Measuring exposure, and what the numbers do not say

Where the big percentages come from

Every article about AI and jobs quotes a percentage. Eighty per cent of workers. Forty per cent of employment. Three hundred million jobs. The figures come from a small number of studies, all built the same way, and once you know the method you can read every one of them correctly.

The method is this. Take a catalogue of occupations, each broken into its tasks: a public database maintained by the US labour department lists around a thousand occupations and nearly twenty thousand tasks. For each task, ask whether a model could do it, or help do it, at least fifty per cent faster with no loss of quality. Then add up, per occupation, the share of tasks that got a yes. That share is the occupation’s exposure.

Who does the asking matters. In the most-cited study, from 2023, the task ratings were made both by people and by a model, and the two agreed closely. That is either reassuring or circular, depending on your view, and it is worth knowing that a model was partly rating its own usefulness.

The figures, stated precisely

The 2023 study found that around eighty per cent of the US workforce could have at least ten per cent of their tasks affected, and around nineteen per cent could have at least half. Exposure rose with wages and with education: the jobs most exposed were not the lowest paid but the office-based, credentialed ones.

The International Labour Organization, the same year, repeated the exercise across countries and found the most exposed occupational group to be clerical support (secretaries, data-entry clerks, administrative assistants) with about a quarter of clerical tasks highly exposed. Two consequences followed. Women’s employment was more than twice as exposed as men’s, because clerical work is disproportionately women’s work. And the share of employment in highly exposed occupations was around five and a half per cent in high-income countries and under half a per cent in low-income ones, because a larger share of work in poorer countries is physical and informal.

The International Monetary Fund’s 2024 estimate put around forty per cent of global employment as exposed (sixty per cent in advanced economies and roughly a quarter in low-income ones) and split the exposed half and half between jobs where the tool complements the worker and jobs where it substitutes.

What “exposed” does not mean

Here is the part the headlines drop.

Exposure is not replacement. A task a model could do faster is a task that might be automated, augmented, ignored, or forbidden. Radiology, from the lesson you reused at the start of this module, is a highly exposed occupation that grew.

It measures the task, not the job. A job is a bundle of tasks plus responsibility, and the studies count the tasks. An occupation can be sixty per cent exposed and be, in practice, entirely reorganised around the forty.

It assumes adoption. Whether a firm actually uses the tool depends on cost, regulation, liability, unions, customers and inertia, none of which is in the number.

It measures the model of 2023. The most-cited figures were computed before the models that can see, hear and act, and before the reasoning models. The exposure of a task is a moving quantity and the studies are photographs.

Why exposure concentrates where it does

The mechanism, and it is the one the whole course has built. A model works on language and images. The tasks that are language-shaped, routine and abundantly documented (drafting, summarising, classifying, translating, first-pass code, answering standard questions) are exactly the tasks of clerical, administrative, junior professional and knowledge work. The tasks that are physical, situated and unpredictable (nursing, plumbing, construction, childcare) are not. So the studies invert the assumption of the previous forty years, that automation came for the factory floor and left the office alone. This time the office is the exposed part.

How to read the next one

When you see a percentage, ask five questions. Task or job? Could-do or will-do? Rated by whom? Which model, and when? And what does the study say about the difference between complementing and substituting, because that difference is where the entire question of whether the number is good news or bad news lives, and the next lesson is about the experiments that tried to measure it.