What the University Is Still For

Unbundle it first

The problem is structural: most universities are organised, staffed, timetabled and priced around function one. Functions three and four — the ones gaining value — are largely treated as pleasant side effects of the arrangement rather than the product. That mispricing is where the next decade of disruption lands.

Function one really is going

It is worth being blunt, because a lot of institutional commentary is not. Delivering standard knowledge to a room of a hundred people at a fixed pace was never a good way to teach. Benjamin Bloom’s 1984 paper argued that one-to-one tutoring moved the average student roughly two standard deviations above conventional instruction — a gap so large it defined a research problem: how do you get tutoring-quality outcomes at classroom cost?

We now have something with tutoring’s core properties — infinitely patient, available at 3 a.m., adapts to your level, never makes you feel stupid for asking. It is not as good as an excellent human tutor. It is available to everyone, which an excellent human tutor is not. For transmission, that trade is decisively favourable.

The catch is the one from cognitive fitness: a tutor that removes all friction teaches worse than one that calibrates it. The technology enables tutoring at scale. Whether it delivers learning at scale depends on pedagogy nobody has finished designing.

The credential argument, staged honestly

Bryan Caplan’s The Case Against Education (2018) makes the strongest version of the sceptical case: the large wage premium for graduates is mostly signalling rather than human capital. Employers pay for the diploma because it certifies intelligence, conscientiousness and willingness to conform — not because the holder learned much that is useful. His evidence includes the enormous amount of curriculum content that graduates cannot recall, the weak evidence for far transfer, and the “sheepskin effect”: the disproportionate wage jump for finishing the final year rather than for accumulating years.

The counter-evidence is also real. Studies exploiting compulsory-schooling law changes — where extra education was imposed rather than chosen — generally find genuine earnings returns, which pure signalling struggles to explain. The truth is a mix, and reasonable economists put the split in different places.

What matters here is what AI does to each component:

• The signalling component erodes. If the signal certifies capacity for cognitive labour, and cognitive labour is being repriced downward, the signal buys less. Employers have begun dropping degree requirements — though the announcements have run well ahead of actual hiring behaviour.
• The human-capital component splits. The part that was procedural knowledge loses value. The part that was judgment, taste and disciplinary depth gains value, because it is what lets you evaluate machine output at all.

What breaks immediately: assessment

The take-home essay is finished as an assessment instrument. So is the unsupervised problem set, the coding assignment, and the literature review. Not because students are dishonest but because the instrument measured the production of an artefact, and artefact production is now free.

What replaces it already exists, mostly in the older and more expensive parts of the university:

• Oral examination. The viva is unfakeable. Ten minutes of follow-up questions establishes what someone understands with a fidelity no written submission ever had. It is expensive in staff time, which is why it was abandoned, and it is coming back.
• Supervised production. Work made in the room, under observation, with the process visible — the studio model that art, architecture and music schools never gave up.
• Critique. Defending your work to peers and a master. This assesses judgment directly rather than through the proxy of output.
• Portfolio over transcript. A body of real work with a documented process and a person willing to vouch for how it was made.

Every one of these is higher touch, lower scale, more human. The economics of higher education have run in the opposite direction for forty years. That tension is the story of the next decade.

So: should you go?

Here is the rubric, stated plainly. It depends far more on your specific situation than on any general claim about AI.

Note the deferral option, which is underused. A gap year spent building something real, then applying with a portfolio and an actual reason to be there, produces better outcomes than arriving at eighteen with no question you want answered.

What to learn: the barbell

This is the more important question, and it has a clearer answer than the institutional one. Distribution of effort should be barbell-shaped — heavy at both ends, nothing in the middle.

The deep end: one domain, far enough in to detect error

Pick something and go deep enough that you can tell when a confident machine output is subtly wrong. This is the verification gap applied to a life plan. Shallow breadth across many fields — which AI makes very easy to acquire and very pleasant to possess — leaves you unable to evaluate anything. You become a highly informed person who cannot detect nonsense, which is a worse position than ignorance because it comes with confidence.

Depth does not require a degree. It requires years, real problems, and feedback from reality or from someone who knows more than you.

