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 do not want BTT AI to win arguments with parents.
I want it to help them see more clearly.
There is a difference.
A parent may arrive saying:
“I think she needs tuition.”
Perhaps she does.
Perhaps she does not.
The AI’s job is not to push the parent toward the answer we prefer.
It is to make the decision easier to inspect.
Good advice should make the reasons visible
If the AI recommends tuition, I want it to explain why.
“The same conceptual gap is appearing across several topics, and she cannot currently repair it independently.”
That is something a parent can examine.
If the AI recommends waiting, explain that too.
“The marks are stable, the errors are narrow, and her independence is increasing. I would try a smaller intervention first.”
Again, inspectable.
The parent should be able to understand the recommendation well enough to disagree with it intelligently.
I do not think persuasion is the same as help
This matters especially because the AI sits on a tuition website.
There is an obvious temptation.
Every educational concern can be made to sound like a reason to buy more education.
I do not want that.
If a family chooses tuition, I want the choice to survive daylight.
It should still make sense after the fear has settled.
After the parent has looked at the marked paper.
After the alternatives have been explained.
That is a stronger relationship than conversion by urgency.
More control can mean fewer options
This sounds contradictory.
It is not.
If the parent begins with fifteen possible explanations and the AI helps reduce them to two plausible ones, she has more control.
Not because choices disappeared.
Because the remaining choices mean something.
If one route is clearly strongest, the AI can recommend it.
But it should still make the assumption visible:
“I’m recommending this because the evidence currently points here. If that detail is wrong, the recommendation changes.”
The AI should never make the parent feel trapped by its confidence
Fluent language can sound authoritative.
That is useful when the answer is solid.
It can be dangerous when the evidence is thin.
I want phrases like:
“From what you’ve shown me…”
“The current evidence suggests…”
“I would start here, but this is the detail that could change my view.”
Not as nervous disclaimers.
As accurate boundaries.
I think a good BTT AI conversation should leave the parent with less noise.
A clearer picture.
A smaller number of meaningful choices.
A reason for the recommended next step.
And the freedom to say:
“No. I know my child. That part does not fit.”
If the AI cannot survive that sentence, I would not trust it very much.
The goal is not obedience.
It is better judgment in the hands of the person who still owns the decision.
More control should show up as better observation
After a useful conversation, the parent should know what to watch next.
Maybe it is whether the child can now begin unfamiliar questions without a prompt.
Maybe it is whether the same algebra error appears again under timed conditions.
Maybe it is whether the student can explain the method before using it.
The next step should be small enough that the family can observe the return.
Then the parent knows what evidence would justify staying the course, changing the plan, or asking for more help.
That is what I mean by agency: not merely being offered choices, but being given a clearer way to judge them.