What future for teachers, in the age of AI?

The question keeps coming back, in staff rooms and in corridors alike: if an AI can explain, generate exercises, mark, re-explain with infinite patience and at three in the morning, what is the teacher still for?

I am unusually well placed to ask it. I teach computer science in an engineering school: algorithmics, object-oriented programming, software architecture, software engineering. In other words, precisely the subjects generative AI is upending the fastest. What I teach my students on Monday, a model can produce for them on Tuesday, in three seconds. And on my side of the desk, I am no better: I lean heavily on AI to prepare my materials, my exercises, my exam topics. I go faster, further. And yet my workload hasn't shrunk, and I sometimes have the strange feeling of having become the quality controller of an assembly line that runs without me.

And yet, let me say it straight away: the question doesn't frighten me all that much. I have a small, rather unexpected advantage over it that few of my colleagues share, and which I will reveal later in this piece. Patience.

I want to take the question seriously, then, without catastrophism or denial. Because the two reflex answers are both wrong. "Nothing changes, AI is just one more tool": wrong. When the tool does precisely what used to be the visible heart of the job, something changes. "The teacher is doomed": also wrong, but only once you understand what really was the heart of the job, and what was merely its wrapping.

What AI actually eats

Let us be clear-eyed about what is disappearing: the part of the job that consisted in being a channel for transmitting information. Explaining a concept, providing an example, answering a factual question, producing a handout, spinning an exercise into ten variants: on all of this, a model is faster, more available, often clearer, and infinitely more patient.

We have to be honest enough to admit it: part of our courses has always been information transmission dressed up as teaching. That part, yes, AI eats.

And there is a symptom every teacher will recognise: silence. Students used to ask questions. It was noisy, disorderly, sometimes wildly off the mark, but it was gold: every question told me where the cohort stood, what had landed, what was still resisting. Now we no longer know. Silence has settled in, and it is not the silence of understanding. The question hasn't disappeared, it has migrated: why risk looking like the class idiot in front of forty people, when you can ask a model in the evening, alone, without judgement?

I recently got a demonstration of this that, I confess, annoyed me considerably. I put a question to the class, out loud, to get them thinking. I watch students type it straight into an AI, right in front of me. And a few seconds later, one of them gives me the answer back. Or rather, no: he reads it to me. The loop was closed, and I had become, in my own classroom, the voice interface of a question-and-answer system. Yet the question I was asking wasn't there to get an answer: I know the answer. It was there to make them think. That is the whole misunderstanding of our times: they optimised getting the answer, when the answer was never the point.

That is where the real damage lies. The problem is not that my students no longer understand. The problem is that I no longer know what they don't understand. Their questions were my probe, my navigation instrument. That instrument is being unplugged.

But a profession does not reduce to its automatable part. The doctor did not disappear when the medical encyclopaedia became available to everyone. The doctor refocused. That is exactly what awaits the teacher.

And what if the machine replaced the engineer too?

Before getting to that refocusing, we have to face the most radical objection, the one my students don't always dare to voice but which I read in their eyes: what's the point of learning computer science, if the code produced by machines becomes better than ours? If tomorrow the machine replaces not only the teacher, but the very profession the teacher prepares them for?

I won't wave the question away, because it is a serious one. Generated code is improving fast, and part of what occupied yesterday's developers is already automated. Anyone who claims to know where this will stop is lying, one way or the other. But three things seem solid to me.

The first is historical. Computer science is the profession that has never stopped automating itself. The compiler replaced writing assembly; libraries replaced the perpetual rewriting of the same algorithms; frameworks, package managers, on-demand infrastructure: each generation of tools ate the work of the previous one. Each time, the profession did not disappear, it climbed a rung in abstraction. Generative AI is one more rung, a brutal one, but a rung. The centre of gravity shifts from writing code to specification, architecture, validation: saying precisely what you want, checking that you got it, understanding why when you didn't.

The second is a matter of responsibility. A system produced by a machine will always need someone to answer for it. When the software drives a train, a ventilator or a power grid, "the AI wrote it" will never be an admissible defence, neither before a judge nor before the families. And you can only take responsibility for what you are able to understand. My research work is precisely about the explainability of automatic systems: making the results produced by algorithms interpretable and verifiable. The more machines produce, the greater that need grows. Tomorrow's engineer is perhaps less the one who writes the code than the one who can hold the code to account, whatever its origin.

