AI Is Not a University. But It Can Build You One.

In this article8 min read

AI has made access to knowledge almost embarrassingly easy. That has not made learning easy. Building a personal AI university takes rather more than access.

We can ask ChatGPT to explain a concept, use NotebookLM to interrogate a set of sources, and have Gemini draft a study plan in seconds. Yet most of us still collect more than we complete. Access was never really the obstacle. The hard parts were deciding what mattered, finding the gaps in our own understanding, and doing enough deliberate work to make the knowledge usable.

One useful way to organise that problem comes from Sandeep Swadia, who describes building a “university in a box” and gives AI five roles: adviser, librarian, tutor, editor and roommate. He arranges them by their initials, which spell ALTER — roles that alter how you think, learn and behave.

It is a memorable model, and it does something more valuable than being memorable. It moves AI away from being a clever answer machine and towards being part of a learning system — a personal AI university you assemble deliberately rather than accumulate by accident.

The phrase “university in a box” needs one warning label, though.

AI can assemble the support structure. It cannot do the learning for you.

What a personal AI university has to solve

For most of history, learning was limited by access. Books were expensive, experts were distant, and specialist courses were bounded by geography, time and money.

That constraint has moved. However, we now have more material than we can evaluate, more courses than we can finish, and more AI-generated explanations than we can verify.

The bottleneck is no longer finding information. It is converting information into judgement.

That conversion needs five different kinds of help, which is where your personal AI university starts to earn the name.

A — Adviser: decide what the learning is for

Most weak learning plans begin with a topic.

Teach me corporate finance.

That sounds reasonable, but it gives the system almost nothing to work with. Corporate finance for what purpose? At what level? Over what period? What evidence would demonstrate competence?

The adviser role settles five things before any curriculum appears:

• Destination: what should you be able to do at the end?
• Baseline: what can you already do without assistance?
• Sequence: what has to be learned first?
• Cut list: what can safely be ignored for now?
• Milestones: what output will prove you are ready to move on?

This is the first meaningful shift. Do not ask AI to produce a curriculum before it has interviewed you. A generic six-week plan feels productive because it is neat. A useful plan is messier, because it reflects your actual gaps, your available time and the thing you intend to apply it to. That interview is where a personal AI university stops being a metaphor and starts being a plan.

L — Librarian: control the evidence

Once the plan exists, the next risk is source quality. A model can produce a confident explanation built on weak, outdated or fabricated foundations.

Fluency is not evidence. A polished answer can still be wrong.

The librarian role narrows the environment. Rather than letting AI roam across everything it may have absorbed, give it a deliberate source pack:

• one strong introductory text;
• one authoritative reference;
• one current course or lecture series;
• a small number of high-quality case studies;
• primary research where the topic demands it.

Tools such as NotebookLM help here, because the conversation stays grounded in material you selected. In a personal AI university, this is the step most people skip. That does not remove your responsibility to assess those sources, but it does make the boundary visible.

Finding the most material is the easy part. The librarian’s real job is protecting your attention from material that does not deserve it.

T — Tutor: expose the gap rather than repeat the lesson

This is where AI gets more interesting, and where most people use it least well.

The common pattern is to treat it as a patient lecturer: another explanation, a simpler analogy, a tidier summary. That can help. But explanation alone produces a dangerous feeling of familiarity. Something makes sense while we are reading it, so we assume we understand it.

A tutor should test that assumption. Two instructions do most of the work:

Teach me.

Test me.

The second matters more than it first appears. Ask for one question at a time, ask the model to diagnose the reasoning behind a weak answer, and keep working the gap until you can explain the idea without leaning on its words. That turns consumption into retrieval practice, which is one of the better-evidenced findings in learning research. Therefore the tutor is the role a personal AI university most often gets wrong.

The system should make thinking harder at the right moment, not make every moment frictionless.

E — Editor: make the work survive contact with reality

Learning becomes valuable when it changes an output, a decision or a behaviour. That is why the editor belongs in the model. Ask AI to attack what you have produced:

• Where is the argument weakest?
• What have I assumed without evidence?
• Which section repeats rather than advances the point?
• What would an informed sceptic challenge first?
• What is precise in my head but vague on the page?

