What maths AI teaching needs to solve

The hard part of learning maths often sits not in whether the final answer is right, but in one step along the way. Two students get the same equation wrong: one mishandled a sign while moving terms across, the other never grasped factorisation. For a teacher, working through thirty students' calculations one by one to find where each got stuck is the most time-consuming task, and also the one with the most teaching value. The bigger the class, the harder it is to attend to each student's specific breaking point.

The AI for Empowering Learning and Teaching Funding Programme asks schools to run AI-assisted teaching in at least three subjects. Maths is one of the suitable subjects, but the entry point has to be chosen well: not handing problem-solving over to AI, but using AI to widen a teacher's ability to diagnose student mistakes.

Step diagnosis: AI diagnoses, it does not give the answer

The right framing is to let AI go through a student's working step by step, point out which step the solution first goes wrong at, and what kind of error that step is, rather than doing the problem for the student. What the teacher gets from this is not a pile of right-or-wrong marks, but an error map of the whole class: who is stuck on the same concept, and which mistakes are just slips in calculation.

This line has to be clear: diagnosis points at the error, it does not reveal the answer. Students still have to work through the solution themselves. The role of AI is to tell the teacher and the student where the break is, not to do it for them. If a tool is built as a solving machine that answers every question, the use in maths shifts from supporting learning to replacing thinking.

Building varied practice from real mistakes

After diagnosis, the next step is practice aimed at the break. The traditional approach gives the whole class the same worksheet, which wastes time for students who already have it and may miss the point for those who are stuck. Generating varied practice based on students' real mistakes lets the same concept come back again and again with different numbers and different contexts, until the student truly has that step.

The value of varied practice is that it keeps the structure of the question type the same while changing the surface conditions, forcing students to recognise the underlying solving step rather than memorising how one particular question is done. A teacher can pair a set of variations with each of the few most common breaks in the class and hand them out by level, without rewriting every question by hand.

How teachers reteach from error patterns

Error patterns are the basis for reteaching. When diagnosis shows a third of the class stuck on the same step, that is not thirty separate problems but one concept that needs to be taught again. A teacher can design a targeted explanation around this shared break, which is far more effective than correcting each student one by one. Individual, scattered mistakes go back to individual follow-up.

This is also a teaching demonstration you can write into a programme report: use AI diagnosis to find the shared mistake, then respond with varied practice and targeted reteaching, forming a complete teaching cycle. The report should focus on how far the work has progressed and on the teaching design, not on a feature list of the tool itself.

Rules for using AI in class

  • AI diagnoses the break, it does not reveal the answer: students still finish the solution themselves, and the tool only points out which step is wrong
  • Try first, check the diagnosis after: require students to complete a full solution once before comparing it with the AI diagnosis, to avoid dependence
  • Aim varied practice at the break: hand it out for the shared mistakes that diagnosis found, not the same worksheet for the whole class
  • Teachers review the diagnosis: the AI's judgement of a step can still be wrong, so important decisions about extra teaching should be confirmed by the teacher
  • Hold the line in assessment: tests and exams stand on the final judgement of hand marking, and AI is used only for everyday diagnosis and practice