers, SAFE and under $45,0002 occupations above a million
Safer work does pay more on average — the median SAFE occupation pays $68,080 against $48,500 for the
median COOKED one. But the association is weak, and it gets weaker where it
matters most: weight every occupation by how many people are actually in it
and r falls from 0.35 to 0.248. The link between being
hard to automate and being well paid is loosest at the exact end of the
distribution where the most people work.
The reason is visible in the chart. The work current AI cannot do is
disproportionately physical care work — in homes, at bedsides, with
people who cannot be scheduled — and physical care work is badly paid. That
is a fact about the labour market, not a fact about AI, and no amount of
automation resistance changes it.
What this number is not
Not causal, in either direction. This describes how two
published medians co-vary across occupations. It is not evidence that
automation resistance depresses pay, or that low pay protects work from
automation.
Occupations, not people. These are occupation-level
medians, so every distribution inside an occupation is invisible. Each
occupation counts once whether it holds 300 workers or 4.3 million —
which is exactly why the employment-weighted figure is published beside
the plain one.
r = 0.35 is a weak-to-moderate association. It leaves 88% of the variation in pay to everything
else: licensing, unionisation, credential length, hours, who does the
work and what the country has historically been willing to pay them.
The wage axis is top-coded. The BLS publishes annual
medians above $239,200 as a ceiling rather than a figure, so the highest
earning occupations are compressed against the top of the chart.
5 occupations are missing from the 825
plotted, because the BLS suppressed their wage or employment. Suppressed
is not zero, so they are dropped rather than imputed.
Using this data
Take it. The full dataset is at /reports.json — the same file this page
uses. Check our arithmetic; publish a contradiction if you find one.
As of the last build, not live. Scores change when
evidence accumulates, not on a schedule, and employment updates once
a year when the BLS publishes.
Employment is BLS; scores are ours. Occupations
missing a year were suppressed by the BLS or didn't exist under that
code — shown as gaps, never as zero.
Attribution: cite the register and link the page, so
a reader can check the number against its reasoning.