AI Question Generator

A good question is a measuring instrument, and like any instrument it can be built wrong: too easy and it measures nothing, too hard and it measures the reader's willingness to guess. The useful distinction is not difficulty but what the question actually requires — recalling a fact, understanding a relationship, or applying knowledge to a situation that was not in the text.

Three levels, three different jobs

Recall questions verify that something was read. They are cheap to write, easy to mark, and almost useless as evidence of understanding, which is why an exam made only of recall questions rewards memorisation and punishes thinking. Their legitimate place is a warm-up or a check on vocabulary.

Understanding questions ask for a relationship: why this happens, what would change if that were not true, how two ideas connect. They are harder to write because a wrong answer has to be plausible. Application questions put the idea in a new situation the reader has not seen — the same principle in a different context — and they are the only kind that predicts whether knowledge transfers.

A balanced set is roughly one recall question to start, six or seven understanding, and two or three application. That mix also reads better as an assessment, because the difficulty climbs instead of jumping around.

Distractors are the real work

In multiple choice, the correct answer is the easy part. The wrong options decide whether the question measures anything: if three options are obviously absurd, the item can be answered without knowing the subject. Good distractors come from real misunderstandings — the ones you have heard in class, in review comments, or in your own inbox.

The tool writes four options and is instructed to make each distractor plausible. Check them anyway: an option that is technically also true makes the item ambiguous, and an option that is silly makes it free.

Open questions need a marking rule

An open question without a defined answer is not assessment; it is a discussion. Before using one, write what a good answer contains, in three or four points, and how many of them a pass requires. This takes two minutes per question and protects you from the far longer conversation that follows a mark somebody disputes.

The answer key produced here is exactly that: the elements a complete answer needs, not a model paragraph. Use it as the marking rule rather than as something to read out.

Where generated questions fail

They are drawn from what the model knows about the topic, not from the text you actually taught. If your lesson used a particular diagram, dataset or example, the questions will drift to the general case and stop matching your assessment. Paste the material you want measured into the subject field, in your own words, rather than naming the topic.

They also tend to be grammatically uniform, and a paper where every question has the same shape gives away its structure. Change three questions by hand — invert one, make one a comparison, add a “why” — and the set stops looking generated without changing what it measures.

Reusing a set

Do not reuse the same paper twice; students talk, and the second sitting measures the conversation rather than the learning. Rotate the order, replace half the items each term, and keep a note of which questions were answered badly across a whole class. Those items are either the hardest material or the worst written questions, and both deserve a second look.

Marking what you wrote

After a test, count how many students got each question wrong. A question everyone answered correctly measured nothing; one that everyone failed is either the hardest idea in the unit or a badly written item, and those two need different responses. Read the failed question aloud and check whether the wording could be read two ways. If it can, the item was the problem. Keep a note of both kinds each term: the list becomes a map of what your teaching actually landed, which is more useful than any single set of results.

Questions for a topic you do not teach

If a colleague asks for a set on a subject you only half know, generate the questions and then read the answer key against a textbook chapter rather than trusting it. The model is confident about plausible details, and a wrong key in someone else’s classroom is a favour that costs them a lesson. Sign the set with your name only after you have checked every answer, or pass it on with a line saying which items you verified.