Blog

How Will AI Affect Jobs? What the Research Says, in Plain Language

JJesse Orrico, founder, Tarva, LLC·November 9, 2026

"Will AI take my job?" If you're 16, the question is really "will there be a job for me?" If you're a parent, it's "what do I tell my kid?" If you're 45 and your field is changing, it's both.

The honest answer is less exciting than the headlines. A lot of jobs will change. Some will shrink. Nobody knows the timing. This post walks through the research that gets quoted most, explains the one distinction that clears up most of the confusion, and shows what has actually happened so far.

Two words that get mixed up: exposure and replacement

Most of the scary numbers you see are about exposure. A job is "exposed" when AI could help with, or do, some share of its tasks. That's all it means.

Take a family doctor. AI can now draft visit notes. That's a real task, and it's exposed. But writing notes was never the job. The job is the patient in the room. In the map, our one-sentence note for family medicine physician puts it this way: "AI scribes write visit notes and flag lab trends; examining patients, judging messy cases and long family relationships stay human."

Replacement is a different question. It asks whether there will be fewer openings, or worse ones. A job can be highly exposed and still hire more people, because the tool makes each person more useful. A job can also be only partly exposed and still shrink, if the part AI takes is the part beginners used to do.

Keep those two words apart and most headlines get easier to read.

What the big studies actually say

These are the studies that get quoted most. I've tried to say only what each one found, and what it doesn't claim.

OpenAI and the University of Pennsylvania: "GPTs are GPTs" (2023)

This paper by Eloundou, Manning, Mishkin and Rock was one of the first big attempts to measure exposure across US jobs. It has more than one version, and the later one adds new numbers, so it matters which one you quote.

The first version (March 2023) estimated that about 80% of US workers could have at least 10% of their tasks affected, and about 19% could have at least half affected. The later version (August 2023) keeps those figures and adds another way to put it: about 15% of tasks could be done significantly faster at the same quality with a language model alone, rising to 47% to 56% once other software is built around the model.

That's a measure of tasks, not jobs lost.

Brookings (2024)

Kinder, de Souza Briggs, Muro and Liu at Brookings found that more than 30% of workers could see at least half of their tasks affected by generative AI, and about 85% could see at least 10% affected. They say directly that exposure isn't job loss. They also found exposure concentrated in cognitive, mid-to-higher-paid work.

International Labour Organization (2025)

The ILO, working with NASK, built a refined global index. About one in four jobs worldwide has some exposure to generative AI. About 3.3% of global employment falls in the highest-exposure group. Clerical work is the most exposed. Their summary is that generative AI looks likely to transform more jobs than it replaces.

World Economic Forum, Future of Jobs Report 2025

This one is a survey of more than 1,000 employers covering 14 million workers in 55 economies. Employers projected 170 million jobs created and 92 million displaced by 2030, for a net gain of 78 million, with about 22% of jobs affected.

One caution. Those numbers come from several big trends at once, not AI alone. It's a picture of churn, not an AI forecast.

Anthropic Economic Index (2025 and 2026)

This comes from a company that makes an AI assistant, and it looks at how people use that assistant. The first report (early 2025) found that about 36% of occupations used AI for at least a quarter of their tasks. About 57% of use looked like augmentation (a person working with the AI) and 43% like automation (the AI doing the task). Use was concentrated in mid-to-high wage jobs. The January 2026 report adds a new exposure measure weighted by how often the AI actually succeeds at a task.

The caution here is big. It measures one assistant's own users, not the job market.

What they have in common

Most of these studies measure what AI could touch, or what people use it for. That's useful. It still isn't a count of jobs lost. For that, you have to look at what has actually happened.

What has actually happened so far

Stanford: "Canaries in the Coal Mine?"

Brynjolfsson, Chandar and Chen at the Stanford Digital Economy Lab used payroll data from ADP to look at real employment. For workers aged 22 to 25 in the most AI-exposed occupations, employment now stands about 19% below where it would be if it had kept pace with less-exposed peers. That's the August 2026 version of the paper, with data through June 2026. Earlier versions reported about 13% (August 2025) and 16% (November 2025), using a different measure. The declines were concentrated in jobs where AI automates the work rather than helps people do it. Older workers in the same jobs were stable.

It's a working paper, and the authors don't claim whole occupations are disappearing. That's the idea in the title. A canary is an early signal, not a verdict.

Yale: no clear disruption yet

At the same time, Gimbel, Kinder, Kendall and Lee at the Budget Lab at Yale found no discernible disruption to the broader labor market in the 33 months after ChatGPT came out. They also found that exposure measures weren't yet related to changes in employment. They flag that the samples for recent graduates are small.

Why both can be true

The two findings measure different things. One looks at a narrow group, young workers in the most exposed jobs. The other looks at the whole market. A small group can feel a real change long before the big averages move.

That's the canary. It isn't the whole mine. It's the first to notice the air.

