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 kind of answer that looks very helpful and leaves the parent more tired than before.
You ask one question.
The AI gives you twelve options.
Read this.
Try that.
Maybe speak to the teacher.
Maybe get a tutor.
Maybe revise fundamentals.
Maybe do timed papers.
All plausible.
And now you have a new problem.
Which one?
Comprehensiveness can be a way of avoiding judgment
I understand why systems do this.
If the AI lists every reasonable possibility, it is less likely to omit something.
But education does not become useful merely because nothing was omitted.
Sometimes the difficult part is choosing.
If the evidence strongly supports one route, I want the AI to take responsibility for saying so.
Good curation is not the absence of options. It is knowing when most of them can safely stay out of sight.
Imagine the parent is already carrying too much
This is easy to forget.
The parent may already be thinking about marks, school demands, subject combinations, schedules, money, confidence and whether she is overreacting.
Then the AI adds twelve more decisions.
That is not always empowerment.
Sometimes it is outsourcing the system’s uncertainty back to the parent.
I would rather the AI say:
“From what you’ve told me, I would start here. I’m not recommending the other routes yet because this one tests the most important uncertainty first.”
That feels much calmer.
One route does not mean one permanent answer
This matters too.
Choosing one next step is not the same as pretending every alternative is false.
It simply means sequencing the work.
Try the smallest useful test first.
See what comes back.
Then update.
If the first route does not explain the situation, the second route can appear later.
The parent does not need the entire decision tree on the first screen.
There are times when two options are genuinely useful
Suppose the evidence does not clearly distinguish between two possibilities.
Then yes, show both.
But explain the fork.
“If she cannot recognise the method even without time pressure, read this page.”
“If she recognises it but cannot finish under exam conditions, start with this other one.”
Now the two options are doing different jobs.
That is not clutter.
That is a useful distinction.
I want the AI to spend complexity internally
This may be the simplest way I can explain the design I want.
The system can compare many possibilities.
It can search many pages.
It can hold several explanations in mind.
It can check which detail would change the route.
Fine.
But after all that internal work, the parent may need only:
one sentence of orientation;
one useful page;
one next action.
The complexity should be doing work for the reader, not appearing as work the reader now has to do.
So if our AI knows one route is clearly better, I want it to say so.
Calmly.
With a reason.
And without pretending the future has become certain.
Parents should still be free to disagree.
But they should not have to do the AI’s curation for it.
Sometimes twelve options are twelve ways of saying:
“I haven’t decided.”
I would rather the system do a little more thinking first.
One route should produce one observable return
If the AI chooses one route, it should also make clear what happens next.
For example: practise one small set under the exact condition we think is weak, then look at what changes.
If the student now starts independently, good. If she still cannot start, the first route has not yet solved the problem.
That gives the parent a clean loop: choose one route → do one thing → observe the result → update.
The point of reducing options is not merely calm. It is to make the next piece of evidence easier to read.