AI for Students & Learning

AI as a learning tool

For a student, AI can work like a patient tutor available at 2am — explaining any topic at the level you ask for, quizzing you, and never getting bored of the same question. But there is an enormous difference between using it to learn better and using it to skip learning, and the gap does not feel like a gap while you are in it. New to AI? Start with What is AI.

The trap: a clear explanation feels like understanding

This is the thing to internalise before any of the tips below, because it is what makes AI more dangerous for studying than a textbook is.

When you read a well-written explanation, your brain registers the experience as comprehension. It is smooth, nothing snags, everything follows. And these tools are extremely good at producing smooth — that is close to what they are optimised for. So you finish the paragraph feeling like you have got it.

Then in the exam, with nothing in front of you, it turns out you could recognise the explanation but not produce it. This is a well-known effect in learning research — recognising material and being able to retrieve it are different abilities, and re-reading something clear trains the first one while feeling like it trains the second.

The implication is specific and it shapes everything else on this page: reading what the AI wrote is the least valuable thing you can do with it. The value is in what it makes you do.

Make it ask, not tell

The single highest-return change is to stop requesting explanations and start requesting questions. Effort at recall is what builds durable memory; smoothness is what defeats it.

• "Ask me ten questions on this chapter, one at a time. Wait for my answer before the next one." One at a time matters — a list of ten lets you skim and feel productive.
• "Don't tell me the answer. Give me a hint, then ask again." The struggle is the part that works.
• "Here is my answer. What did I get wrong, and what does the mistake suggest I have misunderstood?" The second half of that is the useful half.
• "Ask me the same material again tomorrow, harder." Spacing beats cramming, and an AI is the cheapest way to get a second pass at spaced intervals.

Compare the two modes honestly. Asking for a summary takes thirty seconds and feels great. Being quizzed for fifteen minutes is unpleasant and is where the learning actually happens. If your study session was enjoyable throughout, you were probably reading, not learning.

Understanding & learning

• Explanations at levels — "explain X like I'm 12," then "now at an academic level." The gap between the two versions is usually where your confusion actually sits.
• Ask follow-ups without embarrassment until it fully clicks. This is the genuine advantage over a lecture: there is no social cost to the fourth follow-up question.
• Examples and analogies — ask for a real-world example of any abstract concept, then ask where the analogy breaks down. Every analogy breaks somewhere, and that boundary teaches you the concept's actual shape.
• Ask it to contrast two things you keep mixing up. "What is the difference between X and Y, and what is a case where someone would confuse them?"

The Feynman technique, done with a tool that can push back

Explain the material out loud, in your own words, as if teaching it — and the places where you go vague are exactly the places you do not understand. That technique is old and it works. What AI adds is an audience that interrupts.

Paste your explanation and ask: "Where is this vague or wrong? Ask me the question a sceptical examiner would ask." Then answer it. Then do it again.

This is the most valuable single use of these tools for studying, and it is almost the opposite of how most students use them. You are producing and it is critiquing, rather than it producing and you reading.

Research & summarizing

• Summarize long material into key points — as a map before you read, or a check after, never as a replacement for reading the thing you will be examined on.
• Research with sources — Perplexity gives answers with links you can verify.
• Learn from your own documents — tools like NotebookLM answer from material you upload.
• Always verify. Not "usually" — always, for anything you will write down.

Fabricated citations — the one that ends badly

This deserves its own warning because of how specifically it damages students.

Ask a general chat assistant for references and it may produce citations that look completely real — plausible authors, a plausible journal, a plausible year, formatted correctly — for papers that do not exist. It is not lying; it is producing text shaped like a citation, and a citation's shape is very learnable. The output is more convincing than a wrong answer, because it has the texture of scholarship.

Hand in a bibliography containing one of these and you are not explaining a mistake, you are explaining why you cited a paper you never read. That conversation goes badly regardless of your intent.

The rule is simple and has no exceptions: open every source before you cite it. Find it in your library catalogue or the publisher's site, confirm the authors and year, and read enough to know it says what you are claiming. If you cannot find it, it may not exist. Tools that search and link are much safer here than tools that generate from memory, but the rule still applies to both.

Working from your own material is the safer mode

There is a meaningful difference between asking a model what it knows and asking it about a document you gave it. The second grounds every answer in text you can check, which removes most of the invention problem and makes verification a matter of looking at the page rather than searching the internet.

So upload the lecture slides, the paper, the textbook chapter, and ask questions about those. "Where in this does the author address the objection about X?" "Summarise the method section." "What does this chapter assume I already know?" — this last one is excellent for finding the gap that is making everything else hard.

Writing help (done right)

The line here is clearer than people pretend, and it has nothing to do with how much you used the tool. It is about whose thinking is on the page.

• Fine: brainstorming angles before you write; asking what is weak in your argument; asking what a critic would say; grammar and phrasing on your sentences; asking it to explain a piece of feedback you did not understand.
• Fine: "here is my messy paragraph, what am I actually trying to say?" — you then write the clear version.
• Not fine: generating paragraphs and submitting them; generating an argument you did not reach; producing a structure you cannot justify.

A practical test: if your marker asked "why did you make this point here rather than there?", could you answer? If the honest answer is "that is how it came out", the thinking is not yours yet.

What you should know about AI detectors

Many institutions run submissions through detection tools, and it is worth understanding what they can and cannot do — in both directions, because both affect you.

