- 13 Oct, 2026
- 0 Comments
- 27 Mins Read
When an AI Tutor Is Enough, and When Your Child Needs a Human Tutor
AI tutor or human tutor
When an AI Tutor Is Enough, and When Your Child Needs a Human Tutor
An AI tutor is free, awake at midnight and endlessly patient. It is also very good at the one thing a South African examination does not pay for: handing over a finished answer.
For parents of Grade 8 to 12 learnersWhen is an AI tutor enough, and when does a child need a human tutor?
An AI tutor is enough when the job is practice and explanation: drilling routine procedures, re-explaining a concept in different words, testing definitions, or generating more questions of a type the learner already understands. A child needs a human tutor when the job is diagnosis, method, or motivation: finding out why the mistakes keep happening, learning to write a solution the way an examiner marks it, and being held to the work when nobody feels like doing it.
The short version: an AI tutor is excellent at producing answers and poor at producing learners. That sounds like marketing from a company selling human tutoring, so what follows is built on published evidence and the Department of Basic Education’s own marking rules rather than on opinion.
What is an AI tutor actually good at?
Four things, genuinely well. It is available at eleven at night when a question comes up. It will explain the same idea five different ways without losing patience. It will generate an unlimited supply of practice questions of a given type. And it will mark a factual recall answer instantly. For a learner who already understands a topic and simply needs volume and repetition, those four things are worth a great deal.
That covers more of a paper than tutoring companies usually admit. In a 150 mark Mathematics paper the Department of Basic Education weights Knowledge at 20 per cent and Routine Procedures at 35 per cent, roughly 52 to 53 marks, described in its guidelines as questions “generally similar to those encountered in class”. Half the paper rewards work a learner has done many times before, and drilling it is exactly the job an AI tutor can take.
A quick disambiguation. If your child takes the International Baccalaureate, “Maths AI” on their timetable means Applications and Interpretation, one of the two IB Mathematics routes. It has nothing to do with artificial intelligence.
What happened when researchers gave AI tutors to a thousand maths students?
Performance went up sharply while the tool was available, then fell below the control group once it was taken away. In a field experiment with nearly a thousand high school mathematics students, practice performance improved 48 per cent with a standard chatbot interface and 127 per cent with a version built with learning safeguards. When access was removed, the students who had used the standard chatbot scored 17 per cent worse than students who never had it at all.
The study is Bastani and colleagues, published in the Proceedings of the National Academy of Sciences in 2025. The authors’ explanation is worth reading twice: students used the chatbot as a “crutch” during practice problem sessions, and when successful, performed worse on their own. The practice looked like it was going extremely well. That was the problem, not the evidence against it.
Two honest qualifications, because this gets quoted badly. It is one experiment in one school with one model. And the version built with safeguards largely avoided the harm, so the finding is not that AI damages learning. It is that an AI which hands over answers damages learning, and how the tool is set up decides which one you have.
Why does this matter more for a matric learner than for anyone else?
Because the National Senior Certificate examination is precisely the condition the study measured: the moment the tool is taken away. A matric paper is three hours, handwritten, closed book, with no device on the desk. Whatever a learner can do only with help is worth nothing in that room. A Grade 10 can absorb that gap and recover. A matriculant writing in October cannot.
This is why a productive evening with an AI tutor can mislead: homework gets finished, daily marks hold up, and the first honest measurement arrives in a trial examination. A learner leaning on an AI tutor and one who has mastered the work look identical until a timed past paper is marked against the memorandum.
What do South African examiners actually award marks for?
Not the answer. In Physical Sciences the Department of Basic Education’s examination guidelines state that marks are awarded for the correct formula, the correct substitution, and the correct answer with its unit. The same guidelines say plainly that if a final answer to a calculation is correct, full marks will not automatically be awarded. A chatbot produces the finished answer, which is the component carrying the fewest marks.
They are the clearest statement anywhere of the gap between what an AI tutor is good at and what a South African paper pays for:
| Rule | What it says |
|---|---|
| 5.1.1 | “Marks will be awarded for: correct formula, correct substitution, correct answer with unit.” |
| 5.1.2 | “No marks will be awarded if an incorrect or inappropriate formula is used, even though there may be relevant symbols and applicable substitutions.” |
| 5.1.10 | “If a final answer to a calculation is correct, full marks will not automatically be awarded.” |
| 5.2.3 | A candidate forfeits the answer mark for a “correct answer + wrong unit”, a “wrong answer + correct unit”, or a “correct answer + no unit”. |
Read 5.1.10 again. A learner can reach the right number and still lose marks, while one who shows a correct formula and substitution then fumbles the arithmetic still banks marks. The paper buys a method, written in a particular order and notation, with units, and that is a performance learned in front of someone who can see what you did wrong. An AI tutor asked for the answer gives the answer. Asked to show working, it produces working that is often correct and almost never laid out the way a marker expects, and it will not notice that your child has been dropping units for a month.
These are the Department of Basic Education’s rules, not the Independent Examinations Board’s. IEB schools follow the same CAPS curriculum, but the IEB sets and marks its own papers to its own conventions. If your child writes IEB, the principle holds and the specific clause numbers do not. Check your own board’s published guidelines at the Department of Basic Education or with your school.
When is an AI tutor genuinely enough?
When the learner can already tell a right answer from a wrong one in that topic. That is the single test. If they understand the method and need volume, repetition, a second explanation or a definition checked, an AI tutor is a good and free use of an evening. If they cannot judge the output, they are not being tutored, they are copying from something that is confident whether or not it is correct.
Concretely, an AI tutor is enough for:
- More practice of a type already taught. Twenty more quadratic equations, another comprehension passage.
