Artificial intelligence

AI for studying

A model can explain a confusing point three different ways and answer a silly question patiently, without judgement. It can also hand you a finished assignment that leaves you unable to do anything new.

Updated
In this article

First, about cheating — without moralising#

Three separate facts, and it is worth not mixing them up.

The skill does not appear. The ability to solve problems grows out of trying to solve them — out of getting stuck, trying options, finding the mistake. Reading a finished solution gives you recognition: the text feels clear because it is smooth, not because you understood it. The difference shows up at the first exam without internet access and at the first work task where nobody has written the answer for you.

Teachers notice. Not through automatic AI detectors — those cannot be relied on; they make mistakes in both directions and confidently flag carefully written human text as machine-made. They notice differently: the work is stronger than anything the student has done before; the text is smooth but does not draw on the course; things covered in class are missing, and terms that were never used in class are present. Then two questions out loud about your own work, and everything becomes clear. At many schools and universities, submitting text you did not write falls under academic integrity rules, and the consequences are set by those rules.

Sources can be invented. A reference list generated from scratch is a set of plausible titles with plausible authors and years. Checking it takes ten minutes, and it wrecks the work along with the teacher's trust.

The conclusion is not "don't use it" but "use it so that your own head does the work". Below are four ways that achieve this.

Explaining what you do not understand#

This is the best use of a model in studying. Its advantage over a textbook is that you can ask again ten times in a row.

It works like this: state your level and the exact point where you are stuck, ask for an explanation without formulas, then with formulas, then with an example.

I know what a derivative is, but I don't understand why
the product rule has two terms. Explain the idea first,
then derive it, then show it on f(x) = x²·sin x.

Then comes the understanding check, and it is not optional: close the answer and retell it in your own words, out loud or in writing. As long as you cannot retell it, you have recognised it, not understood it. This is the same principle as active recall.

Checking your own solution#

The order matters: first you solve it yourself, then you show it. The other way round is pointless.

Here is my solution to the problem. Do not give me the right answer.
Find the first place where I went wrong and ask a leading
question so that I can find the mistake myself.

The request not to give the answer is the key part. Unless you forbid it, the model will happily solve everything for you.

Check arithmetic and calculations separately: a model predicts plausible text rather than computing, and it slips up in calculations easily. Asking it to calculate with code, or using an ordinary calculator, is more reliable. Why it works this way is explained in what is an LLM.

Generating practice problems#

The textbook runs out, and you need twenty more examples of the same type. Here a model is useful and nearly risk-free: you solve the generated problem yourself and check the answer separately.

Write 10 integration-by-parts problems of increasing
difficulty. Put the answers in a separate list at the end
so I don't see them straight away.

The same works for languages — sentences practising one specific structure; for programming — exercises on one technique; for history — questions on dates and causes. A useful extra: ask it to explain why the wrong options are wrong.

Analysing your mistakes#

After a quiz or test, take your mistakes and work out the type of each one: did not understand the question, did not know the rule, knew it but did not apply it, made an arithmetic slip. A model can help you sort them and pick what to review.

The value is that each type has a different fix. Not knowing a rule is fixed by reading, carelessness by how you work, and misreading questions by practising restating the problem in your own words before solving it. If an exam is coming, fold this into your exam preparation plan.

Limits worth keeping in mind#

Models are confidently wrong. On standard school and university material they are usually right, but on narrow topics they produce wrong definitions and non-existent theorems in the same even tone. Check against your textbook and notes: your teacher grades the course, not the internet.

Check links and quotes one by one. Open the page, find the quote in the original, look the paper up in the library catalogue.

Clarity is not understanding. The "it all makes sense" feeling after reading an explanation is deceptive; the only honest test is to reproduce it without looking a day later.

Your data goes to someone else's service. A draft of your thesis, other people's personal data in a project, closed course materials — better not to send them there.

Psychological, medical and legal questions go to real people. If studying is costing you sleep or bringing on panic, that is a conversation for a person, not a chat. A model's answer does not replace a doctor, a lawyer or a financial adviser: where a mistake costs health, money or rights, an accountable person makes the decision.

And a practical point: the rules of your specific course matter more than general arguments. Some courses allow AI use if you declare it in the work, others forbid it entirely — find out from your teacher in advance, not after you submit.

Step-by-step plan

  1. Solve it yourself — to the end or until you are stuckOnly your own attempt is worth showing to a model; otherwise there is nothing to check.
  2. Ask for an explanation of the stuck pointState your level and where you got stuck, and ask for three passes: idea, derivation, example.
  3. Retell it without lookingClose the answer and explain it in your own words; if you cannot, you have not understood it yet.
  4. Ask for a leading question, not the answerExplicitly forbid it from giving the solution: finding the mistake is your job.
  5. Work through generated problemsTen problems of one type with answers at the end — by hand, without hints.
  6. Sort your mistakes by typeCome back to the topic a day later and repeat it without the model — that is the real test.

Start learning this in your own space

The plan goes into your repository: tick off stages, keep notes — the change history shows how far you have come.

Start the plan

Check yourself

1.A model explained a theorem and cited a textbook with a page number. What should you do before putting that reference in your work?

2.Why should you recheck calculations in a model's answer with a calculator?

3.Which way of working does not help you learn to solve problems on your own?

Sources

Was this helpful?

More articles

Artificial intelligence How to use AI Getting started is simpler than it looks: you need one service, one real task from your own work and the habit of checking the result. Everything else is detail you pick up along the way. Artificial intelligence How to write a prompt A good prompt differs from a bad one not in length or politeness but in having everything the work needs — the task, the material, the shape of the answer and what a usable result looks like. Artificial intelligence AI agents An agent is not a special, smarter model but a harness around an ordinary one: the model gets a list of tools and is allowed to call them in a loop until the task is finished — or until patience runs out. Artificial intelligence How neural networks work A neural network is simpler than its name suggests. Inside there are no thoughts or images — there are tables of numbers multiplied by the input values, and a procedure that keeps nudging those numbers until the answers get more accurate. Artificial intelligence What is an LLM An LLM (large language model) is a program trained to continue text. Everything it does comes down to one operation: look at what has been written so far, predict the next small piece, and repeat. Artificial intelligence Prompt engineering Prompt engineering is choosing the input text so that a model more often produces the result you need. It is not a set of magic spells but ordinary engineering work — a hypothesis, a test on a set of examples, rolling back changes that did not help.

More solutions

App for study notes Good study notes have to be written, but also reread, extended and found a year later. Plain markdown files with formulas and diagrams, sorted by subject and keeping their edit history, do that job well. Here is how to keep your notes with us. Company knowledge base A company knowledge base rarely dies for lack of tools. It dies because nobody owns the pages and they quietly go out of date. Here is how to set up a base that has owners, a history and a clear way to change it. Confluence alternative Confluence keeps documentation as pages inside spaces. In Skillok documentation is a set of markdown files in a repository, and you work with it the way you work with code — edit, history, review. That is stricter and more transparent, but there is no visual page builder and no real-time co-editing.