Which Jobs Is AI Taking First? Not the Obvious Ones

The degree arrived six months ago. Two hundred applications went out. Four replies came back, three of them rejections.

Stories like this fill career forums right now, and they all sound the same: someone does everything right, and the door still does not open. The 2026 data explains why. The explanation is not the one most people expected.

AI has fired almost nobody so far. It stopped hiring. Among workers aged 22-25 in jobs where AI does the task instead of a person, employment is about 19% lower than it would otherwise be. That number comes from the Stanford Digital Economy Lab, which analysed payroll records covering millions of US workers through June 2026. Experienced workers in the very same jobs show no such gap.

Which means the familiar lists of "10 jobs AI will kill" miss the point. The hit did not land on whole professions. It landed on the entrance to them.

AI is not firing people, it just stopped opening doors

The biggest shift in the labour market runs through hiring that never happened, not through layoffs. Researchers Erik Brynjolfsson, Bharat Chandar and Ruyu Chen tracked anonymised payroll data from November 2022, when ChatGPT launched, through June 2026.

What they found:

• no widespread job displacement across the economy, and overall employment holds up;
• workers aged 22-25 in the most AI-exposed jobs sit about 19% below the expected level, and the gap is widening (it was 15% a year earlier);
• older colleagues in those same jobs show no gap at all;
• the decline comes mainly from reduced hiring, not from more firing;
• companies adjust through headcount rather than pay, so base salaries stay roughly intact.

That fourth point deserves a second read. Firms are not marching people out of the building. They simply stop posting the roles they used to hand to recent graduates. From the company side it looks like saving money. From the applicant side it looks like silence after two hundred applications.

The authors call their own numbers "canaries in the coal mine", an early signal rather than proven cause. Their caveats matter more than they first appear, and they get their own section below. First, the mechanism: why newcomers took the hit before anyone else.

Why beginners got hit first

Beginners went first because generative AI (software like ChatGPT that produces text, code or images on request) is best at exactly the work people used to start their careers with.

Think about what the first rung looked like in almost any field:

• a junior lawyer read through standard contracts and pulled out the relevant clauses;
• a trainee accountant moved receipts into spreadsheets;
• a junior developer wrote simple code and fixed small bugs;
• a marketing intern produced product descriptions and routine copy;
• a support agent answered the same customer questions over and over.

Every item shares one trait: the task is clear, repeatable and easy to check. That is precisely what AI now does in seconds for almost nothing.

Losing that work costs a beginner more than a salary. Those tasks were the training ladder. Six months of reading contracts turned into pattern recognition, and a year later that person handled harder cases. Remove the bottom rungs and the ladder becomes a wall: the top still needs experienced people, and there is no longer a way up to it.

That is why the gap shows up at 22-25 and disappears past thirty. Experience banked before 2023 is suddenly worth more, because it is far harder to accumulate now.

AI-touched jobs do not all behave the same way, though. The difference between them decides whose work disappears and whose gets more valuable.

Replacement or amplification: the distinction that matters

Everything depends on whether AI does the job instead of the person or makes the person better at it. In the Stanford data these two paths split in opposite directions: where AI substitutes for human tasks, employment falls; where AI works as a tool in an expert's hands, employment stays flat or grows, and that holds for every age group.

Two situations make the difference concrete:

• Replacement. A customer messages support, a bot answers. Nobody else is needed in that chain.
• Amplification. A radiologist reads a scan alongside an algorithm that flags suspicious areas. The doctor still decides, but gets through more cases per shift.

An everyday parallel: the calculator did not put accountants out of work, it put the abacus out of work. The profession survived; one manual method died. The automated switchboard, by contrast, really did shrink the number of telephone operators: there the machine replaced the service itself, not a tool inside it.

The practical takeaway for anyone choosing a field or a first job: the question is not whether the work "involves AI", but who makes the call. If judgement stays with the person, AI makes that person more valuable. If the judgement can be written down as an instruction, sooner or later it will be.

Which jobs are actually exposed

Office work with documents and data carries the highest exposure, not the futuristic professions from the headlines. The International Labour Organization (ILO), the UN agency for work, measured this worldwide with Poland's NASK research institute and published the results in May 2025.

The headline numbers:

• 25% of global employment sits in jobs where generative AI could handle a meaningful share of tasks;
• 34% in high-income countries, where office work makes up more of the economy;
• clerical staff rank first: paperwork, data entry, processing applications;
• next come certain digitised knowledge jobs in media, software and finance.

