Glossary: the words on this site, in plain English

Our headline score for each job, from 0 to 100. Higher is safer: AI can do less of the work, or does it less well.

It combines how much of the job AI can do with how well AI does it compared with a qualified person, and it is always shown beside those answers.

How the score works

Printed after a score to show which way is good: a higher number means the job is safer from AI.

So 85/100 ↑ safer is safer than 60/100 ↑ safer.

Our one-word answer to “Will AI replace this job?”, read from the score: from Largely. for the lowest scores to Nah. for the highest.

Largely. 0–19, Mostly. 20–39, Partly. 40–59, A little. 60–79, Nah. 80–100. The bands are fixed: we do not grade on a curve.

The five answers

Our answer to “Will AI replace this job?” when the job scores 80 or more out of 100: the work mostly needs a person.

It means AI can do little of this work today, not that the job will never change. Nah is also the initials of NeedsAHuman.

One of five fixed score ranges, each with its own one-word answer. The top band, 80 to 100, holds the safest jobs.

The bottom band, 0 to 19, is for jobs AI could largely do (Largely.).

One piece of work a job involves, such as “Process invoices for payment.” We score every task, then add them up.

The task lists come from O*NET, the US government’s job database.

How a job’s working hours divide between its tasks. Tasks that matter more and come up more often get a bigger share.

A job’s working time sorted three ways: tasks AI does, tasks AI helps a person with, and tasks that need a human.

The three shares add up to 100%. Amber is always the human share.

Tasks AI can already do mostly by itself today, with a person checking the result.

Every task gets an AI score from 0 to 100; tasks scoring 50 or more go here.

Tasks a person still does, with AI speeding up parts of them, such as looking things up or writing a first draft.

Tasks with an AI score from 25 to 49 go here.

Tasks AI can do little of today, such as hands-on work, in-person care, or decisions someone has to answer for.

Tasks with an AI score under 25 go here.

How much of a job’s working time AI can handle today, from 0 to 100. Higher means AI can do more of the job.

Also called coverage. It is not the same as the task split: a task AI only helps with still counts for part of its time.

How we measure it

How AI’s work compares with a qualified person’s on this job, from 0 to 100, where 50 means just as good.

Also called quality parity. Most jobs have no direct test yet, so their pages say “Not yet measured.”

How we judge quality

Nobody has yet tested AI against people doing this job, so we show no quality number rather than guess one.

These jobs have evidence grade D.

How strong the proof is, from A to D. A is a blind, expert-marked test of AI against people; D means none yet.

B is a test on closely related work; C is a benchmark with no human comparison.

The four grades

The span of years when AI could largely do this job, covering 8 in 10 of our simulated futures. A range, not a forecast.

The big year is the middle one: half the futures come sooner, half later. “2060+” means later than 2060, where our model stops.

How the timeline works

Our strict test: AI doing at least 90% of a job’s working time, as well as a person, with at least half of employers using it.

Pages also say “largely automated” for the same test. It is not a prediction that the job ends.

The middle 80% of our results: one simulated future in ten comes sooner than the range, and one in ten later.

One possible future in our timeline model, with its own guesses about how fast AI improves and how fast employers take it up.

Each job’s range comes from 10,000 of them. The year-by-year charts show ten, spread across the range.

The middle value: half are above it and half below. Unlike an average, a few extreme values cannot drag it.

Our published checklist for rating what AI can do on each task, from “no meaningful part” to “the whole task, at a qualified worker’s standard.”

An AI model applies it to every task; the levels are T0 to T3.

Read the rubric

A dated edition of all our scores, such as 2026-Q4. Scores can change from one release to the next, so cite the release.

What changed in each release

The version number of our scoring method, such as 1.2.0. It goes up when the method changes; new data alone does not change it.

Every method change

The things that keep work with people, such as licensing, legal liability, clients wanting a person, and physical work, each scored 0 to 100.

Licensing, regulation, liability and client preference also slow our timeline a little.

The share of a job’s time spent on hands-on work. Software cannot do it, so AI only counts there as far as robots can.

The kind of robot needed runs from fixed machines to mobile robots to humanoids with hands.

The cost of AI doing the hours of work it could handle, next to a person’s wage for those hours. Both are ranges.

