AI's first measurable effect on work is on hiring, not layoffs. In the US, 22 to 25 year olds in highly AI-exposed jobs were about 19% below trend in employment by June 2026; experienced workers in the same jobs showed no such gap. Official forecasts still add jobs in the US and EU to 2035. The squeeze is at the way in.
Every figure below is a published one, linked to its source: payroll data, job postings, official forecasts and Eurostat. Together they point to one pattern that I think deserves more attention than the layoff headlines.
Is AI taking entry-level jobs?
Not in the way the headlines suggest, but the entry level is where it shows up first. The clearest measure comes from payroll records. Researchers at the Stanford Digital Economy Lab compare young workers in the occupations most exposed to AI with young workers in less exposed ones. By June 2026, employment of 22 to 25 year olds in the highly exposed group was about 19% below where it would have been had it kept pace, and experienced workers in the same occupations showed no comparable gap (Stanford Digital Economy Lab). A reading nine months earlier, measured a different way, had already found a 16% relative decline for the youngest workers (Brynjolfsson, Chandar and Chen).
Job postings show the same thing from the employer's side. On Indeed, US entry-level postings in May 2026 were down 7.5% on a year earlier, while senior postings were up 14.7%. Until early 2026 the two had moved together (Indeed Hiring Lab).
Two cautions before going further. These are American figures, and the comparable European evidence is thinner. And exposure is not destruction. The International Labour Organization estimates that one job in four worldwide has some exposure to generative AI, one in three in high-income countries, with clerical work the most exposed (ILO). Its index measures how much of a job's tasks the tools overlap with. It does not count the jobs that will go.
Why would AI hit beginners before experienced staff?
Every job, at its core, is the same loop. You set an objective, you break it into smaller objectives, and you judge whether the result is good enough. I have argued before that AI is very good at the middle of that loop and much weaker at its edges.
Now think about how a first job is built. The graduate gets the middle: the first draft, the reconciliation, the test cases, the summary of forty documents. The senior keeps the edges: what the client actually needs, and whether the draft is good enough to send. For decades this was a fair trade. The junior did the execution and, by doing it, slowly learned the judgement.
You probably see where this goes. AI takes the middle first. The experienced worker becomes faster, and the case for hiring someone new to do the middle gets weaker. Software shows both sides at once. Employment of US developers aged 22 to 25 fell nearly 20% from its late-2022 peak (Stanford Digital Economy Lab), while the Bureau of Labor Statistics still projects developer jobs to grow 10% from 2025 to 2035 (BLS). Both can be true. The profession grows. The way into it narrows.
There is a second reason, less discussed and harder to measure. A company that is unsure what AI will do to its work tends to pause hiring long before it lets anyone go. A pause costs nothing visible this quarter. The people it costs are the ones who were never hired, and they do not show up in any layoff count.
Will AI mean fewer jobs in Europe by 2035?
The official forecasts say no. Cedefop, the EU agency that forecasts skills and jobs, projects EU-27 employment to grow by over 4% between 2023 and 2035 (Cedefop). In the US, the BLS projects 5.9 million more jobs in 2035 than in 2025, a rise of 3.5% (BLS). Neither is a promise, and neither was built for a scenario in which AI agents take over whole pieces of work quickly. They are still the best baselines we have, and they point up.
What would change the picture is agents doing complete projects rather than tasks. The Remote Labor Index tests AI agents on real, paid freelance projects. At its launch in October 2025 the best agent completed 2.5% of them (Scale AI); by July 2026 the best result was 15.8% (Center for AI Safety). That is still a small minority of the work, and a benchmark is not a labour market. It is the number I would watch. I sketched where that could lead in the future of jobs.
Not every lost job is AI's, either. German car-industry employment fell 6.3% in the year to autumn 2025 (Destatis), and I have found no published source that names AI as the cause.
What does this look like in Romania?
Here the story turns, and it is worth being honest about how. Romania has the EU's highest share of young people neither in work nor in education or training: 19.2% of 15 to 29 year olds in 2025, against 11.0% across the EU and 5.3% in the Netherlands, the lowest (Eurostat). Inside the country the share is 7.2% in cities and 27.7% in rural areas, almost four times higher (Eurostat).
Romania is also the EU country where firms use AI least: 5.21% of enterprises with ten or more employees in 2025, against 19.95% across the union (Eurostat). Put those two facts side by side and the conclusion is uncomfortable. Whatever keeps young Romanians out of work, it is not AI. The door was narrow before the tools arrived, for reasons that have more to do with geography, schooling and where the jobs are. AI adds a second pressure on top, on the office work that has been one of the ways in for graduates. For the IT sector specifically, see Romania as an IT delivery location in 2026.
What can a young person, or an employer, do about it?
For someone starting out, the evidence that holds up best is unglamorous. Training aimed at one sector with real local demand pays: in randomised trials of several US sectoral programmes, earnings rose 11 to 40% after training ended (NBER). And the skill worth building is the edge of the loop, not the middle. Framing a problem, checking an answer, noticing when a confident output is wrong. The tools make the middle cheap. They make judgement worth more.
For an employer, the question is less comfortable. If the middle of the work goes to AI and juniors stop being hired to do it, where do your seniors come from in 2035? One answer is to hand juniors the edges earlier: review, contact with the client, deciding what the agent should do and checking what it did, with the AI doing the drafting. It is slower at first, and it asks seniors to explain decisions they used to make silently. It may also be the only version of the apprenticeship that survives the tools. In my view it is the part of AI adoption that deserves the most time in team training, because the tool is the easy half.
We spent a century building careers on the idea that you learn to judge by doing the work first. The tools now do the work. Nobody has yet said who teaches the judgement.
If you are rethinking how your team learns to work with AI: vladtudor.com/workshops.