Students using AI — detection, policy, conversations · AI for Teachers

A Class-10 English teacher in Lalitpur opens her laptop on a Sunday evening and pastes a student essay into a free online “AI detector.” The bar fills red. Eighty-seven percent AI-written, the screen says. The student is a quiet girl from Khotang who has been improving steadily all year. The teacher now has to decide what that number means, what to do with it on Monday morning, and what to write to the girl’s mother — who, like the teacher, has been told by the newspapers that schools “have ways of checking.” This section is about the gap between what the newspapers say and what is actually true. It is also about what a teacher does, professionally, in that gap.

AI detectors are weak signals — read them like one

The most important fact about every commercially available AI-text detector in 2026 is the one the marketing pages do not put on the front page: they are unreliable, and they are most unreliable on writing by second-language English speakers. A 2023 Stanford study found that the leading detectors flagged the writing of non-native English speakers as AI-generated more than half the time, while flagging native-speaker writing correctly almost all the time. That bias has not gone away. Second-language English writing tends to use shorter sentences, more common vocabulary, and more predictable structures — exactly the signals the detectors trained themselves on as “AI-like.” Nepali students writing in English fall directly inside that pattern. So do students who paid attention in their tuition class and learned a careful, formal voice. So do students who copied a phrase pattern from a textbook because the textbook is what they had.

A detector that says “87% AI-written” is not telling you the essay is AI-written. It is telling you that the essay’s surface features match what the detector thinks AI looks like. That is one input. It is not evidence. It is certainly not a conviction. Treat it the way you would treat a single anonymous complaint about a child’s behaviour — worth checking, never worth acting on alone.

The classroom AI policy that ages well

The two policies that fail are the blanket ban — “no AI ever” — and the vague punt — “use your judgement.” The first is unenforceable from the first week; every student has a phone in their bag and a chatbot in their browser, and they know it. The second gives a child nothing to follow and gives a parent nothing to argue with when their child is accused. Both leave the teacher exposed when something goes wrong.

The policy that holds together is specific. It says, for each kind of assignment, what is allowed and what is not. A workable shape, that you can put on the inside of the exercise book at the start of the year:

Allowed without disclosure. Using AI to look up a word, to check spelling, to ask “is this sentence grammatical,” or to brainstorm ideas before you start writing. Treat the model as a dictionary and a thinking partner.

Allowed with disclosure. Using AI to give you an outline you then write from, to generate practice questions you answer yourself, or to explain a concept the textbook explained poorly. On the assignment, write one sentence at the top — “I used ChatGPT to outline this essay and to check my conclusion paragraph.” That is the disclosure.

Not allowed. Pasting an AI-generated essay, paragraph, or paragraph-chunk into your final draft as your own writing. Pasting AI-generated answers into homework. Using AI on any in-class assessment.

The teacher’s check. I will sometimes ask you to write the next paragraph in front of me, to talk me through how you wrote a sentence, or to explain a word you used. This is not because I do not trust you. It is because the line above is the line we are holding together.

A policy of this shape does three things at once. It assumes the technology is in the room rather than pretending it is not. It gives the student a clear line to honour. And it gives the teacher a defensible position when a draft looks wrong: “the policy says you must disclose; you did not; let us talk.”

Assessment that makes learning visible

A policy by itself is paper. What enforces a policy is the design of the work. If every assessment is a take-home essay, every assessment is a place AI can do the work invisibly. If your assessments include surfaces AI cannot stand in for, the question of “did the model write this” stops mattering as much, because the learning has already been made visible.

Four practical surfaces, in order of how much they change for how little extra work.

1. In-class drafts. Twenty minutes at the start of the next class, writing the opening of the essay by hand, in front of you. You collect the drafts. The final essay must be recognisably an evolution of the draft. A child who writes a strong in-class draft and submits a polished final has clearly done the work. A child who writes a confused in-class draft and submits a polished final has a conversation coming.

2. Oral defences. For any major project, the student must spend three minutes explaining their work and answering two questions about it. A student who wrote a piece can talk about why they made the choices in it. A student who pasted from a model usually cannot.

3. Process portfolios. Drafts, notes, photos of the planning page, the marked-up rough version. The portfolio is graded alongside the final. The shift in incentive is real: you are now grading the path, not only the destination.

4. “Show your working.” Older than AI, more important than ever. The maths student who shows the steps, the science student who writes the experimental reasoning, the social-studies student who outlines the argument before the paragraph — each of these is producing evidence the model cannot produce convincingly for them without giving itself away.

None of these requires technology. All of them have worked in Nepali classrooms for decades. AI makes them more, not less, important.

The conversation when a student admits AI use

Sooner or later, a student will tell you, or you will be reasonably sure, that they used AI on something they should not have. The instinct is to treat the moment disciplinarily — sign in the register, note to the parents, marks deducted. Sometimes that is right. Often there is a better move available, especially the first time.

The conversation that teaches has a shape:

What did you write, and what did the model write? Get the student to draw the line themselves. Most students, asked honestly, can. The act of saying “this paragraph was mine; this one I took” is itself the beginning of integrity.

What did you learn from doing it this way, and what did the model do for you that you wanted to do yourself? This is the question that changes the situation from punishment to teaching. A student who says “I learned how to structure the argument, but I let the model write the conclusion because I was tired” has located the problem precisely. A student who says “I learned nothing because the model did all of it” has located a different problem precisely. Either is more useful than a marks deduction.

What would you do differently next time? End on agency, not shame. The student leaves with something they own.

This is not softness. It is the harder pedagogy. A student who feels they have been heard, who has named the line they crossed, and who has decided what to do differently is more likely to be on the right side of the line for the rest of the year than a student who has been processed by the discipline register.

The honest framing

AI in students’ hands is not going away. The chatbots are free, they are on every phone, and the same older cousin who once helped with homework now has Claude open in another tab. Pretending otherwise will not protect a single child from the harms of unthinking AI use. What protects them is a teacher who has thought clearly about it — who has a specific policy, an assessment design that makes learning visible, the patience for the conversation when something goes wrong, and the wisdom not to treat a detector score as evidence. The job is to shape the use, not to wish it away.

Check your understanding

Quick check

—

An online AI detector flags a Class-10 student’s English essay as 87 percent AI-written. The student has been improving steadily through the year. What is the most professionally sound response?

Treat the score as evidence, deduct marks, and inform the parents the same day.
Treat the score as one weak signal, talk to the student about how she wrote the essay, and look at her in-class drafts before deciding anything.
Ignore the detector entirely — they are useless, so the essay should be graded normally without further thought.
Report the student to the head teacher for disciplinary action based on the detector score.

Quick check

—

Which kind of written classroom AI policy is most likely to actually work over the course of a school year?

A blanket ban — students may never use AI for any school-related work.
A vague principle — students should use AI responsibly and exercise judgement.
A specific policy that names, for different assignment types, what AI use is allowed, what must be disclosed, what is not allowed, and what the teacher will check.
No policy at all — these questions should be left to each student and family.

What comes next

The policy and the conversation are how you shape the use. The deeper question is what teaching is for once AI is in the room — what parts of your work, if any, the model can absorb, and what parts it cannot touch. The next section is an honest account of what AI cannot replace in a classroom, and why that part is the part that matters most.