What AI does well for teachers — and what it does not · AI for Teachers

You did not come to this course because you needed another article about how AI will revolutionise education. You came because your headmaster announced a “smart classroom” initiative on Monday, because two of your Class 9 students handed in suspiciously fluent essays last week, and because you have forty-two exercise books to mark before the parent-teacher meeting on Friday. The honest question is small and practical: what can this technology actually do for your week, and where will it embarrass you — or worse, harm a child — if you trust it too far?

The honest taxonomy

Strip out the hype, and current generative AI — the ChatGPT and Claude and Gemini models, the chatbots embedded in Google Classroom and Microsoft Teams, the copilots showing up inside Canva and Khan Academy — is reliably useful for a narrower band of teaching work than the brochures suggest. The band is still wide. It covers a great deal of what a government school teacher in Sindhupalchok or a +2 lecturer in Pokhara actually does between 6 a.m. and 10 p.m. on a school day. But it has edges, and walking past those edges is how a teacher ends up in front of an angry parent, an unimpressed principal, or worse, a student who learned the wrong thing because the model said it confidently.

The clearest way to sort tasks is by one question: how cheaply can you check the output, and what does it cost if a quiet error reaches a student? If verification is fast and the downside is small, AI is a force-multiplier. If verification is slow and the downside touches a child’s grade, a parent’s trust, or your professional standing, AI is a liability dressed as a shortcut.

What AI does well

Seven categories, in roughly descending order of confidence.

1. Drafting lesson plans from a NEB unit page. Open the CDC curriculum, point the model at the unit on photosynthesis or the unit on linear equations in two variables, give it your class size, period length, and available materials, and it will produce a competent first-draft lesson plan in under a minute. Learning objectives, hook activity, board work, group task, formative check, homework. You edit for your students — the model does not know that Sushmita sits at the back because she cannot hear well, or that your projector has not worked since Dashain — but the skeleton is there.

2. Producing varied worksheets and quizzes. The same arithmetic concept, written ten different ways. Ten short comprehension passages at Class 6 reading level on topics your students actually care about — a bus route, a momo shop, a cricket match. A set of multiple-choice items on the water cycle calibrated for SEE-style questioning. This is the work that used to eat your Saturday morning. The model produces drafts in minutes; you check answer keys against the curriculum.

3. Differentiating practice for mixed-ability classes. Most Nepali classrooms hold a wider ability spread than the textbook assumes. The same Class 7 maths period contains children who can multiply fractions and children still consolidating place value. Hand the model your topic and ask for three tiers of practice — foundational, on-level, stretch — and you get the material that a single teacher with one chalkboard could not realistically prepare alone.

4. First-draft Nepali translations of English material. A science explainer you found online, a primary-source passage from a history book, a worked example from a Khan Academy video. The model produces a serviceable Nepali rendering quickly. Nuance is uneven — technical vocabulary in particular drifts toward calques — so you read before you print. But the productivity gain over translating from scratch is real, and it widens what you can bring into the room.

5. Structured grading against a rubric, with teacher review. Give the model the rubric — five criteria, four levels each, with descriptors — and a student response, and it will produce a draft assessment with reasoning. You do not accept the grade. You read the model’s notes against your own judgement, change what the model missed, and use the time saved on the thoughtful feedback that actually moves the student. The grade remains yours.

6. Drafting parent communication in Nepali. A note home about a child who is struggling. A class-wide message about the upcoming field trip permission form. A delicate message to a parent whose son has been disruptive. The model produces respectful, appropriately formal Nepali quickly. You read it carefully — tone in Nepali, especially in honorifics across community lines, is where the model is weakest — and send.

7. Administrative templates. Attendance summaries, meeting minutes, year-end reports to the headmaster, the perpetual paperwork the system demands. The model does not care that this work is dull. It produces serviceable drafts faster than you can type them.

