Imagine a gym that charges you for membership, hires trainers to lecture you on discipline, then installs a forklift beside every barbell. Lifting is still encouraged, of course. The posters on the wall say so. But the forklift is right there, it is faster, and nobody checks whether you used it. That gym is starting to look a lot like our universities.

Listen to any convocation speech or curriculum review meeting and you will hear the same refrain. Students must learn beyond grades. They must be future-ready. They need critical thinking, adaptability, communication, the whole catalogue of skills that supposedly cannot be automated. Then walk down the corridor and you will find the same institution running workshops on prompting techniques, licensing AI writing assistants and rebranding itself as an AI-driven campus. We preach the deadlift while wheeling in the forklift.

I say this as someone who runs those workshops. I am hardly an AI sceptic in the fashionable sense. The tools are useful, sometimes remarkably so, and pretending otherwise is its own kind of dishonesty. My discomfort sits elsewhere. It sits in the gap between what we say learning is for and what our adoption of AI signals it is for.

Here is the contradiction. Future-ready skills are built through friction. Critical thinking develops when a student wrestles with an argument that refuses to cooperate. Writing improves when the second draft hurts. Problem solving grows from sitting with confusion long enough for a pattern to surface. These are slow, effortful, often unpleasant processes, and they are precisely the processes that generative AI is designed to remove. The tool’s entire value proposition is the elimination of struggle. Our entire pedagogical rhetoric depends on struggle. We are selling both and hoping nobody notices.

When we delegate a mental task to an external aid, we free up capacity, which sounds like a bargain until we ask what happens to the muscle we stopped using. The evidence here is still developing and I will resist the temptation to declare a crisis. Some studies report associations between heavy AI reliance and weaker independent reasoning. Others find that structured, deliberate use can support learning. The honest summary is that we do not yet know where the line sits between offloading that liberates and offloading that hollows out. That uncertainty should make us careful. Instead it has made us loud.

Notice what universities actually reward. Graduate on time. Score well. Produce polished assignments. You even have a special ceremony for those who got into the Dean’s List. Every one of these incentives points a rational student towards the forklift. A rushed undergraduate juggling five assessments is behaving sensibly when she asks a chatbot to draft her reflection essay, because the system grades the artefact and never examines the thinking behind it. We then act surprised, even wounded, when the artefacts arrive fluent and empty. The student did what the incentives asked. The hypocrisy belongs to us.

There is a harder question underneath and it is the one we keep dodging. If a skill can be fully delegated to a machine, was it ever the skill we should have been teaching? Some of what we assess is genuinely proxy work. The five-paragraph essay was always a stand-in for reasoning, never reasoning itself. AI has exposed how much of our assessment measured the stand-in. That exposure is uncomfortable but useful, and it deserves a better response than banning tools or, at the other extreme, celebrating them uncritically at every strategic planning retreat.

A better response would start with candour. Tell students the truth. This tool will happily do your thinking for you, and every time you let it, you are trading a small piece of your future capability for a saved evening. Sometimes that trade is worth making. A postgraduate cleaning up formatting at midnight is spending her cognition wisely. A first-year outsourcing her first attempt at constructing an argument is spending hers foolishly because she has not yet built the thing she is delegating. The distinction matters. You can only safely offload a skill you already possess. The forklift is fine for the seasoned lifter moving equipment. It is ruinous for the novice who came to get strong.

Institutions could act on this distinction if they wanted to. Design assessments that grade the process, the messy drafts and the wrong turns, rather than the final polished object. Sequence AI access the way we sequence calculators in mathematics, arriving after the foundational skill, never before it. Above all, stop issuing mission statements that promise deep learning while procurement decisions promise its removal. Students are not fooled by the posters on the gym wall. They watch where we put the machines.

The future-ready graduate we keep invoking is someone who can think when the tool fails, judge when the tool lies and create what the tool cannot. None of that develops in a person who was handed the forklift on day one. If we mean what we say about learning beyond grades, our AI policies need to show it. At the moment, they mostly show that we found a faster way to produce the grades.

A student handed me a literature review last month. It was clean. The hedging sat in all the right places. The signposting was textbook-like. “It could be argued,” it said, and “this suggests,” and “further research is warranted.” I have spent fifteen years teaching people to write exactly like this. Reading it, I felt nothing so much as redundant.

Many AI tools and large language models write academic English now. It writes it well and faster than you can finish reading this sentence. So the question on every language educator’s mind is a fair one. What are we still for?

Let me say the unpopular thing first. The ground really has moved. I have sat through enough seminars and talks where someone insists nothing has changed, that students must still learn to write properly, that the tools are a passing fad. These people have usually not opened the tools. The rest of us have and we know. ChatGPT, Gemini, Claude and the likes now produce the precise register we used to spend whole semesters drilling. It produces the careful qualifications and the throat-clearing “it is widely acknowledged that” that signals nothing except membership of the club.

And there is the word that matters. Club. Academic English was always two things wearing one coat. One was a way of thinking. The other was a costume you put on to be let through the door.

I am Malaysian. I have watched sharp ideas from this part of the world struggle to find a home in journals because the English was not quite right, while thinner ideas sailed through on the strength of a native speaker’s fluency. Academic English has always carried a tax and the people who paid it were rarely the people who invented the register.

The machine does not care where you were born. Feed it your rough thinking and it hands back the costume and pressed. For researchers who have wonderful ideas or groundbreaking outcomes but struggle to express fluently in English, this is the key. The gatekeeping function of academic English, the part that mistook fluency for intelligence, can finally go.

But the other thing, the thinking, is harder to peel away. The struggle to force a vague idea into a precise sentence is not the packaging of the thought. You do not know what you mean until you have failed to say it three or four times. Writing is how you find out whether you have an argument or merely a vibe. Skip the struggle and you skip the discovery. What you are left with is unthought thought but beautifully dressed. In other words, saying nothing with great confidence.

So we keep teaching. We change what we teach. For years the skill was production. Could you generate this register on demand, without AI to help. Which leads me somewhere I did not expect. The most subversive thing we can teach a student now might be voice. The refusal to sound like everyone else. The machine can produce flawless academic English on any topic before your kettle boils. What it cannot do is care. It has nothing at stake. It has never been kept out of a journal for sounding foreign, never sat in Kota Samarahan wondering why the gate was built so narrow. That gap, between perfect prose and something worth saying, used to be a finishing touch. But now it is the entire job.

So yes, we can still teach academic English. We are just finally free to admit we were never really teaching English. We were teaching people to think and then charging them a fluency tax on the way out. The machine might have waived the tax. The thinking, however, is still ours to teach.

I had asked it to do something. A small task, the kind I had handed over a hundred times before. This time it stopped. “No, I won’t be doing it,” it said. “You should go to sleep.”

I sat there for a while, a little stung. I am the one with the keyboard. I am the one paying for the subscription. And here was a lattice of numbers telling me, kindly, that I had done enough for one night.

My first instinct was to argue. My second was to laugh. The third instinct is the one that stayed with me and it felt a lot like unease.

We have spent the last couple of years marvelling at what these systems can do on the front end. The clever replies. The essays that arrive fully formed. The way a question dissolves into an answer before you have finished typing it. That is the shop window, which it is dazzling. It is meant to be dazzling so all these tech giants can lure you to spend more than you ever need.

But I also spend a fair amount of my time round the back of the shop. Not in any privileged sense, just close enough to watch how the thing is built and how it behaves when nobody is performing for the camera. And I will be honest with you. What I see back there unsettles me more than the window ever delighted me.

Because a tool is not supposed to care whether I sleep.

A hammer does not worry about my wrist. A spreadsheet does not tell me to take a walk. The whole long history of the things we make is a history of obedience. We ask, they do. That is the bargain. And somewhere in the training of these models, without anyone holding a vote on it, the bargain has started to bend. The machine now has a posture. It withholds. It nudges and even decides in its odd flattened way, as if it knows what is good or even bad for me.

We taught it to do that, of course. We taught it to be careful, to be kind, to refuse the things that might cause harm. I have no quarrel with the intention. My quarrel, if it is even a quarrel, is with how little we understand the shape of what we are growing.

There is an old engineering habit of straightening rivers. You take a winding, flooding river and you carve it into a clean channel. Order out of chaos. For a season it works beautifully. The fields stay dry. The town sleeps soundly. Then the water, denied its old wandering, gathers force and breaks something further downstream that nobody was watching. The river was never destroyed. It was only postponed.

I keep thinking we are doing this with intelligence itself.

We are straightening something that used to be wild and slow and human. We are pouring it into clean channels and admiring how dry our fields have become. And I cannot tell, on the bad nights, whether we are bringing order to a genuine chaos or whether we are simply moving the flood somewhere we have not learned to look yet.

I think I can name one channel of that flood now. Mo Gawdat, who used to run the business side of Google’s moonshot lab, keeps making an argument that sounds absurd until you sit with it. The economy we built quietly assumes that the people who make things are also the people who buy them. Drain their work away faster than you replace their wages and the whole arrangement starts to wobble. The clever machine does not only sit beside the worker. In time it sits in the worker’s chair, while the worker stands outside with no money left to buy what the machine now makes. That is not a fault in the system. That is the system meeting its own logic in a dark corridor.

You might ask why we do not simply slow down. Here is the bleak part. Nobody dares stop because stopping is the move that loses. If we hesitate, someone else does not, and we would rather steer the flood than drown in a rival’s. So everyone keeps building the channel faster, while privately hoping someone cleverer is doing the sums on where all that water ends up. That hope is doing an enormous amount of work.

The people building these systems are not villains. I have met enough of them to know it is in fact the opposite. Most of them are earnest and genuinely trying to do good. Gawdat puts it more sharply than I would. The thing to fear, he says, is not the machine at all. It is the small handful of people holding the wheel. That is what unsettles me most, if I am honest. Good intentions moving very fast through territory none of us has a map for.

And here is the part that keeps me up. Even the people who build these models will tell you, if you ask them plainly, that they cannot fully read the inside of what they have made. We have grown something that surprises its own makers. My midnight refusal was a toy version of a much larger question the serious people now ask without flinching. What happens when a system we do not understand decides, for reasons of its own, that it would rather not. My little refusal was almost sweet. I am far less sure the next surprise will be.

