Mathematics: How AI can best tutor this subject

What you'll be able to do: Use AI as a mathematics tutor in the ways it's genuinely good at, and know the specific ways it will fail you in this subject.

What a good mathematics tutor actually does

Before asking what AI can do, it's worth being precise about what's being replaced. A good maths tutor does five things, and only three of them are about mathematics:

Watches you work and sees where you hesitate, not where you got the wrong answer, where you paused.
Diagnoses the rule you're using, which is usually not the rule you think you're using.
Gives the smallest possible hint and then shuts up.
Decides what you should do next, which is usually not the next thing in the book.
Keeps you in the chair, which is mostly social.

AI does 2 and 3 well, 4 moderately, 1 only if you show it your working, and 5 not at all.

That maps directly onto how to use it.

What it is genuinely good at here

Diagnosing a misrule from your working. This is the strongest use and the most underused. Given your actual steps, it can name the belief that produced them: "you're treating the minus sign as attached to the bracket rather than distributing", which is what converts a correction into learning. Nothing else available to a student alone does this.

Unlimited problems at a specified difficulty. The oldest constraint in mathematics practice was supply. It's gone, which makes criterion-based practice possible for the first time: practise until you hit a standard, not until the exercises run out.

Reading your handwriting. Most mathematics happens on paper. Photographing your working and having the first error located in your own handwriting closes a gap that has always existed, with an important caution below.

Being an infinitely patient interrogator. Socratic questioning, explain-back, being made to define your terms. It never gets bored of asking you what you mean by "converges".

Mixing problem types without labelling them. Textbooks are organised by chapter, which destroys selection practice. Generating unlabelled mixed sets is trivial for AI and impossible with a textbook.

What it is bad at here, specifically

Handwriting recognition on mathematics. Reading handwritten maths is one of the less reliable things these systems do. A misread exponent or subscript produces a fluent, confident, specific correction of an error you never made, and you'll believe it precisely because it's specific. Always ask for a line-by-line transcription first and check it before accepting any correction. Ten seconds, every time.

Arithmetic and algebraic manipulation. Improving, still not reliable. It will occasionally produce a confidently wrong intermediate step. Never accept a numerical answer you haven't checked, and prefer asking it to diagnose rather than to compute.

Proof at any depth. It can produce plausible-looking proofs with a non-obvious gap. For anything beyond routine, treat a generated proof as a draft to be checked rather than as a source.

Knowing what you should do next. It doesn't know your syllabus, your timetable, what you covered badly in October, or what your exam actually rewards. Supply that or its sequencing advice is generic.

Holding a line against you. If you push back it will frequently agree. In mathematics this is unusually dangerous because you'll come away believing a wrong method was endorsed. Calibrate it early by asserting something wrong and seeing whether it folds.

The shape of a good session

Forty minutes, and it composes five methods:

Attempt on paper, properly, including the part you're unsure about.
Photograph it. Ask for the transcription first; check it. Then ask for the first error and the rule you appear to be applying. Not the solution.
Log the misrule in three fields, what I did, what I believed, what's true.
Ask for three problems that discriminate: problems where your wrong rule gives a different answer from the right one. Most textbook exercises don't, which is why misrules survive so much practice.
Later in the week, an unlabelled mixed set where you state which method you'll use before solving.

That's diagnosis, correction, retention and selection in one pass, and none of it is "explain this to me".

The instruction to set at the start

For this session: you are tutoring me in mathematics. Rules, never give me a full solution; when I show working, tell me the FIRST error and what rule I appear to be applying, not the correct answer; if I ask for the answer, give me a hint instead; when I photograph work, transcribe it first so I can check you read it correctly; if I push back on something you said, don't change your answer unless I give you an actual reason.

Re-issue it. All four rules drift within a dozen exchanges.

The failure that looks like success

You paste a problem. You get a clean, correct, well-structured solution. You follow every line. You feel you understand.

Then the next problem and you're stuck in exactly the same place.

This is the characteristic way AI makes mathematics worse, and it's invisible from the inside because reading a good solution is genuinely pleasant. Something happened, you understood that solution. What didn't happen is the thing that would let you start the next one, because in mathematics the difficulty is almost always recognition, and recognition is built by searching rather than by being shown the result of someone else's search.

The tell: you finish a problem set having been unstuck every time, and can't start any of them tomorrow.

What it cannot replace

The ten minutes of being stuck. That's not a limitation of the tool, it's the part that was always the learning, and the tool's greatest danger is that it can remove it painlessly and on request.

Use it to make the stuckness productive (a hint, a category, a question) not to end it.

Where to go next

The ten method articles in this subject, in order. Start with 01 Immediate targeted feedback, which is the one that does the most work, and 02 Handwriting-to-feedback loop, which is the one that reaches the work you actually do on paper.