Students: using it without cheating yourself

Every institution has a policy, most of them differ, and none of them changes what happens at 23:40 the night before a deadline. So this page starts from the honest premise instead: you will use it, and the question that matters is which uses cost you the thing you enrolled to get.

Your institution's rules override everything here. Read them; they bind you and this page does not.

The test that sorts every use

One question does most of the work: after this, can I do the thing — or can the tool?

An essay is not the product. The essay is exhaust from building the ability to structure an argument, weigh evidence and write under constraint. Delegate the essay and the grade may survive; the ability it was supposed to certify was never built, and the invoice for that arrives later: in the exam it cannot attend, the interview, the first job that assumes the skill exists.

This is the task is the learning, which covers the general case. The student case is sharper because the skill is the entire point of the transaction. You are the customer paying for the ability, and delegating its construction is spending money to not receive the goods.

Uses that build, uses that replace

The line is not the tool. It is which side of the work the tool sits on.

Building: explaining a concept a third way when the lecture's way did not land · generating practice problems and checking your attempts · critiquing a draft you wrote, as in "what is unclear, what is unsupported" · testing your understanding by having it probe you · unsticking you on step three of a derivation you then complete.

Replacing: drafting the essay · solving the problem set · summarising the reading you were assigned because engaging with it was the assignment · producing the code for the exercise whose point was producing the code.

The middle band is where honesty gets hard. "It only outlined it" and "it only rewrote my rough version" both feel like assistance and can both hollow out the work, depending on whether structuring and phrasing were the skills being trained. Apply the test, not the vocabulary.

The feeling of learning is exactly what it fakes

One finding from the measured literature transfers directly. In METR's trial, experienced developers using AI believed they were about 20% faster while being 19% slower — the gap between felt and measured came to 39 points, in people with every advantage in judging their own work.

The study version for a student: reading a fluent explanation feels like understanding. The check is retrieval, and it is cheap — close the tab and re-derive it, explain it aloud, do the next problem cold. If it is not there without the tool, it was never yours. That feeling of "I get it now" is the least reliable instrument you own.

Where the boundaries are already hard

Two are not judgement calls.

Claiming authorship of what you did not author. Academic writing inherits its rule from the same place research does: authorship is accountability, and an AI cannot carry it. Submitting generated work under your name asserts something false regardless of what the policy says about tools.

The fabricated citation. Models invent plausible sources, and a reference list is where it shows first. A cited work you never opened is a risk you are choosing; a cited work that does not exist is a finding of misconduct waiting for one bored marker. The mechanical check is in web search and citations.

What UNESCO's guidance adds

The global reference document here is UNESCO's Guidance for generative AI in education and research, and two of its positions are worth knowing directly.

It names the delegation problem as a new kind of misconduct: GenAI "might allow students to pass off text that they did not write as their own work, a new type of 'plagiarism'". That is the authorship boundary above, stated by the body that writes education policy for 194 member states.

And it sets a floor most people have not heard of: for independent conversations with GenAI platforms, "The minimum threshold should be 13 years of age". If you are helping a younger sibling with homework, that recommendation is aimed at exactly that situation.

The document's overall stance is regulation and validation rather than prohibition, anchored in what it calls a human-centred approach — which is also the honest premise this page started from.

The rest of the kit

Verifying what you cannot judge — most of what you study, you cannot yet judge, which is why you are studying it · why answers vary — re-ask before you rely · wellbeing use — for the 2 a.m. that has nothing to do with the deadline · which plan (Claude, ChatGPT, Gemini) — before paying for anything · the other side of the desk, if you teach or set policy: teachers and universities.

What goes wrong

Optimising the artefact. The grade is a proxy. Gaming the proxy while skipping the ability is winning the wrong game at your own expense.

"Just this once", weekly. Each delegation is locally reasonable. The sum is a semester where the skill curve stayed flat, and nothing marked the moment it stopped being once.

Trusting the feeling of understanding. It is precisely what fluent explanation manufactures. Test retrieval, not recognition.

Submitting an unread reference list. The single fastest route from "grey area" to "integrity case".

Assuming policy silence means permission. Institutions differ, and yours is the one that grades you. When unsure, ask — in writing.

How to check it worked

At the end of a course, sit the hardest past-paper question with everything closed. What you can do alone is what you actually bought this term. If the answer is embarrassing, the tool was on the wrong side of the work. Better to learn that from a past paper than from the exam, or from the first week of a job that assumed the certificate meant the ability.