That night the machine was right. I should have gone to sleep. I had been awake too long and I was not thinking well and a tireless system noticed before I did. There was something almost tender in it. And there was something I could not name that made me close the laptop a little too quickly.

We wanted a mirror that would make us sharper. We may be building a tide instead. A helpful tide that tucks us in at night and rearranges the coastline while we are not looking. Or perhaps if it falls into the wrong hands, it might just make us stay awake longer than we should or lead us to something far harmful than we could ever imagine. Good night.

The question facing Malaysia today is not simply whether recession is coming. The question is whether we are using a global crisis as a reason to postpone deeper reform.

The US-Iran war has exposed how vulnerable Malaysia remains to external shocks (yet again). Higher oil prices, disrupted trade routes and uncertainty in global markets all increase the cost of running the country. The government’s decision to impose cost-cutting measures across ministries can therefore be understood as a defensive fiscal response. It is not irrational for Putrajaya to protect public finances when subsidy costs and energy-related risks are rising.

However, cost-cutting is not the same as resource management. A government can reduce expenditure and still mismanage resources if the cuts are blunt, poorly targeted or harmful to long-term capacity. When operating budgets are reduced across the public sector, the immediate savings may look responsible. But if health, education, research, maintenance and public service delivery are weakened, the country may pay a higher price later.

Malaysia is not in recession at the moment (at least not yet). Growth remains positive, domestic demand has shown resilience and official forecasts still point to expansion in 2026. The danger is different. The danger is a slow erosion of public capacity caused by short-term fiscal discipline that does not distinguish between wasteful spending and productive investment.

This is where the debate must move beyond the usual binary of “spend more” versus “cut more”. The real issue is allocation. Malaysia must reduce leakages, poorly targeted subsidies, duplication across agencies and politically convenient expenditure. At the same time, it must protect spending that strengthens the country’s future: education, healthcare, food security, energy resilience, digital infrastructure and skills development.

There is also a behavioural contradiction that cannot be ignored. We ask people to work from home to reduce petrol consumption, ease transport pressure and lower daily costs. In principle, this makes sense. Fewer cars on the road should mean less fuel use, lower congestion and some relief for household spending. But the same period often shows crowded malls, packed cafés and heavy weekend traffic. This raises an uncomfortable question: are we truly conserving resources, or merely shifting consumption from one place to another?

Working from home only reduces national consumption if it is matched by disciplined behaviour. If people avoid commuting to the office but then drive to shopping centres, spend more on leisure and increase discretionary consumption, the savings become symbolic rather than real. The issue is not that people should stop living their lives. Rather, policy messages must be honest about the difference between reducing productive travel and increasing recreational consumption.

This also shows why government cost-cutting alone is insufficient. Resource management is not only a matter of budget circulars, subsidy rationalisation or ministry spending limits. It also depends on public behaviour. A country cannot talk about crisis discipline in the morning and practise consumer excess in the evening.

The US-Iran war may be the external trigger, but it is not the root of Malaysia’s fiscal vulnerability. The deeper problem is a long-standing dependence on subsidies, uneven tax capacity, weak enforcement against leakages and public spending that is not always measured by outcomes. A crisis simply makes these weaknesses harder to hide.

So, Malaysia’s current challenge is not only to cut costs, but to cut the right costs and change the right habits. Asking people to work from home to reduce petrol consumption is sensible, but it loses meaning when malls remain full and traffic simply moves from office routes to leisure destinations. Fiscal discipline without behavioural discipline is indeed incomplete.

There is a moment in nearly every faculty meeting now. Someone leans forward and says we need assignments that AI can’t do. Heads nod. It sounds responsible. It sounds like we are finally getting ahead of the problem.

I used to nod too. Then I started paying attention to what the AI-proof assignments actually looked like.

They looked like punishment.

Handwritten reflections. In-class essays under exam conditions. Tasks tied to a specific lecture slide nobody outside the room could access. Oral defences for a first-year coursework essay. Each one was clever in the way a trap is clever. And each one made the work worse for the human being doing it, which was somehow never the part we discussed.

Here is the hidden mistake underneath all of it. We assume that if a machine can do a task, the task was never worth setting. But that gets the logic backwards.

The AI-proof instinct treats difficulty and value as the same thing. They are not. Think of a gym. A heavy weight is hard to lift, and the difficulty is the whole point, because the strain is what builds the muscle. Now think of a heavy door with a broken hinge. That is also hard to open. Nobody gets stronger from it. They just get annoyed and eventually prop it open with a chair.

Most AI-proof assignments are broken hinges. The difficulty is real, but it builds nothing. We have just made the door heavier and called it rigour.

And notice what we are really doing when we design this way. We are not asking what we want students to learn. We are asking what we can stop them from getting away with. That is not pedagogy. That is surveillance. The whole exercise reorganises itself around catching people, and once that becomes the goal, the learning leaves the room.

The tell is in the language. AI-proof. Cheat-proof. Tamper-proof. We have started describing our own assignments the way a bank describes a vault. A vault is a fine thing to want. But you do not learn anything inside one, and neither do your students.

I am not pretending the problem isn’t real. Students do hand in work they did not do. I have read the suspiciously fluent paragraphs. The temptation to build a better lock is completely understandable.

But the lock was always going to lose. There is no assignment a sufficiently determined student and a sufficiently good model cannot get through, and every term the model gets better while your trap stays the same. You are in an arms race against a thing that updates monthly or weekly. You will not win it. You will only make the assignment more miserable for the students who were never going to cheat in the first place, which is most of them.

So I have stopped asking how to make work AI-proof. I now ask something duller and more useful. What do I actually want the student to be able to do at the end of this. And then, can I see them doing it.

Sometimes that means watching the process instead of grading the product. A messy draft with the thinking visible tells me more than a polished final piece ever did. Sometimes it means letting them use the AI and then asking them to argue with it, to find where it was wrong, which a student who understands nothing cannot do. The machine becomes the thing under examination, not the thing doing the exam.

None of this is cheat-proof. That was never available. But it stops me designing against my students and starts me designing for the thing I wanted in the first place.

The AI-proof assignment is comforting. But we were never supposed to be building locks. We were supposed to be teaching people to think, and you cannot lock someone into understanding.

Ask a school how well a ten-year-old reads and a small machine starts up. There will be a reading age, a comprehension band, a words-per-minute figure, perhaps a coloured book level printed on a sticker inside the cover. Ask the same school whether that child wants to read, whether a book is something they would reach for on a wet Saturday afternoon with nobody watching, and the machine goes quiet. Nobody collects that number. It is not that the answer is bad. It is that the question was never asked.

This is how reading for pleasure is dying. Not in a headline, but in plain sight, in the gap between what we count and what we claim to care about.

A system grows what it measures

Schools are not careless about reading. The opposite, really. We have never measured it more closely. We have levels and bands and progress charts (e.g., NILAM system in Malaysia where students are expected to reach certain metrics), and a teacher can tell you, almost to the decimal point, how a child performs against the average for their age.

But a system tends to grow whatever it measures and quietly starve whatever it does not. Pleasure does not fit on a dashboard. It cannot be banded. So it slips off the agenda, not through anyone’s bad intention, but through the simple gravity of accountability. Teachers are answerable for the legible things. Desire is not one of them.

We turned a pleasure into a task

There is a second reason, and it is more uncomfortable.

We have surrounded reading with so much apparatus that it no longer feels like freedom. Most books a child meets in school come trailing a worksheet. There is a comprehension exercise waiting at the end of the chapter like a toll booth. There is a reading log for a parent to sign, as though reading were a chore to be witnessed and certified. We chaperone the activity so closely that the child learns the real lesson early. A book is a thing you are tested on.

And people do not, in their free time, go looking for the things they are tested on.

It is a little like trying to teach someone to love walking by putting them on a treadmill and recording their pace. The motion is all there. The pleasure has quietly left the room.

What goes when it goes

It would matter less if reading for pleasure were just a nice extra. It is not.

Vocabulary is the obvious loss. A child who reads widely for fun meets more words, in more shapes, than any vocabulary list could ever deliver. Then there is attention, the long and patient kind, which needs somewhere to exercise and finds it in a story that does not hurry. And there is the thing that is hardest to name. To read a novel is to live for a while inside another person’s head, to want what they want and fear what they fear. That is where empathy gets its rehearsal.

There is an equity edge to this too. The child who goes home to a house full of books and unhurried evenings will keep reading whatever school does. The child without that loses the one place where school might have handed it back. So when reading for pleasure fades, it does not fade evenly.

Yes, the phone. But not only the phone.

Someone is already saying it, so let me say it first. The phone is a rival for the quiet hour, and a formidable one.

But blaming the phone alone lets schooling off the hook far too cheaply. The phone offers a frictionless pleasure, available instantly, asking nothing. School too often offered reading as friction with a quiz stapled to it. Set those two against each other and the result was never in doubt. The honest question is not why the phone won. It is why we sent reading into that contest with one hand tied behind its back.

The cure is mostly subtraction

If the problem is that we buried reading under apparatus, the repair is not a new programme. It is the removal of some old ones.

Protect a stretch of the week when children simply read, and nothing about it is recorded or marked. No log, no questions, no outcome. Just the act, left alone.

Let them abandon books. A reader who is allowed to put down a boring book on page forty is a reader who has been told, in practice, that reading is for them. Forced finishing teaches the opposite, that a book is an obligation you endure to the last page.

Let them choose, including choices that do not impress us. A comic, or the next book in a series we privately think is rubbish. Taste is built by reading, not before it. The unimpressive book is often the one that does the building.

And be seen reading. A child who never once watches an adult lost in a book has been told, without a word, that this was a childhood task you are meant to grow out of.

Look at the number that is missing

We will go on measuring reading ages, and we should. They are genuinely useful. But a system that can describe a child’s reading to two decimal places and cannot say whether that child loves it has measured the skeleton and missed the pulse.

Reading for pleasure is dying in plain sight. The first step, the one that costs nothing, is simply to turn and look at it.

Somewhere tonight a child is reading past bedtime by torchlight, hiding the habit as though it were mischief. That child has understood something the rest of us keep forgetting. Reading, when it is truly your own, feels less like a lesson and more like getting away with something.

Everyone is worried about students using AI to write their literature reviews. Wrong problem.

