What a good AI tutor refuses to do

Every AI product in education right now is racing to show what it can do. We're more interested in what a good one should refuse to do.

Ask Numera is the AI math tutor inside Numera, the Ontario MPT prep app we build. It's small, scoped, and a little stubborn. The stubbornness is the design — what follows is what it refuses to do, and why each refusal is a feature rather than a missing one.

Refusal 1: it refuses to do the work for you

Type a math problem at most AI chat tools and you'll get a clean answer in three seconds. That looks like a win. For a student actually trying to learn the material, it's the opposite — a right answer with no understanding is worse than a wrong answer arrived at honestly. The first one trains the student to outsource thinking; the second one is where actual learning happens.

So Ask Numera's most-used prompt isn't "solve this." It's "diagnose this" — you paste your wrong attempt, not just the question. The reply takes the path you tried, finds the place your reasoning broke, and rebuilds the concept from there. The correct answer comes with the step-by-step working and the curriculum reasoning attached, never stripped out.

A tutor that does the work for you isn't a tutor. It's a homework laundry service.

Refusal 2: it refuses to talk about anything other than math

Ask Numera is scoped. Paste a history question, an essay prompt, or "write me a cover letter," and it declines. The product surface is one subject, one curriculum, one exam family.

That's a real design choice and it costs us a marketing point: every competitor's pitch starts with "your AI tutor for everything." Ours starts with "your AI tutor for one thing, done well."

We think the maximalist version is the worse product. A general-purpose AI tutor has more ways to hallucinate, more ways to be misused for cheating in subjects the AI wasn't designed for, and a lower pedagogical floor because it can't be specialised to a curriculum the student is actually being tested on. The tutor that can do anything tutors nothing well.

Refusal 3: it refuses to invent a curriculum

When Ask Numera explains a problem, it cites the Ontario curriculum expectation behind it — codes like B3.8 — prime factorization or C1.2 — patterns and expressions. The student sees the same scaffold the test itself was built on.

That's not decoration. It's a refusal to do the thing most AI education products quietly do: generate "personalised practice problems" out of thin air, trained for a test the student isn't actually taking. An AI that wanders off-curriculum produces fluency in a subject that doesn't exist on the exam.

The Numera question bank is human-authored and tagged to the Ontario expectations the MPT actually probes. Ask Numera works from inside that grounding. The boundary is intentional — we'd rather be useful for one real test than impressive at a generic one.

Refusal 4: it refuses to be the only source of truth

A surprising amount of what Ask Numera says is "this isn't the right tool for this — go do that instead."

Asked for an overall progress check, it suggests taking a mock test, because a mock will show the strand-by-strand breakdown the chat can't. Asked the same conceptual question three different ways, it eventually says the gap is structural and points the student at the underlying curriculum rather than running another loop. Asked "am I ready?" — it won't say yes, because that's not a question an AI tutor has standing to answer.

A tutor that's willing to say "I'm not the right tool for this" is more trustworthy than one that always has an answer. The refusal is the trust signal.

What this costs us

These four refusals are not free.

The general-purpose AI tutor demo is more impressive than ours. The competitor that solves a calculus problem in two seconds and then writes a five-paragraph essay about Romeo and Juliet looks like the future of education. Ours looks restrained next to that — sometimes deliberately slow, sometimes saying no.

Refusing to do the homework is harder to sell. A parent who wants their kid's worksheet finished by 8pm will pay a competitor before they pay us. A student who wants the answer fast doesn't immediately appreciate "here's where your reasoning broke."

Scoping to math and one specific Ontario exam means we can't credibly expand to "all subjects, all curricula" tomorrow without violating refusal three. Our addressable market is bounded by choice.

The trade we're making: a smaller, slower-growing product that produces students who actually learn the material. We think that's the right side of the trade. It's also the side that's harder to fund and slower to brag about.

The shape of the bet

AI in education is forking into two products. One is the homework-doer: maximalist scope, faster output, ads-funded or data-funded, optimised for the metric of completed assignments per minute. The other is the explainer: scoped, slower, paid for honestly by the people using it, optimised for the student understands the thing now.

We're building the second one. The four refusals are how you can tell which version of AI in education you're using.

Try Ask Numera free at app.numeracode.com. It only answers math, it won't write your essay, and it shows you the reasoning every time. That's the point.

#philosophy

What a Good AI Math Tutor Should Refuse to Do