The third is almost an obvious point we forget: a society whose entire infrastructure rests on systems no one understands any more is a society in danger. Even in the extreme scenario where machines wrote everything, we would still need to train humans able to supervise them, audit them, unplug them when it's warranted. That would still be computer science, and it would still take teachers to teach it. We don't train students to compete with the machine on production. We train them to stay above it on understanding.

So yes, I may be training the last engineers who will have coded much by hand, just as the last navigators who could take a bearing with a sextant were once trained. But we still teach navigation, precisely because GPS breaks down, gets it wrong, or is fed lies. The lesson holds for everything else.

My colleagues are no better off

You might think this is a computer scientist's problem. I am lucky to have a wider vantage point: my wife is an English teacher, and one of my colleagues teaches mathematics. Our three subjects have almost nothing in common, except the essential: each faces, in its own words, the same question. What's the point of learning X, if the machine does X?

For English, the version is almost more brutal than mine. Machine translation is excellent, real-time voice translation is arriving in earbuds, and a model writes an essay in cleaner English than most pupils manage. The written assignment no longer proves anything: a flawless piece can come out of a phone in ten seconds, and correcting the grammar of a text the pupil didn't write simply no longer makes sense. So my wife sees her job pushed towards what doesn't go through the machine: live conversation, speaking, the accent, the hesitation overcome, the nerve to speak. Because a language is not a content you transfer, it is a capacity you embody. The earbud will translate the message; it will never make it you who speaks English. And there will always remain that difference between understanding someone through a machine and understanding them yourself: the second is called an encounter.

For mathematics, my colleague could almost look at us with a veteran's irony: his discipline has been facing "the machine does X" for half a century. The calculator, computer algebra software, apps that solve a photographed equation: maths had a forty-year head start on our problem. And its historical answer is illuminating for everyone: we never taught calculation to get results, we taught it to structure a mind. No one has needed a human to multiply two numbers since 1975; yet we still learn to do it, because it is by doing it that you understand what a number is. But generative AI crosses a threshold the calculator never crossed: it no longer produces only the result, it writes the reasoning. The proof — the last bastion, the thing that showed you had thought — is now generable too. Homework is dead, and my colleague knows it: his assessment is tipping towards the live, the whiteboard, the oral, the "show me how you think, here, now."

Three subjects, three versions of the same earthquake, and, if you look closely, three times the same answer taking shape: production was never the point. The code, the essay, the calculation were only traces; what we were aiming at was what their making built inside a head. Now that the traces are generable, all that's left is to assess, and to cultivate, the building itself. Hence a remarkable convergence: in all three of our disciplines, the future runs through the return of the live, the in-person, the oral, the person. This is not a nostalgic retreat. It is a refocusing on the only place where learning ever happened.

The temptation of laziness, or: optimise, but what for?

There remains one adversary to name, and it is not the machine. It is us.

Because let's be honest: the real engine of everything above is laziness. Ours, the teachers': delegating one more task, then another, until proofreading becomes a ritual and you wake up as the quality controller of your own course. I know that slope, I slide down it like everyone else. And the students', which is worse because it is perfectly rational: why struggle for an hour on a problem when the answer is one prompt away? No one inflicts on themselves an effort that seems pointless. The tragedy is precisely there: the effort seems pointless, because we have confused the product of learning with its mechanism.

And yet effort was never the price to be paid in order to learn. It is the learning. Struggling with a problem is exactly the operation by which a concept settles into a head; difficulty is not an obstacle on the path, it is the path. That is why no one ever invented a machine to do your push-ups for you: everyone sees immediately that it would be absurd, that the point of push-ups is not for the push-ups to be done. Curiously, no one sees the same absurdity when a machine does the exercises in the student's place. Yet it is the same.

And behind laziness stands its respectable alibi: optimisation. All our current vocabulary comes from it: save time, go faster, be more productive, more efficient. I recognise myself in it, I have adopted this vocabulary myself: faster, further, deeper. But "optimise" is a transitive verb that has lost its object. Optimise, yes: but what for? Optimisation knows how to answer "how to go faster"; it is structurally incapable of answering "where to go" and "why." My students typing my question into a model had perfectly optimised getting the answer. They had just forgotten to ask whether the answer was the point. It wasn't. It never is, in education.