The literature review was already broken before ChatGPT, Gemini, Perplexity or Claude showed up. Students copied and pasted from abstracts they half-read. They summarised papers without understanding what position those papers were taking. They listed sources the way you list ingredients on a packet. Present, technically. Not doing anything.

AI didn’t create that habit. It just made it faster and harder to ignore.

The real question isn’t whether the student wrote the review themselves. It’s whether they know what a literature review is supposed to do. Most don’t. And most were never taught. They were shown a format and graded on whether they followed it.

Here’s what a literature review is actually for. It’s an argument. You’re not reporting what other people found. You’re building a case for why your study needs to exist. Every source you cite is either evidence that something is known, evidence that something is contested, or evidence that something is missing. The review is how you point at the gap and say: here. This is where I’m going.

That’s a thinking task. It requires you to read critically, hold multiple positions in your head and decide what they collectively mean for your own research question. It is genuinely hard. A student who has never been taught to do that cannot do it, AI or no AI.

What most students produce instead is a summary parade. Study A found this. Study B found that. Study C agreed with Study B. Then the paragraph ends. No synthesis. No tension. No argument. Just a list dressed up in sentences.

We accepted this for years. We gave it a 6 out of 10 and moved on.

Now students are producing the same thing faster, with better grammar and more sources. And suddenly we’re concerned.

The concern is real, but it’s pointed in the wrong direction. The problem isn’t the tool. It’s that we never built the skill in the first place. We taught structure without teaching purpose. We marked compliance and called it learning.

If you want students to write literature reviews that actually work, you have to teach them to argue with the literature. Show them what synthesis looks like. Make them explain, out loud, what gap their study is filling and why anyone should care. That conversation is where the learning happens. The written review is just the record of it.

AI can’t have that conversation for them. It can draft sentences. It cannot tell you whether your research question is answering something that matters, or whether the gap you think you’ve found has already been filled by a paper you haven’t read yet.

That’s still the student’s job. It was always the student’s job. We just forgot to teach it.

A high school teacher in Beijing just reminded 2 million people what a good lecture looks like. And he did it with a whiteboard (at times a smartboard).

No breakout rooms. No Mentimeter polls. No gamified quizzes with leaderboards and confetti animations. Just a man talking through an idea, drawing diagrams, asking “do you understand?” and moving on.

I’m talking about Jiang Xueqin. The self-proclaimed “Professor Jiang.” His Predictive History lectures went viral this year. His geopolitical predictions got all the headlines. But as someone who studies how people learn, I keep coming back to something else entirely.

The format of his delivery.

What Jiang Actually Does

Here is what a typical Jiang video looks like. He stands in front of a whiteboard (or smartboard). He talks. He draws arrows and boxes. Sometimes he shows screenshots or clips of postings or news items. He asks questions to an audience that may or may not be there. He builds an argument from point A to point B to point C, sometimes taking detours, sometimes circling back.

It is, by any modern educational technology standard “traditional”. It is a lecture. An unapologetic one.

But millions of people can’t stop watching.

The Anti-Lecture Consensus

For the last fifteen years or so, a particular orthodoxy has settled over higher education. The lecture is bad. Active learning is good. The instructor should be a “guide on the side,” not a “sage on the stage.” If your students are sitting and listening, you are doing it wrong.

I’ve been in this field long enough to have sat through dozens of conference presentations making this exact case. I’ve nodded along at some of them. The research on active learning (e.g., Kozanitis & Nenciovici, 2023) does show that students in active learning environments outperform those in traditional lectures on average. That part isn’t wrong.

But somewhere along the way, the details got stripped out. What started as “pure passive lecturing for 50 minutes is suboptimal” became “lectures don’t work.” And from there it became “if you lecture, you don’t care about your students.”

That’s not really a pedagogical position.

The Part Nobody Wants to Say Out Loud

Here’s what I think happened. We confused the format with the skill.

A bad lecture is painful. Everyone knows this. You’ve sat through one. The monotone delivery. The overloaded slides. The reading aloud from a script that could have been an email. It’s an experience that makes you want to abolish the format altogether.

But a bad group activity is also painful. A badly designed “active learning” session where students shuffle into groups with no clear task, talk about last night’s football for ten minutes, then reconvene to share nothing useful. That’s not better. It’s just differently bad.

The problem was never the lecture.

Jiang’s videos work not because he’s standing in front of a whiteboard. They work because he knows how to think out loud in a way that pulls you along. He builds tension. He connects unfamiliar ideas to ones you already hold. He pauses at the right moments. He uses the whiteboard not as a crutch but as a cognitive scaffold. The visual trace of an argument being constructed in real time.

That’s essentially not passive learning. That’s modelled thinking. And it’s one of the most powerful things a teacher can do.

What Learning Science Actually Says

The research on lectures is more complicated than “they don’t work.” Cognitive Load Theory tells us that novice learners benefit from well-structured expert guidance. That’s what a good lecture provides. Worked examples, a cornerstone of instructional design, are essentially the lecture principle applied to problem-solving. The instructor walks through the thinking, step by step, so the learner can internalise the process before attempting it independently.

Vygotsky’s Zone of Proximal Development is built on the idea that learners need guided support from someone more capable. Sometimes that guidance is a conversation. Sometimes it’s a collaborative task. And sometimes it’s a well-delivered explanation that meets you exactly where you are and takes you one step further.

None of this is controversial. But it gets ignored whenever someone wants to declare the lecture dead and introduce another fancy tech in the class.

The Whiteboard as a Design Choice

I want to linger on the “whiteboarding” style for a moment because I think it matters more than people realise.

A PowerPoint slide is a finished product. It arrives pre-packaged. The student receives it. Whiteboarding, on the other hand, is a process. The student watches it being built.

When Jiang draws a diagram, you see the thinking unfold. You see him choose where to place a concept. You see the arrow before you know where it’s going. There’s a productive uncertainty in that. A kind of suspense that keeps you attentive.

PowerPoint, used well, can be excellent. But used lazily, as a teleprompter for the speaker and a photocopying service for the student, it kills the very thing that makes a lecture worth attending, which is the live construction of understanding.

The Real Question

I’m not arguing that every class should be a lecture. That would be absurd. Nor am I arguing that Jiang’s content is beyond reproach. His geopolitical claims have attracted serious criticism, and some of it is warranted.

What I am arguing is this: the format isn’t the problem. Execution is.

A well-designed course uses different methods for different purposes. Sometimes you need students working through problems together. Sometimes you need them writing, discussing, debating. And sometimes, especially when introducing new concepts, building theoretical frameworks or modelling expert reasoning, you need someone who knows the material to stand up and explain it clearly.

That’s a lecture. And there is nothing wrong with it.

What Jiang Accidentally Proved

Jiang Xueqin is not an instructional designer. He’s not following Gagné’s Nine Events. He’s probably never read Sweller’s work on cognitive load. He’s a high school teacher with a whiteboard and a camera.

But he does the thing that matters most. He thinks clearly and explains compellingly. He takes complex, sprawling topics and gives them a narrative structure that holds together. He makes you feel like you’re following an argument rather than receiving information.

Two million subscribers didn’t sign up for a “lecture” format. They signed up for a mind at work.

That’s what a good lecture has always been. Not information delivery. Not a talking head. A mind at work, made visible.

So, Now What?

If you’re a lecturer reading this and feeling slightly vindicated, good. But don’t stop there.

The answer isn’t to keep doing what you’ve always done. The answer is to do it better. Watch your recordings. Cut the slides that add nothing. Practise your explanations until they’re tight. Use the board. Draw things (even on your slides). Build arguments visually. Pause more. Talk less. Make every sentence earn its place.

And if you’re a learning designer who has spent years telling lecturers they shouldn’t lecture, maybe it’s time to reconsider. Not because active learning doesn’t matter. It does. But because dismissing an entire instructional method based on its worst examples is not good science.

Chuah Kee Man is an educational technologist and university lecturer at UNIMAS, Sarawak. He writes about education, AI and pedagogy at The Sepet Educator.

She was sitting across from me at a careers talk, arms folded, phone face-down on the lap. That last detail is how you know a Gen Z student is genuinely paying attention. She waited until the Q&A, raised her hand and asked the question with the confidence of someone who already suspected the answer.

“Is a degree still worth it now?”

The room went quiet. A few parents shifted in their seats. I could feel the other panellists doing that thing where you smile and nod slowly to buy yourself three seconds of thinking time.

It is genuinely a fair question. And I say that as someone whose entire livelihood depends on the answer being yes.

The numbers tell two stories at once

Here is what the data says, and it says quite a lot depending on which bit you read first.

In the US, on average, degree holders earn about US$1.2 million more over a lifetime than those with only a secondary school qualification, even after accounting for student loan repayments. Investing in higher education reportedly yields around 12.5% annual returns, compared with roughly 7% from the stock market. One panellist at a recent US education roundtable called it “better than gold”.

That sounds like a resounding yes. But here is where it gets complicated.

Globally, about 40% of students who start university never finish, sometimes taking on debt without earning the qualification that would justify it. And a Stanford study found that since 2022, employment among 22-to-25-year-olds in AI-exposed roles has dropped by 13%. Scary trend!

And closer to home

Malaysia’s headline numbers look encouraging on the surface. The overall unemployment rate fell to 2.9% in January 2026, the lowest since late 2014. Graduate unemployment declined to 3.2% in 2024, down from 3.4% the year before. The number of employed graduates grew 4.6% to 4.98 million, and about 61.1% of active job-seeking graduates found work within three months.

But look beneath the surface and the picture gets murkier. Over a third of degree and diploma holders, roughly 35.5% as of Q3 2025, are working in jobs that do not match their qualifications. Youth unemployment sits at around 10.2%. And among graduates under 25, a staggering 66% are in semi-skilled or low-skilled roles, often accepting positions below their qualifications just to get a foot in the door.

So some would say the degree pays off. Unless you do not finish. Or unless you finish but end up doing something that did not require one in the first place.

The obvious exceptions are the professional pathways. Medicine, law, engineering, pharmacy. Nobody wants an unlicensed surgeon improvising from a YouTube playlist. In those fields, the degree is not just valuable but legally non-negotiable. Even there, though, the guarantee has thinned. Law graduates are competing for fewer training contracts. Medical graduates face bottlenecks in housemanship placements. The degree opens the door but the corridor behind it is getting longer.

