The number was right. The name was wrong.

Every job on this site carries a number out of 100. For a long time the number beside crop farm labourers was 90.9, and the word above it was resilience.

The number is still 90.9. The word is gone, and the story of why is the most useful thing we can tell you about how to read anything else here.

Where 90.9 comes from

The figure is not a model output. It is an average and a subtraction, and you can check it yourself.

The ILO’s Working Paper 140 scores individual tasks for exposure to generative AI. Crop farm labourers — ISCO unit group 9211 — have eight tasks in that dataset. Every one of them scored in the ILO’s lowest band:

Task	ILO band	Score
Planting and harvesting field crops, such as rice, by hand	Very Low	0.063
Raking, pitching and stacking straw, hay and similar material	Very Low	0.073
Digging and shovelling to clear ditches or for other purposes	Very Low	0.075
Watering, thinning, and weeding crops by hand or using hand tools	Very Low	0.085
Picking fruit, nuts, vegetables and other crops	Very Low	0.093
Loading and unloading supplies, produce and other materials	Very Low	0.099
Performing minor repairs on fixtures, buildings, equipment and fences	Very Low	0.120
Grading, sorting, bunching and packing produce into containers	Very Low	0.125

The mean of those eight scores is 0.091. Take it away from 1, render it as a percentage, and you have 90.9. That is the whole calculation. Nothing is weighted, smoothed or blended, because with one measured input there is nothing to blend.

And the ILO is right. A language model cannot pick fruit. It cannot rake hay, dig a ditch, or stack a pallet of sorted produce. Score those eight tasks for exposure to generative AI and you should get very low numbers, and the paper does.

What we did with it

We printed 90.9 under the heading “resilience”, near the top of a list of 427 occupations.

Read that as a sentence and it says: picking fruit is one of the safest jobs from the arrival of machines. Which is not what the source measured, is not what the number means, and is not true. Fruit picking, weeding and harvesting are exactly the jobs engineers have been pointing robots at for years: orchard pickers, weeding rigs, self-driving harvesters. The one claim our label made was wrong in precisely the occupations where the opposite case is strongest.

The paper never made that claim. It measured exposure to generative AI and said so in its title. The error was entirely ours, and it was an error of naming, not of arithmetic: we took a narrow, true statement and rendered it as a broad, false one by choosing a friendlier word for it.

A number derived from one measured input beats a number derived from four, three of which were invented to make the formula look complete. The same discipline has to apply to the label. If the source measures one thing, the heading says that thing.

What changed

Three things, none of them the number.

1. The name. The score now says what it is: how much of the everyday work still needs a person, measured against AI software. Every big score on the site carries the same line underneath it: not the odds you keep your job. No band on the scale says “safe”.
2. Two measures beside it. For 419 of the 427 occupations there are now two more figures drawn from separate sources: how hands-on the work is, and how fiddly and cramped that hands-on work gets. They are shown beside the main score and never folded into it. Together they mark where a high score is too hopeful.
3. The ILO’s own words, kept verbatim, and ours labelled as ours. Every score shows the source’s category word for word — Not Exposed, Minimal Exposure, Exposed: Gradient 1 through 4 — as “the ILO’s official verdict”. Our plain-English reading (“mostly holds up against AI”) sits beside it, clearly as our reading, with cut-offs we chose for readability. Disagreeing with a source’s emphasis is not a licence to relabel its categories, but it is fine to translate them, as long as you say you did.

Crop farm labourers still score 90.9 against generative AI. They now also carry a hands-on figure high enough that the page says, in words, that machines are the part this score cannot see.

What to take from it

Two things, if you read nothing else here.

Read the label, not just the number. Every score on this site answers one specific question, and the question is written next to it. A 90 that answers “how much of this could a language model do?” tells you nothing about a robot arm, and a page that lets you forget which question you asked is a badly built page.

Watch what a site does when it is wrong. Ours is in the changelog, dated, with the reason. Plumbers sit near the top of this index for the same reason crop farm labourers do — most of the work is physical — and the honest version of that sentence needs the second and third measures to be worth anything at all.

The number was never the problem. We were.