The empty lecture halls

On 17 September the University of Antwerp published its first enrolment figures for the year that opened four days later. Total enrolment rose 5.2 per cent. First-year computer science fell 29 per cent, after a fall of almost 35 per cent the year before. Generation students, the first-time entrants, fell from 3,433 to 3,208, and the university says this is not a smaller birth cohort. Pharmaceutical sciences rose 13 per cent.

The spokesman, Peter De Meyer, told VRT: "We hebben nood aan gediplomeerde informatici, maar die boodschap komt blijkbaar niet over bij de huidige generatie studenten." Students, he suspects, think "dat AI alles vanzelf zal doen en dat het niet meer nodig is om te kunnen programmeren". Three days earlier the city's college, AP Hogeschool, had reported 10 per cent fewer new students in its applied computer science bachelor and 40 per cent fewer in its programming graduaat, and named the same suspect.

On 25 September Wim Casteels, who coordinates IT and AI at that college, gave the pattern a headline in Trends: AI does not threaten the job, it threatens the first job. His column is not the event scored here; the figures are. The fear, Casteels writes, is partly justified: the first job will be harder to find.

The claim on the table

Chapters 1 and 3 argue that automation decouples output from employment and that entry-level cognitive work is the current front line. Chapter 3 rests on two studies. In the call-centre experiment of Brynjolfsson, Li and Raymond, agents with two months' tenure and the AI assistant matched agents with more than six months without it: "six months of accumulated tacit knowledge, the kind that justified the seniority differential, were now reproducible by a model trained on chat logs." Then Stanford's "Canaries in the Coal Mine", by Brynjolfsson, Chandar and Chen on ADP payroll data: workers aged 22 to 25 in the most AI-exposed occupations saw a 16 per cent relative employment decline between 2022 and mid-2025, while the same exposed occupations held steady for workers over 30. The chapter joins the two: AI raises the value of any individual junior, the firm responds by hiring fewer juniors at lower starting wages and the surplus flows to the firm.

Two European bodies have since found the same. In January, three Rabobank economists reported in ESB that Dutch employment of 15 to 24 year olds in the twelve most GenAI-exposed occupational groups fell 13 per cent between late 2022 and the third quarter of 2025, against a 3 per cent rise elsewhere: about 19,000 fewer young workers. In February, Belgium's High Council for Employment reported that opportunities for younger workers in highly exposed occupations appear to be declining, "waardoor het moeilijker wordt om de eerste stappen op de carrièreladder te zetten". The Council has no Belgian payroll data behind that sentence. Page 44 rests on the same Stanford paper the chapter uses and on one more American study, Lichtinger and Hosseini, who find that firms using generative AI hire fewer young workers and as many experienced ones as before. The Council keeps a conditional the headlines dropped: if these findings hold, AI could make the first steps harder. Its account of why is the chapter's: the young bring codified knowledge and do the routine tasks AI automates first; the experienced hold tacit knowledge specific to the job.

September added a step the book did not place. Chapter 3 put the squeeze at the door of the firm. In Antwerp it stands at the door of the lecture hall. The eighteen-year-olds have read the signal and are not applying to become the juniors nobody is hiring.

The honest complication

Enrolment is not employment and the figures are provisional. Two institutions are named with numbers; De Tijd's 10 to 60 per cent range across Flemish colleges sits behind its paywall and this ledger has read only its headline. Generation students fell across the board while total enrolment rose, which argues for causes wider than fear of a language model, and the university itself lists some: the pull of the colleges' job-specific programmes, Antwerp's diverse student population and its drop-out rates. The fear of AI is what a spokesman and a coordinator suspect, not what any survey of students measured. And the Council finds no clear aggregate employment effect in Belgium yet, with 35 per cent of firms using AI in 2025. If enrolment recovers next September, or if Belgian junior hiring in exposed occupations rises with adoption, this entry gets rescored.

Verdict

Complicates. The register's clause is entry-level contraction in AI-exposed occupations, and that is employment. The book predicted the contraction from US payroll data, the Dutch data show it next door and the Council expects it in Belgium. The enrolment figures do not measure it; they measure the expectation of it, arriving one step earlier than the book placed the squeeze, at the door of the lecture hall. Chapter 3 calls it the pipeline that turns juniors into seniors, and what Antwerp shows is that pipeline thinning at its source, before the firm has hired anyone. The ledger's rule for a shape the criteria do not price is to say so rather than stretch the clause, so this scores as a complication: the chapter's mechanism is running ahead of the chapter, and the book has a step to add.

Casteels closes with the remedy that "bedrijven moeten starters een kans geven", because "wie vandaag geen starters aanneemt, heeft morgen geen seniors". The mechanism the book traces says why that appeal is weaker than it sounds: the firm that trains a junior pays for experience the model already serves to everyone for nothing. The column reaches the coming shortage of seniors. It does not reach the question of who owns the machine that made the juniors' experience surplus to requirements.

Who owns the machine?