Is AI taking jobs? What the data and history show

A layoff is easy to see. Fewer new hires are easier to miss.

A company can keep its existing team and still offer fewer opportunities to the next person trying to join. That is one reason the question “Is AI taking jobs?” needs more than a count of layoffs.

The clearest warning so far concerns younger workers and weaker hiring. AI use is spreading quickly, but the employment data cannot tell us how many jobs it caused to disappear.

I wanted to understand this by looking at earlier technologies alongside today's numbers. Did faster work eventually create more work? Who benefited, and who struggled?

For anyone running a team, those questions lead to a decision we can actually make: what should happen to the time AI saves?

How fast is AI adoption compared with PCs?

To compare technologies, user counts are a poor starting point. Populations change. What helps is the share of people using each technology, measured from a clear starting date.

Around 21 months after ChatGPT launched, 44.6% of US adults ages 18-64 reported using generative AI. PC use at work or home was 19.7% roughly three years after the IBM PC launched. The AI figure includes a later survey correction. St. Louis Fed adoption research.

AI reached people quickly. It could arrive through a browser on a computer they already owned, rather than through another hardware purchase.

But opening a tool and replacing someone's work are very different steps. To understand the jobs question, we need to see how much work people use it for.

Using AI is not the same as replacing a job

Someone using AI for one email counts as a user. So does someone using it throughout the day. The adoption rate treats them alike; their working days look very different.

In Q2 2026, 45.2% of employed US adults ages 18-64 reported using AI for work. Across all work hours, including non-users, the survey estimated 6.3% were AI-assisted and workers reported time savings equivalent to 2.2%. RPS survey data via FRED.

So widespread use can coexist with a much smaller change in total working time. Even the hours saved might become more output, extra checking or a shorter backlog.

That still leaves an important question: are the people doing work AI could handle finding fewer jobs?

Are entry-level jobs being hit harder?

Stanford and ADP publish employment data by age and by how much of an occupation's work AI could help perform. That is what AI exposure means here. It describes the tasks, not whether an employer actually uses AI.

To compare the groups, set each one's employment to 100 in November 2022. By August 2026, among ages 22-25, the most exposed group falls to 87.23 and the least exposed rises to 108.10. Employment fell 12.8% in one and grew 8.1% in the other. Stanford and ADP public data.

Employment in the most exposed group is 19.3% lower than if it had kept pace with the comparison group. Here is the calculation:

That 19.3% is a comparison between groups. It does not mean AI eliminated 19.3% of these jobs.

The relative gap is 8.6% for ages 26-30 and much smaller for older workers. Age is only a rough guide to career stage, but the difference is hard to ignore.

It also fits a concern that layoffs alone would miss: companies may need fewer new recruits even when they keep experienced staff. To see whether that is happening, we need evidence beyond this one payroll sample.

Is AI causing the change in hiring?

A US Census study also finds weaker hiring among ages 22-24 in highly exposed industries. They estimate employment for this age group fell about 12% from late 2022 through Q2 2025, after allowing for other factors. It uses different records from ADP, so the hiring concern shows up in another dataset.

A Danish firm study looks at actual AI adopters. By late 2025, their employment is about 11% below their earlier growth path relative to non-adopters. Smaller firms and reduced recruitment drive the difference. Those firms could still be growing, just hiring less than their earlier trend suggested.

That makes weaker hiring a concern worth watching. It does not settle the cause: a company choosing AI may also be changing its hiring plans for other reasons.

Even the starting date changes what the numbers tell us. In our youngest-worker comparison, moving it from November 2022 to November 2023 reduces the gap from 19.3% to 11.7%.

AI arrived during a changing job market. Research using career histories and unemployment claims finds problems in exposed occupations before ChatGPT. New York Fed researchers also link remote work to difficulties training and hiring less experienced workers.

And widespread use does not always produce a measurable employment change straight away. An earlier Danish chatbot study, following workers through December 2024, finds no detectable average change in earnings or recorded hours, ruling out effects larger than roughly 2% in its setting.

There is a hiring problem to take seriously. Blaming the whole gap on AI would hide the other things we may need to fix.

To understand why faster work and worse job prospects can coexist, it helps to look further back.

What past automation tells us about AI and jobs

Earlier technologies changed both the amount of work and who got to do it. The results were often good for the wider economy and painful for particular workers.

Mechanised textile production expanded output while some workers faced years of weak wages and less control over their work. Computers changed demand for routine clerical tasks. Industrial robots reduced employment and wages in more exposed US areas.

New work appearing elsewhere did not make the transition easy for someone whose skills or location no longer matched it.

That is why the history matters here. A productivity gain can benefit the business while the worker loses an opportunity. Looking only at total employment, or only at the company's output, misses part of the story.

History also leaves us with a useful question for today: if work gets cheaper and faster, will there be more of it?

Does higher AI productivity mean fewer jobs?

In a study of 5,172 customer-support agents, AI assistance increased issues resolved per hour by about 15%. Less experienced workers benefited more. Generative AI at Work.

So AI can help a newer worker do the job better. Whether the company needs more workers is a separate decision about workload.

Here is a fictional example using 15% as the starting assumption. A team needs 100 hours to deliver 100 units of work. After the improvement, it can deliver 1.15 units per hour.

At the same demand, the work needs 87.0 hours. If demand rises to 115 units, it still needs 100 hours. At 130 units, it needs 113.0 hours.

The same productivity gain produces three different outcomes.

This is the decision behind the AI jobs question inside a company. The saved time could mean fewer hires. It could also mean serving more customers, clearing overdue work or reducing overtime.

The demonstration tells you the work can be faster. The workload and the company's choices determine what happens next.

When does time saved become a real business benefit? →Follow the saved hours into capacity, output and spending before calling them cash savings.

What should leaders check before cutting headcount?

If you are introducing AI into a team, follow the result beyond the number of people using it.

Question	What to measure
Did we finish more useful work?	Completed work, errors, rework and review effort
Did it take less total effort?	Paid hours, including checking and support
What did we do with the time?	More output, shorter queues, less overtime or work moved elsewhere
Who gained or lost an opportunity?	Hiring, departures and moves into other roles
How will new people learn?	Supervised tasks and opportunities to build experience

The last question belongs in the plan. If AI takes over the tasks juniors used to learn on, someone needs to decide how they will build that experience instead.

AI may change a team through fewer hires long before it produces a visible round of layoffs. And faster work may increase demand for people rather than reduce it.

Before celebrating the time saved, decide what it is for.

Sources and calculation notes

The charts use data saved on 4 October 2026. Employment data run through August 2026; the work-use survey runs through Q2 2026. Open “Source and data” beside a chart for the underlying values and methods.

• St. Louis Fed adoption comparison and survey correction. Work-use series via FRED: people, last-week use, hours assisted, reported time saved.
• Stanford and ADP: public employment data and August 2026 research paper. Our chart compares the highest and lowest exposure fifths, unlike the paper's headline grouping. The sample follows firms over time, not individual workers' careers. Education controls substantially reduce estimates in the paper; the chart's calculated ratios have no statistical uncertainty estimates.
• Census early-career hiring study, Danish AI-adopting firms and Danish chatbot study. These studies examine different people, measures and periods; several are working papers.
• Competing explanations: changes before ChatGPT, remote work and younger workers and Federal Reserve research on coder employment.
• Historical sources are linked inside each case. The customer-support productivity result comes from Generative AI at Work, published in the Quarterly Journal of Economics in 2025.