Go deeper · Use, question, build
How does AI change jobs, and what can you do about it?
Mostly by changing tasks, not by swapping a whole job for a machine. A job is a bundle of tasks: AI can speed some of them up, take over a few, and create new ones, such as labeling data, testing models and keeping them running. You can prepare by learning how your field works, how to check AI output, and where people stay responsible.
What you will be able to do
Break a job into tasks and judge where AI could help, explain the hidden human work that AI systems depend on, read a jobs forecast by asking who made it and what it measures, and choose a small career project.
Before you start
Nothing required. It helps to know what a model is and where training data comes from: see Is every AI a chatbot? and What did we teach it by accident?
1 · You have already seen this
The chat box that answers before a person does
Many websites now answer simple questions in a chat box.
Ask a library or a shop what time it opens, and a chat box may reply in a second. A person still wrote that answer and has to change it when the hours change. The staff are still there too, with more time for things a chat box handles badly, such as helping someone find a source or calming down an argument.
That is one common way AI changes work. A job is a bundle of tasks, and AI can speed some of them up, take over a few and create new ones, such as checking what the tool produced. This page shows how to think that through for a real job, lets you try it, and shows where common claims about AI and jobs go wrong.
2 · What happens, step by step
Thinking through a job, one task at a time
Four steps for working out what AI changes in a job. The next section goes deeper into steps 1 and 3, and the lab lets you do step 2 yourself. After the lab, a section on hidden human work goes deeper into step 4.
How to think it through
Four steps
- 1 · List the tasksWrite down what the person actually does in a day. A librarian answers repeated questions, creates records for new books, teaches a class and keeps the library a welcoming place.
- 2 · Sort each taskFor each one, ask whether AI could help, could do much of it, or whether it is mostly human work.
- 3 · Weigh more than the toolWhether a task changes hands also depends on cost, how bad a mistake would be, rules and laws, whether customers want a person, and who is held responsible.
- 4 · Look for the new workSomeone has to set up the tool, check what it produces and handle the cases it gets wrong. Those tasks belong to the job too.
Going deeper · steps 1 and 3
Jobs are bundles of tasks
Researchers who study AI and work usually start by breaking each job into tasks.
Think about a librarian. In one day they might answer the same question about opening hours twenty times, create records for new books, help a student find a source, teach a class, and calm down an argument. Those are very different tasks. An AI tool might handle the first two quite well and be of little use for the last.
That is why serious studies, such as those by the OECD and the International Labour Organization (ILO), score tasks and skills, not whole job titles. When a few tasks in a job become automatic, the job often changes shape: people spend more time on the remaining tasks, and new ones appear, such as checking the tool’s output.
Whether a task is handed to AI also depends on more than what the technology can do. It depends on cost, on how bad a mistake would be, on rules and laws, on whether customers want a person, and on who is held responsible. A tool that is right 95% of the time may be fine for drafting a caption and unacceptable for a medical decision.
The ILO’s 2023 study of generative AI concluded that most jobs are more likely to be augmented, meaning a person does the job with AI help, than fully automated. Clerical work, such as data entry and scheduling, was the most exposed.
Predict first
If AI tools can do much of 2 of the 6 tasks in a job, does that job disappear?
Show the answer
Not necessarily. The other 4 tasks still need doing, and someone has to set up and check the AI’s work. The job may change, or fewer people may be needed for some tasks, or the time saved may go into work that was neglected. Which one happens depends on the employer, the cost, the risks of mistakes and the demand for the service, not only on the technology. Try it with real tasks in the lab below.
3 · Try it
Break a job into tasks
Pick a job, sort its six tasks, then compare with ARQIV’s suggestions.
Lab
Task breakdown
For each task, choose “AI could help,” “AI could do much of it,” or “Mostly human.” Then press Compare to see ARQIV’s suggestion and the reasoning.
