Every school we speak to has some version of the same conversation running internally. One group wants AI in the classroom immediately, because students are already using it and the school looks slow. Another group wants it kept out, because they can see what happens to a student’s writing when the machine does the difficult part. Both groups are responding to something real.
The way out of that argument is not to pick a side. It is to be much more specific about what “AI in school” actually refers to, because at least four different things are hiding inside the phrase.
Four different questions wearing one coat
AI as a subject. Students learning how these systems work: what a model is, what training does, why a system can be confidently wrong. This is curriculum, and it belongs in the technology pathway like any other domain.
AI as a tool for students. Students using assistants to draft, summarise, debug, translate or generate. This is a question about assessment design and academic honesty.
AI as a tool for teachers. Lesson preparation, differentiation, first-pass feedback, administrative load. This is a workload and quality question, and it is the one with the clearest near-term upside.
AI as infrastructure. Adaptive systems that decide what a student sees next. This is the one that deserves the most caution, because it quietly moves pedagogical authority from a teacher to a vendor’s optimisation target.
Most school debates go badly because two people are arguing about different items on that list.
Sequence matters more than enthusiasm
The most common mistake we see is introducing AI to young children as a magic box that produces things. Not because it is dangerous in some dramatic sense, but because it teaches exactly the wrong first lesson: that the interesting part of thinking is the output.
A better sequence starts earlier and lower. Before a child meets a generative model, they should have built something that follows rules they wrote. That experience — I told it what to do, it did exactly that, including the parts I got wrong — is the foundation for understanding what a machine is. A student who has that experience meets a language model as a much more sophisticated version of something they already understand in principle. A student without it meets an oracle.
In the Gurukul One pathway this is why artificial intelligence is introduced around Grade 6 rather than Grade 2, and why generative AI comes later still. It is not caution for its own sake. It is that the concept lands differently when there is something underneath it.
Teach the failure modes first, not last
Most introductory AI material follows the same shape: here is what it can do, and then, near the end, a section on limitations and ethics. Reverse it.
A student’s first serious encounter with a model should include watching it be wrong — confidently, fluently and in detail. Ask it about something the class genuinely knows well: a local historical detail, the school’s own timetable, a fact about their town. Watch what it produces. Then check.
This single exercise does more for a student’s judgement than a term of instruction about responsible use. It converts an abstract warning into an observed property of the system. Students who have seen a model confabulate about something they knew are permanently different users of that model.
What the human element actually is
The phrase gets used loosely, so it is worth being concrete. In a classroom running technology projects, the things a teacher does that a system cannot are specific:
They notice that a student has been quiet for two sessions. They know which pair should not work together this week. They can tell the difference between a student who is stuck and a student who is avoiding. They decide when to let a group fail productively and when the failure has stopped teaching anything. They hold the standard — the sense that this could be better — which is the single hardest thing to encode.
None of that is at risk from AI. All of it is at risk from a teacher’s time being consumed by preparation, marking and administration. Which is why the most defensible early use of AI in a school is usually the least visible: reducing the load on teachers so more of their attention lands on students.
Practical guardrails that hold up
A few positions we would defend in any school:
Disclosure over prohibition. Students should be expected to state what they used and how. A rule that can be followed is worth more than a rule that cannot be enforced.
Process artefacts alongside outputs. If a project requires a design note, a decision log and a demonstration, the assistant becomes a tool inside a process rather than a replacement for it.
No unsupervised model access for young learners. Not because of catastrophic risk, but because the material value of the interaction depends on an adult being able to say “look at what it just did there”.
Student data stays boring. Any system in a school should be assessed on what it retains and where that goes, before it is assessed on what it can do.
The version of this that goes wrong
A school buys an AI product, announces an AI programme, and two years later has a subscription, a press release, and students whose relationship with the technology has not changed at all — because nothing about the curriculum, the assessment or the teaching changed.
The version that goes right is less announceable. Students spend years building things. Somewhere in the middle of that, they meet models as one more powerful and untrustworthy tool among several. They learn to direct them, check them, and know when not to use them. By the time they leave, they can build with these systems and they can tell you where the systems fail.
That is not a smaller ambition than the announceable version. It is just harder to photograph.