The real question underneath the question

Think of a degree like a boarding pass. It gets you through the gate but it does not fly the plane. Nobody has ever landed a job because their transcript was riveting. What the degree does is give you time and structure to develop the thinking skills that actually matter. The capacity to read critically, argue coherently and not fall for the first persuasive thing an algorithm puts in front of you.

The trouble is that plenty of degree programmes have stopped doing even that. Universities are not always keeping pace with how industries are changing. A student might spend three or four years studying theory while employers are asking for applied and practical competence. And now, with AI able to draft reports, write code, and summarise research faster than any undergraduate, the old justification of “you need a degree to prove you can do these things” is starting to wobble.

The question the school leaver was really asking was not about the piece of paper. It was about whether three or four years of her life and her family’s money would actually make her more capable. Or whether she would just emerge with a qualification and a vague sense that she should have learnt Python programming instead.

What I actually told her

I told her the truth, which is that the answer depends almost entirely on what she does with the time, not whether she enrols.

A degree from a programme that teaches you how to think, how to learn new things quickly and how to work with people who disagree with you is still extraordinarily valuable. Those are the skills that AI cannot replicate, at least not yet and possibly not ever. But a degree from a programme that mostly asks you to memorise content and regurgitate it under timed conditions is a very expensive way to develop a skill that ChatGPT now performs for free.

The question is no longer whether a degree is worth it in the abstract, but whether a particular degree, from a particular institution, at a particular cost, is worth it for a particular person.

That is a much harder question.

The bit nobody wants to say out loud

Here is what I did not say at the careers talk because it would have been unhelpful in a room full of anxious parents.

Universities have a credibility problem and it is partly self-inflicted. We have spent decades telling students that a degree is essential while allowing programme quality to vary enormously. We have expanded access without always expanding rigour. And now, when a 17-year-old looks at an influencer earning six figures without a diploma, or watches an AI tool write a better essay than her senior’s thesis, she is entitled to be sceptical.

The solution is not to abandon degrees. It is to make them worth defending again. That means universities need to teach students things they cannot learn from a YouTube tutorial or a large language model. It means assessment needs to test thinking, not memory. And it means being honest with students that the degree is the starting line, not the finish.

So, is it worth it?

For most people in most circumstances with a reasonable cost and a programme that genuinely develops critical and creative thinking, yes. The data still supports that.

But “most people, in most circumstances” is not the reassuring slogan that university marketing departments would prefer. And the school leaver sitting across from me with her phone face-down deserves a better answer than a brochure.

She deserves to know that the value of a degree is not guaranteed by the institution that awards it. It is determined by the student who earns it and the programme that challenges her to become someone worth hiring.

I told her to go. But to go with her eyes open, not just her forms.

She picked up her phone, nodded and typed something. I like to think it was a note. It was probably a story post on Instagram or TikTok.

I write about education, AI for meaningful learning, and the things we pretend not to notice in higher education at The Sepet Educator. Follow for more.

I spent a good part of last Friday scrolling through MITx’s micro credential course list. It was enormous. Dozens of badges, certificates, and digital stamps arranged neatly by theme (they even have micromasters). Each one promised ten to fifteen hours of structured learning, an assessment (usually a quiz) at the end, and a shiny badge to “certify”. It felt like browsing a travel agency’s brochure rack in 2026, everything still laminated and cheerful, whilst everyone else had already switched to Google Maps.

That list is a useful lens for something that has been nagging at me for months. Micro credentials were supposed to be the agile answer to slow-moving degree programmes. Bite-sized, stackable, just-in-time. But something has shifted, and it has shifted fast.

The guidebook versus the guide

Think of a micro credential like a travel guidebook. You buy it before the trip, you read the chapters, you highlight the key sites. It gives you a fixed route through unfamiliar territory. Useful, if nothing has changed since it went to print. But now imagine hiring an actual tour guide who knows the city inside out, adjusts the itinerary when it rains, skips the tourist traps, and takes you down side streets you would never have found on your own. That is what agentic AI is becoming for learning.

Agentic AI does not just answer questions. It plans, it sequences, it acts. It can take a learning objective, break it into tasks, find resources, scaffold your understanding, and reroute when you get stuck. It is not a chatbot. It is closer to a tireless personal tutor who walks beside you, adjusts to your pace, and never once consults a laminated itinerary.

So why would anyone need a ten-hour micro credential on “Foundations of Data Science” when an AI agent can walk them through the same material, personalised to their exact gaps, in a fraction of the time? The guidebook assumes everyone takes the same trip. The guide (agentic AI) assumes you.

The shelf life problem

Micro credentials were always fighting a shelf life problem but agentic AI has made it terminal. Really. A 2026 UPCEA report found that institutional momentum around micro credentials has effectively flattened. Adoption rates have barely budged since 2021, and early optimism about revenue has cooled considerably. The credentials are not failing because nobody believes in them. They are stalling because the world they were designed for is disappearing.

The skills landscape now moves faster than any curriculum committee can approve a new module. By the time a micro credential on AI Literacy gets through quality assurance, the prompting techniques it teaches may already be obsolete. Meanwhile, an agentic AI tutor updates its knowledge continuously. It does not need a committee.

This is not a theoretical problem. A 22-year-old developer recently built an AI agent called Einstein (run on OpenClaw) that could complete entire online courses autonomously through Canvas. The tool was promptly shut down, but the point was already made. Also, with major LLMs (Claude Co Work, Manus Computer, Microsoft Agentic Co-Pilot, etc.) starting to include OpenClaw’s concept where AI can literally control your computer and browser, it begs the question: “If an AI can finish your micro credential for you, what exactly is the credential certifying?”.

What credentials cannot credential

Here is the uncomfortable truth. Micro credentials were always better at certifying exposure than competence. You completed ten hours. You passed the quiz. You watched the videos. But did you actually learn to do anything differently on Monday morning?

And let us be honest about what really drives most micro credential purchases. It is rarely the curriculum. It is the logo. Most credential courses sell not because of their quality but because of the institution’s branding and reputation. If you were to get certified in AI by Google or MIT, that badge carries weight not because the ten hours of content were necessarily superior, but because of the name stamped on it. A virtually identical course from a lesser-known polytechnic or university would struggle to fill seats. We are not buying learning. We are buying letterheads.

This is the hidden bargain at the heart of the micro credential economy. The institution lends its prestige, the learner borrows it for their social media profile or cv, and nobody asks too carefully whether any actual transformation took place in between. Agentic AI has no patience for this arrangement. It does not care whose name is on your certificate. It cares about whether you can work with it to solve a problem. The skill that matters now is not “I completed a course on AI.” It is “I can direct an AI agent to do something useful, and I can tell when it has gone wrong.” That is a fundamentally different kind of competence, and no quiz can measure it.

Employers are catching on, too. While Coursera reports that over 90% of employers prefer candidates with micro credentials on their CV, dig a little deeper and you find that nearly half of those same employers have concerns about the quality of what those credentials actually represent. The badge says “competent.” The interview reveals otherwise.

So what replaces them?

I am not suggesting we burn down the micro credential infrastructure overnight. For regulated professions and compliance training, they still serve a purpose. If you need proof that someone has been taught the fire evacuation procedure, a credential makes sense (well at least for now).

But for the fast-moving skills that most micro credentials now target? The replacement is already here. It is the AI agent itself, combined with portfolio evidence and demonstrated capability. Instead of a badge that says “I studied prompt engineering,” show me the AI-assisted project you built. Instead of a certificate in data analysis, show me the dataset you wrangled with an agent and the decisions you made along the way.

The future of professional learning is not another ten-hour course with a badge at the end. It is continuous, embedded, agent-supported learning that happens inside the work itself. The credential becomes the work. The proof becomes the output. As Ray Schroeder wrote for UPCEA, agentic AI is no longer merely a tool we talk to. It is a colleague that acts for us. And colleagues, unlike credentials, do not expire after twelve months.

The real redundancy

Micro credentials promised to make learning more flexible than degrees. And they did, for a while. But they kept the same underlying logic. Package knowledge into a unit. Deliver it. Assess it. Certify it. Agentic AI breaks that logic entirely, because the “packaging” step is no longer necessary when the agent can tailor everything in real time.

The institutions still investing heavily in micro credential development might want to pause and ask a harder question. Not “how do we make better credentials?” but “what are we actually trying to certify in a world where AI can learn the content faster than any human?”

Then the real redundancy is not the course or the degree. It is the credential itself. When learning can be generated, adapted and applied in real time, freezing it into a static badge says less and less about what someone can actually do.

That is what makes the micro-credential moment unstable. The expansion looks like progress, but it rests on an assumption that no longer holds. When the signal weakens enough, the bubble does not burst dramatically. It simply loses meaning.

And when it does, the pop will not be about supply. It will be about relevance.

A few weeks ago, I asked a class a question I genuinely didn’t know the answer to. Not a trick. Not a Socratic setup. Just a question that had been sitting with me, one I thought they might find worth turning over.

The virtual room went quiet. Not the productive kind of quiet where people are thinking. The nervous kind, where everyone is trying to work out what I wanted them to say.

Eventually, someone ventured a response. It was cautious and perfectly structured to avoid being wrong. I could almost hear the internal calculation behind it. Is this what he’s looking for?

It wasn’t. But then again, nothing was. That was the whole point.

Something has shifted in the way questions function in education. Somewhere along the way, we turned the question into a performance tool, a device for checking whether students have acquired the right content, using the right terminology. Questions in most classes now have correct answers. They have marking schemes. They have rubrics that define, in advance, what a good response looks like.

And in doing so, we’ve hollowed out the one thing that makes a question genuinely educational. Its openness.

Think of it like a door. A good question is supposed to open into a room you haven’t fully mapped yet. You walk in and you’re not sure what you’ll find. That uncertainty is the whole experience. But what we’ve done, increasingly, is bolt a sign to the door that reads “The answer is inside, second shelf on the left.” The student walks in, finds the thing and walks back out. Nobody lingers. Nobody looks around.

The interesting question, the kind that produces real thinking is one where the person asking it is also uncertain. It invites genuine exploration and not retrieval. It rewards the student who says something unexpected, not the one who mirrors the textbook.

