You've probably seen headlines claiming that AI will take over a big share of jobs. They sound precise, with a percentage and a study behind them. Here, we show you how those numbers may be produced, so you can judge them for yourself. Then you could try to make some of your own.
Where the numbers come from
Many of these studies start in the same place: O*NET, the US government's database
of jobs. It groups them into occupations, such as “Registered Nurses”, each covering many job titles,
and lists the tasks each occupation involves. Researchers assess how AI might affect those tasks, then
combine the results to produce a score for the whole occupation. In June 2026, a review written
for O*NET examined 19 major studies of AI's impact on jobs. Sixteen of them used O*NET data.
So far, so solid. But the review also found that the studies differ considerably in what they
measure and how they measure it. Those choices make a big difference to the number that appears in
the headline.
Four questions to ask of any “AI and jobs” headline
What does “affected” mean? Studies use different definitions of AI's impact. A task is
exposed if AI could affect it, whether by automating some of the work or helping a
person do it. For example, one study considered a task exposed if AI could cut the time needed to
complete it by at least half.
A task is automated if AI could do it with no person involved, and augmented if
AI helps a person do it. The review stresses that automation and augmentation are “not simply opposite
ends of a single spectrum but distinct forms of AI impact with different implications for workforce
planning and worker experience”.
A headline about AI exposure, therefore, is not a headline about jobs lost.
What did they look at? A job's task list, the skills it requires, job advertisements, or
people's interactions with chatbots. Each sees a different part of the job. The review warns
that job advertisements can be “more of a ‘marketing’ device” than a description of the work involved.
Chatbot conversations, meanwhile, reveal only how people use a particular AI tool, not everything they do at work.
Who made the judgement? Someone who knows the job well may not know what AI can do, and an
AI expert may not know the job. Some studies ask an AI model to assess the tasks instead. That's
efficient, but remember what a chatbot is doing: it predicts what's likely, and likely isn't the same
as true. The review proposes using multiple AI models and aggregating their assessments to guard against
the limitations of any single model.
What was left out? Many studies count only tasks. But a job is more than its task
list. It also includes helping a colleague who's stuck, training someone new, noticing that something looks wrong
and saying so. The review argues that by overlooking this kind of work, existing studies may have
overstated how much of a job AI could do. A job isn't simply the sum of its listed tasks.
Try this: Consider a made-up headline: “New study: AI can now do 60% of a lawyer's work.” Apply the four questions.
Affected: does “do” mean do independently, or help a lawyer work faster? Those are
very different claims.
Looked at: 60% of what? The tasks on a list, the hours in a working week, or
the duties described in job advertisements?
Judged by: lawyers, AI researchers, or an AI model?
Left out: what about the parts of a lawyer's day that aren't on a task list,
like reassuring a worried client or taking responsibility for the advice?
Until you know these, “60%” could mean very different things.
Make your own headline
Now you be the researcher. Pick a job and rate the 8 tasks O*NET lists as most important
for it. For each task, decide whether AI could do it independently, help a person do it, or
whether it still needs a person. There are no definitive answers here. Studies don't agree either, which is
the point.
Registered Nurses: the 8 most important of the 15 core tasks O*NET lists,
in its own words. Considerable preparation (usually a four-year degree). Task list last updated 2021.
Electricians: the 8 most important of the 20 core tasks O*NET lists,
in its own words. Medium preparation (vocational school, on-the-job experience or a two-year degree). Task list last updated 2022.
Graphic Designers: the 8 most important of the 17 core tasks O*NET lists,
in its own words. Considerable preparation (usually a four-year degree). Task list last updated 2023.
Software Developers: the 8 most important of the 11 core tasks O*NET lists,
in its own words. Considerable preparation (usually a four-year degree). Task list last updated 2025.
Chefs and Head Cooks: the 8 most important of the 21 core tasks O*NET lists,
in its own words. Medium preparation (vocational school, on-the-job experience or a two-year degree). Task list last updated 2025.
