Occupations Most Exposed to Generative AI

This wave points at the opposite end of the labor market

This wave points at the opposite end of the labor market from the last one, and almost everyone reasoning from the last one gets it backwards. The intuition built on robotics points the wrong way here.

Robotics and earlier software automation concentrated on routine manual and routine clerical work. That work was predictable, physical or rule-following, and generally lower paid. Physical predictability was the property that mattered then.

Exposure to large language models runs the other way entirely. The research finds it increases, weakly, with the difficulty of job preparation: the more credentialed the information work, the more of its tasks a model can materially speed up. Preparation difficulty and exposure move in the same direction.

Why the reversal makes sense

The technology is good at producing and transforming language and structured reasoning. That is precisely what credentialed information work consists of. The overlap is close to exact rather than approximate.

Memos, analyses, code, recommendations and summaries of other documents are the daily output of a great many well-paid jobs. All of them are text with structure, which is the format the models handle best. Structured text is the native material of these systems.

The same technology is bad at physical presence in unpredictable environments, which is what a great deal of lower-paid work consists of. The exposure profile follows the technology’s actual competence rather than any story about which jobs are valuable. Capability decides the profile and nothing else does.

What high exposure looks like in practice

It looks like work whose product is a document, an analysis, a piece of code or a recommendation. Anything heavily mediated by text sits in this category. If the deliverable is a file, the exposure is high.

It looks especially like work where producing a competent first draft is most of the effort and reviewing it is the rest. That shape is common across law, consulting, marketing, analysis and software. Those fields share a shape rather than a subject matter.

What it does not look like is anything requiring physical presence, manipulation in unstructured environments, or being the person legally accountable for an outcome. Those three properties are close to a complete description of what remains unexposed. Presence, manipulation and accountability are the durable three.

The definition doing the work, and its limits

Exposure here means something narrow and precise, and the precision matters. A model could cut the time a task takes by at least half without reducing quality. Quality preservation is part of the definition rather than an afterthought.

That is a statement about speed rather than about replacement. Everything downstream of it depends on what an employer decides to do with the time. The technology creates the option and people make the choice.

An occupation where every task gets twice as fast might need fewer people, the same number doing more, or more people because the work became worth doing at greater volume. The measure does not distinguish those three outcomes, and they are entirely different situations for you. One costs you a job and another hands you leverage.

Why the official projections do not show this yet

The projections are built from long-run patterns and observed adoption, and this technology is recent. That is a structural lag rather than a disagreement with the exposure research. Both can be correct about different time periods.

The declining occupations in the current release are still the previous wave: data entry, telemarketing, order clerks and payroll clerks. Nothing in the published decline list reflects language models at all. The current release is describing the previous wave entirely.

If the exposure research is right, the next several releases will look materially different. If it is wrong, they will not, which makes this a genuine test worth watching rather than a rhetorical one. Successive releases will settle the question either way.

Seniority cuts across this

Within the same occupation, the exposed share is usually higher at the junior end. That is a finding people miss because exposure gets discussed at the level of whole professions. Seniority is a dimension the profession-level view hides.

Junior work in information occupations is disproportionately drafting, summarizing, first-pass analysis and research. That is the exact profile the models handle best. Junior work and model capability overlap almost perfectly.

The implication for how these fields train people is genuinely awkward. The tasks that traditionally taught juniors the domain are the tasks most likely to be automated, and nobody has a good answer for what replaces that apprenticeship. The training problem is real and largely unaddressed.

What high exposure has actually produced so far

It has not produced visible displacement in the employment data. That is the current state of the evidence and it deserves stating as plainly as the alarming version. Absence of evidence deserves the same prominence as evidence.

Software developers are projected to add 267,700 jobs over 2024 to 2034, among the largest numeric gains of any occupation. They are also one of the most exposed occupations by any index you care to pick. High exposure and strong growth are currently coexisting.

That is not proof the exposure research is wrong, since projections are built on observed adoption and the technology is recent. It is the honest description of where things stand today. Tomorrow may differ and today is what we can observe.

What to do about it

Assume the drafting, summarizing, routine analysis and first-pass reasoning in your week gets substantially faster. That assumption is well supported and costs nothing to act on. Acting on it improves your position either way.

Then ask honestly whether your value sits in producing those things or in deciding what to do with them. Most people know the answer and would rather not examine it closely. Examining it is uncomfortable and considerably cheaper than waiting.

If it is the first, the move is toward the second, and it is available now inside most jobs. Take the exceptions, own an outcome rather than a task, and get closer to whoever the work is actually for. Proximity to the decision is what survives the automation of the draft.

The honest position to hold

Exposure is well measured and displacement is not yet observable. Holding both of those at once is harder than picking a side and it is the only defensible position. Two uncertain facts are still better than one confident error.

Anybody claiming certainty in either direction is ahead of what the data can currently support. That includes the people saying this changes everything and the people saying it changes nothing. Both camps are asserting more than anybody currently knows.

The defensible move is to act on the task-level finding, which is solid, rather than on the employment forecast, which does not yet exist. Moving toward judgment and accountability is good advice whether or not the wave arrives. No-regret moves are the right response to genuine uncertainty.