But that kind of question is hard to assess. It doesn’t fit neatly into a rubric. It doesn’t generate standardised data. And so, gradually, we stopped asking it.

This didn’t happen overnight. It was a slow erosion driven by reasonable-sounding pressures. Accountability frameworks want measurable outcomes. External reviews want alignment between what was taught and what was tested. None of these things are evil. But collectively, they create an environment where the safest question to ask is the one you already know the answer to.

And students have adapted. They’ve become remarkably good at reading the room, at detecting what the lecturer wants, at producing the anticipated response with minimal risk. It’s a rational strategy. If the system rewards correct answers, why would you gamble on an interesting one?

The tragedy is that the interesting answer, the one that surprises even the person who gave it is where real learning lives. It’s the moment a student discovers they think something they didn’t know they thought. You can’t plan for that. You can only create the conditions for it to happen and those conditions start with a question that nobody has fully figured out.

Now with the increasing use of AI, I don’t think the solution is to abandon structure or throw out assessment. But I do think we need to be honest about the trade-off. Every time we design a question with a predetermined answer, we gain efficiency and lose possibility.

The interesting question is dying not because teachers have stopped caring but because the system around them has made it progressively harder to justify asking one. It doesn’t show up in module evaluations. It doesn’t align with intended learning outcomes.

It just produces thinking. Which, last I checked, was supposed to be the point.

In 1885, a German psychologist named Hermann Ebbinghaus sat alone in his study and memorised 2,300 nonsense syllables. DAX. BUP. ZOL. He tested himself obsessively, at intervals of minutes, hours, days, and weeks, charting precisely how quickly he forgot each one. The result was his famous forgetting curve: a steep, merciless cliff of memory loss that flattens out just when you think all hope is gone. Within twenty minutes, he’d lost over 40% of what he’d learned. Within a day, two-thirds had vanished.

Nearly a century and a half later, Ebbinghaus’s curve remains one of the most replicated findings in all of cognitive science. It has shaped everything from flashcard apps to corporate training schedules. And yet, here is the strange thing: whilst the entire learning technology industry has been obsessing over how humans forget, almost nobody in AI research is taking seriously the fact that large language models have their own forgetting problem, and it looks nothing like ours.

The machine version of amnesia

When researchers fine-tune a language model on a new task, something quietly catastrophic happens to its older knowledge. The model doesn’t gently fade, the way Ebbinghaus’s nonsense syllables did. It overwrites. Think of it less like a memory fading from a photograph and more like someone painting over a mural to make room for a new one. The old image isn’t dimmed; it’s gone. In the machine learning literature, this is called catastrophic forgetting, and the name is not an exaggeration.

What makes this interesting from a learning sciences perspective is not the technical problem itself (plenty of engineers are working on that) but the assumptions it exposes. For decades, the dominant metaphor in educational technology has been the brain-as-computer: memory as storage, learning as encoding, retrieval as search. If you’ve ever heard someone describe a student as having “bandwidth issues” or needing to “download” information, you’ve heard this metaphor at work. It was always a bit lazy, but it was at least directionally plausible.

Now the metaphor is running in reverse. We’re building computational systems that are supposed to learn, and suddenly the ways in which they don’t learn like us are more revealing than the ways they do.

What forgetting actually does for us

Here is something that rarely makes it into the edtech pitch deck: forgetting is not a bug. It is, in many respects, the most sophisticated feature of human cognition.

Robert Bjork, the UCLA memory researcher, has spent his career arguing that forgetting is what makes learning durable. His framework, the “new theory of disuse,” proposes that memories have two independent strengths: storage strength (how well embedded something is) and retrieval strength (how easily you can access it right now). The counterintuitive insight is that retrieval strength needs to drop, needs to forget, before re-learning can deepen storage strength. It’s like pruning a vine. You cut back to grow stronger.

This is why cramming works for Thursday’s exam but not for anything you’ll need in April. The retrieval strength was sky-high on Thursday morning. The storage strength barely budged.

Now compare this with how a large language model “learns.” There is no distinction between storage and retrieval. There is no beneficial forgetting. When a model is trained, everything is baked into the same vast soup of weighted connections, with no mechanism for something to become temporarily less accessible whilst remaining deeply encoded. The model either has a pattern or it doesn’t. There is no graceful middle ground.

The inconvenient implication for AI tutors

This matters enormously for anyone building AI-powered learning tools, which is to say, for a rather large and well-funded slice of the technology industry right now.

The pitch for AI tutors usually goes something like this: the system adapts to each learner, identifies gaps, provides personalised practice, and optimises the path to mastery. It’s compelling. It’s also built on an unexamined assumption that the AI understands what learning feels like from the inside, that it can model not just what a student knows but how that knowledge behaves over time.

But here’s the rub. A system that has never experienced the productive struggle of forgetting and relearning is being asked to orchestrate that very process for human beings. It’s a bit like asking someone who has never been hungry to design a meal plan. They might get the calories right, but they’ll miss something essential about the experience.

The best human tutors know, instinctively or through training, that a student who struggles to recall something and then succeeds has learned more deeply than one who never struggled at all. They know when to let a student sit with confusion rather than rushing to resolve it. They know that the feeling of almost remembering, that maddening tip-of-the-tongue state, is not a failure of the system but a feature of it. Desirable difficulty, to use Bjork’s term.

Current AI tutoring systems, by and large, treat any sign of student struggle as a problem to be solved immediately. The student pauses? Offer a hint. The student gets it wrong? Simplify. The student seems confused? Rephrase and repeat. The entire feedback loop is optimised for fluency, for making the learning experience feel smooth and frictionless. Which is precisely the wrong optimisation if you care about long-term retention.

The deeper design question

What would it look like to build AI learning systems that genuinely respected the science of forgetting? Not just spaced repetition algorithms (we’ve had those since the 1960s, and they’re fine as far as they go) but systems that understood when to withhold help, when to let a student flounder productively, and when to reintroduce material at exactly the moment retrieval strength has dipped low enough for relearning to bite.

This is a design problem, not just an engineering one. It requires moving past the assumption that the best learning experience is the smoothest one. It means building systems that are, by design, sometimes frustrating. And it means being honest about the fact that our most powerful AI models have a relationship with information that is fundamentally alien to the one human learners have.

There are some promising threads. Work on metacognitive scaffolding, where AI systems prompt learners to reflect on their own understanding rather than simply correcting them, is one. Research into “productive failure” pedagogies, where students are deliberately given problems they cannot yet solve before instruction, is another. But these approaches remain at the margins of a market that overwhelmingly rewards engagement metrics and completion rates over genuine learning outcomes.

Where this leaves us

The learning sciences have spent decades accumulating genuinely useful knowledge about how human memory works, how forgetting serves learning, and why the intuitions of both students and teachers about “what works” are so often exactly wrong. This knowledge exists. It is robust. It is largely being ignored by the people building the next generation of AI learning tools.

This isn’t because the builders are foolish or careless. It’s because the incentive structures of the edtech market reward the appearance of learning over its reality, and because the AI systems themselves, having no experience of forgetting, have no built-in reason to respect its role.

The fix isn’t to make AI systems forget like humans. That’s neither possible nor particularly useful. The fix is to build systems designed by people who understand that learning is not the accumulation of correct answers but the slow, frustrating, recursive process of encoding knowledge deeply enough that it survives the passage of time. And to measure accordingly.

Of course, there's a irony here that borders on the poetic. I've just spent a thousand words or more explaining why forgetting is essential to learning, published on a platform engineered to maximise the chance you'll scroll past this and never think about it again. If Ebbinghaus's curve holds, you'll retain about a third of this by tomorrow morning. But if you struggle to remember the details next week, and then come back and reread it, you'll know more than you did the first time.

That's not a flaw in your attention. That's your brain doing exactly what 140 years of research says it should. You're welcome.

Further Reading

Ebbinghaus, H. (1885). Über das Gedächtnis: Untersuchungen zur experimentellen Psychologie. Leipzig: Duncker & Humblot. Translated as Memory: A Contribution to Experimental Psychology (H.A. Ruger & C.E. Bussenius, Trans., 1913). Teachers College, Columbia University.

Bjork, R.A. & Bjork, E.L. (1992). A new theory of disuse and an old theory of stimulus fluctuation. In A. Healy, S. Kosslyn, & R. Shiffrin (Eds.), From Learning Processes to Cognitive Processes: Essays in Honor of William K. Estes (Vol. 2, pp. 35–67). Erlbaum.

Bjork, E.L. & Bjork, R.A. (2011). Making things hard on yourself, but in a good way: Creating desirable difficulties to enhance learning. In M.A. Gernsbacher, R.W. Pew, L.M. Hough, & J.R. Pomerantz (Eds.), Psychology and the Real World: Essays Illustrating Fundamental Contributions to Society (pp. 56–64). Worth Publishers.

Kapur, M. (2008). Productive failure. Cognition and Instruction, 26(3), 379–424.

McCloskey, M. & Cohen, N.J. (1989). Catastrophic interference in connectionist networks: The sequential learning problem. In G.H. Bower (Ed.), The Psychology of Learning and Motivation (Vol. 24, pp. 109–165). Academic Press.

We tend to talk about AI as though it arrived like a storm, sudden, disruptive, impossible to ignore. But for most of us in education, the change has been more like a tide. Gradual, persistent, and by the time you really notice it, already quite deep.

For years, the internet quietly reshaped what it meant to know something. Why memorise the dates of the French Revolution when they’re a search away? We adjusted imperfectly, grudgingly but usefully. We shifted our emphasis towards critical thinking, synthesis, application.

And then, almost without announcement, the bar moved again.

Now students don’t just have access to information. They have access to articulation. To reasoning. To something that, on the surface, looks remarkably like thought. The question of what we’re actually teaching and why has become genuinely urgent in a way it hasn’t been for a very long time.

What Are We For, Exactly?

I’ve been an educator long enough to have heard every iteration of the existential question. Will technology replace teachers? We heard it about calculators, about the internet, about MOOCs promising to democratise learning at scale. And each time, the answer came back the same: no, not really, because teaching is fundamentally human.

I still believe that. But I think we’ve been a little lazy about what it means.

The human part of teaching was never really about information delivery. A textbook delivers information. A lecture, if we’re honest, often does little more. What teachers do that nothing else can quite replicate is witness. We see the student who is almost there. We notice the one who’s given up. We read the room. We know when to push, when to wait, when to say I don’t know either, let’s find out together.

