AI Job Risk Lists Score Titles, but Exposure Lives in Tasks

The short answer: the research that ranked lists borrow from rates tasks, not titles. Collapsing an occupation's two dozen tasks into a single number assumes everyone with that title spends the week the same way. They do not, so the list hands two people with nothing in common the same score, printed to one decimal place.

What an occupation actually contains

O*NET, the Labor Department's occupational database, breaks each occupation into the tasks people in it actually perform. Across the 1,016 occupations we score, that is 18,796 tasks, or about eighteen apiece. A paralegal has around two dozen. Drafting correspondence sits next to interviewing clients, which sits next to serving subpoenas.

The exposure ratings in Eloundou et al. (2024) attach to those tasks, because that is the level at which the question can be answered. "Could a language model halve the time on this task without losing quality?" has an answer for drafting correspondence. It does not have a single answer for "paralegal."

What collapsing does

To get from two dozen task ratings to one number for the title, a ranked list has to average them, and the only average available without asking anyone is the unweighted one. That average describes a person who spends equal time on every task in the occupation. Nobody has ever had that week.

Two paralegals illustrate it. One works litigation support: document review, legal research, deposition summaries, most of it text transformation that the ratings score high. The other runs a small firm's calendar, meets clients, and files with the court, most of it acts in the world that the ratings score low. Same title, and almost no overlap in daily work. A ranked list gives them the same number.

The fix is not complicated

Ask how the week actually splits. Weight each task rating by the hours the person reports on it. Report the result as what it is: the share of working time spent on tasks the research rates as highly exposed.

That number is smaller and considerably more useful than a rank. It moves when the week moves, it can be checked against the task list, and it tells the person which hours are doing the work. A rank can do none of that. The AI Job Risk Check methodology is built around that weighting step, and it is the reason two people with one title can get scores twenty points apart.

Why the lists persist anyway

Ranked lists are easy to produce, easy to headline, and easy to share. "Paralegals rank seventh" is a sentence. "Your exposure depends on how many hours you spend on document review versus court filings" is a conversation. The first travels further. The second is the one that could change what somebody does on Monday.

There is also a data problem. Producing a task-weighted number requires asking each person about their week, which a list cannot do. So the list substitutes an assumption for a question and hopes nobody notices.

Try it on your own occupation

Pull up your occupation's task list on O*NET and count how many of the tasks you touched last week. Most people are surprised how few. Then look at which ones took the most hours. If the hours are concentrated in reading, drafting, summarising, and formatting, your week is more exposed than your title's rank suggests. If they are concentrated in meetings, site visits, negotiations, and sign-offs, it is less. Either way you now know something the list could not tell you.

Frequently asked questions

Are "jobs most at risk from AI" lists accurate?

They are accurate about the tasks they borrow from and inaccurate about people. The underlying research rates tasks. A list averages those ratings into one number per title, which describes nobody's actual week.

How many tasks does O*NET list per occupation?

Across the 1,016 occupations AI Job Risk Check scores, O*NET lists 18,796 tasks, or roughly eighteen per occupation. Some have far more; paralegals have around two dozen.

What is a better measure than a job ranking?

The share of your working hours spent on tasks the research rates as highly exposed. It requires asking how your week splits, then weighting each task's rating by those hours.