Weighting by Hours Turns a 30% Title Score Into 10% or 65%

The short answer: the task lists come from O*NET and the exposure ratings from Eloundou et al. in Science. Neither is mine. What the tool adds is the weighting, and that step is where the number acquires its meaning. Without it, the score describes a person who spends equal time on every task in their occupation, and nobody has ever had that week.

The two inputs that are not mine

The task list for each occupation comes from O*NET, the Labor Department's catalogue. The exposure rating for each task comes from Eloundou, Manning, Mishkin and Rock (Science, 2024), who asked whether a language model or software built on one could halve the time on the task without losing quality. Both are public. Anyone can check them.

A title-based tool stops there. It takes the ratings for an occupation's tasks and averages them. The AI Job Risk Check methodology adds one step, and the step is the point.

The arithmetic

Say your occupation has twenty tasks and the research rates six of them highly exposed. A title-based tool computes six over twenty and reports 30%.

Now look at a real week. Those six tasks take you four hours. The other fourteen take thirty-six. The honest number is four over forty, which is 10%.

The person at the next desk, same title, spends twenty-six hours on those same six tasks. Their honest number is twenty-six over forty, which is 65%.

Wrong in a specific way

The unweighted number is wrong for almost everybody, and it is wrong in a particular direction: toward the middle. It describes a person who spends equal time on every task in their occupation. Anyone with an unusual week, which is most people, gets pulled toward a centre that describes nobody. The people it misleads most are the ones furthest from the average, and they are exactly the people who most need to know it.

Two marketing managers at 55 and 31 both get 52 from the unweighted method. Two paralegals, one in document review and one running a small firm's front office, get the same number. The precision is real. The measurement is not.

What it costs

The cost of doing it properly is that you have to think about your own week for two minutes. Which tasks did you actually do? Roughly how many hours went to each? The tool cannot know that without asking, and no dataset on earth contains it, which is why the tools that skip the question have to substitute an assumption.

Most people tell me that part was more informative than the score. Writing down where forty hours went, task by task, is something almost nobody has done for their own job. The number at the end is a summary of an exercise that was worth doing on its own.

Try the estimate now

You do not need the tool to start. Estimate, right now, what share of last week went to tasks that were mostly reading and writing: drafting, summarising, researching, formatting, reporting. That share is a rough proxy for the exposed side of your week. If it is under a quarter, a title-based score is probably overstating your exposure. If it is over half, it is probably understating it. Either way you now know something the title could not tell you, and the free score turns the estimate into a number with the rating on each task.

Frequently asked questions

Why does AI Job Risk Check ask how I spend my week?

Because the score weights each task's exposure rating by the hours you report on it. Without the hours, the only available number is an unweighted average that describes someone who spends equal time on every task, which is nobody.

Where do the task list and exposure ratings come from?

The task lists come from O*NET, the U.S. Department of Labor's occupational database. The exposure ratings come from Eloundou et al., published in Science in 2024. Both are public.

How different can two people's scores be under the same title?

Very. In the example above, one person with four exposed hours out of forty scores 10%, and a colleague with twenty-six exposed hours scores 65%. A title-based tool gives both of them 30%.