Will AI take my job, or change it?

If you're wondering whether AI will take your job, "AI could affect 1 in 4 jobs" may sound like an answer. Usually, it measures tasks a tool might touch. Job loss needs different evidence. Each figure below names what its study measured. Sources were checked on October 28, 2026; newer results may change the picture.

Consider a fictional volunteer coordinator at a neighborhood book swap, written for this guide. Their job is a bundle of tasks. They write the monthly notice, answer emails, keep the volunteer schedule, sort donated books, decide what to do with damaged ones, greet visitors, and carry boxes.

An exposure index scores how much a generative AI tool could do each task, then combines those scores for the occupation. Writing the notice and answering routine emails would score high. Carrying boxes would score zero. Deciding whether a water-stained book is still worth shelving sits somewhere between.

The International Labour Organization's refined global index (Working Paper 140, May 2025) does this at scale. It scored tasks using a worker survey, expert review, and an AI model trained on their judgments. It found that 1 in 4 workers worldwide are in an occupation with some exposure. Only 3.3% of global employment falls into the highest of its 4 exposure levels. Clerical work is the most exposed group. Exposure in the top level is higher for women's employment (4.7%) than men's (2.4%), and overall exposure is 34% of employment in high-income countries against 11% in low-income ones.

The ILO concludes that "transformation of jobs is the most likely impact," because most jobs still contain tasks that need a person. The index measures likely change, not job loss. Whether a job shrinks, grows, or changes shape depends on cost, demand, and employers' choices.

The Anthropic Economic Index maps conversations with Anthropic's own assistant to occupational tasks, using an automated system that reports aggregated patterns without revealing individual conversations. In its March 2026 report, coding remained the most common use: tasks linked to computer and mathematical occupations made up 35% of conversations on its consumer product. Its June 2026 report added a survey. Of its respondents, 10% rated losing their own job in the next year as likely or very likely.

These figures describe one company's users on one product. The June survey says it is not representative. Computer and mathematical occupations were about 30% of respondents against about 4% of US employment. Usage data shows which tasks people hand over first, not what happens to people who never open the tool.

Field experiments compare workers given a tool with similar workers without it. They can show cause and effect for the task, tool, and people studied.

One widely cited study followed a staggered rollout to 5,172 customer support agents at one software company. In Generative AI at Work, published in the Quarterly Journal of Economics in 2025, an assistant that suggested replies raised issues resolved per hour by 15% on average. Newer and less skilled agents gained 30%. Experienced, higher-skilled agents saw small speed gains and small declines in quality.

In a 2025 trial by METR, a research group, 16 experienced open-source developers took 19% longer on 246 real tasks when AI tools were allowed. They had expected to be 24% faster, and afterward still believed they had been about 20% faster. METR now marks that result as out of date. Its February 2026 update found estimates leaning toward a speedup with newer tools. The authors call the new data weak evidence, though, because developers who most wanted AI often declined to work without it.

Gains depend on the user and task. Measure time saved; people's estimates can be wrong.

Employment data asks whether there are fewer jobs. Early evidence is mixed.

Canaries in the Coal Mine?, from the Stanford Digital Economy Lab (revised August 12, 2026), uses payroll records from ADP, a payroll company, through June 2026. It finds no sign of economy-wide job loss. Employment of workers aged 22 to 25 in AI-exposed occupations, though, stands 19% below where it would be had it kept pace with less-exposed peers. The gap comes mostly from less hiring, not layoffs. The authors present it as an early descriptive signal, not proof of cause, and note it is larger in their sample than in national surveys.

A study of Denmark, Still Waters, Rapid Currents (revised March 2026), linked adoption surveys to official employment records. It rules out effects larger than 2% on earnings or recorded hours in exposed occupations, 2 years after ChatGPT's launch. It did find employers reorganizing tasks.

Stanford HAI's 2026 AI Index compiles surveys in its economy chapter reporting that 88% of surveyed organizations used AI in 2025.

When you meet a forecast of jobs lost or created, ask 4 questions. Is it measuring exposure, usage, an experiment, or actual employment? Whose data is it, and who is missing? What period does it cover? And what does it assume about how fast companies adopt the tool? Reading AI news covers the same habit for product claims.

For 1 person or a team, start with a task list. The book swap coordinator could test drafting the monthly notice, checking its date, time, and book limit in a minute. Routine emails also suit a draft, read before sending. Deciding about damaged books needs a person who knows what the swap accepts. Carrying boxes needs neither.

Count the time to a finished, checked result, including repairs. Counting drafts overstates the gain: heavy repairs can take longer than writing from scratch. A team can compare a few weeks of the same task with and without the tool, noting errors that slipped through. A manager choosing where to use a tool should start with tasks that are frequent, easy to check, and cheap to get wrong. Hold back where mistakes are costly or hard to see.

Here is a fictional job, written for this exercise: the front desk assistant at the community hall that hosts the swap. The tasks are:

1. Answer phone questions about opening hours.
2. Write the weekly room booking summary from the booking sheet.
3. Decide whether a group can use the hall for free.
4. Unlock the building each morning.
5. Translate a parking notice into Spanish for regular visitors.

Mark each one: an assistant could draft it, a person must do it, or neither applies.

Answer: tasks 2 and 5 suit an assistant draft. The summary can be checked against the sheet, and a Spanish-speaking colleague or volunteer should review the translation before it goes up. Task 3 needs a person, because it applies the hall's own policy and someone has to answer for the decision. Task 4 needs neither. Task 1 is the interesting one. An assistant could draft a standard answer sheet, but a live caller with an unusual question still needs a person. If you marked it "person," that is a reasonable reading of the phone call. The useful split is between preparing the answers and giving them.

An index like the ILO's would call this job partly exposed. Your marks show what the score can't: which changes save time and which only move the checking elsewhere.

• The ILO's refined global index explains its task-scoring method, with results by income group and sex.
• Generative AI at Work, in the Quarterly Journal of Economics, is the customer support experiment.
• The Anthropic Economic Index publishes its reports and datasets on assistant use by task. It describes one vendor's users.
• Canaries in the Coal Mine?, from Stanford's Digital Economy Lab, tracks early-career employment in exposed occupations.
• The AI Index economy chapter, from Stanford HAI, gathers adoption, investment, and labor figures in one place.
• On this site, how AI is used across fields gives one documented use per field, with what gets checked first.