AI Exposure Rises With Wages, Reversing the 2013 Automation Map

The short answer: the jobs that rate highest for AI exposure are not the low-wage, routine ones the last decade of automation research pointed at. They are the desk jobs that required a credential. That does not mean those jobs vanish. It means the safe-harbour advice of the last decade was aimed at a different technology.

What the paper reports

Eloundou, Manning, Mishkin and Rock (Science, 2024) rated every task in the O*NET catalogue for whether a language model, or software built on one, could halve the time it takes without losing quality. When they aggregated those ratings by occupation and lined them up against wages, the pattern was clear: higher-wage occupations generally show higher exposure. Not lower.

The occupations with the largest share of exposed tasks are concentrated among roles that typically require a bachelor's degree and pay above the median. The paper is careful to describe this as exposure, which is a measure of what the technology could plausibly speed up, and not as a forecast of displacement.

The map most people are still carrying

In 2013, Frey and Osborne estimated that about 47% of US employment was at high risk of computerisation over the following decade or two. Their model put the risk at the bottom of the wage distribution: routine manual work, clerical processing, transport. The advice that followed was consistent for ten years. Get a credential, get behind a desk, do work that involves judgement and language, and the machines will go around you.

Language models moved the target. What they do cheaply is read, summarise, draft, translate, and produce a first version of a structured document. That is a fair description of a large share of professional work, and it is exactly the work the 2013 model treated as safe.

What stays near the floor

The same ratings leave three kinds of work close to the bottom, and none of them is defined by wage.

• Tasks that need physical presence. Somebody has to be in the room, the plant, the clinic, or on the site. A system that reads and writes has no hands.
• Tasks where a named person is legally accountable for the outcome. An audit opinion, a prescription, a court filing, an engineering stamp. The accountability is the product.
• Tasks whose whole point is someone deciding in front of someone else. Negotiation, delivering hard news, reading a room and changing course inside the same conversation.

Most occupations are a mix of these and their opposites. A physician has high-exposure documentation tasks and low-exposure examination tasks. A lawyer has high-exposure research and low-exposure advocacy. The score for an individual depends on the proportions, which is why the same title can produce very different numbers.

Exposure is not displacement

I want to be direct about what this does not say. A high-exposure title is not a title on its way out. Exposure measures what is technically possible on individual tasks. Whether an employer adopts the tools, whether regulation still requires a licensed human, whether the output clears a real quality bar, and whether demand for the occupation grows anyway are all separate questions, and the rating does not answer any of them. Software development carries some of the highest measured exposure in the research, and BLS still projects employment growth for the decade.

So the message for someone in a high-exposure professional role has nothing to do with panic. Stop assuming the credential is a moat, look at the actual task list, and see which hours of the week are on which side of the line.

A practical read

Take your own occupation. Sort its tasks into the three floor categories above and everything else. Then estimate how many hours a week you spend in each pile. If most of your week is in the second pile, you are in the group the 2013 map called safe and the 2024 ratings call exposed. That is worth knowing early, while there is time to move hours across the line on your own terms.

Frequently asked questions

Are high-paying jobs more exposed to AI than low-paying ones?

According to Eloundou et al. (2024), yes on average. Occupations that require a degree and pay above the median tend to have a larger share of tasks a language model could speed up, because those tasks involve reading, drafting, and summarising.

Does this mean high-paid professionals will lose their jobs?

No. Exposure measures technical possibility on tasks. Adoption, regulation, quality standards, and demand for the occupation all sit between exposure and any change in employment.

How is this different from the Frey and Osborne 47% figure?

Frey and Osborne (2013) modelled computerisation risk at the occupation level and found it concentrated in routine, lower-wage work. The 2024 Science study rates individual tasks for language-model exposure and finds it concentrated in higher-wage, desk-based work. They measure different technologies at different units.