By Maria Demertzis

The familiar debate asks whether AI will destroy jobs or create new ones. While we await the outcome, the most alarming labour-market effect of artificial intelligence so far is that it is making too many young people unemployable. 

The latest edition of the Stanford Human-Centred Artificial Intelligence (HAI) index shows that AI is beginning to destroy entry-level jobs, those jobs that might be repetitive but through which novices become experienced. The report argues that AI’s labour-market effects are “uneven” and appear to destroy hiring pipelines for entry-level jobs in exposed occupations. By September 2025, employment for US software developers aged 22 to 25 had fallen close to 20 per cent from its 2022 peak, even as headcount for older groups continued to grow. Among workers aged 22 to 25, employment in the most AI-exposed occupations had fallen by roughly 16 per cent relative to the least-exposed occupations.   

A McKinsey survey shows that 30 per cent of organisations expect a decrease in workforce size due to AI in the coming year. This goes up to 35 per cent for companies with revenue above $1 billion. 

The HAI report concludes that this is not a general unemployment panic, as the evidence does not yet point to broad, uniform displacement. But can a labour market that is healthy in aggregate remain healthy while its apprenticeship system is quietly being hollowed out from within? 

The threat to entry-level jobs generates three problems. The first is the obvious one, and that is that young people become harder to hire.  

AI is proving very good at handling the repetitive parts of junior jobs. The HAI index report shows that productivity gains are the largest in structured, measurable work. For example, studies cited report gains of 14 to 15 per cent in customer support, 26 per cent in software development, and 50 per cent in marketing output. Given the pressure to cut costs, it is no surprise that the entry-level layer is the easiest place to harvest efficiency gains.  But creating efficiency gains is one thing; dealing with a cohort of displaced young people is another, with grave societal implications. 

Second, this also sows the seeds for another problem, which is slower to manifest itself, but potentially more dangerous. Juniors cannot become seniors if they are never allowed to be juniors. A working environment is where inexperienced labour converts into expertise. If AI now threatens to interrupt and possibly stop that conversion process, how will expertise be built?  

Take coding, for example. Ironically, just like Shelly’s Frankenstein, computer programmers appear to have created their own monster by killing the need for computer programming graduates, as AI can code faster and better. But if the trainee no longer debugs the simple module, she loses the friction that builds competence. Today’s firms may believe they are becoming more productive by increasing efficiency at the lower level. In reality, however, they also jeopardise their ability to hire in the long term by reducing the supply of senior labour. 

That leads to the third problem: AI does not abolish the need for seniors. It increases it. The HAI report talks about a “jagged frontier”, in other words, progress that is both uneven and unpredictable. Gemini Deep Think reached gold-medal performance at the International Mathematical Olympiad, yet the top model read analogue clocks correctly only 50.1 per cent of the time. AI agents in OSWorld, a comprehensive benchmarking and research environment designed to evaluate their ability to operate computers and navigate various software, improved to about 66 per cent task success, but still failed roughly one in three attempts.  

Like any technological progress, AI cannot be left alone. Someone will be required to check the code, defend any diagnosis, and most importantly, assume accountability for its consequences. Those who do that need to have long working experience. If AI replaces entry-level jobs, the labour market risks becoming a barbell with a missing middle. More demand for experienced reviewers, fewer entry-level roles, and a shrinking path from one to the other.  

It would, of course, be futile and economically damaging to try to protect every junior task from automation.  The answer is to treat entry-level jobs as infrastructure. Firms should measure not only output per employee, but expertise created per year.  

If AI is to remove drudgery, then it has promise. But when it conflates drudgery with on-the-job training, it will create shortages of senior talent down the line. There will always be a demand for talent, and the earlier companies can find it and train it, the more they will enjoy its benefits. Innovation in no way implies replacing expertise. 

Maria Demertzis is Professor of Economic Policy at the European University Institute, Florence. The article is reposted from the Blog of the Cyprus Economic Society https://cypruseconomicsociety.org/blog/blog-posts/