Expertise is rarely created in a single decisive moment. It accumulates through repeated encounters with real work: drafting, checking, revising, explaining, being corrected and gradually taking responsibility for harder decisions.
For a long time, organisations obtained much of this development as a by-product of production. Junior professionals completed bounded, lower-stakes tasks. The task created value, but it also exposed them to examples, exceptions and feedback. Over time, patterns became knowledge and knowledge became judgement.
Generative and agentic AI can now perform a growing share of precisely that clean, checkable work. This creates an important educational problem. If AI removes the first rung of the ladder, how will people acquire the expertise required to supervise what happens higher up?
Expertise can no longer be treated as a by-product
The immediate attraction of AI is efficiency. A capable system can summarise, classify, draft, compare and organise at a speed that makes manual practice look wasteful. In production settings, that efficiency may be entirely rational.
Learning follows a different logic.
A novice does not become capable merely by receiving a high-quality output. Capability develops through the cognitive work of attempting a task, detecting an error, understanding why it occurred and trying again. The learner needs structured exposure to variation and consequence—not only access to the answer.
This is why the same AI workflow can be productive for an expert and developmentally thin for a novice. The expert can use the output as an object of judgement. The novice may lack the knowledge needed to recognise what is incomplete, misleading or wrong.
The central issue is therefore not whether students should use AI. They will. The question is whether the learning environment preserves the work through which they become qualified to direct it.
Deep knowledge returns to the centre
The claim that students no longer need substantive knowledge because a system can retrieve it confuses access with judgement.
To evaluate an AI-supported answer, a person needs a model of the domain: its concepts, causal relationships, standards of evidence and typical failure modes. Without that structure, fluency is easily mistaken for accuracy and confidence for warrant.
AI literacy must therefore include deep disciplinary learning. Students need to know enough to ask consequential questions, notice omissions, compare alternatives and recognise when the system is operating outside an appropriate boundary.
Knowledge is not the alternative to higher-order skill. It is part of the infrastructure that makes higher-order judgement possible.
Productive friction is not inefficiency
If education optimises only for the speed and polish of the final artifact, it can optimise against learning.
Some difficulty is productive. Writing an imperfect explanation, receiving critique and revising it can build a mental model that accepting a polished answer does not. Searching for evidence, resolving contradictions and defending a choice can develop judgement that a frictionless workflow conceals.
The goal is not to ban helpful tools or romanticise unnecessary struggle. It is to locate the effort that builds capability and protect it deliberately.
That may mean asking students to formulate a position before consulting AI, document how evidence changed their reasoning, compare multiple generated approaches, identify where they overrode a system or explain why a seemingly strong answer should not be trusted.
The relevant question is not “Did AI make the task easier?” It is “Which human capability did the learning design make stronger?”
Assessment must make performance observable
Generative AI weakens the value of unsupervised written products as evidence of individual capability. A polished submission may still be useful work, but it reveals less about what the student understands and can do independently.
Assessment therefore needs a broader evidence base:
- oral defence and structured dialogue;
- live or time-bounded problem solving;
- authentic projects with visible roles and decisions;
- source trails and human-intervention records;
- iterative drafts with feedback and revision;
- individual reflection connected to specific choices;
- opportunities to respond when assumptions or conditions change.
These forms of assessment are more demanding to design and facilitate. That cost is not incidental. If credible evidence of learning becomes harder to produce, institutions must decide that it is worth producing.
Apprenticeship moves inside the curriculum
When workplaces automate more entry-level cognitive production, universities cannot assume that graduates will acquire professional judgement later through routine work. They need to create more deliberate forms of apprenticeship inside the educational experience.
This does not require imitating a workplace superficially. It requires the conditions through which responsibility grows:
- bounded tasks with clear feedback;
- repeated practice across varied situations;
- access to peers and more experienced judgement;
- authentic problems with competing constraints;
- increasing agency as capability becomes visible;
- accountability for decisions, not only deliverables.
Project-based and inquiry-driven learning become especially valuable when they are properly scaffolded. Students can begin with guided preparation, move into facilitated application and eventually take responsibility for an open-ended problem whose framing and evaluation criteria are not supplied in advance.
The university’s job changes
AI makes access to capable performance more abundant. It does not make the formation of capable people automatic.
The university’s distinctive contribution is increasingly to create the environments in which learners build the knowledge and judgement that organisations may no longer develop incidentally. That means treating curriculum, assessment and feedback as infrastructure for expertise.
It also means redefining AI literacy. Prompt technique and tool familiarity are useful, but the deeper capabilities are:
- building knowledge before delegating;
- framing the problem before solving it;
- evaluating evidence before accepting a conclusion;
- supervising AI-supported work rather than consuming it passively;
- recognising exceptions and contextual limits;
- defending decisions and retaining accountability.
AI can make performance easier. Education must still make expertise possible.
This essay develops the teaching philosophy behind my work on AI literacy and adaptive expertise in higher education. See the teaching portfolio and the framework on adaptive expertise for human–AI work.