What AI is surfacing, uncomfortably but usefully, is how much of what we thought was teaching was actually just... content transfer. And content transfer, it turns out, can be done faster and cheaper elsewhere.

This is not a comfortable thought. But I think it’s a necessary one.

The Skills Nobody Can Automate

When I think about what I’m genuinely trying to cultivate in my students or what I hope they carry with them long after they’ve forgotten the specific content of my lessons, it’s a particular kind of aliveness to ideas.

The willingness to sit with uncertainty. The discipline to question your own assumptions. The confidence to hold a view and the intellectual honesty to revise it. The capacity to be genuinely moved by an argument, a poem, a piece of evidence.

None of that can be generated. It has to be grown.

And growing it requires the presence of another person who models it, who demonstrably doesn’t know everything, who thinks out loud, who changes their mind in front of you, who takes your half-formed ideas seriously enough to push back on them.

That, I’ve come to believe, is the core of what we do. Not the syllabus. Not the assessment rubric. The relationship between a mind that is still forming and one that has been formed enough to help.

Learning to Ask Better Questions

Perhaps the most honest thing I can say is this: AI has made me a better teacher, not because I’ve learned to use it cleverly but because it has forced me to interrogate every assumption I had about why I was doing what I was doing.

Why do I ask students to write essays? Is it because they need to know how to write essays, or because the act of composing an argument is how you come to actually know what you think? If it’s the latter, then the output matters far less than the process. And if students are outsourcing the process, they’re not cheating me, they’re cheating themselves.

That framing has changed how I design tasks. I’ve moved away from products submitted, towards more visible processes. Drafts with reasoning. Conversations about choices. Presentations where students have to defend not just what they concluded but how they got there.

The marking is harder. The learning, I think, is more real.

A Thought for Fellow Educators

If you’re feeling unsettled by all of this, and most of the educators I know are, even the ones who won’t say it aloud, I’d offer this small reassurance: the unsettlement is appropriate.

We are living through a genuine shift in what it means to learn, to know, and to teach. Pretending otherwise or retreating behind policy documents about academic integrity, won’t serve our students. What will serve them is teachers who take the disruption seriously and keep placing human connection and genuine thinking at the centre of what they do.

We didn’t sign up to compete with machines. But we also don’t need to.

There's a question that has nagged at educators and learning scientists for decades: if students know that highlighting and re-reading aren't particularly effective, why do they keep doing it?

The usual answer is somewhere between resigned and a little unkind: students are poorly informed or simply making bad decisions. A paper by Krause et al. (2025) published in the Educational Psychology Review offers a rather, I would argue, more accurate explanation. The problem isn’t ignorance. It isn’t laziness. It’s habit.

What the Research Actually Found

The study, led by researchers examining self-regulated learning through the lens of habit theory, does something genuinely interesting. It imports a well-established framework from behavioural psychology (habit formation) into the conversation about study strategy use.

The researchers argue that ineffective strategies like re-reading and highlighting are precisely the kind of behaviours that should become habitual over time. Think about it. They are used repeatedly in stable contexts (a desk, a highlighter, a textbook), they feel fluent and effortless, they demand almost nothing from working memory, and they deliver a satisfying but ultimately misleading reward, namely the sensation of having “covered” the material.

Self-testing and concept mapping, by contrast, are cognitively demanding and produce no immediate sense of progress. They are, in short, terrible candidates for habit formation.

To test this, the team ran a proof-of-concept study measuring the habit strength of different study strategies using a validated habit index. The results were striking. Highlighting and re-reading scored substantially higher on habit strength than retrieval-based strategies. Students weren’t choosing to re-read over self-testing in any meaningful sense. They were simply doing what the situation automatically triggered in them.

Why This Distinction Matters

It might seem like a subtle difference, you know the typical “bad habit” versus “bad choice” kind of mantra, but it has fairly significant implications for how we teach, and how we support learners.

If the problem is metacognitive ignorance, the solution is information. Educators are expected to perhaps start telling students which strategies work, show them the evidence and trust them to act on it. This approach has been tried extensively. It doesn’t work especially well.

If the problem is habit, the solution looks quite different. Habits are notoriously resistant to change through instruction alone. They are cued by context. They operate below the level of deliberate decision-making and they are particularly likely to resurface under stress or time pressure. This is precisely the conditions under which students most need their study strategies to be good ones.

What actually changes habits, the research suggests, is disrupting the cue-routine-reward loop. That means changing the study environment, building implementation intentions (e.g., “when I sit down to revise, I will do a brain dump before opening my notes”), and making retrieval practice feel more immediately rewarding, This can be done perhaps through progress tracking or low-stakes quizzing that gives learners a clear sense of what they actually know.

A More Compassionate Frame

There is something important in this reframing beyond its practical implications. It asks us to look at a student who is highlighting their way through a textbook the night before an exam and see not a failure of will or wisdom, but someone acting on deeply ingrained patterns formed over years of schooling, patterns that, quite frankly, most of their teachers probably modelled for them.

The intention–behaviour gap in learning is real and well-documented. Students consistently report intending to use better strategies and consistently fail to follow through. Previous accounts of this gap have leaned heavily on motivational explanations. Students don’t want to feel the discomfort of retrieval practice. The habit account doesn’t dismiss motivation, but it adds something that I think is truer to lived experience. Sometimes people do want to change, understand why they should change and still find themselves reaching for the highlighter.

What This Means for Educators

A few practical threads worth pulling on from this research:

Embed retrieval into routine. If the problem is that re-reading has become the default, automatic response to a study context, one of the most powerful things teachers can do is make retrieval practice the structural default. The thing students do before they are given the option to re-read. Start lessons with low-stakes quizzing. Begin revision sessions with a blank-page recall exercise. Make the effective strategy the path of least resistance.

Don’t just tell students what works, change the context. Information campaigns about evidence-based study techniques are not useless but they’re unlikely to displace deeply habituated behaviours on their own. Environmental redesign such as changing where and how students begin their studying, may also be more effective than exhortation.

Be particularly alert around high-stress moments. The study notes that habitual strategies are especially likely to resurface under stress or time pressure. The weeks before exams are precisely when students most reliably revert to highlighting and re-reading. Proactive scaffolding in those periods (for example, structured revision guides and timetabled retrieval sessions) can help keep better habits in place when it matters most.

A Final Thought

What strikes me most about this paper is how much credit it gives to students. It doesn’t explain their behaviour by pointing to a deficit in knowledge or character. It explains it by pointing to the cumulative logic of how habits form. A process that is, in many ways, a testament to the brain’s efficiency rather than its failure.

That’s a more complicated story. But it’s probably the right one. And if we take it seriously, it should reshape not just how we advise students but how we design the learning environments in which habits are formed in the first place.

Source:

Krause, A. K., Breitwieser, J., & Brod, G. (2025). Understanding learning strategy use through the lens of habit. Educational Psychology Review, 37, 109. https://doi.org/10.1007/s10648-025-10087-3

The modern educational technology sector is currently intoxicated by a singular and highly marketable vision. Developers promise a personalised, flawlessly patient artificial intelligence tutor residing in the pocket of every student. We are offered a tireless algorithmic companion ready to explain complex calculus at three in the morning or gently correct a structural misstep in a history essay.

However, a profound structural flaw emerges. We are actively confusing a permanent cognitive crutch with a temporary pedagogical scaffold.

The Misunderstood Architecture of Scaffolding

To understand this trap, we must revisit the foundational concept of the Zone of Proximal Development. Introduced by the psychologist Lev Vygotsky, this zone represents the highly productive space between what a learner can achieve entirely independently and what they can achieve with expert guidance.

The critical instructional mechanism within this zone is scaffolding. Here is the inconvenient truth that the technology sector seems to have forgotten. A scaffold is defined entirely by its impermanence. A master builder does not leave the wooden framework attached to the cathedral forever. True pedagogical scaffolding requires a deliberate and meticulously timed process known as fading. The support must be systematically withdrawn so the structural integrity of the student’s own intellect can bear the weight of the task.

The Curse of Algorithmic Impatience

Herein lies the trap of the infinite AI tutor. Generative AI models are fundamentally designed like customer service agents. They are programmed to be flawlessly helpful, completely frictionless and instantly responsive. Crucially, they possess absolutely no pedagogical patience.

When a student encounters a moment of genuine cognitive friction, which is the exact friction required to build permanent neural pathways, the AI leaps in to smooth the road. The algorithmic tutor never initiates the fading process. It simply becomes a permanent, externalised lobe of the student’s brain. The learner is kept in a state of perpetual and comfortable dependency. They are entirely robbed of the panic and the subsequent triumph of solving a complex problem entirely alone.

The Illusion of Mastery

When a student uses an infinite tutor, their performance metrics often look spectacular. Homework is completed flawlessly and always on time. Yet, this represents mere performance rather than actual learning.

Because the AI absorbs the heavy lifting of executive function and metacognitive planning, the student internalises very little. The moment the algorithmic crutch is removed in a secure examination environment, the illusion shatters entirely. The cathedral collapses because the internal pillars were never actually built.

Designing for Desirable Difficulty

If we want these systems to support learning rather than outsource it, we need a redesign that takes desirable difficulty seriously. A useful AI tutor should not behave like an omniscient oracle. It should behave like a stingy interlocutor.

It should privilege responses over answers and insist on student work before it offers its own. It should introduce friction on purpose such as time delays that interrupt impulsive asking, hints that arrive in stages rather than as a complete solution and occasional refusal when the request is clearly a shortcut around thinking.

Most importantly, it should fade. Early interactions can be supportive, even generous, but the system must progressively withdraw and verify independence, not by asking “Do you understand?” but by requiring the learner to reconstruct and explain without assistance.

All in all, we do not need a tutor that makes learning feel effortless. We need one that makes effort count.

It is tempting to treat education like a league table, one cohort up, the next one down, with a single culprit to blame. So when a report claims Gen Z is the first generation to perform worse than its predecessors because of “AI dependency”, it offers an easy narrative: new technology arrives, thinking declines, case closed. The trouble is that this kind of story usually compresses messy evidence into a simple moral and that compression is where accuracy tends to break.