The meta end: what generalises across every domain

• Question formation. The scarce skill. Answers are free; knowing which question to pose, and noticing when the question itself is wrong, is not.
• Verification and epistemics. How to check a claim. What counts as evidence. Calibrating confidence. This is now a core literacy in the way arithmetic was.
• Taste. The ability to tell good from merely competent. Machines generate competent output in unlimited quantity; the bottleneck moves entirely to selection.
• Systems thinking. Second-order effects, feedback loops, incentives. Where most consequential error now lives.
• Communication and persuasion. Not prose generation — the machine does that. Holding a room. Changing a mind. Being trusted.
• Negotiation, ethics under uncertainty, and the ability to sit with an unresolved problem. Rarely taught, universally load-bearing.

The hollowed middle: routine cognitive procedure

Producing a competent standard artefact to spec — the summary, the boilerplate, the first-pass analysis, the formatted report, the template code. This was the substance of enormous numbers of well-paid careers. It is where machine capability is strongest and improving fastest. Do not build a career on it.

This is correct, unresolved, and the single hardest problem in this whole subject — the apprenticeship problem in its economic form. The only honest response available to an individual: pass through the middle deliberately and unaided, treating it as training rather than as output. Do the routine work yourself, knowing a machine could do it faster, because you are not producing an artefact — you are building the perception that will later let you judge one. That is expensive, it will feel irrational, and there is currently no institution that will pay you for it. Which is precisely why it has to be chosen.

How to learn, when the tutor is infinite

Practices

01 Unaided first, then adversarial

Attempt it yourself. Then hand your attempt to the machine and ask it to find everything wrong. This inverts the default and converts the tool from an answer source into a tutor — which is the only configuration where it actually teaches.

02 Ask for the Socratic mode explicitly

“Do not give me the answer. Ask me questions until I find it.” The tool will do this well and will never do it unprompted, because every interface is built to resolve rather than to prolong.

03 Build things with real users

A project nobody uses teaches you what you already believed. Reality is the only feedback source that is not trying to be agreeable, and agreeableness is the machine’s dominant failure mode as a teacher.

04 Apprentice to a person, not a curriculum

Find someone better than you who will look at your work and tell you what is wrong with it. This single relationship outperforms any course. It is also the hardest thing on this list to arrange, which is why it stays scarce and valuable.

05 Learn in public, and teach it

Explaining forces the retrieval and exposes the gaps that recognition hides. Publishing does the same with added stakes. The Feynman method survives contact with AI intact — because the constraint was never information access.

06 Space it and interleave it

Return to material after you have begun to forget. Mix problem types instead of drilling one. Both feel worse and work better — the desirable-difficulty finding is one of the most robust in the field.

If you are forty-five, not eighteen

The reskilling conversation usually gets the asymmetry backwards. The mid-career learner’s disadvantage is time and identity, not capacity — the research on adult learning does not support the idea that the mind closes. Their advantage is enormous and consistently undervalued: they have domain context, which is exactly what determines whether machine output can be evaluated. A twenty-five-year veteran of logistics with six months of fluency in these tools is more valuable than a technologist with no logistics.

The move is almost never to retrain into a new field from zero. It is to become the person in your existing field who understands the tools deeply. The scarce combination is domain depth plus tool fluency, and you already hold the half that takes twenty years.

The genuine obstacle is identity. If your professional self-concept is “I am the person who does X,” and X is being automated, the threat is not economic first — it is a demotion of the self, which is the subject of the meaning page. That is the work to do first. The skills are the easy part.

Where learning sits on the ladder

1. 1

   Cede Machine

   The machine does it. You never see the work.

   Having the machine produce the work you were assigned. The artefact is fine; you are unchanged. This is the default and it is the trap.

2. 2

   Approve Machine-led

   The machine drafts. You review and sign.

   Editing generated work teaches editing. It does not teach the thing the assignment was for.

3. 3

   Collaborate Shared

   You and the machine work the problem together.

4. 4

   Critique Human-led

   You do the work first, unaided. Then ask the machine to attack it.

   Unaided attempt, then adversarial critique. The highest-yield learning configuration currently available to anyone.

5. 5

   Reserve Human

   You do it alone, on purpose, knowing help exists.

   The deliberate reps through the hollowed middle — inefficient by design, because you are building perception rather than output.