Which Jobs Are Actually Safe From AI?

This is the question that brings a lot of parents to this site, and it is also the question where the internet is least reliable. The confident answers run in both directions: AI is coming for everything, or AI cannot touch anything that needs hands. Both are guesses wearing a statistic as a costume.

Here is what is actually known, what it measures, and where it stops.

What is the best evidence, and what does it measure?

The strongest per-occupation research available is a 2025 Microsoft Research study, Working with AI: Measuring the Applicability of Generative AI to Occupations. The researchers took roughly 100,000 anonymised Bing Copilot conversations from January to September 2024, plus a second set of about the same size used for feedback analysis, and mapped what people were asking the chatbot to do onto the work activities that make up occupations. The study covers 785 occupational codes.

Two things about that description matter more than any number in it.

First, this is not a government source. It is a company studying its own product’s logs, published openly with its data, which is more transparency than most research on this topic offers, and it is still one company’s product over nine months.

Second, the headline measure is an AI applicability score, from 0 to 1. It combines whether the chatbot completed the work activity successfully and how much of that activity it covered. It measures overlap between an occupation’s activities and what a chatbot was seen doing. It does not measure hiring, wages, job loss, or the future.

What are the actual scores?

These are the published values from the authors’ own dataset, on a scale of 0 to 1.

Occupation	AI applicability score
Interpreters and translators	0.492
Customer service representatives	0.408
Computer user support specialists	0.334
Solar photovoltaic installers	0.160
Electricians	0.151
Wind turbine service technicians	0.125
HVAC and refrigeration mechanics	0.117
Electrical power-line installers	0.112
Welders, cutters, solderers, brazers	0.075
Plumbers, pipefitters, steamfitters	0.074
Roofers	0.009

Source: the author-published dataset, file ai_applicability_scores.csv.

The honest headline is roofers at 0.009 against interpreters and translators at 0.492, a gap of roughly fifty five times. That is a striking finding and it does not need inflating.

Are the trades uniformly low exposure?

No, and this is the part that gets flattened into a slogan. Look again at the middle of that table. Roofers, plumbers and welders sit near the floor. Electricians at 0.151 and solar photovoltaic installers at 0.160 are more than fifteen times the roofing score, and they sit above every other trade in the table.

That is not a warning about electricians losing work. It is a reminder that skilled trades contain different amounts of the kind of activity a chatbot handles: diagnosis, code lookup, layout, estimating, scheduling, documentation. A trade with more of that on the office side scores higher. Anyone summarising this research as “the trades are safe” has stopped reading at the roofing row.

What do the authors themselves say?

They are careful, and reprinting their caution is more useful than another chart. In their words, “downstream consequences of new technologies are very hard to predict and often counterintuitive,” and “AI’s effects on employment and wages will depend on hard-to-predict business decisions.” They explicitly caution against reading a high score as automation causing job losses.

There is a second caveat worth stating that is structural rather than modest. The data is chatbot conversation logs. It measures what people brought to a chatbot, which massively over-represents people who work at a keyboard. You cannot type a question to Copilot while holding a nail gun. So part of the low score for manual trades is a measurement artefact of how the data was collected, not a pure statement about whether the task could be automated. That does not make the finding worthless. It makes it a finding about observed use, which is what it says it is.

Finally, a warning about numbers you will see elsewhere. The study reports two different things: a binary coverage flag on individual work activities, and the 0 to 1 applicability score on occupations. Coverage figures are always much larger. Presenting one next to the other roughly doubles the apparent scale of everything, and it is the most common error in coverage of this paper. This site prints applicability scores only.

Which jobs does the government itself project to decline?

This is firmer ground, because it comes from the agency that has been making these projections for decades and publishes its assumptions. From the BLS fastest declining occupations table for 2024 to 2034:

• Word processors and typists, down 36.1 percent
• Roof bolters in mining, down 34.2 percent
• Telephone operators, down 27.5 percent
• Switchboard operators, down 26.3 percent
• Data entry keyers, down 25.9 percent
• Foundry mold and coremakers, down 25.9 percent
• Patternmakers in metal and plastic, down 24.4 percent
• Loading and moving machine operators in underground mining, down 22.3 percent
• Telemarketers, down 22.1 percent
• Grinding and polishing workers working by hand, down 21.2 percent
• Engine and other machine assemblers, down 21.1 percent

Those are the eleven steepest. A little further down the same table sit the office jobs people usually expect to see at the top of it: order clerks, down 17.2 percent, payroll and timekeeping clerks, down 16.7 percent, and file clerks, down 15.9 percent.

Two features of that list are worth sitting with. BLS attributes these declines to its own industry and technology assumptions and does not label any of them an AI effect, so neither will we. And six of the eleven fastest declining occupations are in manufacturing or mining: roof bolters, foundry mold and coremakers, patternmakers, underground mining loading and moving machine operators, hand grinding and polishing workers, and engine assemblers. The decline story is not a white-collar story with a couple of exceptions. It is more than half a factory and mine story. Pretending otherwise is the mirror image of the error this page is trying to avoid.

What this looks like in practice

Theo is sixteen and good at school, and his mother has spent a year worrying that whatever he picks will be automated before he finishes training. She wants a list of safe jobs. There is not one.

What she can do instead is put two things side by side for any option he raises. On one side, the BLS projection for the occupation: openings per year, growth for 2024 to 2034, and whether it appears on the declining list at all. On the other side, the Microsoft Research applicability score, read as what it is, an overlap measure from 2024 chatbot logs.

For electrical work that pairing reads: BLS projects 9 percent growth and about 81,000 openings a year over 2024 to 2034, and the applicability score is 0.151, low but clearly not zero. That is not a promise. It is a much better basis for a conversation than either of the headlines Theo’s mother started with, and the same two-column exercise works for any job he names.

So what does a parent do with this?

Stop looking for safety and start looking for optionality. The questions that survive uncertainty are the ones this site keeps returning to: how many openings a year, what does entry cost, what do the first two years look like, and what does this person do next if the work changes.

Two related pages apply the same method: data centre jobs, what is real and what is not, where the marketing has run far ahead of the federal data, and are the trades getting oversaturated, which answers the mirror-image worry with the same projections.