What AI cannot replace in the classroom

Knowing your students

A model does not know that Sushmita at the back sat through last night’s argument between her parents and has not slept. It does not know that Rabin, who is failing maths, lost his father to a heart attack four months ago and has not been the same since. It does not know that Anjali has been quietly figuring out fractions for three weeks and is one good explanation away from understanding them properly. It does not know that the boy in the third row whose Nepali is excellent learned it from a grandmother who is now in her last winter. It does not know, because nobody told it, and nobody could.

The teacher who walks the classroom knows. Not all of it, not always, but enough. She knows which child to call on first this morning and which to leave alone. She knows that the question about Rabin’s homework needs to be asked privately, after class, with the door half-closed. She knows that Anjali is about to break through and that the next ten minutes of teacher attention will be worth more than any worksheet ever produced. This kind of knowing is not information. It is a relationship maintained over months in a small physical room. No model has access to it. No model will, even when the models are much better, because the relationship is not data — it is the accumulated trust of being present, repeatedly, in the same room with the same children.

Holding the room

There is a craft to keeping forty children, or sixty, in productive attention for forty-five minutes. It is not a craft any textbook fully describes, and certainly not a craft any AI tool can take over. It is built out of a thousand micro-decisions per period — who needs a question now, who needs five seconds of eye contact, whose pencil has stopped moving and why, which two boys at the back are about to start an argument and need a separate task each. The new teacher learns this by failing at it. The experienced teacher does it without consciously deciding.

A model cannot hold a room. A model cannot read the temperature of a class on the morning the SLC results came out, or on the day after the load-shedding kept everyone awake, or on the Friday before Tihar when half the children are already mentally home. A model cannot decide, in the third minute of a lesson, that the planned activity is wrong for the energy in this room today and improvise something better. A model cannot project, with eye and voice and the simple physical fact of standing at the front of a room, the signal that says: what we are doing here matters, and I am taking it seriously, so you can too. That projection is itself most of teaching. A chatbot at every desk is the opposite of it.

Building confidence — the relational work

A great deal of what determines whether a Nepali child eventually learns maths, or English, or science, is not whether the lesson was well-designed. It is whether, at some point, a specific teacher kept showing that child that they could do it — when they had been told by themselves and their family and the last teacher that they could not. The child who finally believes she is the kind of person who can solve a quadratic equation believes it because a specific adult, with a name and a face, told her so often enough and accurately enough that she started to believe it. This is the most quietly important thing teachers do. The textbook does not capture it. The lesson plan does not capture it. The marks register does not capture it. It happens in the small moments — at the desk, after the bell, in a corridor — and it is what turns a curriculum into a life.

No model produces this. No model can. The model is not a person; the child knows it is not a person; the child cannot believe a model believes in them, because there is nothing in the model that believes anything. The relational confidence-building of teaching is, by its nature, between two people who are both real. AI does not approach it from a distance, even.

Moral and civic formation

A school is not only a place where children learn the chapters in their books. It is, whether teachers like the responsibility or not, a place where children learn what counts as fair, what it means to disagree without contempt, how to share something scarce, what to do when someone is being treated badly, how to be the kind of person other people can rely on. These lessons are mostly not taught explicitly. They are caught from the adults the children are around. A teacher who is consistently fair shapes a class with a sense of fairness. A teacher who treats the cleaner with respect teaches an entire row about who counts as a person. A teacher who does not lose her temper when she is provoked teaches forty children what it looks like to be an adult.

These lessons are also the lessons that survive longest. A student forgets the dates of the Anglo-Nepal war within five years of leaving school. They do not forget the teacher who took the time, in front of the class, to apologise when she had made a mistake — because it taught them what an adult is. AI does not deliver this. Content cannot deliver it. It is delivered, slowly and continuously, by a specific person in a specific role, who is being watched by children who are deciding what kind of adult they are going to be. The textbook calls this the “hidden curriculum.” It is the curriculum that matters most.

Judgement at the margins

The final piece is the hardest to name. It is the set of judgement calls a good teacher makes every day that no algorithm and no policy fully captures. When to push a struggling child and when to let them have an easy day. When to call the parent and when to wait a week and see. When to mark a borderline answer up because the child has been trying, and when to mark it down because they need to learn that effort and outcome are not the same. When to keep a class behind for five minutes and when to let them go. When to overrule the policy because the policy does not anticipate this particular child. When to follow the policy because the policy is wiser than the impulse of the moment.

A model cannot make these calls. A model has no skin in the outcome, no relationship with the child, no professional reputation at stake, no view of the village this child will go back to and what their life will look like at eighteen. The model produces confident output regardless. The teacher reads the room, the child, the moment — and acts. The acts may be wrong. The acts are answerable to the person who took them. That answerability, that ownership of a judgement call about a child, is the centre of the profession. AI does not replace it. AI does not even approach it.

What this means as AI improves

The misreading of this section would be: AI is irrelevant, then. That is not true and the previous chapters made the opposite case. AI is going to absorb a serious share of the mechanical work that fills a teaching day. The right reading is the harder one: as AI gets better at the mechanical parts, the parts of teaching that AI cannot touch become relatively more of what your day is for and relatively more of what you are valued for. The teacher who has always defined herself as “the person who produces good worksheets” will find her self-image quietly hollowed out. The teacher who has always defined herself as “the person who sees these children and judges well what each one needs” will find herself, in 2030, more necessary, not less. The profession is moving toward the relational and judgement-laden core that was always its actual centre. AI is, in this sense, returning teaching to itself.

Check your understanding

Quick check

—

True or false: as AI tools improve, the mechanical parts of teaching (typing first drafts of plans, producing worksheets, grading routine items) will become MORE valuable to a teacher’s career, while the relational and judgement parts (knowing your students, building confidence, holding the room) will become LESS important.

1. ATrue — automation always makes the automated tasks more valuable to the humans who can still do them.
2. BTrue — schools will pay teachers more specifically for the parts AI also handles, because of regulatory requirements.
3. CFalse — it is the opposite. As AI absorbs the mechanical parts, those parts become cheaper and less central to a teacher’s value. The relational and judgement parts, which AI cannot replace, become relatively more of what teaching is for and what teachers are valued for.
4. DFalse, but only for senior teachers; junior teachers will see the reverse pattern.

Check answer→

What comes next

If the relational and judgement-laden core of teaching is becoming more, not less, of what the work is for, then the question for the closing section is what teaching actually looks like as AI scales — what changes, what stays, and what the new teacher craft is. That is the final section of the course.