No AI runs here. ARQIV’s suggestions are a judgment about today’s tools (October 2026), not a prediction
For each job, decide for every task whether AI could help, could do much of it, or whether it is mostly human. Then compare with ARQIV’s suggestions below. They are a judgment about today’s tools (October 2026), not a prediction.
| Task | ARQIV’s suggestion | Why |
|---|---|---|
| Answer “What are the opening hours?” and “Where is the printer?” | AI could do much of it | Repeated questions with fixed answers suit a chatbot or a help page. A person still writes the answers and keeps them up to date. |
| Create catalog records for new books | AI could do much of it | Software can draft records from publisher data and suggest subject headings. A librarian checks them, because one wrong record hides a book from everyone who searches. |
| Suggest a book to a student who says they hate reading | AI could help | A recommender can list titles similar to things the student liked. Noticing what the student really means, and changing course mid-conversation, is the human part. |
| Decide which new books to buy with a small budget | AI could help | Loan data can show what gets borrowed. Deciding what this community should have, including books few people ask for yet, is a judgment about values. |
| Teach a class how to check whether a source is reliable | Mostly human | AI can help prepare examples, but teaching means reading the room, answering follow-up questions and being responsible for what students learn. |
| Keep the library a safe, welcoming place for everyone | Mostly human | This is about relationships, fairness and handling conflicts face to face. |
A librarian’s day mixes quick questions, careful record-keeping, teaching and looking after people. Tools can take on the routine questions and first drafts of catalog records, which can leave more time for teaching and for readers who need help.
| Task | ARQIV’s suggestion | Why |
|---|---|---|
| Explain the scan to a nervous patient and help them hold still | Mostly human | Reassurance, touch and judgment about someone in pain are human work, and the technician is responsible for the patient’s safety. |
| Position the patient and the equipment | Mostly human | Bodies differ and patients may be injured. Software can show guides, but a person moves and checks. |
| Pick scan settings for a clear image with as little radiation as possible | AI could help | Machines suggest settings by body size and scan type. The technician adjusts for the patient in front of them. |
| Check that an image is sharp and shows the whole area, and retake if not | AI could help | Software can flag blurred or cut-off images. The decision to repeat a scan, which means more radiation, stays with a trained person. |
| Clean up noisy or grainy scan images | AI could do much of it | Image reconstruction and noise reduction are done by software, and some scanners now use trained models for it. People still check the results. |
| Follow radiation-safety rules to protect patients, visitors and staff | Mostly human | Machines log doses and show warnings, but checking who is in the room, shielding and stopping a scan when something is wrong are a trained person’s responsibility. |
A radiology technician (also called a radiographer or radiologic technologist) works with machines all day, but the core of the job is a nervous patient on the table. Software increasingly helps with settings and image quality; the physical, caring and safety work stays with people.
| Task | ARQIV’s suggestion | Why |
|---|---|---|
| Spot weeds among crops and spray only the weeds | AI could do much of it | Camera-guided sprayers do this in some fields today. The farmer chooses where and when to use them and checks for misses, especially in new light or weather. Try it at the farming lab. |
| Fill in field records and paperwork | AI could do much of it | Machines can log what was planted and sprayed automatically. The farmer signs off and fixes gaps. |
| Decide when to water, from weather forecasts and soil sensors | AI could help | Forecasts and sensor dashboards help a lot. The farmer weighs them against what they see and the cost of water. |
| Choose what to plant next season | AI could help | Yield maps and price data inform it. The decision includes risk, family plans and long-term care of the soil. |
| Notice that an animal is unwell and care for it | Mostly human | Sensors can flag changes in movement or feeding. Care, treatment decisions and responsibility for animal welfare stay with people. |
| Repair a broken machine in the middle of a field | Mostly human | An app can help find a manual, but improvising a fix with the tools at hand is skilled physical work. |
Farming already uses GPS steering, sensors and camera-guided sprayers. The farmer’s job shifts toward setting up, checking and repairing these tools, while decisions about land, animals and money stay human.
| Task | ARQIV’s suggestion | Why |
|---|---|---|
| Turn a recorded interview into a written transcript | AI could do much of it | Speech recognition does most of this. The reporter checks names, numbers and quotes before publishing anything. |
| Write a short, routine report from a table of sports scores | AI could do much of it | Some newsrooms have generated routine data stories with software for years. An editor sets the templates and checks the output. |
| Summarize a long report to find what is new in it | AI could help | A model can draft a summary quickly, but it can miss or invent details, so the reporter reads the parts that matter. |
| Check whether a viral photo is real | AI could help | Reverse image search and other tools help. Tracing the original source and judging the evidence is careful human work. |
| Build trust with a source and interview a witness | Mostly human | People decide whom to trust and what to share face to face. This is relationship work. |
| Decide what is newsworthy and fair to publish | Mostly human | Editorial judgment is where responsibility lies. Someone has to answer for the choice. |
Journalism has long used software for transcripts and routine data stories. The tasks that define the job, finding out what is true and deciding what the public needs to know, are the ones where people are accountable.