Because there are activities you do not optimise, because they are their own end. You don't call a hike inefficient because a helicopter would reach the summit faster: the helicopter doesn't optimise the hike, it abolishes it. You don't ask a musician why they don't just listen to the record instead, which plays better than they do. Learning belongs to that family. And so does teaching. We learn for pleasure, the pleasure of finally understanding, of feeling something click into place; we teach for pleasure, the pleasure of showing, of watching a face light up. That pleasure is not the reward that comes after the work: it is what makes the work possible, its fuel. If we optimise away everything that is slow, difficult, frictional, we will not have made learning more efficient. We will have optimised the pleasure out of the circuit, and with it the learning itself. That, I think, is the exact source of the feeling I described at the start: going faster than ever, and yet being less present. I optimised. And I lost something along the way, without any metric telling me what.

This is where I reveal the advantage announced at the start, and it will seem paradoxical: I am disabled. I live with a body that has its breakdowns, its slownesses, its off days, and a machine in my head that compensates without erasing. In other words, I know, with a certainty few people possess, that I will never be optimised to a hundred per cent. No tool, no discipline, no AI will get me there. And against all expectation, this is a wild stroke of luck. That goal so many people pursue to exhaustion, the fully efficient version of themselves, is barred to me from the outset: I am thereby excused from the race. When a hundred per cent is out of reach by construction, you stop aiming at it, and you finally ask the only question that matters: what good can I do with what I have? Disability handed me, by default, the lucidity this piece is trying to build through reasoning. And there is something delicious in watching perfectly able-bodied people voluntarily inflict on themselves a quest for total optimisation from which I am, by nature, exempt. I want to be the one who stays clear-eyed about this: the un-optimisable is not a flaw in the system. It is the place where you live.

The right question, for a teacher as for a student, is therefore not "what can AI do in my place?" It is: "what do I lose if I stop doing it myself, and do I want to keep it?" Laziness always answers no. Pleasure, on the other hand, knows where it lives.

What the teacher can do, and the machine will not

1. Teach judgement, not production

My students can now generate a complete software architecture in one prompt. Fine. But knowing whether it was the right one, spotting the plausible-but-wrong, the solid-but-unsuited, the elegant-but-shaky, demands exactly what we teach: the fundamentals. A student who has never written a data structure by hand will not see that the model has just proposed a bad one. The parallel holds elsewhere: the pupil who has never built a sentence in English will not see that the translation is wrong in nuance; the one who has never written a proof will not see that the generated reasoning has holes. A precious paradox: the freer production becomes, the more valuable the assessment of what is produced.

Concretely, this means shifting the course's centre of gravity: less "produce this," more "here is a production: what do you make of it?" Have them critique a generated piece of code, a model's answer, a plausible text. The well-chosen error becomes the best teaching material of our times: critiquing an anonymous production is infinitely less risky than asking a question, and it makes even the most silent cohorts speak.

2. Create productive friction

A model is compliant by construction: it gives the answer. The teacher, by contrast, can refuse it. That is exactly why my earlier anecdote annoyed me: my questions are not requests awaiting an answer, they are resistances deliberately placed on the path. Asking the awkward question, letting a student struggle for just as long as needed, neither too little nor too long. In lab sessions it's an art I practise daily: seeing a pair stuck, knowing I could unstick them in ten seconds, and choosing to ask a question instead. Learning passes through that measured discomfort, and dosing it is a craft. It may even be the shortest definition of the job: AI optimises the comfort of the one who asks; the teacher optimises their progress. These are two different objective functions, and the second requires someone who isn't trying to satisfy you.

3. Make the course a collective event

One-on-one tutoring is the ground where the machine has the structural advantage: total availability, infinite patience, instant adaptation. Betting the survival of the profession on that ground is choosing the match you've already lost. What the machine will not provide is the embodied collective: the moment when a whole room discovers something together, when one person's question unblocks the other five, when a demonstration that goes wrong makes the entire cohort laugh, and teaches them more than the handout would have.

Anyone who has attended one of my courses knows I cultivate this register: the ritual catchphrases, the "it worked on my machine," the little theatre of the teacher who's out of arguments. This is not decoration. It is what makes a course a place where you are together, and not a stream you could watch at double speed. The future of the in-person course is not the lecture (that is replaceable) but the event: what is only worth living together, live, with someone orchestrating.

