Which jobs and skills are hardest for AI to replace?

Three things protect a job most. The first is hands-on work in places that change from one job to the next, like an electrician or a nurse. The second is work built on trust, care or persuasion, like counselling, social work or winning large sales. The third is work the law says a qualified, named person must sign off, like a doctor, solicitor or auditor. Jobs that mix two or three of these tend to change the slowest.

Will AI replace software developers, writers and lawyers?

Not in one go. AI tools are cutting the time that routine parts of these jobs take: standard code and tests, first drafts of copy, document review and standard contracts. The parts that hold up are working out what to build or say, checking what the tools produce, and the judgement and client trust that people pay for. Expect fewer people doing the routine versions of these jobs, and more weight on the senior skills.

What is the difference between a task being automated and a job being replaced?

A job is a bundle of tasks, and software usually takes over some of the tasks long before the whole job. When software does an accountant's reconciliations, the accountant tends to spend more time on advice and review rather than disappear. Your result page shows this task by task for your own job.

What do the exposure levels mean?

Your level comes straight from your score, which compares your work with 923 jobs: 73 and above shows among the most exposed jobs, 50 to 72 shows more exposed than most jobs, 28 to 49 shows less exposed than most jobs, and under 28 shows among the least exposed jobs. Work done mostly on a screen with writing, figures or messages tends to land higher. Hands-on work, care and work with real responsibility tend to land lower. A high level means your work is likely to change, not that your job will go.

What should I do if my score is high?

Start small and start now. Try one AI tool used in your field on a real task each week. Move your time towards the parts of the job the tools find hard: deciding what needs doing, handling the exceptions and dealing with people. If your work can lead to a licence or a sign-off of your own, working towards it is one of the strongest protections there is. Your result page lists next steps based on your answers.

Free Career Risk Test • No Signup

Will AI Replace Your Job?

Q: What data does this test save?

We never ask for your name or email. The score is worked out in your browser and checked again on our server. When you finish, we save your job choice, your answers and your score with a random ID kept in your browser, plus your rough location (country or time zone), your browser type and a code made from your IP address that changes every day. We use this to count results and improve the test, and we never sell it. If you leave before the end, we only count which question you stopped on. Full details are in our privacy policy.

Find out how much of your day to day work AI tools could take on, how your job compares with other jobs, and which skills and courses are worth your time.

Free, no signup

How exposed is your job to AI?

A score based on your real tasks, compared with 923 jobs, in about 3 minutes.

Or pick your field to start

No name or email needed. When you finish, we save your job choice, your answers (including how you split your week and how much time your main tasks take) and your score with a random ID kept in your browser, plus your rough location (country or time zone), your browser type and a code made from your IP address that changes every day. We use this to count results and improve the tool, and we never sell it. If you leave part-way, we only count which question you stopped on. Privacy policy

Methodology

How the score works

Your score shows how exposed your work is to AI tools compared with other jobs. It runs from 5 to 95. A score of 50 means your work is more exposed than about 50% of the 923 jobs in the US O*NET database. It measures how much of the work AI tools could speed up or take on. It does not say whether you will lose your job, or when.

1. Your job's usual picture
O*NET 31.0 lists the tasks of each job and how often they are done. The GPTs are GPTs study (Eloundou and colleagues, 2023) rated whether AI tools could cut the time each task takes. We sort the tasks into seven kinds of work, such as writing, figures and hands-on work, and see how exposed each kind is and how much of the week it takes.
2. How you split your week
You build your week from 10 blocks, so each block is about 10% of your working week. You give the blocks to the kinds of work your job has. Your split counts for 70% and the usual split for your job counts for the rest, because guesses about time are rough.
3. Your main tasks
We show up to 8 of your job's main tasks, one card at a time, and you answer rarely, sometimes or most days. We read these as half, the usual and double time for that task. Your answers decide which tasks count most in each kind of work, so the result is closer to your own week. No single task can count for more than 40% of its kind of work.
4. How often you do things
Questions such as how often you deal with the public come from O*NET's work context survey. They show how much your work needs you there in person, needs people or carries responsibility, and these protect part of the exposed work. How much each answer counts was fitted across all 923 jobs, not set by hand.
5. Using AI tools
If you already use AI tools for a kind of work that takes at least 10% of your week, we count that as a sign the work is exposed. Using AI tools does not lower your score.

The answers do not show which one lowers your score, because there is no simple rule to show. The same answer can raise one score and lower another, depending on the job and how the week is split. Answer for your real week and the result will be closest to the truth.

Your result comes with a 90% range: given your answers, your true score is likely to be in it. We work the range out from what the model is unsure about, then widen it by 7%. That is the amount that made the range hold the true score 90 times in 100 when we tested the quiz on thousands of made-up people whose true scores we knew. "How sure we are" is high when the range is narrow or one exposure level is at least 80% likely.

