Two voices worth hearing
Cognitive scientist Gary Marcus, professor emeritus at New York University, calls it the succession problem. In an interview with Gustavo Entrala in November 2026, he put it like this:
“I worry less about whether young people will find jobs than about the succession problem. We hire fewer junior programmers, but we still need seniors to fix what these systems break. And we are not training the seniors of the future.”
Engineer and science communicator Javier Ideami, on the “Tengo un Plan” podcast, points to the same gap from another angle. A junior’s typical tasks are repetitive pattern-matching, exactly what AI does best. His advice to young people is to train what machines do not yet master:
“Exercise slow thinking and depth more and more. That is what will save you. Today’s juniors are tomorrow’s executives: if we cut them, where will they be trained?”
Quotes adapted and translated from the transcripts of both interviews.
What the data says
In August 2025, a Stanford team led by Erik Brynjolfsson analysed payroll data from the largest US provider. Since generative AI became widespread, employment of workers aged 22 to 25 in the most AI-exposed occupations fell by 13% in relative terms. In those same occupations, employment of more experienced workers held steady or grew. The decline is concentrated where AI replaces tasks rather than complementing them.
Venture firm SignalFire reaches a similar conclusion in tech: in 2024, new graduates made up only 7% of hires at the 15 largest tech companies, 25% fewer than in 2023 and less than half the 2019 level.
At IESE, economist Enrique Ide names the mechanism. Juniors learn by working alongside experts, and that practical knowledge is in no manual. If automation takes that work away, they learn less, become weaker mentors, and the passing of good practice between generations slows down. In his analysis, what we see so far with generative AI looks more like displacing juniors than making experts more productive.
These are US and tech-sector figures. We have not yet found comparable published series for Spain, but the trend deserves attention.
Learning to think, not just to delegate
The other side of the problem is how AI is used. A study by Michael Gerlich with 666 people, published in 2025 in the journal Societies, found that those who delegate most to AI score lower on critical thinking. Those aged 17 to 25 delegated the most. It is a correlation and does not prove cause, but it points the same way.
A study by Microsoft Research and Carnegie Mellon with 319 professionals, presented at CHI 2025, adds nuance: the more people trust AI, the less they check its output; the more they trust their own judgement, the more they think.
AI does not replace judgement: it demands it. And judgement is built through practice, mistakes and someone who corrects you.
What we do at Caelum
That is why we work with universities, polytechnics and vocational schools: internships, final-year projects and AI challenges on real projects, Preluma and CBLX. Students work with AI, but they also learn to check what it produces. In Preluma, for example, every data point is checked against its official source before it reaches a client, and newcomers learn to do that review with someone experienced beside them.
Giving juniors opportunities is not just social responsibility. It is making sure that, ten years from now, there are professionals able to direct AI.
Sources
- Gary Marcus, interview on Gustavo Entrala’s podcast (YouTube), November 2026
- Javier Ideami, interview on the “Tengo un Plan” podcast (YouTube), 2026
- Brynjolfsson, Chandar and Chen, “Canaries in the Coal Mine?”, Stanford (SIEPR), August 2025
- SignalFire, “State of Tech Talent Report 2025”, 20 May 2025
- Enrique Ide (IESE), “Automation, AI, and the Intergenerational Transmission of Knowledge”, arXiv, 2026
- Michael Gerlich, “AI Tools in Society”, Societies, January 2025
- Lee et al. (Microsoft Research and Carnegie Mellon), CHI 2025