Can ChatGPT make a question paper for school?

It can produce a list of questions on a topic in seconds, and for some purposes that is enough. It cannot reliably produce a board-pattern question paper, because it has not read your prescribed textbook, it does not hold the board's section structure and marks split, and it does not remember which questions it gave you last time. The questions it invents look correct whether or not they are, so the checking cannot be skipped.

Are AI-generated question papers reliable?

It depends entirely on where the questions come from. A tool that generates questions from a model's memory of a subject will produce plausible questions that are sometimes outside the syllabus, sometimes worded differently from the textbook, and occasionally wrong. A tool that draws questions from the prescribed textbook and has a person approve them before use is a different proposition. Ask any tool which one it is.

Can ChatGPT set your question paper? An honest answer

Plenty of teachers have already tried it, and the honest answer is not “no”. A chat assistant will give you twenty questions on photosynthesis in fifteen seconds, and there are jobs for which that is genuinely enough. The problem is that setting a board-pattern question paper is not one of them, and the reasons are structural rather than fixable with a better prompt.

We have not dressed this up as an experiment with screenshots. You can run it yourself in five minutes on a chapter you know well, which is more convincing than anything we could publish. What follows is what to look for when you do.

1. It has not read your textbook

It has read a textbook — a great many of them, plus notes, guides, forum posts and other people’s question papers, blended together. Ask for five questions from Class 10 Science Unit 4 and you will get five questions about that topic. Some of them will not be in your syllabus. Some will use terminology your prescribed book does not use. One will be quietly wrong in a way that survives a quick read.

This is the part that costs the most, because the errors are not obvious. A wrong question looks exactly like a right one. You will find out while marking, in front of a class that has already sat the paper.

The tell to check for: ask it which chapter a question came from, and which page. A tool that took the question from a book can answer. A model that generated it will produce a confident chapter number that means nothing.

2. It does not know the shape of the paper

A real board paper is not a list. It has named parts. It has a marks split that adds up. It has internal choice in certain places and not others, a fixed count of sub-questions in particular sections, a header, a register-number strip, and a print layout.

What comes back from a chat window is questions. Turning that into a paper — ordering it, assigning marks, placing the choice, formatting it, getting it onto two sides of a sheet without a section breaking across a page — is most of the work you were trying to skip. In our experience that formatting is the bulk of the hour, not the question-writing.

3. It cannot remember what it already gave you

Ask for a second unit test on the same chapters next month and you will get substantially the same questions. Not identical — reworded, reordered, and recognisably the same set, because you asked the same thing of the same model about the same chapter.

For a subject where you set a weekly test, this is the failure that actually bites. Rotation is a bookkeeping problem, and a chat window has no books.

4. The language subjects break first

If you teach Tamil, try it and watch what happens to the question types that have no English equivalent, to the அ ஆ இ ஊ option lettering, and to the header. Generic tools tend to produce an English-shaped paper with Tamil words in it. That is not the same object.

Where a chat assistant genuinely is the better tool

It would be dishonest to write this without saying so. Reach for it when you want to:

Reword a question you already have — make it easier, harder, or shorter.
Draft a comprehension passage at a particular reading level.
Translate or explain something for your own preparation.
Write the covering instructions at the top of a paper.
Think out loud about how to assess something you have not assessed before.

All of those are tasks where you are the one checking the output against something you already know. That is exactly the condition under which a model is useful and exactly the condition a generated question paper does not meet.

The question worth asking any tool

Not “does it use AI” — nearly everything does now, and it tells you nothing. Ask where the questions come from.

If the answer is “our AI generates them”, you are the proofreader. That is a real job, it takes real time, and it is the job you were trying to avoid. If the answer is “from the prescribed textbook, and a person approved each one”, you are checking a paper rather than auditing it — which is a different evening.

That is the bet Examaker is built on: questions taken from the book your students are holding, every one of them approved by a person before it can appear, assembled into the real board pattern rather than a list. How the question bank is built →