The AI cost covers the AI service only: not setup, licenses, oversight or the time a person still spends checking.

The kinds of ability a job’s tasks lean on, such as writing, analysis or care, with how good AI is at each today.

How much of a job’s work AI could affect, by doing tasks or helping with them. Exposed does not mean the job will disappear.

What exposure really means

A job as the US government defines it, such as “Accountants and Auditors.” Your own role may mix tasks from more than one.

We score about 900 of them. Your employer and your own tasks may differ from the average.

A group of related jobs in the official US list, such as Legal or Healthcare Support. We show each family’s average score.

Browse the families

An industry, such as banking or restaurants. Its score is the average of the jobs people there actually do.

Every sector

An average where each job counts in proportion to how many people do it, so a common job counts more than a rare one.

A place’s position when every state or sector is ordered by its average score, such as 23rd of 51. First is the least exposed to AI.

Averages sit close together, so a few places in rank can mean very little.

A job that makes up a bigger share of work here than across the US. “3.5 times” means three and a half times as common.

People whose job scores below 60 out of 100 on Still needs a human, where AI can already do much of the work.

The share of businesses that told a US Census Bureau survey they had used AI in their work in the past two weeks.

The adoption tracker

How much people in a state use Anthropic’s Claude AI assistant, for the number of working-age people, where 1.00 is the US average.

The US government’s estimate of how much a job will grow or shrink over ten years, from the Bureau of Labor Statistics.

How many jobs the government expects to need filling each year, from growth plus people retiring or changing careers.

The middle yearly wage for the job: half the workers earn more and half earn less.

The US government’s free database of about 1,000 jobs, listing each job’s tasks and how much each one matters.

Run for the US Department of Labor. Every score here starts from its task lists (version 31.0).

onetcenter.org

The US government department for work and workers. Its Employment and Training Administration (ETA) pays for O*NET.

dol.gov

The Bureau of Labor Statistics: the US government’s official source for how many people do each job, what they earn, and job outlook.

bls.gov

Occupational Employment and Wage Statistics: the government survey that counts how many people do each job, and their pay, by state.

Run by the Bureau of Labor Statistics.

The Bureau of Labor Statistics’ official ten-year outlook for each job, now 2025 to 2035: expected growth or decline, and openings each year.

The official US code for a job, such as 13-2011 for accountants and auditors. The code ties each page to government data.

SOC is the Standard Occupational Classification. O*NET adds two digits for some specialties, such as 29-1141.01 (O*NET-SOC).

The UK’s official job codes, such as 2421 for chartered and certified accountants. We match each job to its closest UK codes.

The Office for National Statistics, the UK government’s official statistics body. Our UK job and pay figures come from it.

ons.gov.uk

A large UK household survey by the ONS, used to estimate how many people in the UK do each job.

The ONS survey of UK pay, job by job. It is where our UK pay figures come from.

The official codes for industries in North America, such as 5221 for banks. We use them to say which jobs belong to a sector.

The US Census Bureau’s Business Trends and Outlook Survey, which asks businesses every two weeks, among other things, whether they use AI.

Data published by Anthropic, the company that makes the Claude AI assistant, on which work tasks people actually use Claude for.

It shows where AI is used today, not only where it could be. Its robot data also caps hands-on tasks.

anthropic.com

A Microsoft Research study of real conversations with its Copilot assistant, showing which kinds of work people use AI for.

A test by OpenAI in which experts compared AI’s work with professionals’ work on real tasks from 44 jobs, without knowing which was which.

It is run by an AI company, so we weight it as a vendor study.

An independent group that tests AI tools for legal work against lawyers answering the same questions.

Microsoft’s AI system for medical diagnosis, tested against doctors on difficult published patient cases.

A 2023 study by Eloundou and colleagues that rated every O*NET task for exposure to AI. We use it only to check our scores.

A data company we use for Google search counts and for collecting answers from AI assistants. Its figures are estimates.

A free license: anyone may reuse and republish the data, even commercially, as long as they credit the source with a link.

Our scores are open under it; so is O*NET.

Download the open data

The UK government’s free license for its data: anyone may reuse it as long as they credit the source.