Universities are racing to teach students how to use artificial intelligence. That is necessary to prepare future professionals for the workplace but it doesn’t address a harder problem: if employers outsource entry-level tasks to AI, will graduates be expected to arrive ready for more complex work? There is good reason to worry about this gap. Universities might be teaching students to prompt better but the value of those skills is reduced if graduates enter the workforce without enough experience to know when the powerful systems they can operate are wrong.
Generative AI has the potential to deliver the same kind of upward shift in efficiency as calculators, spreadsheets and high-level programming languages – but only if universities distinguish performance from learning. A student can produce an impressive answer with AI while learning very little. The labour market is making this problem urgent. Stanford Digital Economy Lab researchers have reported that employment among workers aged 22 to 25 in highly AI-exposed occupations stood about 19 per cent below where it would have been had it kept pace with similarly aged workers in less-exposed occupations.
Although the authors of the study stress that they do not see economy-wide displacement, the mechanism matters for universities. The revised analysis highlights the advantage of tacit knowledge acquired through practice, mentorship and repeated exposure.
So, AI is creating an apprenticeship problem. Entry-level tasks are where professional judgement is often built; they expose novices to variation, mistakes, exceptions and feedback. Junior analysts clean data before they learn which anomalies matter. Young lawyers review documents before they learn which facts change a case. But this work looks inefficient once a machine can do it. When the work disappears, and AI short-circuits acquisition of foundational skills, the training path can vanish, too.
- We promise our graduates will be employable. Are we keeping that promise?
- Video: To AI or not to AI? That is the question
- Connecting the institutional dots for workforce readiness
Universities should therefore manufacture experience that employers could stop providing. That means designing sequences in which students encounter a problem with enough friction to build a mental model, then use AI to extend performance and, finally, explain, audit or recover the work without relying on the tool. A finance student, for example, might build a small valuation manually, then use AI to scale the analysis, and finally identify assumptions that would make the model fail.
The same approach can use AI aggressively without romanticising struggle. AI can sometimes help novices learn faster. A field study of 5,179 customer-support agents found that access to a generative AI assistant raised productivity by about 14 per cent on average and by 34 per cent for novice and lower-skilled workers. The researchers also found suggestive evidence that the tool helped transmit practices used by more capable workers. Universities should want AI to serve as compressed access to expert practice rather than a substitute for developing expertise.
The educational danger appears when AI improves the artefact while concealing weakness in the person. A student who has never wrestled with a difficult first draft can ask a model to produce polished prose and still be unable to judge whether the argument is coherent.
That said, AI can extend the meaningful practice universities can offer. Instructors can use these tools to create opportunities for students to practise interpersonal skills, for example. And the principle is broader than communication courses. AI can generate cases, adversarial scenarios, alternative explanations, simulated clients and edge cases. Used this way, it increases exposure to variation instead of removing it.
Practice alone is insufficient, though. Students also need to learn calibration; metacompetence is a another name for the ability to know which tasks can be delegated, which require direct understanding, how much confidence to place in an output, what evidence should trigger doubt, and when to descend below the abstraction. Expertise in an AI-rich workplace increasingly includes knowing where human attention remains valuable and where automation has become brittle.
Work-integrated learning, internships, co-ops and degree apprenticeships all become more important to the university offer as employers automate entry-level work. A practical starting point to boost these is an experience audit for every degree with a clear professional pathway. Departments should list the recurring tasks that graduates used to encounter in their first two years of work, identify the capability each task helped build, and ask whether AI is likely to shrink or eliminate that exposure. The key column identifies the learning function that would disappear when a task no longer provides practice. A journalism programme might find that basic reporting assignments taught students to distinguish a confident source from a reliable one, for example. Once the learning function is visible, faculty can design a better route to the same judgement.
The redesign will require resources and partnerships. Good simulations, supervised projects, oral defences and feedback-rich assignments demand more faculty attention than simply assigning a paper and grading the final product. Employers also have a stake because they will eventually need the judgement that universities are being asked to cultivate. Universities should treat employer projects, alumni mentoring, clinics and apprenticeships as part of the learning infrastructure for an AI economy. AI itself can generate varied practice cases and simulated stakeholders, but human experts still need to decide which experiences matter and what competent performance looks like.
An experience audit also exposes a hidden subsidy in the old university-to-work transition. Employers often expected universities to provide knowledge while assuming workplaces would provide experience. Universities could reasonably expect some forms of professional judgement to arrive later on the job. If AI reduces the number of junior roles or strips those roles of routine practice, that bargain breaks. Telling students to become more AI literate does little good if the labour market offers fewer chances to become seasoned.
Universities have spent the past few years debating whether students should be allowed to use AI. The more consequential issue is what students must experience before they can safely let AI do the work. If automation removes the first rung of professional learning, higher education will have to build a new one.
Gleb Tsipursky is a behavioural scientist and CEO of Disaster Avoidance Experts. His most recent book is The Psychology of AI Adoption at Work: From Resistance to Results (Georgetown University Press, 2026).
If you would like advice and insight from academics and university staff delivered direct to your inbox each week, sign up for the Campus newsletter.
comment