The ILO repeats one caveat throughout: exposure is not job loss. Fully automating an entire occupation stays rare, because nearly every job keeps tasks that need a human. What usually changes is the content of the role, not its existence.

The data also holds an imbalance that gets little airtime. In high-income countries, 9.6% of jobs held by women fall into the highest-risk category, against 3.5% of jobs held by men. The reason is not ability but structure: women have long been concentrated in clerical and administrative roles, exactly where AI takes the most tasks.

So far this reads like a verdict on one generation and one type of work. Country-level numbers show the shift moving at very different speeds, which is the first clue that technology is not the only thing going on.

How it looks around the world

The entrance narrowed almost everywhere, but for different reasons and by different amounts. Here is what official statistics and research bodies report by region.

• Euro area. Youth unemployment (under 25) stood at 15.1% in the first quarter of 2026, up 0.6 percentage points from 2023. Young workers in IT and communications fell hardest: down 18.6% over three years (European Central Bank).
• European Union. Youth unemployment hit 15.5% in June 2026, with a huge spread: Germany 7.1%, Spain 23.4%, Sweden 25.9% (Eurostat).
• United States. Unemployment among degree holders aged 22-27 ran near 5.6% in 2026, against roughly 4.3% nationally. The sharper number: about 42% of recent graduates work in jobs that never required a degree (Federal Reserve Bank of New York).
• India. Tech sector headcount grew just 2.3% in fiscal 2026, according to the industry body NASSCOM. The era of mass graduate intake (30,000 hires a year at a single IT services firm) is over, and entry-level IT roles have shrunk by an estimated 20-25%, per consultancy EY.
• Singapore. The Ministry of Manpower told parliament in February 2026 that graduate employment rates have stayed broadly stable for a decade and that AI's specific effect on entry-level professional jobs remains uncertain. The government still launched a paid traineeship scheme for graduates, in case the signal turns out to be real.
• Russia and the CIS. Competition for starting roles is the fiercest on the market: roughly 19 applications per junior vacancy versus 2.5 per senior one (hh.ru). After the talent shortage of 2022-2024, employers turned picky: far less willing to hire someone to grow into a role, far more focused on people who deliver immediately.

Put the rows together and one thing stands out: the gap between "experience required" and "nowhere to get it" widened worldwide at the same time. Which raises the obvious question. Is AI really the culprit?

Or maybe this is not about AI at all

Nobody has proven AI did this, and the serious researchers say so plainly. That does not make the problem go away, but it does change what you should conclude from it.

Start with the people who produced the numbers.

Stanford calls its findings descriptive indicators rather than causal estimates. The gap shrinks noticeably once education levels are accounted for. Part of the divergence started before generative AI spread. And the effect is stronger in their payroll sample than in national statistics.

The European Central Bank goes further: the available evidence does not allow the weakness in youth employment to be pinned on generative AI. The ECB's main explanation is a cooling economy. Young workers always take the first hit in any slowdown: shorter tenure, more temporary contracts, and it is far easier to skip a hire than to dismiss someone with ten years of service.

Two more factors sit on top. High interest rates: expensive money pushes companies to cut spending on growth, and hiring beginners is an investment in future output rather than this quarter's revenue. And the hangover from the 2021-2022 over-hiring spree, when tech firms took on staff far beyond need and spent years digesting it. Economic pressure on young adults shows up outside the job market too. this piece on housing costs and delayed marriage traces how it reshapes personal decisions.

So what should you do with that? Treat AI as one cause among several, but do not wave it away. Economies recover, and some vacancies come back. The open question is whether the specific tasks already handed to machines come back with them. The money trail suggests an answer.

What is happening to people already inside

Workers with AI skills are seeing pay rise faster than the market, and the spread keeps growing. Consultancy PwC analysed more than a billion job ads across 27 countries and published its findings in June 2026.

The key results:

• the wage premium for AI skills reached 62%, up from 57% a year earlier;
• the range across sectors is enormous: from 16% in government to 118% in consumer markets;
• roles where AI amplifies expert judgement grow twice as fast in headcount and 42% faster in pay than roles where AI simplifies the task down to "anyone can do this".

That last line describes the mechanism. AI acts as a difference amplifier. If your value comes from doing something an ordinary person cannot, AI multiplies it. If your value comes from performing a routine step more carefully than others, AI erases the advantage, because now everyone performs it carefully.