In programme delivery, this is the difference between knowing a framework and being able to use it when the evidence is incomplete, the stakeholders disagree and the deadline is real.

AI can tighten the work. Responsibility does not transfer with the redraft. Swadia puts it plainly: AI can be a powerful ally when editing your work, “but you still have to fly the plane.” That is the one rule a personal AI university cannot bend.

R — Roommate: borrow a different way of seeing

The final role is the least obvious and possibly the most useful. A good roommate brings knowledge from outside your field. Random novelty misses the point. What you want is something that disturbs a familiar pattern.

Questions of this shape tend to work:

• What could programme governance learn from air-traffic control?
• How does a jazz group coordinate without a fixed script?
• What does queueing theory reveal about an overloaded approval process?
• How would a behavioural psychologist explain repeated failure to adopt a new tool?

Cross-disciplinary prompts can produce nonsense, so the comparison still needs testing. For example, an air-traffic-control analogy breaks down the moment nobody has authority to ground a flight. They can also surface assumptions that stay invisible inside one profession’s language. The roommate gives you a stranger’s lens. However, you decide whether the view is useful, because a personal AI university has no opinion of its own.

Where a personal AI university goes wrong

ALTER is not a licence to outsource judgement. Each role in a personal AI university has a failure mode, and they are worth knowing before you lean on any of them.

Role	Useful when…	Risky when…
Adviser	the destination and constraints are explicit	the model invents a neat plan around the wrong goal
Librarian	sources are deliberately chosen and traceable	generated citations get treated as verified evidence
Tutor	questions expose gaps and force retrieval	hints arrive so fast that the struggle disappears
Editor	critique strengthens work you still own	the model quietly replaces your judgement and voice
Roommate	an analogy opens a new line of inquiry	novelty gets mistaken for truth

Removing friction from learning is the wrong goal. Put friction where it creates value instead: choosing, recalling, applying, defending and revising.

Start your personal AI university with one course

The temptation is to turn this into an elaborate dashboard with fourteen integrations and a colour-coded curriculum that updates itself.

Resist it, because the system is not the point.

Choose one capability that matters now. Give it a six-week boundary. Define one real output. Select a small source pack. Schedule only the work you can genuinely complete, and review the system after the first week rather than after a weekend spent perfecting it.

I have watched this pattern repeat across four platform shifts — mainframe to client-server, on-premise to cloud, waterfall to agile, and now AI. Each one arrived with its own wave of courses, certifications and confident explainers. The people who came out of it genuinely capable were rarely the ones who consumed the most. They were the ones who built something real while the shift was still happening.

Download the Personal AI University Builder

I have turned the ALTER framework into a practical worksheet you can use with ChatGPT, Gemini or Claude. It helps you define the outcome, control the source material, set up the tutor and editor prompts, and choose an outside perspective worth borrowing.

The six-page workbook includes:

• a one-page course brief covering destination, baseline, constraints, cut list and final evidence;
• the five ALTER role prompts, written to be pasted straight into a model;
• a six-week planner built around evidence rather than activity;
• a weekly retrieval, application and revision check;
• a final test of whether the course changed what you can actually do.

Six pages, no sign-up. Print it or work through it on screen.

Swadia closes with a parable worth borrowing. A student in a hurry to cross a river finds no bridge, only boats of every shape and size lying on the bank. She asks the old boatman which one will get her across fastest. He tells her: the one you get into and start rowing.

The best learning system is the same. Not the one you designed most carefully. The one you actually start.

Choose one capability, and build your first six-week course.

Sources used

1. Sandeep Swadia, “How To Become Dangerously Self-Educated With AI (for free)”, YouTube. Origin of the “university in a box” framing, the five roles and the ALTER acronym.
2. Henry L. Roediger and Jeffrey D. Karpicke, “Test-Enhanced Learning: Taking Memory Tests Improves Long-Term Retention”, Psychological Science, 2006.
3. Google, NotebookLM. Used in the librarian role to keep a conversation grounded in sources you selected.

Related reading

• The Learning Triage: How to Decide What to Learn Next
• Not Everyone Needs to Learn Everything
• The 80% Problem: Why Your Brain Is Outsourcing Its Judgement to AI