When a new technology reaches education, the first move is usually to buy it, or to announce it. Tablets, smart boards, learning apps, and now AI assistants arrive with a launch event. The teacher is invited to attend, shown the screen, and sent back to class.
I have run AI training for government school teachers, and the lesson I took away is that the tool is never the hard part. The hard part is the teacher's day. A teacher in a public school may have a large class, children at very different levels, a thick syllabus and little time to prepare. A tool that ignores that day will be switched off by Friday.
So my argument is this. Before Nepal's teachers need an AI tool, they need three things: a clear reason to use it, the skill to use it safely, and someone to call when it goes wrong.
Start with the Monday morning problem
In my sessions I do not begin with what AI is. I begin by asking teachers what takes their time. The answers are practical and honest: preparing lesson plans, making practice questions, marking, explaining one idea to children who learn at different speeds, finding examples that the children can relate to.
Each of those is a task AI can help with, and each can be shown in ten minutes. A teacher can paste a textbook chapter and ask for fifteen practice questions at three difficulty levels. She can ask for the same idea explained for a weak reader and for a fast one. She can ask for examples set in a village market instead of a London street.
When a teacher sees her own problem shrink on the screen, the training changes. She stops being a listener and starts being a user. That moment cannot be created by a slide about what machine learning is.
Teach the limits as carefully as the uses
A tool used without knowing its limits is a risk, and in a classroom the risk lands on children. I teach four limits in every session.
First, AI makes mistakes and states them with confidence. A teacher must read every generated answer key before using it. Second, AI works unevenly in Nepali. It is often better in English, so teachers should check the language, the grammar and the cultural fit of anything produced. Third, generated material may not match the national curriculum. The teacher is the one who knows the curriculum, so the teacher stays in charge. Fourth, never paste children's names, marks or personal details into a public chatbot.
These four rules are not complicated. They take twenty minutes. But without them, a well-meaning teacher can hand out a worksheet with a wrong answer printed on it, and thirty children will learn the wrong thing.
Why one-day workshops are not enough
A single workshop gives a teacher an experience. It does not give her a habit. Habits form through repetition, and repetition needs a reason to return.
What works better, in my experience, is a short, hands-on session followed by a small task to try in real classes, then a second meeting to share what happened. Teachers who tried something and saw it work in front of real children carry that confidence back to the staff room. Peer teachers are far more persuasive than any trainer who visits for a day.
It also helps to build around small groups inside one school. When two or three teachers in the same building learn together, they can help each other next week. A lone teacher in a lone workshop has nobody to ask.
School leaders matter here too. A head teacher who says 'try this, and tell me what worked' gives teachers permission to experiment. A head teacher who says nothing signals that the training was a formality. The cheapest support a school can give is time: thirty minutes a month when teachers share what they tried. Without that time, even a good workshop fades within weeks.
I also tell teachers that they do not need to be experts to begin. They need curiosity, a clear task and a habit of checking. Everything else can be learned along the way, and the first small win usually changes how a teacher feels about the whole subject.
The conditions outside the classroom
Honest training also names what teachers cannot fix. Many schools have limited devices, patchy internet and unreliable power. A plan that assumes every child has a screen will fail in exactly the schools that most need help.
The better design puts the technology in the teacher's hands, not the children's. The teacher uses AI at home or in a staff room to prepare, then teaches with chalk, paper and voice as before. This is realistic for a very large number of schools, and it respects what is already there.
It also means Nepali-language support matters more than any new feature. If the best tools work well only in English, the gap between English-medium private schools and Nepali-medium public schools grows. That would be a quiet and expensive mistake.
What a good session looks like
Let me describe the shape I aim for, because the shape matters more than the slides. The first twenty minutes are talk: what AI can and cannot do, shown with one live example that sometimes goes wrong on purpose. Then teachers open the tool and work on a real task from their own subject and grade. I walk around and answer questions one by one.
The most useful part is the sharing at the end. One teacher shows a quiz she made for her science class. Another shows an explanation she rewrote for a weaker group. Others see what is possible in their own subject and ask for the prompt. No trainer can match the force of a colleague saying 'I used this yesterday'.
I always end with one small commitment. Each teacher picks a single task to try before we next meet. One task is enough. Teachers who try ten things at once try nothing.
Who trains the trainers
There is a larger question behind all of this. Nepal has far too many teachers for any small team of outside trainers to reach. If AI skills are to spread, they must live inside the institutions that already train and support teachers: teacher education campuses, resource centres, local education offices and the colleges that prepare new teachers.
That means training the trainers, and giving them materials in Nepali, in forms they can adapt. A guide that fits on two pages and uses examples from the Nepali curriculum will travel further than a large report in English.
It also means being honest about what we do not yet know. We have little solid evidence from Nepal on which AI uses actually improve learning for children. That is not a reason to wait. It is a reason to start small, record what happens and share the results, good and bad.
Measure what changes
Training programmes love to count how many teachers attended. That number tells us almost nothing. A better question is what changed three months later. Are teachers still using the tools? For which tasks? Did preparation time fall? Did the quality of questions and explanations improve?
These are simple questions to ask in a short follow-up survey, and they allow honest adjustment. If a part of the training did not stick, we should know and fix it. Public money spent on teacher training deserves that discipline.
The choice in front of us is not between AI and no AI in Nepali classrooms. AI is coming either way. The choice is whether teachers meet it as a tool they understand and control, or as one more thing handed down from outside. I know which one I am working for.