4. Exploit the lock that just broke

AI finally makes possible what has always been impossible: the multi-speed course. The historical lock was simple: one teacher, one pace. Impossible to explain three things to three groups at once; everyone suffered the average, the fast ones were bored, the fragile ones dropped off, and the teacher spoke to the statistical ghost of the middle. I lived it every year: sessions calibrated for a median student who does not exist.

If explanation becomes self-service, the teacher no longer needs to be at the front unrolling it. They can build tiered pathways (a compulsory base, extensions in a staircase, an open challenge for those who always finish early) and become the one who circulates. Five minutes of a teacher sitting next to a stuck student weigh more than thirty minutes of artificial tutor: not because they are cleverer, but because they are rare, and because they come from someone to whom it matters.

Better still: differentiation can become invisible. In a group project where each team owns a different component of a shared system, you can hand out the components by level without ever announcing it. No one is in the "weak group": everyone holds a piece the whole thing needs.

5. Stay the one who knows what's going on in the heads

There is a real risk, and it must be named: the closed loop. An AI produces the material, an AI does the exercise, an AI grades it, and no one learns anything any more, in a system that is nonetheless running at full throttle. My anecdote of the question read out loud is the miniature version of it: I was myself a link in the chain. I watch this loop take shape from both my windows at once: on the teacher's side, when I generate my content; on the students' side, when their submissions carry the obvious trace of the same tool. The teacher's irreducible role is to stop this loop from closing: to remain the one who knows where the cohort stands, what has landed, what is resisting, and who adjusts.

This means deliberately preserving channels of signal, now that the traditional ones are drying up: moments of oral work, live restitutions, work done in one's presence, attention paid to the process and no longer only to the final product, because the final product now proves very little. This is, as we've seen, the conclusion the English teacher and the maths teacher also reach, each by their own path. When three disciplines as different as these converge on the same solution, it is probably because it is structural.

6. Embody enthusiasm

Finally, the most important, and the least measurable. When I'm asked what I still enjoy in this job, my answer hasn't changed: discovering something new and being eager to show it to my students. That pleasure I have never delegated to the machine, and I notice I never even considered it. So the laziness I spoke of earlier is better targeted than it looks: it bears on what could be mechanised, and spontaneously spares what cannot. It may be the best test we have: what you never wanted to delegate is the heart of the job.

Because the most precious thing a teacher transmits is not information: the machine supplies better information, faster. It is living proof that after years in the job, one can still find one's field wonderful. An English teacher who loves the language transmits that love first; a maths teacher who finds a proof beautiful teaches first that a proof can be beautiful. A model's enthusiasm costs nothing, so it proves nothing. A teacher's costs them their evenings, their fatigue, their years: that is why it is contagious.

I speak of embodiment from experience. The machine in my head, the one I mentioned earlier, is a deep brain stimulation implant I have carried since 2025: I am, literally, a human who runs with a machine. I have already written here about what that changes in the teaching relationship. It has taught me at least one thing: it helps me be there, it does not replace me. That is a fairly good definition of what AI should be for a teacher. And there is a second asymmetry, deeper still: a machine adapts to you, but it expects nothing of you. It is not disappointed if you give up. You owe it nothing. The teaching relationship, by contrast, commits you: someone in front of whom you don't want to hand in shoddy work, someone who will remember you in two years. It is precisely that commitment that makes you learn.

Conclusion: a refocusing, not a disappearance

The teacher who is threatened, whether they teach code, English or mathematics, is not the one who uses AI, nor even the one who delegates a lot to it. It is the one whose course is nothing but a stream of content, because a stream of content is something an AI replaces tomorrow morning.

The one who remains, and will remain for a long time, is the one who makes their room a place where something happens: where you judge, where you confront, where you discover together, and where someone — a tired, enthusiastic, fallible human — still wants to show you something. And this holds even in the world where machines wrote all the code, translated all the languages and drafted all the proofs: we will always need humans who understand, therefore humans who learn, therefore humans who teach.

AI frees up time for us and breaks old locks. The question is not whether the profession will survive. It is whether we will spend that time on one more task, in the name of an optimisation that cannot say what it optimises towards, or on the one side of the scale the machine cannot fill: the side where we learn, and where we teach, for pleasure.