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 think one of the hardest things for an educational AI to say should also be one of the most important.
“I don’t think your child needs more tuition right now.”
That sentence matters because it tells you what the system is for.
If every parent question eventually ends with “add another class”, then the AI is not really helping you decide.
It is selling with extra steps.
Sometimes the evidence points somewhere smaller
Imagine you tell the AI:
“My daughter is getting around 70 for Mathematics. She understands most topics, starts homework independently, and the recent mistakes are mostly from rushing near the end of timed papers.”
There are many things the AI could recommend.
More tuition.
More worksheets.
More classes.
A new programme.
But from what you have described, I would first want to ask whether the problem is actually instructional.
If the concepts are secure, the methods are selected independently, and the errors appear mainly under time pressure, perhaps the next useful step is narrower:
two weeks of timed mixed practice;
a simple checking routine;
then look again.
The intervention should not become larger than the evidence.
I would want the AI to distinguish support from dependency
Tuition can be extremely useful.
I teach tuition.
I have seen what good support can do when the problem is correctly identified.
But support is not automatically better because there is more of it.
If a child can increasingly begin, choose methods, check, recover and learn without help, I do not want an AI to mistake that growing independence for an opportunity to add another layer of support.
The direction matters.
Are we helping the child carry more herself?
Or quietly moving more and more of the work outside her?
There are also times when I would absolutely recommend help
I do not want this article to become another simple slogan.
“Tuition is unnecessary” is no more intelligent than “everyone needs tuition”.
If a student repeatedly cannot begin questions because the concepts are missing, help may be useful.
If school work is moving faster than the learner can repair accumulated gaps, help may be useful.
If the student understands during lessons but cannot retrieve or transfer the learning later, structured support may be useful.
If parent and child are spending every evening in conflict over Mathematics, outside support may be valuable for reasons beyond the worksheet.
The point is not “never add tuition”.
The point is: earn the recommendation.
I would especially watch what the child can do when nobody is beside her
Suppose the marks are 75.
One child gets there by working independently, checking errors and recovering from unfamiliar questions.
Another gets the same 75 because a parent sits beside her, a tutor previews every topic, notes are always open, and someone identifies the method whenever she hesitates.
The score is the same.
The support burden is not.
That is why the AI should not treat marks as the whole child.
I want it to ask what the mark costs to produce.
And whether that cost is rising or falling.
A trustworthy answer may be temporary
“I would not add tuition yet” does not mean “nothing can ever change”.
A child can move into harder work.
School demands can change.
A new pattern can emerge.
So the AI might say:
“From what you have described, I would not add another Mathematics class yet. Try this smaller intervention for two weeks, then look at whether the same errors are still appearing.”
I like that answer because it preserves both humility and movement.
It does not pretend to know the future.
It gives you a reasonable next step.
And it leaves you free to decide.
This may be where I judge the AI most harshly
Not when the answer is easy.
When the honest recommendation conflicts with the easiest commercial outcome.
If the evidence says the child probably needs help, say so calmly and explain why.
If the evidence says wait, say wait.
If the evidence says the school teacher is the right person to speak to first, say that.
If the evidence says the problem is not yet clear, do not manufacture certainty.
That is how I would want a trusted human adviser to behave.
I do not see why we should ask less of the AI.
There is a very simple test for the kind of AI I want on Bukit Timah Tutor.
After the conversation, does the parent understand the situation better?
Are the options clearer?
Is the next step proportionate?
And does the parent still feel free to choose?
If yes, I think the AI has done something useful.
Even when the useful answer is:
“Not more tuition. Not yet.”