These tools do not detect authorship. They estimate how predictable the text is, and predictable text gets flagged. That means they produce false positives on perfectly honest writing, particularly for students writing in a second language, for anyone with a plain factual style, and for heavily edited formal prose. It also means genuinely generated text that has been reworded can pass.

Two practical consequences. Do not assume a detector protects you from an accusation, and do not assume it will catch someone else. And keep your drafts — version history in your word processor, or dated files. A document with a visible editing history, with your false starts and reorganisations in it, is the strongest evidence of your own process if you are ever asked. That is worth doing whether or not you touched an AI.

Lectures and notes

Transcription is accurate enough now that recording a lecture and getting usable text is trivial. Two things before you do it.

Check whether recording is permitted — many institutions require the lecturer's consent, and some require it in writing, particularly where students contribute to discussion. The lecturer is not the only person in the room. Second, know that a transcript is not notes. An hour of speech is thousands of words of filler, tangents and repetition, and having it does not mean you have learned anything; it means you have deferred the work.

What makes a transcript useful is turning it into something you have to answer. Ask for the five claims the lecturer argued for and the evidence given for each. Ask what was said that contradicts the textbook. Ask it to produce ten questions from the transcript and quiz you — which is the same move as everywhere else on this page, because it is the move that works.

If you take notes by hand, photograph them and ask what is missing compared with the slides. The gaps are usually the bits you did not follow at the time.

Studying in a language that is not your first

This is where these tools are most straightforwardly transformative, and it is worth being deliberate about it.

Reading a dense academic paper in a second language means paying a comprehension tax on every sentence before you get to the actual difficulty of the subject. Asking for a plain-language restatement of a paragraph removes the tax and leaves the subject, which is the part you are there for. That is legitimate and nobody sensible thinks otherwise.

Writing is where to be careful, for two reasons. Asking a model to "make this sound more academic" will often replace your voice with a generic register and quietly change your meaning — check that the rewritten sentence still says what you meant, because fluent and correct are not the same thing. And heavily polished second-language writing is disproportionately likely to be flagged by a detector, as above. Editing your own sentences for grammar is safer than having new ones written, and it keeps the drafts that prove the work was yours.

Practice & exam preparation

• Practice questions — ask for a short quiz and test yourself, closed-book.
• Explain mistakes — "I got this wrong, why?" — errors are the highest-information part of studying.
• Past-paper style — give it a real past question and ask it to generate three more in the same style and at the same difficulty. Generating variants is something these tools do genuinely well.
• Ask for the marking scheme. "What would a top answer to this include that an average one would miss?" This reframes revision around what is rewarded.
• Mixed practice. Ask it to interleave questions from several topics rather than drilling one. Harder, and closer to the exam.

Maths and anything with numbers

Language models are unreliable at arithmetic in a way that surprises people, because they are so fluent everywhere else. They are predicting plausible next tokens, and a long calculation has many places for a plausible-but-wrong step to slip in — often presented with complete confidence and tidy working.

Use them for the method, not the answer: "what approach would you use for this kind of problem, and why that one?" Then do the computation yourself, or in a calculator or a spreadsheet. If you do ask for a worked solution, check each step rather than the final line — the final line is exactly where an error is least visible.

Where they are genuinely strong in STEM: explaining what a formula means, why a method works, what a result implies, and where a definition you half-remember actually comes from.

Academic integrity

• Check your institution's policy, in writing, and check whether your individual module has its own. Some allow AI as an aid, some restrict it to specific uses, some prohibit it. Policies differ between departments at the same university. Do not assume, and do not rely on what a friend told you.
• Don't submit AI's work as yours. It is an integrity violation, and it also costs you the learning you are paying for.
• Declare it if required, and keep the declaration specific: which tool, for what — "used for grammar checking and for generating practice questions" is a real statement, "AI was used" is not.
• When the policy is ambiguous, ask your tutor before you submit. Asking costs nothing. Asking afterwards is not asking.

What a week of this actually looks like

The advice above is only useful if it survives contact with a real timetable, so here is the shape of it.

1. After each lecture, ten minutes. Explain the main idea out loud to the tool and let it find the vague bits. This is the highest-value ten minutes in the week, because the material is still fresh enough to fix cheaply.
2. Once a week, closed-book quiz on everything so far, not just this week. Mixed, not blocked. Note what you missed.
3. Before a reading, one question: "what does this chapter assume I already know?" Fill that gap first and the reading takes half as long.
4. When writing, use it at the start and the end — angles before you draft, criticism after — and not in the middle, where the thinking has to be yours.
5. Never the night before. By then only practice helps, and explanations are just comfort.

When to close the laptop

• When you have not tried the problem yet. Reaching for help before you have struggled removes the part that does the teaching. Give it ten real minutes first.
• When you are revising the day before. At that point, practice under exam conditions beats any explanation. You need retrieval, not input.
• When the task is the skill. If the assignment exists to teach you to write, or to read a paper closely, outsourcing the difficulty defeats the assignment — and the skill was the point, not the mark.
• When you are using it to feel productive. Generating summaries you never revisit is a very comfortable way to spend an evening without learning anything.

Recommended tools

• Explaining & learning: ChatGPT, Claude.
• Research with sources: Perplexity.
• Learning from your own documents: NotebookLM.
• More: the best AI tools (including free ones).

The free tiers are genuinely sufficient for studying. Before paying for anything, check whether your institution already provides access — many do, and a university licence usually comes with better data handling than a personal account. While you are checking, find out whether the tool trains on what you paste, and think about that before uploading unpublished research or anything covered by a confidentiality agreement.