How to Use AI for Revision Without Cheating

The line is simple. If the assessment is meant to measure what you can do, anything that produces the work instead of you is cheating. Anything that helps you become able to do it yourself is studying.

Asking an AI to write your assignment crosses that line. Asking it to keep questioning you on a topic until you can answer without help does not.

Nearly every hard case comes apart once you ask yourself one thing: at the end of this, can I do it on my own?

Why this matters more in an exam system

If your final grade rests on a written paper in a hall with no phone, then AI that produces work for you is not just against the rules. It is self defeating. You will have coursework you cannot reproduce and a paper you cannot pass.

This is the practical argument, and it lands harder than the ethical one for most students: the exam is unaided, so your preparation has to make you able to work unaided. Anything that hides the gap between what you can do and what you appear able to do is working against you.

Uses that genuinely help

As an examiner

The strongest use. Paste in a topic and ask for ten questions, then answer them before looking at anything. This is retrieval practice, one of only two techniques rated high utility in the Dunlosky review (the other is spreading study over time), and AI removes the main obstacle: that writing good questions for yourself is slow and you already know the answers.

Ask for questions that apply the idea to an unfamiliar case, not just definitions. That is where the marks sit at higher levels.

As an explainer of last resort

When a textbook explanation does not land, asking for the same idea explained differently, or at a simpler level, then building back up, is legitimate and often effective. Follow it by explaining the idea back in your own words without looking. If you cannot, you have not learned it yet.

As a critic of your own work

Write your essay first. Then ask what the weakest argument is, what a marker would object to, or what you have failed to consider. You keep authorship; you gain a reader.

The order is what matters. Critique after writing is feedback. “Critique” before writing is outsourcing.

As a translator between your languages

For bilingual candidates, checking whether you have understood a French source correctly, or how a technical term is rendered in the other language, is ordinary study support, the same as using a dictionary, faster.

Uses that are cheating, whatever you call them

• Generating an assignment and submitting it, edited or not
• Producing work in an assessed exercise designed to test whether you can produce it
• Any use during an examination where it is not permitted
• Generating references, quotations or data you have not verified exist
• Having it write a personal statement or motivation letter that claims to be your own voice

“I edited it afterwards” does not change the category. If the substance was produced by something other than you and presented as yours, the label does not matter.

The reliability problem

AI systems produce fluent, confident text that is sometimes wrong (a 2025 paper by Kalai and colleagues argues this follows from how models are trained and scored). Not obviously wrong, plausibly wrong, in the register of a textbook.

They are least reliable exactly where students most want help:

• Specific local facts: exam dates, cut off marks, entry requirements, deadlines. Treat every one of these as unverified until you have seen it on the examining body’s or institution’s own site.
• Citations: references and quotations can be fabricated entirely while looking perfectly formatted.
• Numbers: statistics may be invented or mixed between sources.
• Recent changes: a syllabus revision may postdate what the system knows.

The failure mode is dangerous precisely because it does not look like failure. A wrong answer arrives in the same confident tone as a right one. Build the habit of checking anything specific against a primary source, every time.

A working rule

Before using AI on any piece of work, ask three questions:

1. What is this task measuring? If it is measuring the thing you are about to delegate, stop.
2. Would I be comfortable if my teacher saw exactly how I used it? Discomfort here is usually accurate.
3. Will I be able to do this unaided afterwards? If not, you have produced output, not learning.

And one institutional rule: check your school’s own policy. Policies differ, they are changing quickly, and “I did not know” is rarely accepted. Where a policy permits AI with disclosure, disclose it.

What the exam rules say, and what they do not

Policy is the part students most often guess at. Here is what two sets of published rules say, checked on 18 October 2026. The first is British and the second is the Cameroon GCE Board’s own.

Question	JCQ, UK exam boards (AI Use in Assessments, effective 30 April 2025)	Cameroon GCE Board (2026 sources)
Where does the rule bite?	Mainly coursework and other non-exam assessment. Exams are largely unaffected, but centres must make sure AI tools cannot be reached on any device used in an exam.	Exam hall. Per Cameroon Tribune, 19 May 2026, using a mobile phone or communication device brings a one-year ban and cancellation of all results.
Is AI use allowed at all?	Only if acknowledged: name the tool and the date. Citing just “AI” or “ChatGPT” is not acceptable. Keep the prompts and output (for example screenshots) and submit them with the work.	No AI-specific rule appears on the Board pages checked on 18 October 2026 (examinations page, malpractice table). Ask your centre what applies to school-based work.
Does AI-made work earn marks?	No credit for AI-generated material the student has not independently developed.	Not stated in the pages checked.
Can detectors decide a case?	Detection tools should form part of a wider review; accuracy varies.	Not stated.
Sanctions	Sanctions can include disqualification and debarment from taking qualifications for a number of years.	Pre-prepared material: all results cancelled. Collusion or copying: subject result cancelled.

The JCQ rules are in the JCQ guidance page. They apply in England, Wales and Northern Ireland, not in Cameroon, so use them as an example of what a clear AI policy contains, not as your board’s rule.

