AI design essay: This article sets out the standards we would want educational AI to meet. The examples illustrate proposed behaviour; they do not establish that a deployed BTT AI currently performs these functions. Mathematics advice should be checked against the learner’s actual work and the teacher’s evidence.
“I don’t understand Mathematics” is a very large sentence.
It can mean almost anything.
“I do not understand what the question is asking.”
“I know the topic, but I cannot choose the method.”
“I know the method, but my algebra falls apart.”
“I can do it when someone shows me the first step.”
“I understood this last week and cannot retrieve it now.”
“I actually know how to do it. I am just scared I am wrong.”
All of those can feel identical from inside the child.
Stuck.
But they are not the same teaching problem.
So I do not want the AI to hear “I don’t understand” and start teaching immediately
I want it to help make the sentence smaller.
Perhaps it asks:
“What is the last part of the question or working that still makes sense to you?”
That question can change the tone immediately.
Because now the child does not have to diagnose an entire subject.
She only has to locate the edge of what she still knows.
“I know these are simultaneous equations.”
Good.
“I know I need to eliminate one variable.”
Good.
“I cannot see which equation I should multiply first.”
Excellent.
Not excellent that she is stuck.
Excellent that the uncertainty now has an address.
A smaller problem is usually a kinder problem
I mean that quite literally.
“I am bad at algebra” is emotionally enormous.
“I keep distributing the negative sign wrongly when there is a bracket” is irritating, but manageable.
“I do not understand differentiation” sounds like a whole chapter has collapsed.
“I can differentiate the function, but I do not know how the derivative gives me the tangent” is one missing bridge.
That difference matters for the teacher.
It matters for the parent.
And perhaps most of all, it matters for the child.
A difficult subject becomes less frightening when the difficulty acquires edges.
The AI should not turn every uncertainty into a diagnosis either
This is another boundary I care about.
A child says:
“I always make careless mistakes.”
That is useful evidence.
It is not automatically true.
Maybe the same “careless” error is actually a recurring misunderstanding.
Or a method-selection problem.
Or a weak checking routine.
The AI should treat the student’s self-description as one piece of evidence and compare it with what the work shows.
Not contradict the child for sport.
Not believe every label automatically.
Just keep the picture correctable.
Sometimes the child cannot explain the gap in words
That is fine too.
I do not want an AI that rewards articulate students and frustrates quiet ones.
The system can ask for something observable instead.
Show the last line you trust.
Which part of the diagram is confusing?
Can you start the question if I do not tell you the topic?
Did the mistake happen before or after you chose the method?
Good diagnosis should not depend on a child sounding sophisticated.
Eventually, I want the AI’s question to become the student’s question
This is the part that interests me most.
At first, the AI asks:
“What do you know?”
“Where exactly did you get stuck?”
“What are the two methods you are deciding between?”
Later, the student begins asking those questions internally.
I know this.
I am unsure here.
I have two possible routes.
This result needs checking.
I can continue from the last secure line.
At that point, the AI has done more than answer questions.
It has helped the student acquire a better way to be uncertain.
There is already an article on this site about the sentence that comes after “I don’t understand”. The reason I keep returning to the idea is simple: it tells us something important about what educational AI should be for.
Not a machine that rushes to fill every blank.
A system that helps the learner locate herself before deciding what kind of help belongs.
So when your child says:
“I don’t understand Mathematics,”
I hope the AI does not hear the conversation ending.
I hope it hears the beginning of a much better question.