Which leaves one question: if the bottom rung is gone, is there still a way in?

The entrance did not close. It moved up

Entry-level roles still exist; what changed is what they contain. The same PwC study carries a figure that rarely makes the news: AI-exposed entry-level roles grew 35% since 2019, while other entry-level roles declined 10%.

The catch is real. Those roles are seven times more likely to demand skills once reserved for senior staff: independent judgement, ownership of outcomes, the ability to explain things to people.

In plain terms: a beginner used to be judged on completing an assigned task neatly. The machine now completes the task, so the beginner is judged on noticing whether it was the right task at all, and on catching the moment the machine produced confident nonsense. That is a different job, and it needs different preparation.

Which creates a paradox: competing with AI at the entrance means using AI better than everyone who just asks it to write their cover letter. This guide to ChatGPT, Claude and Gemini covers how these tools work and where they lie.

Time to turn all of that into moves.

Five moves that work in this market

One thing works: proof of a result that cannot be generated in sixty seconds. The steps below run from fastest payoff to slowest.

1. Stop applying in bulk. Starting roles draw hundreds of applications (around 19 per junior opening versus 2.5 per senior one in one large market). Two hundred identical applications lose to fifteen targeted ones that show you read the posting and understood the company's problem.
2. Build proof of work, not a list of skills. "Knows Python" means nothing; anyone can type it. What lands is a thing someone can open and inspect: a finished analysis, a working bot, a documented case, a piece with real numbers in it. One of those outweighs ten completed courses.
3. Learn AI as a work tool, not a toy. The pay premium for those skills hit 62%, so this is present tense, not future. What gets rewarded is not writing prompts but knowing where the tool fails and how to verify it.
4. Grow what machines do not take. The World Economic Forum surveyed over 1,000 major employers across 55 economies: demand rises fastest for AI and data skills, followed immediately by creative thinking, flexibility, resilience and a habit of continuous learning. One planning number from the same report: about 39% of an average worker's core skills will change by 2030.
5. Go where AI amplifies instead of replacing. Jobs with physical presence, real responsibility and human contact keep growing. Healthcare, engineering, energy, care work, repair, teaching. In those fields the algorithm speeds the person up rather than doing the work instead.

None of these guarantees an offer tomorrow. Each one moves you out of "easy to replace" and into "more expensive to replace than to hire".

What comes next

Total employment worldwide is projected to grow, not shrink. The World Economic Forum estimates that by 2030 roughly 92 million jobs will disappear while about 170 million new ones appear: a net gain near 78 million, close to 7% of the employment covered in the study.

Averages are cold comfort for one person, though. Jobs vanish in certain countries and sectors and appear in others. Between those two events lies a transition, and the transition hurts whoever happens to be entering the market right now.

Worth remembering, too, that a figure like a 19% gap is a snapshot, not a sentence. It opened over three years and could change over the next three: if the economy warms up, part of that hiring returns. What does not return is the old bargain, where a beginner got a seat in exchange for being willing to do simple work and learn along the way.

What to do today

Back to those two hundred unanswered applications. The problem is rarely the number and almost always that the CV reads like every other one: a list of duties instead of results.

A five-minute move, right now: open your CV, find one line like "responsible for reporting" or "participated in development", and rewrite it as a result with a number in it. "Rebuilt the monthly report covering 400 line items and cut preparation from two days to four hours." That single line does more than another hundred applications, because it is the one thing nobody else can generate for you.

Sources

• Stanford Digital Economy Lab. Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence (2026)
• Stanford Digital Economy Lab. Data update: the AI employment gap for young workers widened to 19% (August 2026)
• International Labour Organization and NASK. Global index of occupational exposure to generative AI (2025)
• Federal Reserve Bank of New York. The Labor Market for Recent College Graduates (2026)
• European Central Bank. Youth employment amidst cooling labour demand (2026)
• Eurostat. EU and euro area unemployment indicators, June 2026
• PwC. Global AI Jobs Barometer 2026: analysis of over 1 billion job ads across 27 countries
• World Economic Forum. Future of Jobs Report 2025
• Ministry of Manpower, Singapore. Parliamentary reply on AI and fresh graduate hiring (February 2026)
• hh.ru. Hiring trends 2026: market polarisation and competition for entry-level roles