Does my child still need a tutor now that AI can explain everything?

A pattern has become common in Sydney homes over the past two years. Homework comes back finished and mostly correct, the student says they understand it, and then the in class test goes badly. The homework was done with an AI chatbot open beside it. Parents reasonably ask a second question too: if AI can explain any HSC question in seconds, for free, at eleven at night, what exactly is a tutor for now? Both questions have the same answer, and the research on it is now clear enough to act on.

The most useful study was run with nearly a thousand high school maths students. One group practised with a standard chatbot, a second with a version built to behave like a tutor, giving hints rather than answers, and a third with no AI at all. During practice, the chatbot group solved 48% more problems correctly than students without AI, and the tutor group 127% more. Then the AI was taken away for an exam on the same material. The chatbot group scored 17% lower than students who had never used AI. The tutor group scored about the same as them.

Two details matter for parents. First, the students did not notice. Those who used the chatbot did not believe it had cost them anything, and even the tutor group overestimated how well they had done. Your child will say the AI is helping, sincerely, and the homework marks will seem to agree. Second, the tutor version only avoided the harm; it did not lift exam marks above the students who had no AI. Easy practice feels like learning. The effort a chatbot removes is the part that builds the skill.

The same pattern appears in writing. In a small early study from MIT, participants wrote essays with or without ChatGPT. Minutes after finishing, 83% of the ChatGPT group could not correctly quote a single sentence from the essay they had just written, against 11% of those who wrote unaided. A student who cannot recall their own argument has not yet learned how to build one, and an HSC English examination asks for exactly that, by hand and against the clock.

None of this happens because students are dishonest. It happens because general chatbots are built to answer. When one testing group asked six leading AI models to tutor simulated algebra students without special instructions, the models handed over the answer in 97% of conversations. A student has to deliberately set up an AI to make them think, and most never do.

This matters more now than it did a year ago. From the Class of 2027, NESA allows each HSC course at most one take home assessment task, worth no more than 15% of the school based mark, which is 7.5% of the final HSC mark. Year 11 follows from Term 1, 2027. Schools have also been told not to rely on AI detection software and to design tasks so that marks reflect the student's own knowledge. The practical effect is that, outside the courses with major works, at least 92.5% of an HSC mark is now earned under supervision, without AI, and mostly by hand.

This is not an argument against AI. Designed well, it is one of the most promising study tools in decades. At Harvard, students taught physics by an AI tutor built on research based teaching principles learned significantly more than students in an active learning class, with gains of 0.73 to 1.3 standard deviations, in less time. In a UK trial with students aged 13 to 15, an AI tutor whose every message was checked by an expert human tutor helped students solve new problems 66.2% of the time, against 60.7% with human tutors alone. NSW public school students now have NSWEduChat, which is deliberately designed to ask guiding questions instead of giving full answers.

What separates the good results from the bad is not the technology but the conditions around it. When a large American university simply gave 2,379 students access to an AI tutor for their course, final grades fell by 0.37 standard deviations and engagement with the course dropped sharply. The successful trials all had structure: the AI was designed to withhold answers, the work was chosen by someone who knew what the student needed, and in the strongest results a skilled adult was checking every exchange.

That is the real answer to what a tutor is now for. AI has made explanation close to free. Chegg, the homework help company that sold answers to students, has lost about 99% of its market value since ChatGPT launched, because answers stopped being scarce. A tutor whose value is explaining content is now competing with a free tool that never sleeps. But explanation was never what moved marks most. The research on tutoring is consistent that it works when it targets the specific reasons a student loses marks. In one large American programme, tutoring that simply repeated classroom content produced minimal gains, even with trained tutors and plenty of hours.

For families, the useful question is not whether your child uses AI, because most students now do, but whether the first attempt at the work is theirs. Five rules make the difference, and they cost nothing.

If you are choosing a tutor for 2027, ask how they handle this. A tutor who mainly explains content again is offering what your child already has on their phone. Ask instead what work is done under exam conditions, how lost marks are tracked, who checks that homework is the student's own, and what you will be told each week.

AI is not going away, and nor should it. The students who will do best are the ones who can think without it in an examination room and use it well outside one. That combination does not happen by default. It is built week by week, and it starts with the first attempt being theirs.

Edward MaDirector, MA99 Education