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.
There is a sentence I want BTT AI to be completely comfortable saying.
“I don’t know yet.”
Not as an escape.
Not because it cannot be bothered to think.
Because sometimes the evidence really is not enough.
And I would rather see uncertainty stated cleanly than hidden underneath a confident paragraph.
Parents often ask questions that sound more complete than they are
“My daughter dropped from 78 to 61. What happened?”
I understand why you want an answer.
But from those two numbers alone, there are many possibilities.
The paper may have been harder.
One heavily weighted topic may have gone badly.
She may know the Mathematics and have run out of time.
She may have made repeated algebra errors.
She may have understood familiar questions and struggled with transfer.
Perhaps something genuinely has changed.
Perhaps very little has.
The honest answer may therefore be:
“I can see the drop. I can’t tell you the cause from the score alone.”
That is not a weak answer.
It is an accurate one.
But “I don’t know yet” should lead somewhere
This is important.
I do not want an AI that hides behind caution.
If it does not know, it should ask what would help it know.
Perhaps:
“Can you look at the marked paper and tell me whether most of the lost marks came before she had a workable method, or after?”
That question earns its place because the answer can genuinely change the route.
If she could not find a method, we investigate one set of possibilities.
If the method was sound and execution failed, we investigate another.
Uncertainty has now become productive.
I want confidence to track evidence
If the AI has one weak clue, it should sound tentative.
If it has a repeated pattern across several papers, it can speak more firmly.
If the student shows the actual working and the same error appears five times, confidence can increase again.
The language should move with the evidence.
“One possibility is…”
Then perhaps:
“The pattern now points more strongly toward…”
And eventually:
“Yes. This is the recurring issue I would repair first.”
I do not want confidence as a personality trait. I want confidence as a receipt from the evidence.
Sometimes parents need permission not to know immediately either
This may be the quieter benefit.
When a result goes badly, there can be pressure to decide what it means that same evening.
New tutor?
More practice?
Change class?
Something is wrong?
Sometimes the sensible answer is:
We do not know yet.
Let us inspect the paper tomorrow.
That is not neglect.
It is refusing to let urgency outrun understanding.
I do not want BTT AI to say “I don’t know” often.
I want it to say it accurately.
And when it does, I want the next sentence to be useful.
“Here is the one thing that would help us know more.”
That combination—uncertainty plus direction—feels much more trustworthy to me than certainty manufactured for the sake of sounding capable.