Accountants and Auditors: the 8 most important of the 30 core tasks O*NET lists,
in its own words. Considerable preparation (usually a four-year degree). Task list last updated 2026.
News Analysts, Reporters, and Journalists: the 8 most important of the 26 core tasks O*NET lists,
in its own words. Considerable preparation (usually a four-year degree). Task list last updated 2026.
Secondary School Teachers, Except Special and Career/Technical Education: the 8 most important of the 30 core tasks O*NET lists,
in its own words. Considerable preparation (usually a four-year degree). Task list last updated 2025.
Lawyers: the 8 most important of the 15 core tasks O*NET lists,
in its own words. Extensive preparation (usually a graduate degree, such as a master's, medicine or law). Task list last updated 2024.
Customer Service Representatives: the 7 most important of the 7 core tasks O*NET lists,
in its own words. Little to some preparation (a few days to a year of training on the job). Task list last updated 2022.
Pharmacists: the 8 most important of the 20 core tasks O*NET lists,
in its own words. Extensive preparation (usually a graduate degree, such as a master's, medicine or law). Task list last updated 2026.
Veterinarians: the 8 most important of the 14 core tasks O*NET lists,
in its own words. Extensive preparation (usually a graduate degree, such as a master's, medicine or law). Task list last updated 2020.
Architects, Except Landscape and Naval: the 8 most important of the 25 core tasks O*NET lists,
in its own words. Extensive preparation (usually a graduate degree, such as a master's, medicine or law). Task list last updated 2026.
Interpreters and Translators: the 8 most important of the 8 core tasks O*NET lists,
in its own words. Considerable preparation (usually a four-year degree). Task list last updated 2021.
Video Game Designers: the 8 most important of the 24 core tasks O*NET lists,
in its own words. Considerable preparation (usually a four-year degree). Task list last updated 2022.
Hairdressers, Hairstylists, and Cosmetologists: the 8 most important of the 14 core tasks O*NET lists,
in its own words. Medium preparation (vocational school, on-the-job experience or a two-year degree). Task list last updated 2026.
Data Scientists: the 8 most important of the 15 core tasks O*NET lists,
in its own words. Considerable preparation (usually a four-year degree). Task list last updated 2026.
Plumbers, Pipefitters, and Steamfitters: the 8 most important of the 24 core tasks O*NET lists,
in its own words. Medium preparation (vocational school, on-the-job experience or a two-year degree). Task list last updated 2022.
Your ratings make two headlines about the same job:
What your headline left out
Every task counted the same. Real studies use different ways of combining task-level ratings. One weighted a job's core
tasks twice as heavily as the rest; another weighted each task by how much of the working day it takes.
Change the weights and the headline changes, without anyone changing their mind about AI.
The work around the tasks. You rated a list of duties. You didn't rate helping a
colleague, covering for someone who's off sick, or spotting a problem nobody asked you to look for.
Neither do most studies.
Try this: pick the job you'd most like to do. Write down one thing a person in that
job does every week that isn't on any task list. Then ask: Could AI do it? Could it help?
And would people want AI to take over that part of the job?
The official data hasn't caught up
O*NET doesn't measure AI's effect on jobs yet. The June 2026 review is a proposal: it
recommends that O*NET consider building 16 new measures, and says the final set may end up smaller.
Meanwhile, O*NET's lists of the software each occupation uses name an AI chatbot for only about
ten of the 900+ occupations it has software lists for. That doesn't mean AI isn't used in the rest. It means
the official record lags behind what's happening.
That's worth remembering whenever you see a confident number about AI and jobs. Ask when the data
behind it was collected, and how fast things have changed since. It's the same question as the
critical thinking toolkit's Over what period?, and you
can see it used on a real argument about AI and jobs in
the first case study.
The short version
A number about AI and jobs is only as good as the choices behind it: what “affected” means, what was
measured, who made the judgement and what was left out. Ask those four questions before you
let it influence what you choose to study.