- A second explanation at an awkward hour, when a textbook paragraph has not landed and nobody is awake to ask.
- Factual recall drilling. Definitions, vocabulary, dates, formulae. Low stakes and instantly checkable.
- Turning a marked past paper into more questions. Once a human has named the weak topic, this gets you fifteen more of that exact question.
Three of those four begin after someone has worked out what the problem is. That is the division of labour this article argues for, and it is not a grudging one.
When does your child need a human tutor?
When the problem is not information. If the marks are falling and nobody knows why, if the same error keeps returning, if the method is right but the writing-up loses marks, or if the obstacle is anxiety, avoidance or a lost year of foundations, then more explanation will not fix it. Those are diagnostic and motivational problems, and a tool that answers the question it is asked cannot find the question nobody has thought to ask.
Five situations where a person is not a luxury.
- Nobody knows what is actually wrong. A parent sees 41 per cent. A tutor sees that the algebra is fine, the reading of the question is not, and every lost mark is the same mistake.
- The gap is older than the current grade. Grade 11 trigonometry failing on Grade 9 fractions means going backwards deliberately.
- The method is the problem, not the answer. Writing a solution the way a marker reads it is a trained skill.
- The obstacle is emotional. Anxiety, avoidance and lost confidence do not respond to better explanations.
- Nothing is getting done. Accountability to a person at a fixed time on a Tuesday is the mechanism. A chatbot has never noticed that a learner did not show up.
The United Kingdom’s Education Endowment Foundation rates one to one tuition at around five additional months of progress, “moderate impact for moderate cost based on moderate evidence”. That is UK data and an average, not a promise. Two findings beat the headline. Sessions work best short and regular, about 30 minutes, three to five times a week, over up to ten weeks, not one long rescue session. And tuition should be “additional to, but explicitly linked with, normal teaching”, which an AI tutor cannot do at all: it does not know what the class covered on Tuesday.
The same source is honest about something that cuts against the most expensive option. Comparisons with small group tuition are mixed, and groups of two or three are sometimes equally or more effective, worked through in one on one tutoring or group classes. Not sure the problem warrants either? The signs a child needs a tutor settles that first.
How do you use both without building a crutch?
Put the AI tutor after the attempt, never before it. The crutch effect in the research came from learners reaching for help at the start of a problem. If the rule in your house is that nothing gets asked until a genuine attempt exists on paper, the same tool becomes a marker instead of a ghostwriter. The order matters more than the amount.
Four rules from the evidence, not from taste:
- Attempt first, ask second. A written attempt on paper before the question is typed. This single rule is the difference between the two conditions in the study.
- Ask it to find the error, not to solve the problem. “Here is my working, where did I go wrong” teaches. “Solve this” does not.
- Close the laptop for the last pass. Whatever is being learned has to be reproduced once, cold, with nothing open. That is the examination condition, and the method in active recall revision.
- Check the output where it matters. AI tools stay unreliable on multi-step calculation and on anything local: prescribed set works, CAPS sequences and board rules are where a confident wrong answer is likeliest.
Used that way the two are not competitors. The person works out what is wrong and teaches the method. The machine supplies the repetition afterwards, at a volume no family could pay for.
Where Apex fits, and where it does not
Apex Academic Centre provides human tutors, online and in person, across CAPS, IEB, Cambridge, Pearson Edexcel, the International Baccalaureate and GED. It does not sell an AI tutor and is not launching one. If your child just needs more practice questions this week, a free chatbot will do that job and no booking is required.
A verified Apex tutor is for the part a tool cannot reach: working out what is going wrong, rebuilding a foundation from an earlier grade, teaching a learner to write a solution the way the memorandum reads it, and turning up every week. A first session spent diagnosing beats five spent explaining the wrong thing.
Apex is a tutoring provider, not an examination centre. It does not set papers, issue timetables or mark the National Senior Certificate, and no tutoring service can promise a mark. Book a verified tutor on the MyApex app, or look through the subjects offered.
Frequently asked questions
Can an AI tutor replace a human tutor?
Not for the work that decides marks. It is strong at explanation, repetition and generating practice. It cannot diagnose why a learner keeps making the same error, teach the written method an examiner marks, or hold a child accountable for showing up.
Is it bad for my child to use ChatGPT for homework?
It depends on the order. Research in the Proceedings of the National Academy of Sciences found students using a standard chatbot during maths practice scored 17 per cent worse once it was taken away. A version with learning safeguards largely avoided that harm. Require a written attempt before anything is typed.
Why does my child get the right answer but still lose marks?
Because South African papers pay for method. The Department of Basic Education’s Physical Sciences guidelines award marks for the correct formula, the correct substitution and the correct answer with its unit, and state that a correct final answer does not automatically earn full marks. A correct answer with no unit forfeits the answer mark.
How much tutoring does a struggling learner actually need?
The Education Endowment Foundation found short regular sessions work best: about 30 minutes, three to five times a week, over a period of up to ten weeks, for roughly five additional months of progress. That is United Kingdom evidence and an average, not a guarantee, but it argues against the single long rescue session.
Does Apex Academic Centre offer an AI tutor?
No. Apex provides human tutors, online and in person, across CAPS, IEB, Cambridge, Pearson Edexcel, the International Baccalaureate and GED. If your child takes IB Mathematics, the AI on their timetable means Applications and Interpretation, a course name rather than artificial intelligence.
Diagnosis first
A chatbot answers the question. A tutor finds the one nobody asked.
If the marks are falling and nobody can say why, that is the session to book. Everything after it is easier, including the part you hand to a machine.