Disagreeing is part of the activity. These sorts are judgments, and reasonable people differ, especially about tasks in the middle. The useful question is not “Will AI take this job?” but “Which tasks change, who checks the result, and who is responsible when it is wrong?”
Reading the numbers
Read a jobs claim like a scientist
Headlines about AI and jobs often mix very different kinds of numbers. Ask who measured, how, where and what exactly.
| Source | How it was made | What it found | Read it as |
|---|---|---|---|
| OECD Employment Outlook 2023 | Experts judged which skills AI can automate; occupations were scored by how many of their skills are exposed | Occupations at highest risk of automation account for about 27% of employment across OECD countries. Little evidence so far of AI reducing overall employment. | Exposure, not a count of lost jobs |
| ILO Working Paper 96 (2023) | Each task in each occupation was rated for exposure to generative AI | Most jobs are more likely to be augmented (helped) than automated. Clerical work is the most exposed group, and because many clerical workers are women, effects differ by gender. | A task-level estimate of what is possible, not a forecast |
| World Economic Forum, Future of Jobs Report 2025 | A survey of more than 1,000 large employers about their plans for 2025–2030 | Employers expected about 170 million jobs to be created and 92 million displaced worldwide, a net gain of 78 million. | What managers of big companies expect, not a measurement |
| U.S. Bureau of Labor Statistics, Occupational Outlook Handbook (2025–35 projections) | Official projections for U.S. occupations | Data scientist employment is projected to grow 35% from 2025 to 2035, “much faster than average” (the average for all occupations is 3%). | One country, one occupation, a projection |
None of these numbers says “X% of people will lose their jobs.” An exposure score says how many tasks a technology could affect. A survey says what employers expect. A projection extends recent trends for one country. All are useful, and all can be wrong, so check the year and the method before you repeat a number.
New evidence · entry-level jobs
The first effects are landing on young workers
Stanford economists tracked millions of U.S. paychecks through June 2026. Employment of 22- to 25-year-olds in the most AI-exposed jobs is about 19% below where it would be had it kept pace with less-exposed peers; older workers in the same jobs show no comparable gap. It comes mostly from less hiring, and it is concentrated where AI replaces tasks rather than helps with them. The authors call these early signs, not proof of cause, and found no economy-wide job loss.
Documented, early · working paper More in the Future of work hub →
Careers
Where people work with AI
Many AI jobs are not programming jobs. Fields need people who know the domain, the data and the people affected.
A model that spots crop disease is only as good as the agronomists who labeled the photos and the farmers who say whether its alerts make sense. A weather model is trusted because meteorologists compare it with physics models and know when it fails. This domain expertise is hard to replace and becomes more valuable when AI tools are common.
AI in space
Observatory data scientist, astronomer
Teaching models to sort millions of objects, then deciding which discoveries are real.
AI in weather
Meteorologist, emergency manager
Comparing AI and physics forecasts and deciding when to warn people.
AI in health
Radiologist, clinical data scientist
Checking that tools work fairly for every group of patients.
AI in engineering
Robotics technician, safety engineer
Maintaining robots and testing self-driving systems on rare, risky events.
AI in farming
Agronomist, precision agriculture technician
Calibrating camera-guided sprayers and reading drone maps.
For counselors
Career conversations
A 20-minute guide for counselors and advisory teachers: work as tasks, people as accountable, one skill to practice.
Your move
Career projects you can start now
Small projects teach more about AI and work than any forecast. Pick one, do it in a week, and write down what surprised you.
Task diary
Interview an adult about their job. List ten tasks, sort them like the lab, and ask which tasks they would want help with. Compare their view with yours.
Label it twice
With a friend, label the same 40 photos (ripe or unripe fruit, weed or crop) without talking. Count how often you disagree. That is what labelers face. See the weed-spotter activity on AI in farming.
Build and document a model
Follow the seven steps in Build a small model with data you collect yourself, and write a model card.
Test a chatbot at a job task
Pick one task from the lab, write five test questions with known answers, and score two chatbots fairly. Use Test the models.
Be the maintenance crew
Record two weather apps’ forecasts for a week and compare them with what happened, as in the classroom activity on AI in weather. When would you stop trusting one?
Write a career card
Choose a job in your town. Write a card like the ones on the “AI in…” pages: what the job is, how people there work with AI, and which tasks stay human. Ask someone in that job to check it.
4 · Common mix-ups
Four claims about AI and jobs to read twice
Each one comes with a better way to say it.