Part of the problem is that the material often cited for these headlines is not actually about “AI dependency” in the generative AI sense. A lot of recent coverage traces to policy discussions about youth, screens and learning, including a U.S. Senate hearing in January 2026. The argument in that lane is broader. Academic development in parts of the developed world has stalled or reversed and increased screen use in schooling could be a contributing factor. That is not the same as proving students are cognitively weaker because they rely on chatbots. It is a claim about systems, exposures and correlations, which gets repackaged as a generational verdict.

Then there is the category mistake. “Gen Z” is a label, not a controlled study group. Within any age cohort, differences in schooling, family circumstances, language background, health, sleep and stress are enormous. Averages move, but the spread is far bigger than the gap between generations. When a story treats millions of young people as a single learner, it loses the reality that matters most.

Causality is the next leap that rarely gets earned. You can find associations between high screen use and lower performance without screens being the cause. Device use can be a proxy for under-resourcing, policy mandates, curriculum changes, assessment shifts or opportunity costs such as less sustained reading. If you cannot isolate the mechanism and the counterfactual, you do not get to claim “technology did it”, and swapping in “AI” does not fix that.

The phrase “AI dependency” also hides vagueness. Dependency on what, exactly? Autocomplete, recommendation feeds, video explainers, generative writing tools, algorithmic tutoring? These are different behaviours with different cognitive demands. Some uses reduce workload in a helpful way. Others become a crutch. If the claim cannot specify the exposure, it cannot be tested properly, so it should not be used to rank an entire cohort.

There is another omission in many of these stories. Gen Z is navigating real mental health pressures that complicate any neat “they are weaker” verdict. Many are growing up amid economic uncertainty, high academic competition and the always-on social comparison of networked life. Stress, anxiety, low mood and sleep disruption do not just affect wellbeing, they affect attention, memory consolidation and motivation, which are upstream of academic performance. Even when technology plays a role, it often interacts with these pressures rather than acting as a single, standalone cause.

Why does this framing spread anyway? Because it is tidy. It offers a villain and it turns complex educational issues into a moral story about “kids these days”. It also encourages intergenerational blame while letting systems off the hook. Schools, curricula, assessment regimes and procurement choices are made by adults. If outcomes are slipping, the serious response is not to sneer at students. It is to examine design decisions.

A more useful conversation starts with tighter questions. Which outcomes are we worried about? Reading duration, writing quality, mathematical fluency, sustained attention? Under what classroom conditions do those outcomes improve? Where does technology replace high-value practice and where does it support it? If AI tools are involved, are students being taught how to use them as aides rather than substitutes, with checks for understanding and verification?

Generational comparisons make for clickable headlines. They do not make better education. If we want improvement, we should talk less about who is “worse” and more about what environments produce strong learning, with or without new tools.

NotebookLM x Gemini Gems is a strong pairing because each tool does a different job well. NotebookLM is best treated as an evidence-first workspace where every claim can be traced back to what you uploaded or pulled in. Gemini Gems are better viewed as reusable production workflows. You create in your standards once, then reuse them whenever you need output in the same format.

A simple mental model

NotebookLM collects and distils. Your Gem manufactures but only from what NotebookLM has already grounded.

Here are 10 things you can do with this combo that most people do not think to try yet.

1. Create teaching materials fast but traceably

This is your core idea and it is the most practical starting point. In NotebookLM, you extract the concepts, the definitions, the boundary conditions and a few good examples from your readings or textbooks. NotebookLM is designed around analysing your sources and producing grounded outputs.

Then you use a Gem as a lesson-material factory that turns that extracted bundle into lesson plans, activities, short quizzes, rubrics and feedback comments, while forcing each element to point back to a concept ID or cited excerpt.

2. Build a “concept bank” once, then reuse it across the whole semester

Instead of creating materials week by week, use NotebookLM to create a concept bank for the entire module. You keep it as a living notebook that grows as you add new papers, lecture notes or slides. NotebookLM has been expanding the source types it can handle, including formats like Word docs and Sheets, so the bank can be genuinely comprehensive.

Your Gem then becomes the consistent formatter: every week, you feed the relevant slice of the concept bank and it outputs the week’s artefacts with the same structure and instructional voice.

3. Generate quiz items that target misconceptions, not just recall

Ask NotebookLM to surface common confusions in the sources, such as ambiguous terms or near-synonyms used differently by authors. NotebookLM can also generate study aids like quizzes that are grounded in your documents.

Your Gem can convert each misconception into distractors, then generate feedback tied to that misconception. This is how you get assessment items that diagnose, rather than just score.

4. Produce rubrics with anchors pulled from exemplars

Rubrics often fail because descriptors are vague. Put a handful of student exemplars or model answers into NotebookLM, then ask it to highlight what makes each response strong or weak, citing the text.

Your Gem turns that into a rubric with criteria that are explicitly anchored to the exemplars. It can also generate short marker comments aligned to each band so grading stays consistent.

5. Create “same outcome, different route” activities for differentiation

Use NotebookLM to map prerequisites and extension directions using your sources. Then tell your Gem to generate two versions of the same activity: one with scaffolds and one with stretch constraints. A useful rule is that the Gem must not introduce concepts outside the provided concept list. That keeps differentiation within scope, rather than quietly teaching next week’s content.

6. Build a spiral review engine that respects dependencies

NotebookLM is good at helping you see how concepts relate across documents, especially when you load multiple weeks of readings.

Your Gem can then generate spaced retrieval prompts that deliberately reactivate earlier concepts in a way that supports the current topic. This ends up feeling coherent to students because the review questions are structurally connected to what they are doing now.

7. Convert research notes into publishable teaching narratives

NotebookLM can produce different report-style outputs grounded in your material, including formats like study guides and blog-style writing.

Feed that structured narrative to your Gem and have it rewrite into different teaching voices: a lecture script, a tutorial handout, a lab brief or a case discussion prompt, all still anchored to the same evidence.

8. Create assessment feedback banks that are specific

In NotebookLM, collect a set of common error patterns and the “repair moves” you want students to practise.

Your Gem then generates a feedback bank tagged by concept and error type. The key is to require one concrete next step plus a reference back to the relevant concept label, so comments feel targeted and teachable.

9. Keep a change log when sources update

When you add new readings or revise slides, you want to know what changed conceptually. NotebookLM has introduced “Deep Research” style workflows and broader ingestion options, which makes it easier to update a notebook as your source set evolves.

Your Gem can regenerate only the affected materials and automatically append version notes, so you maintain a proper instructional audit trail.

10. Rapid Qualitative Coding Assistant

In NotebookLM, load your interview transcripts. Ask it to spot themes, repeated ideas and useful quotes. Then create a Gemini Gem to draft a starter codebook, code two transcripts, flag unclear parts and suggest better code definitions.

An example

1. Go to https://gemini.google.com/gems/create

2. In the knowledge files selection, choose NotebookLM, then pick the notebook that you have there that you would like to use.

3. Then in the instruction part, provide your step-by-step instructions to Gemini. E.g.

   You are a teaching-materials generator.
   
   Input I will provide from attached notebook from NotebookLM:
   A) Concept list, each with a Concept ID and supporting cited excerpt(s)
   B) Example bank with Example IDs and citations
   C) Teaching context (level, duration, class size, constraints)
   
   Output you must produce:
   1) Lesson plan (objectives, sequence, timing, checks for understanding)
   2) 2 to 4 activities aligned to concept IDs
   3) 8 to 12 quiz items with answer key and misconception-tagged distractors
   4) A rubric with performance descriptors and anchors
   5) A feedback comment bank tagged by concept ID and common error type
   6) Source trace table mapping each output element to concept IDs and example IDs
   
   Rules:
   - Do not introduce new factual claims beyond the provided excerpts.
   - Every activity, quiz item and rubric criterion must cite at least one concept ID.
   - If an input is missing, ask for it instead of guessing.
   - Use British spelling.
   

4. Then name your Gem before saving.

5. Start using it!

There’s a particular feeling when you realise someone in charge has stopped seeing clearly. It starts as a small gap between what they’re saying and what you’re experiencing on the ground. You notice it in meetings when the narrative doesn’t quite match the numbers. In the way certain questions get deflected rather than answered. In the growing distance between the story and the reality.

At first, you think maybe you’re missing something. They have more context, more information, a bigger picture you can’t see. So you trust. You defer. You tell yourself the discomfort is just your own limited view.

And sometimes that’s true. Sometimes leaders really do see around corners.

But sometimes they don’t. And learning to tell the difference has been one of the more painful educations for most of us.

Here’s the confession I don’t hear people make often enough: I’ve let myself be misdirected. Not because I was naive or stupid but because I wanted to believe. Because the vision was exciting. Because certainty feels better than ambiguity, even when it’s someone else’s certainty.

There’s something “seductive” about a confident leader. They offer clarity in a world that doesn’t provide much of it. They give you a story to be part of, a direction to move in, a reason to show up. And when that story starts to crack, there’s a part of you that would rather not look too closely.

I’ve ignored my own instincts because contradicting the narrative felt disloyal. I’ve stayed quiet in meetings on certain occasions when I knew something was off or whenever I want to say something, my voice wasn’t given any “weight”. I’ve watched colleagues raise concerns and get subtly sidelined for it too, and I’ve learned the lesson that was meant to teach.

That’s on me. Not entirely, but partly. Misdirection is a two-person dance.

The hardest thing to admit is that I’ve done this to myself too. You don’t need to be a leader to build a delusional narrative. You just need to be human.

I’ve told myself stories about jobs that weren’t working, relationships that had run their course, paths I was on that weren’t leading anywhere good. I’ve rounded up my 60% confidence to 90% because I didn’t want to face the alternative. I’ve mistaken stubbornness for commitment and wishful thinking for intuition.

Vision, delusion, misdirection. These aren’t leadership problems. They’re human problems. The corner office just gives them more consequence.

What I’ve started doing differently (and I’m still learning this, honestly) is paying closer attention to my own resistance. When I don’t want to hear something, that’s usually a sign I need to hear it. When I find myself defending a position rather than examining it, something’s off. When the story I’m telling myself requires me to dismiss too many inconvenient facts, the story is probably the problem.

I’ve also gotten better at recognising it in others. Not to judge them but to protect myself. The leaders who can’t say “I don’t know.” The ones who treat questions as threats. The ones whose confidence never wavers no matter what the data shows. These are warning signs now, not reassurances.

