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.
If an AI gives you a link, I want you to know why that link belongs to your question.
Not because it contains the same keywords.
Not because it is one of our newest pages.
And certainly not because the AI felt obliged to produce something.
I want the connection to make sense to you.
A link without a reason still leaves you doing the diagnosis
Imagine you say:
“My daughter understands A-Math once the teacher starts the question, but she cannot begin unfamiliar questions by herself.”
The AI could reply with five articles about A-Math.
Technically useful.
But now you still have to decide which one matters.
I would rather it say:
“I’m choosing this page because what you described sounds less like a calculation problem and more like a starting problem: she can continue once the structure is identified, but she is not yet identifying that structure independently.”
Then give you the page.
Now you know what the AI thinks it heard.
And if that interpretation is wrong, you can correct it.
That explanation is part of the trust
I do not need the AI to show me its internal machinery.
I do need enough of the reasoning to judge whether the recommendation fits.
There is a difference.
Parents should not have to accept:
“This is the page.”
as though the system has become an oracle.
A better answer is:
“This is the page, because this particular distinction matches what you described.”
A recommendation becomes more useful when the reader can see the bridge between the evidence and the route.
I also want it to tell you what would change its mind
This is important.
Suppose the AI thinks your child has a starting problem.
Then you say:
“Actually, she usually knows how to start. Most of the marks disappear in the last third of the paper because she runs out of time.”
Good.
The route should change.
The system should not defend the first page because it already recommended it.
It should say:
“That changes my reading. I would look at exam execution and time pressure instead.”
I like systems that can change their minds for a visible reason.
That is much more reassuring than a system that always sounds certain.
The page should solve a specific job
On Bukit Timah Tutor, we already have articles that do different jobs.
Some explain a mathematical idea.
Some help a parent interpret a pattern.
Some help a student recognise a particular kind of difficulty.
Some are better for reassurance.
Some are better for repair.
The AI should know the difference.
If your child can finish but cannot start, for example, the useful article is not simply “an A-Math article”. It is the page about that exact pattern:
Additional Mathematics | The Student Who Cannot Start but Can Finish.
That specificity is the point.
I would rather one well-explained route than five vaguely relevant ones
This may make the AI look less busy.
I am comfortable with that.
If one page is clearly the best fit, give me one page.
If two genuinely different possibilities remain, perhaps give me two and explain the distinction.
But do not turn uncertainty into a menu merely because menus look comprehensive.
A parent who arrives uncertain does not necessarily need more choices.
Often she needs a clearer reason to choose.
If the AI recommends a page, I want the experience to be very simple.
“I heard this.”
“That makes me think this distinction matters.”
“So I am sending you here.”
“If this detail is different, tell me, because I would choose another route.”
That is enough.
You remain able to disagree.
The system remains able to correct itself.
And the page arrives with a reason, not just a URL.
The recommendation still needs a receipt
A page can feel exactly right and still fail to change the learner.
So after reading, I want one small test under changed conditions.
If the page was meant to improve method recognition, give one unfamiliar question without naming the topic. Can she now identify the structure and choose the method herself?
If yes, the recommendation was useful in the way that matters.
If no, the page may still have been relevant, but it did not yet solve the learner’s actual job.
That is when the AI should update rather than simply recommend more reading.