Mix-up: “AI will take 27% of jobs.”
Better: The OECD’s 2023 figure is the share of employment in occupations at highest risk of automation, an exposure measure. It is not a forecast of job losses.
Mix-up: AI jobs are only for programmers.
Better: Labelers, technicians, testers, domain experts, trainers and people who check fairness and safety all work with AI. Knowing a field well is a career asset.
Mix-up: AI systems run by themselves.
Better: They rely on people who label data, review content, evaluate answers and maintain models, often out of sight and sometimes in poor conditions.
Mix-up: If a task can be automated, it will be.
Better: Cost, risk, rules, trust and customer choice all decide whether a task actually changes hands.
5 · Check yourself
Three quick questions
Answer in your head or on paper before you open each one.
A headline says, “Robots will replace journalists.” Using the idea of tasks, give a better description.
Check your answer
Some journalism tasks, such as transcribing interviews and writing routine score reports, can be done largely by software. Others, such as interviewing witnesses, checking facts and deciding what is fair to publish, depend on people who are accountable. The job changes; it does not simply vanish.
Name two kinds of hidden human work that an image-recognition app depends on, and one concern about how that work is done.
Check your answer
For example: people who labeled the training photos, and people who review its mistakes or maintain it after launch. A concern: labeling and content review are often low-paid, outsourced piecework, and reviewers may have to look at disturbing material.
One report surveys 1,000 employers; another rates which tasks AI could do. Why might they give different-sounding answers?
Check your answer
They measure different things. The survey records what managers expect to happen in their companies. The task rating measures what is technically exposed, not what will happen. Check who made a number, how, where and when before comparing them.
For younger learners
A grades 3–5 version
No forecasts and no vocabulary: one sorting game about a familiar place.
20 minutes · whole class
Helper or human?
- List eight tasks people do at school: unlock the doors, cook lunch, take attendance, comfort a crying child, read a story, sort lost property, plan a trip, fix a leaky tap.
- For each one, ask: could a machine help, could a machine do most of it, or does it need a person? Vote with three corners of the room.
- Talk about the tasks where the class split. What would the machine need to know?
Exit question
One sentence
“Name one task a machine could help with and one task that needs a person, and say why.” Good answers mention feelings, safety or checking.
Honest limits
What this page leaves out
What this page leaves out
Wages, unions and labor law, how automation affects different countries and regions, retraining programs, and the long history of earlier automation. We give no salary figures or “best jobs” lists: those change quickly and depend on where you live. The task sorts in the lab are ARQIV’s judgment about today’s tools, not research findings.
Where to go next
Keep going
Build something
Can you build and test a model yourself?
A complete small project, from question to model card.
Who checks?
Who checks that an AI system is fair and safe?
Purpose, affected people, evidence and oversight.
Revisit · 03
What did we teach it by accident?
Where labels come from and why they matter.
Future of work
What is changing, and which skills last?
Entry-level evidence, skills employers name, and labs.
Teach it
Decide & Adapt
Six lessons on deciding with missing facts and learning fast.
Sources
Where this comes from
- OECD (2023), OECD Employment Outlook 2023: Artificial Intelligence and the Labour Market.
- Paweł Gmyrek, Janine Berg and David Bescond (2023), “Generative AI and Jobs: A global analysis of potential effects on job quantity and quality,” ILO Working Paper 96.
- World Economic Forum (January 2025), Future of Jobs Report 2025, based on a survey of more than 1,000 employers.
- U.S. Bureau of Labor Statistics, Occupational Outlook Handbook, “Data Scientists” (2025–35 projections, checked October 2026).
- Partnership on AI (2021), “Responsible Sourcing of Data Enrichment Services.”
- Billy Perrigo, TIME (January 2023), investigation into outsourced content labeling for a chatbot’s safety filter in Kenya.
- Erik Brynjolfsson, Bharat Chandar and Ruyu Chen, "Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence," Stanford Digital Economy Lab working paper, revised August 12, 2026 (first version August 2025).
AI words on this page
Artificial intelligence, Machine learning, Training data, Label, Evaluation, Learning from human feedback (RLHF), Human oversight, Automation bias, Model drift, Generative AI, AI literacy
All definitions live in one place: AI words in plain language →
Review status
Written by ARQIV. Last checked October 2026. Independent subject review and classroom pilots are still to come, so treat the activities as drafts and tell us about anything that is wrong or unclear.