What an AI tutor should and should not do
October 10, 2026
Tunde Okafor
The first time you use an AI assistant for schoolwork, it feels like a small miracle. You paste in a question, and a clear explanation appears in seconds. There is no waiting for a teacher, no embarrassment about asking twice, no limit on how many times you can say "I still do not get it."
That is genuinely valuable. Anyone who has studied at night with no one to ask knows the gap it fills.
But there is a second pattern, and it is less helpful. You paste in the question, the answer appears, you copy it down, and two weeks later you cannot reproduce it. The explanation was correct. The learning did not happen.
An AI tutor should help you do the thinking, not replace it. In practice, that means a few specific behaviours.
It should ask what you have already tried. A student who has attempted the problem and made a wrong turn needs a very different answer from a student who has not opened the textbook. Asking first also slows the interaction down just enough for the student to think.
It should explain the step you are missing, not the whole problem. If you are stuck on factorising because you forgot the difference of two squares, the useful answer is a reminder of that rule and one example. Solving all twelve questions for you is efficient and useless.
It should admit uncertainty. Language models are confident by default, even when they are wrong. A good tutoring interface makes it easy to say "I am not sure, check your textbook" instead of inventing a plausible-sounding fact. We would rather a student leave with a question than with a wrong certainty.
It should not do the assessment for you. There is a place for AI help while learning, and there is no place for it in an exam or a marked assignment. A tool that blurs that line is not doing students any favours. Teachers notice, and more importantly, the student knows.
It should encourage you to close the chat and try. This sounds trivial, but it is the hardest part. The moment an explanation lands, there is a temptation to ask one more question, then one more. The best learning happens in the quiet minute after the chat closes, when you try to reproduce the idea without help.
We are also wary of the opposite failure: pretending AI is never useful. It is useful. It is patient, it can rephrase, and it can give you a second example when the textbook only gives one. Those are real advantages, especially for students who cannot afford extra lessons or live far from a good library.
The rule we keep coming back to is simple. A tutor is doing its job when the student needs it less over time. That is the measure. Not how many questions were answered, but how many the student can now answer alone.
If you use an AI assistant while studying, try this once. Ask for a hint instead of an answer, then put the phone down and finish the problem yourself. The feeling of getting there is the point, and it is the part no model can do for you.
That is genuinely valuable. Anyone who has studied at night with no one to ask knows the gap it fills.
But there is a second pattern, and it is less helpful. You paste in the question, the answer appears, you copy it down, and two weeks later you cannot reproduce it. The explanation was correct. The learning did not happen.
An AI tutor should help you do the thinking, not replace it. In practice, that means a few specific behaviours.
It should ask what you have already tried. A student who has attempted the problem and made a wrong turn needs a very different answer from a student who has not opened the textbook. Asking first also slows the interaction down just enough for the student to think.
It should explain the step you are missing, not the whole problem. If you are stuck on factorising because you forgot the difference of two squares, the useful answer is a reminder of that rule and one example. Solving all twelve questions for you is efficient and useless.
It should admit uncertainty. Language models are confident by default, even when they are wrong. A good tutoring interface makes it easy to say "I am not sure, check your textbook" instead of inventing a plausible-sounding fact. We would rather a student leave with a question than with a wrong certainty.
It should not do the assessment for you. There is a place for AI help while learning, and there is no place for it in an exam or a marked assignment. A tool that blurs that line is not doing students any favours. Teachers notice, and more importantly, the student knows.
It should encourage you to close the chat and try. This sounds trivial, but it is the hardest part. The moment an explanation lands, there is a temptation to ask one more question, then one more. The best learning happens in the quiet minute after the chat closes, when you try to reproduce the idea without help.
We are also wary of the opposite failure: pretending AI is never useful. It is useful. It is patient, it can rephrase, and it can give you a second example when the textbook only gives one. Those are real advantages, especially for students who cannot afford extra lessons or live far from a good library.
The rule we keep coming back to is simple. A tutor is doing its job when the student needs it less over time. That is the measure. Not how many questions were answered, but how many the student can now answer alone.
If you use an AI assistant while studying, try this once. Ask for a hint instead of an answer, then put the phone down and finish the problem yourself. The feeling of getting there is the point, and it is the part no model can do for you.