Exposure to large language models rises weakly with the difficulty of job preparation — the better-credentialed the information work, the more of its tasks a model can materially speed up. That reverses earlier automation, which pressed hardest on routine manual and low-credential work. The occupations most exposed are those whose output is text, code, analysis and structured judgment.
This wave points at the opposite end of the labor market
Robotics and earlier software automation concentrated on routine manual and routine clerical work — predictable, physical or rule-following, and generally lower paid.
Exposure to large language models runs the other way. 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. Anybody reasoning from the last wave will get this one backwards.
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 — memos, analyses, code, recommendations, summaries of other documents.
It 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.
What high exposure looks like in practice
Work whose product is a document, an analysis, a piece of code or a recommendation. Work that is heavily text-mediated. Work where a competent first draft is most of the effort and review is the rest.
What it does not look like: anything requiring physical presence, manipulation in unstructured environments, or being the person legally accountable for the outcome.
The definition doing the work, and its limits
Exposure here means something narrow and precise: a model could cut the time a task takes by at least half without reducing quality. That is a statement about speed.
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, and they are entirely different outcomes for you.
Why the official projections do not show this yet
They are built from long-run patterns and observed adoption, and the technology is recent. The declining occupations in the current release are still the previous wave — data entry, telemarketing, order and payroll clerks.
If the exposure research is right, the next several releases will look materially different. If it is wrong, they will not, and that is a genuine test worth watching rather than a rhetorical one.
Seniority cuts across this
Within the same occupation, the exposed share is usually higher at the junior end. Junior work in information occupations is disproportionately drafting, summarizing, first-pass analysis and research — which is the exact profile the models handle.
That has an awkward implication for how these fields train people. 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.
What to do about it
Assume the drafting, summarizing, routine analysis and first-pass reasoning in your week gets faster. Then ask honestly whether your value sits in producing those things or in deciding what to do with them.
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.
What high exposure has actually produced so far
Not visible displacement in the employment data. Software developers are projected to add 267,700 jobs over 2024-34 — among the largest numeric gains of any occupation — despite being one of the most exposed occupations by any index you pick.
That is not proof the exposure research is wrong. Projections are built on observed adoption and the technology is recent. But it is the current state of the evidence, and it deserves to be stated as plainly as the alarming version.
The honest position to hold
Exposure is well measured and displacement is not yet observable. Anybody claiming certainty in either direction — that this changes everything, or that it changes nothing — is ahead of what the data can support.
The defensible move is to act on the task-level finding, which is solid, rather than on the employment forecast, which does not exist yet. Move toward judgment and accountability; that is good advice whether or not the wave arrives.
Common questions
Which occupations are most exposed to generative AI?
Credentialed information work — text, code, analysis and structured judgment. Exposure rises weakly with the difficulty of job preparation.
How is that different from earlier automation?
It is close to a reversal. Earlier waves pressed on routine manual and low-credential work; this one presses on better-educated information work.
What does exposed actually mean here?
That a model could cut the time a task takes by at least half without reducing quality. It is a claim about speed, not replacement.
Does faster mean fewer people?
Not necessarily. It can mean fewer, the same doing more, or more if the work becomes worth doing at greater volume. The measure does not distinguish them.
Why do official projections not show this?
They are built from observed adoption and long-run patterns, and the technology is recent. The declining list still reflects the older wave.
Which occupations are most exposed to generative AI?
Credentialed information work — text, code, analysis and structured judgment. Exposure rises weakly with the difficulty of job preparation.
How is this different from earlier automation?
It is close to a reversal. Earlier waves pressed on routine manual and clerical work; this one presses on better-educated information work.
Is junior or senior work more exposed?
Junior, generally. Drafting, summarizing and first-pass analysis are exactly what models handle — which is also how juniors traditionally learned the domain.