What AI does badly, or unsafely

The same technology fails — and fails most dangerously when it looks like it is succeeding — on a different cluster of tasks.

1. Final grading on high-stakes papers without teacher review. SEE pre-boards, +2 internal assessments, scholarship qualifiers. A model can produce a defensible-looking score, but a model cannot defend it to a parent, cannot weight a partially-correct method the way the marking scheme intends, and cannot be held to account when a child’s future turns on the number. Use it to draft feedback. Do not let it produce the grade alone.

2. Promotion decisions and placement recommendations. Whether a child should repeat Class 5, whether a student is ready for the science stream at +2, whether a struggling reader should be referred for support — these decisions need the child’s whole picture, the parent’s circumstances, and a professional acting under accountability. The model has none of these. It will produce confident-sounding output anyway.

3. Replacing the teacher in the room. The temptation to let a chatbot “tutor” the back row while you work with the front is understandable and almost always a mistake. Students at every level read your attention as the signal of what matters. A model is not a substitute presence. It is a tool you use to free your attention for the children who need it most.

4. Quoting specific NEB textbook content verbatim. The model has read about Nepali curriculum. It has not necessarily read the current edition of your CDC textbook. It will confidently produce a definition, a worked example, or a historical date that is subtly wrong — a verb tense flipped in the Nepali, a kingdom misdated by twenty years, a formula written with the wrong constant. For anything you write on the board as fact, verify against the textbook.

5. Diagnosing learning disabilities. A child who reverses letters at age 9, a child who cannot sit still, a child whose Nepali is fluent but whose English progress has stalled — these patterns may indicate dyslexia, ADHD, hearing loss, or simply that the child speaks Maithili at home and English is their third language. The model will offer a label. The label may be wrong, and acting on it without a trained specialist can do real damage.

6. Safeguarding issues. A child discloses something at home. A student shows bruises. A girl stops coming to school. These situations need a human in the chain — the head teacher, the parent, the local social worker, sometimes the police. The model is not equipped to handle them and you should not route them through it.

The two-question test

Put the two lists together and a rule falls out. Before you use AI on any teaching task, ask two questions. First: if the output is wrong in a way I do not immediately spot, what does it cost? A clumsy lesson plan costs forty minutes. A wrong fact written on the board costs a class of children. A wrong grade on a board paper costs a student’s future. Second: how long would it take me to catch the error against a source I trust? Catching a wrong worksheet answer takes a minute against the curriculum. Catching a subtle misjudgement on a child’s readiness for the science stream may take months.

If both answers are small — low downside, fast catch — use the model freely. If either answer is large, the model is the wrong tool for that decision, no matter how confident it sounds. This is not timidity. Most teaching work, by hours, sits squarely inside the safe band. The professional act is knowing which hour is which, and never letting the tool absorb the responsibility that ends with your name in the register.

Check your understanding

Quick check

—

Which of the following is the most reliably safe use of current generative AI for a Nepali classroom teacher?

Producing the final grades for your Class 10 SEE pre-board English papers.
Drafting a first version of a Class 7 lesson plan on linear equations, which you then edit for your specific students.
Deciding which Class 8 students should be referred for special-needs assessment.
Quoting the exact definition of "photosynthesis" from the current CDC science textbook onto the board.

Quick check

—

True or false: if a teacher uses an AI tool to produce a worksheet or grade, the responsibility for any errors that reach the students is partly absorbed by the tool.

True — the AI provider shares responsibility for what the model produces.
True — once a school adopts AI tools, the institution carries the risk, not the teacher.
False — the teacher remains fully responsible for everything that reaches the students, regardless of which tool produced the draft.
False, but only for government-school teachers under NEB jurisdiction.

What comes next

Knowing what AI does well in the abstract is half the picture. The other half is knowing where, in the texture of an actual teaching week, those strengths land — which stages of preparation, in-class teaching, marking, and communication change, and which stay the same. The next section walks through a typical week and marks the map.