An AI Quiz Generator That Knows Whether the Answer Is Right

Status: in development. The AI-Math quiz generator is not live yet. This page describes what it will do and, more usefully today, what a good generated math quiz actually looks like — plus the free tools on this site you can practise with right now.

What one generated question should look like

Most AI quiz generators produce questions. Very few produce questions where the answer key is trustworthy, the wrong options are meaningful, and the solution is included. Here is the standard we are building to, as a single item:

Q3 (medium, quadratics / applications). A ball is thrown so its height in metres after t seconds is h(t) = −5t^2 + 20t + 1. At what time does it reach maximum height?
A. t = 4 s B. t = 2 s C. t = 20 s D. t = 21 s

Answer: B.

1. The graph is a downward parabola (a = −5 < 0), so the maximum is at the vertex.
2. Vertex time: t = −b / (2a) = −20 / (2 · −5) = 2 seconds.
3. Check the height if you want it: h(2) = −5(4) + 40 + 1 = 21 metres.

What each wrong option means. A comes from using −b/a instead of −b/(2a). C comes from reading off b directly. D is the maximum height, 21 m, answered in place of the time — a question-reading error, not a math error.

That last block is the part almost no generator produces, and it is the part that makes a quiz worth taking. When a student picks D, they do not have an algebra problem; they have a habit of answering before finishing the question. Those need different fixes, and a quiz that cannot tell them apart just reports a score.

Why general quiz makers struggle with math

The generic tools — the ones aimed at trivia, training modules and reading comprehension — fail on math in three specific ways.

They generate the answer instead of computing it. A language model asked to write a math question and its answer will produce something that reads like a correct key. On multi-step algebra and calculus this is unreliable enough to matter, and the failure is silent: the question looks fine, the key is wrong, and nobody notices until a class has been graded against it. Our approach is to route every generated item through the solver and discard anything the engine cannot verify.

Distractors are random. Generic makers perturb the correct answer — add one, negate it, double it. Sometimes that accidentally matches a real misconception; usually it produces an option no student would ever pick, which makes a four-option question effectively a two-option one and inflates scores.

Notation degrades. Anything with a fraction, an integral, a matrix or a summation tends to come back as flattened text. On a printed worksheet that is not a cosmetic issue: an ambiguous expression is a question with two defensible answers.

There is also a difficulty problem. Ask a generic tool for a "hard" version and you usually get the same problem with uglier numbers. Genuine difficulty in math comes from structure — an extra substitution, a constraint you have to notice, a setup buried in a word problem — not from arithmetic weight.

What you can practise with today

The generator is not ready, but the solvers behind it are live and free, and they already close the practice loop if you bring your own problems from a textbook or past paper: work the problem yourself, then check every line rather than just the final answer.

• Linear equation solver — for drilling the rearrangement steps where most quiz mistakes originate.
• Quadratic equation solver — vertex, discriminant and roots, all derived.
• Integral calculator — substitution and by-parts choices explained rather than asserted.
• Probability calculator — the topic where a plausible-looking wrong answer is hardest to catch by eye.

For exam practice specifically, SAT and ACT math prep with AI feedback loops describes how to turn a past paper into a repeatable drill without a generator. Teachers in the same position — needing items faster than they can write them — will get more out of pulling problems from an existing bank and using the solvers to produce clean worked keys.

Honest status

No date. The blocker is verification, not generation: producing a hundred plausible questions is easy, and producing a hundred questions whose keys are all correct is the actual product. We would rather ship that a year later than hand a teacher a worksheet with three wrong answers in it. The waitlist gets one email when it opens.