Robots run on a different clock than chatbots

Generative AI is software. It spreads fast and reaches desk work first: writing, summarizing, coding, routine analysis, customer chat. Robots and self-driving machines have to be built, paid for and trusted in the real world. They move slower and reach physical work later.

We built that into the map. Our ratings lean harder on robotics at 10 years than at 5, which is why some physical jobs, like long-haul truck driver, sit at low for 5 years and medium for 10.

The split shows up clearly in our estimates. Of the jobs we classed as physically heavy, about 80% are rated low risk at 10 years. Of desk jobs, about 5% are. Of jobs where fully remote work is common, about 1% are. Those are our judgments, not measurements. How we estimated AI risk explains the rubric behind them.

Why the beginner's seat matters most

Here is the part I think matters most, and it's the part the headlines skip.

When AI takes over the junior work in a job, companies can hire fewer juniors even while the senior roles hold up. We call that the entry-level squeeze. It's what the Stanford paper appears to be picking up: the drop was among 22-to-25-year-olds, and older workers in the same jobs were steady.

If you're 16, you're probably looking for your first full-time job somewhere around 2031 to 2036. That's the window our two ratings cover.

If you're changing careers, you're a beginner too, in a new field. I want to be careful here. The Stanford study measured age, not how new someone is to a field, so it doesn't tell us what happens to a 45-year-old starting over. But it's the reason we judged every job from the newcomer's side.

So the question I'd ask isn't "which job is untouchable?" It's "which jobs will still be hiring beginners, and what will beginners do there?"

What we did with all this in the map

We didn't compute our ratings from these studies. They mostly measure exposure. Our ratings ask a narrower question: how likely are AI and robots to shrink the number and quality of openings for someone entering this job, over the next 5 years and the next 10? Every rating comes with one sentence on what changes and what stays human.

Across all 10,000+ careers in the map, here's how our estimates move between the two horizons:

These are our estimates, not forecasts. They show how we applied our rubric, and they would tell you something about the future only if the rubric is right.

It also helps to remember the government's own projections still show growth. The Bureau of Labor Statistics projects about 5.2 million more jobs from 2024 to 2034, with healthcare and social assistance the largest and fastest-growing sector (about 8.4%). Wind turbine service technicians and solar photovoltaic installers are the fastest-growing occupations on its list.

For a practical take on using the ratings, see How to Use AI-Risk Ratings When Choosing a Career. If you've lost a job or think you might, this post is for you.

Questions people ask

What does "AI risk" mean on a job card?

It's our estimate of the chance that AI and robots shrink the number or quality of openings for someone starting that job, over the next 5 years and the next 10. Low means AI is mostly a tool in the job. Medium means fewer or different openings, especially at entry level. High means a big drop in openings looks likely.

Does "low" mean AI won't change a job?

No. Low means we judged that demand for the job holds steady or grows and that AI is mostly a tool within it. It's an estimate, not a promise. Jobs rated low can still change a lot, and local conditions matter. The quick filters say "AI-safe for 5 yrs" and "AI-safe for 10 yrs" as shorthand for "rated low in our estimates".

Why do some jobs go from low to medium at 10 years?

Mostly because robots and self-driving machines arrive slower than software. A job that's hands-on today can be low for 5 years and medium for 10 if machines look likely to take on more of the physical work later. The 10-year rating is never lower than the 5-year one.

Where do the AI-risk ratings come from?

From us. They aren't from the Bureau of Labor Statistics or any of the studies above. They were written with AI research assistants against a strict schema, checked by validation scripts and spot-checked by hand, by a very small team. The method post lays out the rubric.

Did you use AI to make this?

Yes, in places. The careers and the ratings were researched with AI assistants and checked as described above. I use AI tools when drafting posts too, and I read and edit everything before it goes out. If you spot something wrong, email privacy@tarvacore.ai.

Your next step: open the map and tap AI-safe for 10 yrs. Watch which sectors thin out. Then turn it off, open a job that dropped out, and tap "Why" under "Will it last?" to read what changes and what stays human.

Sources: Eloundou, Manning, Mishkin and Rock, "GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models", arXiv:2303.10130, March 2023 (v1) and August 2023 (v5). Kinder, de Souza Briggs, Muro and Liu, "Generative AI, the American worker, and the future of work", Brookings, October 10, 2024. International Labour Organization, "Generative AI and Jobs: A Refined Global Index of Occupational Exposure", Working Paper 140, May 2025. World Economic Forum, Future of Jobs Report 2025, January 8, 2025. Anthropic, The Anthropic Economic Index, 2025, and the January 2026 report. Brynjolfsson, Chandar and Chen, "Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence", Stanford Digital Economy Lab, 2025, revised August 2026. Gimbel, Kinder, Kendall and Lee, "Evaluating the Impact of AI on the Labor Market: Current State of Affairs", The Budget Lab at Yale, October 2025. US Bureau of Labor Statistics, Employment Projections 2024–34, August 28, 2025. Counts and shares are from our own estimates in the map. Parts of the research behind Career Universe were written with AI research assistants and checked by hand.