The Apprenticeship Is Part of the Job
AI may not abolish whole professions. It can still remove the junior work through which people learn to enter them.
The easiest argument about artificial intelligence and employment is also the least useful: either the machines will take every job or technology will create better ones as it always has.
Both claims skip the institution that connects a novice to a profession. A junior analyst learns by checking tables. A young lawyer learns by reading files and drafting imperfect notes. A programmer learns by fixing small errors inside a larger system. These tasks are repetitive, which makes them attractive targets for automation. They are also instruction disguised as production.
If firms automate the task but preserve the occupation, they may still damage the route into it. The employee survives today while the apprenticeship disappears for tomorrow.
Exposure is not disappearance
The evidence does not support a declaration of mass AI unemployment. The International Labour Organization and World Bank distinguish between exposure to generative AI and actual job loss. Many exposed occupations contain tasks that can be assisted without being eliminated. Adoption also varies by infrastructure, wages and management.
Recent firm-level evidence remains early. Anthropic's labour-market analysis found no clear rise in unemployment among highly exposed workers. It did find tentative evidence that job-finding among 22- to 25-year-olds entering exposed occupations had fallen relative to 2022, by about 14 percent in its preferred comparison. The result was barely statistically significant and sensitive to alternative choices.
That is not proof of a general collapse. It is a signal aimed at the exact point where institutions should look: entry.
Headline employment data can remain calm while recruitment narrows, graduate roles shrink and remaining juniors are expected to perform at an experienced level with machine support. A labour market can lose its learning machinery before it loses its payroll.
A private incentive creates a public risk
Each employer has a rational reason to automate junior output. A model can summarise, classify, draft and test at low marginal cost. A smaller team of experienced employees may produce more, faster.
The collective result can be irrational. Senior workers retire. Technology changes. Organisations need people who understand why a process works, not merely how to approve its output. If nobody paid to train beginners, the market eventually discovers that expertise was a shared asset everyone expected somebody else to finance.
This is familiar. Many economies already underinvest in apprenticeships because trained workers can leave. AI sharpens the problem by making the training tasks themselves look disposable.
The answer is not to preserve useless paperwork. It is to preserve deliberate practice while removing drudgery.
Redesign entry rather than subsidise nostalgia
Public policy should ask firms that gain from automation to disclose what happens to junior hiring and training. Large employers need not reveal commercial plans, but sector-level reporting can show whether productivity gains coincide with a broken entry ladder.
Professional bodies should define supervised tasks that remain necessary for competence even when software can complete them. Medical training does not abandon diagnosis because a tool suggests one. Legal, financial, engineering and administrative training should similarly distinguish between work that may be delegated and judgement that must be demonstrated.
Governments can co-fund time-limited apprenticeships in exposed sectors, particularly for smaller employers that cannot absorb the full training cost. Procurement can reward suppliers that maintain accredited entry routes. Universities should assess students without AI as well as with it, so credentials continue to certify independent ability.
The strongest countercase
Protecting junior roles can become a defence of cheap labour and outdated hierarchy. Many entry jobs teach little; they consume talented people's years in formatting slides or moving information between systems. AI can give a young worker access to expertise, raise the level of the first assignment and allow ability to count more than pedigree.
That possibility is important. The objective is not a quota for drudgery. It is measured capability transfer. If AI-assisted workers reach independent competence faster, firms should be free to shorten the old ladder. But they should have to demonstrate that the ladder was replaced, not simply removed.
The test is delayed
The decisive employment statistic may appear years after the productivity gain. It will show up when organisations cannot fill experienced roles internally, when professional error rises or when opportunity closes around a small group that acquired expertise before automation.
Democratic societies should not wait for that delayed signal. They can welcome tools that remove routine work while insisting that the next generation still receives difficult tasks, correction and responsibility.
The apprenticeship is not an inefficiency attached to the job. It is how the job reproduces its knowledge. An economy that automates that function must build another one on purpose.
The Global Federation treats progress as an institutional question: who gains capacity, who keeps agency and which public goods markets will not preserve by themselves.