Everyone builds their week and goes through the task cards first. After that, each question is the one we expect to narrow your range the most. The quiz stops when the likely error in your score is 4 points or less, or one exposure level is at least 80% likely. If at least 85% of people in your job give the same answer to one of the work context questions, we skip it, because it would tell us little. That takes at least 8 and at most 16 questions, about 10 for most people.

While you answer, the accuracy meter shows how close we are to being sure enough to stop. It never shows your score, so it cannot hint at which answers raise or lower it.

Each answer is final once you give it, so you cannot go back and try other answers to move the score. Each browser can take the test twice, and your last two results stay on your device so you can see them again.

The four exposure levels are the quarters of the scale: 73 and above is among the most exposed jobs, 50 to 72 is more exposed than most jobs, 28 to 49 is less exposed than most jobs, and under 28 is among the least exposed jobs. They compare jobs with each other. They are not dates.

Some honest limits. Exposure is not replacement: a job can change a lot without going away, and much of the work AI can speed up still needs a person to check it. The task lists describe jobs in the US and may not match how your job is done where you live. The ratings show what AI tools could do in 2023, and the tools have changed since. Your answers are rough guesses about your own week, which is why we show a range.

This page includes information from the O*NET 31.0 Database by the U.S. Department of Labor, Employment and Training Administration (USDOL/ETA). Used under the CC BY 4.0 license. O*NET® is a trademark of USDOL/ETA. Zebeh has modified all or some of this information. USDOL/ETA has not approved, endorsed, or tested these modifications.

Task exposure ratings: GPTs are GPTs, by Eloundou, Manning, Mishkin and Rock (2023), used under the MIT licence, github.com/openai/GPTs-are-GPTs.

Model built and checked by Ramzi Ahmad at Zebeh. Read about our team and how we use AI.

Reading your result

What each part of the score means

Each row is one thing the test looks at, and what puts you at lower or higher risk.

What we look at	Lower risk	In between	Higher risk
Exposure score The score compares how much of your work AI tools could speed up with the 923 jobs in the US O*NET database.	Under 28: among the least exposed	28 to 72: the middle half	73 and above: among the most exposed
Exposed tasks in your week Each task has a published rating for how much AI tools could cut its time. The more of your week goes on highly rated tasks, the higher your score.	Mostly hands-on or people work	A mix	Mostly writing, figures or messages
What protects you Work that needs you on site, needs trust or care, or carries real responsibility is slower to hand to software.	Being there in person, people and responsibility	Some of these	Little of these
How sure we are Your answers are rough guesses about your week, so the result shows a range. Answering more questions narrows it.	A narrow range inside one level	A range across two levels	A wide range

Occupation Benchmarks (O*NET 31.0)

How sample jobs compare before you answer

The test covers 203 jobs across 14 fields and ranks your answers against 923 occupations in the US O*NET 31.0 database. This table shows starting task exposure and main protection for 12 common jobs before your weekly split adjusts the score.

Job & O*NET Code	Field	Starting Task Exposure	Strongest Protection
Data entry clerk O*NET 43-9021.00 • ISCO 4132	Office, admin & customer service	92/100	Screen & office workflow
Translator or interpreter O*NET 27-3091.00 • ISCO 2643	Marketing, media & writing	90/100	Screen & office workflow
Copywriter O*NET 27-3043.00 • ISCO 2641	Marketing, media & writing	88/100	Screen & office workflow
Software developer O*NET 15-1252.00 • ISCO 2512	Software, IT & data	75/100	Screen & office workflow
Accountant O*NET 13-2011.00 • ISCO 2411	Finance & accounting	72/100	Licence & named sign-off
Lawyer or solicitor O*NET 23-1011.00 • ISCO 2611	Legal & compliance	60/100	Licence & named sign-off
Project manager O*NET 13-1082.00 • ISCO 2421	Sales & business	55/100	People, trust & care
School teacher O*NET 25-2031.00 • ISCO 2330	Education & training	35/100	People, trust & care
Doctor O*NET 29-1215.00 • ISCO 2211	Health & care	40/100	Licence & named sign-off
Nurse O*NET 29-1141.00 • ISCO 2221	Health & care	25/100	Licence & named sign-off
Electrician O*NET 47-2111.00 • ISCO 7411	Skilled trades & construction	15/100	Hands-on work on site
Firefighter O*NET 33-2011.00 • ISCO 5411	Public service & security	8/100	Hands-on work on site

Frequently Asked Questions

Questions about the test

It is a guide, not a forecast. It starts from published data for your job: the tasks O*NET lists for it and a rating from the GPTs are GPTs study of how much AI tools could cut the time of each task. Your answers adjust that picture to your own week, and the result shows a range because your answers are rough. When we tested it on made-up people whose true scores we knew, the 90% range held the true score about 90 times in 100. It measures exposure, not whether you will lose your job, and the ratings reflect AI tools as they were in 2023. The methodology section on this page has the details.