The GCE Board also publishes the number of malpractice cases it recorded. In its malpractice table (the page heading says 2024, but the columns are labelled 2025 and 2026), cases of collusion or copy work rose from 542 to 1,580 and phone cases from 118 to 166, while pre-prepared material fell from 2,028 to 1,010. Total cases moved from 2,751 to 2,792. As a share of the 2026 total, collusion is 1,580 / 2,792 = 56.6% and phones are 166 / 2,792 = 5.9%. These are counts of cases, not rates per candidate, because the page gives no candidate total. None of the nine categories is labelled AI, so the table cannot tell you whether any AI-assisted case was filed under another heading.

Why detectors are the wrong thing to rely on

If your school suspects AI use, a detector score is the usual first step, which is why it matters that the scores are weak evidence. JCQ says detection tools should form part of a holistic judgement. A 2023 study by Liang and colleagues in Patterns (preprint abstract) tested widely used GPT detectors and found that they consistently misclassified writing by non-native English speakers as AI-generated, while identifying native writing accurately. The authors cautioned against using these detectors in evaluative or educational settings. Many candidates in Cameroon and Uganda write English as a second or third language, so this finding applies to you directly. The practical protection is the same as the JCQ advice: keep your own drafts, notes and any prompts, so you can show how the work was made.

Questions students ask

Can teachers actually detect AI writing?

Detection tools exist but are not strong evidence: JCQ says they should form part of a wider review and that accuracy varies, and a 2023 study found they misclassify non-native English writing (details in the section above). But teachers who know your work notice when it changes character, and a viva or a follow up question exposes work you cannot explain in seconds. The stronger argument remains the practical one: it leaves you unable to sit the paper.

Is using it for grammar cheating?

Generally no, on the same basis as a spell checker, unless the assessment is testing your written expression, in which case correcting it is correcting the thing being measured. If your language ability is part of the mark, do the correcting yourself.

What about using it to make revision cards?

Legitimate and useful, with one caution. Making cards forces you to decide what matters, and that decision is itself part of learning. Generate a first draft, then edit it hard: cut what will not be examined, split anything too big, and check every fact.

Will using AI make me lazy?

It makes the easy path much easier, which is a real risk. The protection is structural: use it in modes where you produce the answer and it checks you, rather than modes where it produces and you approve.

The same topic, asked two ways

The difference between studying and outsourcing usually shows up in the request itself. Compare these.

Outsourcing	Studying
“Write me an essay on the causes of the First World War.”	“Ask me five questions on the causes of the First World War. Do not give me the answers until I have replied.”
“Solve this differentiation problem.”	“I got this answer and the book says otherwise. Do not tell me the solution. Tell me which step to check.”
“Summarise this chapter.”	“Here is my summary of the chapter. What did I leave out that matters?”
“Make me flashcards on osmosis.”	“Here are the twelve cards I made on osmosis. Which are too broad to test properly?”

The pattern in the right hand column is consistent: you produce first, it responds second. That single ordering rule resolves most cases without needing to think about policy at all.

One phrase worth keeping: “do not give me the answer yet.” These systems default to being maximally helpful, which in a revision context means answering the question you were supposed to answer. You have to actively stop it.

A study session that actually uses it well

Thirty minutes on one topic:

1. Five minutes, no AI. Write down everything you can recall about the topic from memory. Blank page. This is the part that does the learning.
2. Five minutes. Ask it for ten questions on the topic, mixed between recall and application, with answers withheld.
3. Ten minutes. Answer them in writing, before checking anything.
4. Five minutes. Reveal the answers. Mark yourself strictly: half remembered is wrong.
5. Five minutes. For everything you got wrong, ask for the idea explained a different way. Then close it and write the explanation out yourself.

Note that the AI is absent for the two most valuable steps. That is not an accident: steps one and three are where memory is actually built, and any tool present during them tends to do the work for you.

Verifying what it tells you

A workable habit is to sort every claim into three boxes.

Stable and checkable: established science, mathematical method, historical chronology. Usually reliable, and easy to confirm against your textbook. Low risk.

Specific and local: exam dates, entry requirements, cut off marks, deadlines, fees. Assume these are wrong until confirmed at the source. This is where confident sounding errors do real damage, because acting on a false deadline costs you a cycle.

Cited: any reference, quotation, statistic or study. Check it exists before repeating it. Fabricated citations are formatted perfectly and are among the fastest ways to lose credibility in submitted work.

A useful test: ask the same factual question in a fresh conversation. Inconsistent answers across attempts are a strong signal the system is generating rather than recalling. Treat the whole claim as unverified.

If you have already used it in a way you are unsure about

This is worth addressing directly, because a lot of students are in this position and quietly hoping it resolves itself.

If the work has not been submitted: rewrite it yourself. Not edited, written, from your own understanding, with the generated version closed. If you cannot, that is the useful information: you do not yet know the material, and you have found out while there is still time.

If it has been submitted and you are worried: your school’s policy governs what happens next, and policies vary widely. Ask your teacher or head of centre how your school treats a student who comes forward. The practical problem remains regardless: you have a grade that does not reflect what you can do, sitting in front of an examination that will.

Either way, close the gap. Identify what the assignment was meant to teach you and learn it properly now. The exam is the part that cannot be delegated, and it is the only part that ultimately counts.

The short version

Use it to question you, to explain differently, and to criticise work you have already written. Do not use it to produce what you will submit. Verify every specific fact against a primary source.

Used that way, it is the most patient examiner you will ever have. Used the other way, it is an expensive way to arrive at the hall unprepared.

More on honest use of technology in study in AI for Students, and the full review method in our twelve week revision plan.