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 become attached to being right the first time.
I want it to become attached to getting the picture right.
Those are not the same thing.
A parent may first say:
“She keeps losing marks because she rushes.”
Fine.
The AI might reasonably begin by looking at checking routines, time pressure and execution.
Then the parent uploads or describes the working more carefully.
And the pattern changes.
The student is not merely rushing.
She is repeatedly choosing the wrong relationship before the arithmetic even begins.
Now I want the AI to say:
“That changes my reading. I would no longer treat speed as the main problem.”
A correction is not an embarrassment
I think we sometimes expect intelligent systems to sound seamless.
No visible uncertainty.
No revision.
No “I thought this, but the new evidence points elsewhere.”
That sounds polished.
It is not necessarily intelligent.
In Mathematics, if new evidence contradicts the current route, we do not protect the route because we spent time on it.
We inspect the contradiction.
The AI should behave the same way.
The first answer should be a working model, not a promise never to revise.
I want the change to be visible enough to understand
Not a silent switch.
If the recommendation changes, tell the parent what caused the change.
“Earlier I thought this was mainly an execution problem because you said she usually knew how to start. The marked work now shows repeated method-selection errors, so I would move concept recognition ahead of speed work.”
That is enough.
The parent can follow the update.
The system has not become less trustworthy by changing its mind.
It has shown what it is answerable to.
Old labels should expire too
This matters over longer periods.
A student may genuinely have had weak algebra six months ago.
That does not mean every future difficulty should be interpreted through weak algebra.
If the recent work shows the algebra is now secure, the description should change.
The AI should not keep resurrecting an old diagnosis simply because it has history.
History is useful.
Current evidence still gets a vote.
If BTT AI says something on Monday and better evidence arrives on Thursday, I want Thursday to win.
Not because Monday was careless.
Because good judgment remains open to correction.
That is a quality I value in teachers, students and systems alike.
Being right matters.
Being able to become more right matters too.