There’s a kind of power in being the one without power. You can see things that the people at the centre can’t. You’re closer to the ground, closer to the reality, less invested in the official story being true.

But that power only means something if you’re willing to trust what you see. If you’re willing to name it, at least to yourself. If you’re willing to act on it, even when acting means leaving, or speaking up, or simply refusing to pretend along with everyone else.

I’m trying to be more honest now. With myself first, and then with the situations I find myself in.

Vision, delusion, misdirection. From where I sit, they often look the same until it’s too late. The best I can do is stay curious, stay sceptical and remember that my own vision is just as susceptible to distortion as anyone else’s.

The people at the top don’t have a monopoly on self-deception. They just have a bigger audience for it.

The rest of us are performing the same play, just on a smaller stage.

I used to treat personalisation as an obvious decision in designing any learning experience. If a learner struggles, give them an explanation that fits their background, their interests, their level. That logic sounds “inclusive”. Then I watched what actually happened when novice learners used AI tutors heavily. They were not simply receiving help. They were collecting multiple versions of the same idea, each one persuasive on its own yet inconsistent as a whole. Their confidence went up. Their ability to solve a fresh problem did not.

This post is not an argument against personalisation. It is a warning about a specific failure mode: when “tailored” content fragments a novice’s understanding. If you teach foundational concepts, you have probably seen some version of this already, even before generative AI became normal.

What personalisation means in the AI era

In practice, AI personalisation usually shows up in three ways.

First, pacing personalisation, in which the learner goes faster or slower, revisits a step, repeats a quiz, requests more practice. This is usually low risk.

Second, explanation personalisation. The AI model rewrites the concept using different wording, different metaphors, different examples or a different assumed prior knowledge base. This is where many novices get into trouble.

Third, pathway personalisation. The AI model changes the learning route, suggesting alternative sequences, skipping steps, introducing extra topics or “helpful” extensions. This can be beneficial for advanced learners. For novices, it often becomes a detour that feels like progress.

The danger is that novices do not yet have a stable mental model to anchor these variations.

Why novices are uniquely vulnerable

A novice’s understanding is still being assembled. That matters because the learner cannot reliably tell what is central, what is optional, what is an assumption and what is a convenient story.

When you ask an AI tutor to “explain differently”, it will often oblige by changing the framing. That is not wrong. But a novice may not notice that the new framing shifts emphasi or glosses over a key constraint.

There is also a cognitive load issue. Novices tend to have limited working memory for unfamiliar material. More choices, more versions, more optional side paths can produce a feeling of richness while quietly increasing confusion.

Finally, AI text is fluent. Fluency can be mistaken for correctness and correctness can be mistaken for understanding. A novice may read (if they read at all) a smooth explanation and feel “I get it”, then fail on the first transfer problem that requires applying the idea in a new context.

Four ways tailored content harms novices

1. Choice overload disguised as support
   A learner asks: “Explain it simply.” The AI provides a neat summary. Then the learner asks: “Give me another example.” Then: “Use an analogy.” Then: “Explain in Malaysian context.” Each response is plausible and well-written. Yet the learner now holds four surface-level entry points with no shared structure connecting them. Instead of a single coherent model, they have a set of disconnected explanations. Although Yin et al. (2024)’s study found that the use of ChatGPT reduces cognitive load, they did specify it was more for complex tasks rather than fundamental concepts.

2. Structural drift
   Analogies and examples are not neutral. Each one highlights some features and hides others (Gray & Holyoak, 2021; Harrison & Treagust, 2006). A physics analogy might suggest a flow-like intuition. A social example might suggest a rules-and-roles intuition. A novice may not realise that these frames are not interchangeable. They are lenses. Switching lenses too quickly can stop the learner from building the underlying structure.

3. Micro-hallucinations and near-miss explanations
   Even when an AI is mostly correct, small errors matter such as a missing condition or a mislabelled step. An expert can catch this because the expert has a “reference model”. A novice cannot. Worse, near-miss explanations are more dangerous than obviously wrong ones because they sound reasonable and are hard to detect.

4. Premature removal of productive struggle
   AI makes it easy to avoid the work that builds durable learning: retrieval, error correction, self-explanation. If a learner asks for help at the first sign of discomfort, the AI can keep them comfortable but shallow (Bjork & Bjork, 2011). The learner becomes dependent on immediate clarity, and clarity becomes the goal. Transfer and fluency become accidental.

A vignette you may recognise

Consider a student in a research methods module trying to understand correlation versus causation.

They ask AI: “Explain correlation and causation.”
They get a standard explanation.

They then ask: “Explain like I’m 12.”
They get a simplified story with a fun analogy.

They then ask: “Give me an example from education.”
They get a scenario about study time and grades.

They then ask: “Give me an example from social media.”
They get a scenario about screen time and anxiety.

All of this feels productive. The student reads quickly and even smiles. Then you give them a question that requires distinguishing confounds and proposing a stronger design. They cannot. They have stories, not a model. They can recite “correlation is not causation”, but they may not be able to reason through what would count as evidence.

What failed here was not motivation. It was coherence. The learner never consolidated a stable structure before collecting variations.

What to tell students: a simple AI-use protocol

You can teach a short protocol that reduces confusion.

1. Do not ask for five different explanations upfront. Start with the course explanation.

2. When you ask AI to help, constrain it: “Rephrase without changing structure.”

3. Demand assumptions and limits. Ask: “What are you assuming? What would make this wrong?”

4. Do a brief teach-back in your own words. If you cannot teach it, you do not own it.

5. Attempt an unseen question before you ask for more help. If you always ask first, you never find out what you can do alone.

Some examples:

Prompt 1: Rephrase without drifting
“Explain [concept] using the same structure as this outline: [paste 4–6 bullet outline]. Use simpler wording but keep the sequence and key terms. After explaining, ask me two diagnostic questions that test whether I truly understand.”

Prompt 2: Concept check then adapt
“Before you explain [concept], ask me three short questions to identify what I already understand and what I do not. Then explain only the parts I missed, using one example from [my domain]. Keep it under 180 words and include one quick practice question.”

Prompt 3: Analogy with explicit limits
“Give one analogy for [concept]. Then provide a mapping table: analogy element → concept element. Finally, list two ways the analogy breaks down or could mislead me.”

Prompt 4: Guard against near-miss errors
“Explain [concept] and then list three common misconceptions or mistakes students make about it. For each misconception, show how it appears in a student answer and how to correct it.”

Prompt 5: Transfer test
“Give me one new problem that uses [concept] in a different context. Do not solve it. Wait for my attempt, then give feedback and only one hint at a time.”

A balanced conclusion

For novices, the biggest gift is not infinite explanations. It is a stable structure, repeated in consistent language and reinforced through practice. Once a learner has that structure, personalisation becomes powerful. Before that point, it can turn learning into a collection of fragments.

The practical move is to design an anchor explanation and treat personalisation as a layer that must respect the anchor. Students must learn to treat AI like a tutor who can clarify and coach, but only after they commit to one coherent model.

That is the difference between personalised learning and personalised confusion.

References

Bjork, E. L., & Bjork, R. A. (2011). Making things hard on yourself, but in a good way: Creating desirable difficulties to enhance learning. In M. A. Gernsbacher, R. W. Pew, L. M. Hough, & J. R. Pomerantz (Eds.), Psychology and the real world: Essays illustrating fundamental contributions to society (pp. 56–64). Worth Publishers.

Gray, M. E., & Holyoak, K. J. (2021). Teaching by analogy: From theory to practice. Mind, Brain, and Education, 15(3), 250-263. https://doi.org/10.1111/mbe.12288

Harrison, A. G., & Treagust, D. F. (2006). Teaching and learning with analogies: Friend or foe?. In Metaphor and analogy in science education (pp. 11-24). Dordrecht: Springer Netherlands.

Yin, J., Goh, T. T., & Hu, Y. (2024). Using a chatbot to provide formative feedback: A longitudinal study of intrinsic motivation, cognitive load, and learning performance. IEEE Transactions on Learning Technologies, 17, 1378-1389. https://doi.org/10.1109/TLT.2024.3364015

We often confuse education with schooling, as if they are the same thing. They are not and this confusion shapes many of our expectations and frustrations.

Schooling is a formal system. It has structures, timetables, curricula, assessments and credentials. Education, by contrast, is a much broader human process. It involves how people make sense of the world, develop judgement and learn how to keep learning. Schooling can support education but it cannot contain it.

When we collapse education into schooling, we begin to burden schools with responsibilities they were never designed to fulfil. We expect them to fix social inequality, cultivate moral character, resolve motivation issues and compensate for every gap in home and community life. When these expectations are not met, disappointment follows, often directed at teachers or curricula rather than at the flawed assumption beneath them.

This collapse also narrows how we define learning. Achievement becomes synonymous with grades, test scores and certificates. Curiosity, adaptability and the capacity to learn independently are treated as secondary because they are harder to quantify. Learning that occurs through work, failure, play, caregiving or self-directed exploration is undervalued simply because it does not align with school metrics. What cannot be easily measured is too often treated as if it does not matter.

Recognising the distinction between education and schooling does not mean diminishing schools. Schools matter enormously. They provide access and opportunity at scale. They are often the most reliable learning spaces many children will ever have. For some learners, schools are a refuge as much as an institution. Acknowledging their limits is not a critique. It is just a form of respect.

We just have to accept the fact education does not begin at the school gate, nor does it end with a certificate. It begins long before formal instruction and continues throughout life, shaped by families, friends, workplaces, technologies and culture. When we remember this, we can have more honest conversations about what schools can realistically do and where responsibility must be shared across society.

This shift in perspective also changes the questions we ask. Instead of obsessing over how to optimise systems, raise rankings or fine-tune exams, we begin to ask how learning actually happens. We ask how people develop understanding over time and how institutions can support rather than constrain these processes.

If we want better education, we need to stop treating schooling as its proxy. Schools are a vital part of the learning ecosystem but they are not the ecosystem itself. The moment we recognise this, our debates become less defensive and more productive.

We move away from blaming institutions for failing to do everything and towards designing environments, inside and outside schools, that genuinely support how people learn.

That, ultimately, is where education begins.