The advice students get about AI comes in two varieties. Schools say be careful, do not use it for coursework, and it may be wrong. The internet says here are forty prompts that will get you an A.
Neither is much use, because both skip the thing that actually determines whether AI helps you: not what you ask it, but what you do while it answers.
So here is the version we would want a student we cared about to read. It is not “do not use it”. Most of you already do, several times a week, and telling you to stop is advice nobody follows and everybody nods along to.
The one principle everything else follows from
Learning happens when you produce something effortfully, and struggle to do it. That is not a moral position, it is close to the most replicated finding in the study of memory. Retrieving an answer strengthens the memory of it. Reading an answer barely does anything.
This is why highlighting does not work, why re-reading your notes does not work, and why watching a worked solution feels productive and mostly is not.
AI is extremely good at removing effort. That is what it is for. Which makes it, by default, a machine that converts your revision into reading.
So the whole question is: did I do the thinking, or did the model do it and I agreed with the result?
Every use below is sorted on that basis.
Uses that genuinely help
These share a shape: the model does something you were never going to do yourself, and you still do the thinking.
Ask it to explain the step you are stuck on, not the whole question. Not “how do I do this question” but “why does the sign change when I multiply both sides by a negative number?” You keep the problem. It fills one gap. That is a real use and it is the best one on this list.
Have it generate practice questions, then close it and do them. Ask for eight questions on the same topic in the style of your board. Then leave the chat, do them on paper, and only come back to check. The generation is a genuine service. The doing has to be yours, and the same rule applies to past papers.
Have it quiz you. Ask it to ask you questions one at a time, wait for your answer, and tell you what was missing. This is retrieval practice and it is one of the few things a chatbot does that is structurally better than a textbook, because a textbook cannot wait for your answer.
Ask it to rephrase something you did not understand. Textbooks explain a concept one way. If that way did not land, having it explained four other ways, with an analogy, is genuinely useful and costs you nothing.
Ask what you have failed to consider. “I have written this answer about osmosis, what is missing?” You produced the answer first. That order matters enormously.
Ask it to explain why a wrong answer is wrong. You got it wrong, you have seen the right answer, and you still do not know where your reasoning broke. That is a good question and a hard one to get answered at eleven at night.
Uses that quietly make you worse
Also a shape: it produces the artefact, and you get a feeling of understanding that is not backed by anything.
Asking it to do the question. Obviously. Less obviously: asking it to do the question and then reading the solution carefully and nodding is nearly the same thing. The nod is recognition, not knowledge, and they feel identical from the inside. This is the single most common way a term of AI-assisted revision produces no improvement.
Asking it to summarise a topic for you. Making the summary was the revision. Reading someone else’s summary is reading.
Asking it to write your answer so you can “see the structure”. You will absorb far less structure than you think, and you now have a model answer sitting where your attempt should have been.
Trusting its marking. More on this below, because it deserves its own section.
Using it as your only source. It will state things confidently that are not on your specification, are true for a different board, or are simply wrong. It has no idea which exam you are sitting unless you keep reminding it, and it will not tell you when it is guessing.
Where AI is unreliable in ways that specifically matter for exams
Being concrete, because “it can be wrong” is too vague to act on.
It is confidently wrong at arithmetic and algebra. Language models are not calculators. They are much better at this than they were, and they still produce fluent, well laid out, wrong working. Wrong working that looks right is worse than no working, because you have nothing to trigger your suspicion.
It does not know your mark scheme. Ask it to mark your answer and it will give you a plausible-sounding mark based on its general sense of what a good answer looks like. That is not how marking works. Marking works by finding specific creditable points from a specific scheme, in order. A model that has not been given the scheme is not marking, it is estimating, and it estimates generously.
It agrees with you. Push back on a correct answer and it will frequently fold and apologise. This makes it useless as an arbiter of whether you were right, which is exactly what students most want to use it for.
It does not know what is on your specification. Boards differ, tiers differ, specifications change. It will happily teach you A-level content for a GCSE paper, and content that was removed in the last specification revision.
It cannot see your handwriting or your working. Photograph your page and ask it to mark it and you are stacking an unreliable read on top of an unreliable judgement.
The line into malpractice
Worth knowing exactly, because the consequences are real and the rules are less mysterious than people assume.
In an exam: no AI. Obviously.
In coursework and non-exam assessment: this is where students get into trouble, and the rules are set by JCQ, which speaks for the boards. The principle is that work submitted for assessment must be your own. Submitting AI-generated content as your own is malpractice, and it is treated as seriously as any other form of plagiarism. That includes text you have edited afterwards. If you use AI at any point in producing assessed work, you are generally required to say so and to show where, and your school will have a process for it. Ask your teacher before, not after.
In homework and revision: not a malpractice question at all, because it is not assessed. It is entirely a question of whether you are learning anything. Which is not the school’s problem to detect, it is yours to solve.
The thing worth understanding: schools are increasingly able to notice, but the reason not to do it is not the risk of being caught. It is that a student who submits work they did not produce arrives at an exam hall with nothing, and the exam hall is the only assessment that counts.
The test to apply
One question, and it works for every case:
After the AI helped, could I now do it on my own, from blank, without it?
If yes, it helped. If you are not sure, it did not, and the only way to find out is to close the tab and try.
That is genuinely the whole thing. It is not about which tool, which prompt, or which rule. It is about whether the ability transferred to you or stayed on the screen.
A word on the tempting shortcut
The reason this article is not “do not use AI” is that we build an AI marking system for schools, and it would be a strange position for us to take.
Worth being clear about the difference, though, because it is the same difference as everything above. The teacher-facing version of this argument, including what these tools cannot do, is here. A marking system is given the official mark scheme and asked to find the specific points in a student’s own work. It is judging something the student produced. A chatbot asked to write the answer produces the work itself, and there is nothing of the student’s left to judge.
One of those tells you where you are. The other tells you what a good answer looks like while quietly ensuring you never find out whether you could have written it.
For teachers and parents
A ban tends to convert a manageable behaviour into a hidden one. The alternative worth trying is teaching the test above explicitly, then setting an assessed task in class, on paper, soon after a piece of AI-assisted homework. The gap between the two makes the argument without anyone having to make it.
Marky marks student work against the official scheme, and a teacher can switch on a check that flags signs of answers amended after the exam or copied from the mark scheme, as a prompt rather than a verdict. See how it works.