Explaining a concept is the part of tutoring that AI systems do well. Working out what a specific student has misunderstood is the part they do poorly, and it is the part that matters.
What a tutor is actually doing
A human tutor spends most of a session inferring the shape of a misunderstanding from sparse evidence: an error, a hesitation, a question phrased oddly.
The explanation that follows is chosen for that specific misconception, which is why a good tutor rarely gives the textbook account.
Diagnosis precedes explanation, and it depends on evidence the student never states directly.
Why the diagnosis is the hard step
A wrong answer is compatible with many underlying errors. A student who subtracts incorrectly may have a procedural slip, a place-value misconception, or a misread question.
The written answer alone rarely distinguishes between these, so the tutor probes with a further question designed to separate the possibilities.
Systems that skip the probe select the most common misconception for that error type, which is right often enough to seem effective and wrong often enough to confuse the student it fails.
How a confident correction closes the conversation
When a system responds to a wrong answer with a clear correct explanation, the student's own reasoning is never examined.
The student sees a correct method that does not connect to what they did, and the mistake survives untouched underneath an accepted correction.
Human tutors avoid this by asking the student to explain their thinking first, which is a slow move that most systems are tuned against.
Why worked solutions defeat the purpose
Students under time pressure ask for the answer, and a system that complies removes the productive struggle that produces learning.
Refusing outright drives the student to a general-purpose tool that will comply, so refusal is not a solution either.
The designs that hold up give partial help scaled to demonstrated effort, which requires tracking what the student has already tried rather than treating each request independently.
Where the systems do help
Unlimited patient repetition is genuinely valuable, and no human tutor supplies it at scale.
Practice generation is similarly strong, since producing fifty variations of a problem type at graded difficulty is mechanical work that consumed teacher time.
The realistic role is as an amplifier for a teacher who retains the diagnosis, with the system handling drill, generation and availability while the judgement about what a particular student has misunderstood stays with someone who can watch them work.