Should you use AI to study? The honest answer

Do you use AI in your studies? That is the question running through the study subreddits right now, and it sits on top of two harder ones: am I an idiot for not using ChatGPT to write papers, and why did everyone suddenly start using it? You are not an idiot. And the panic in those two questions is understandable, because the same tool that could save you an evening of work can also quietly save you from doing the thing the exam is testing.

Here is the honest version before the detail. AI can genuinely help you study, and it can genuinely undo your studying, and those are the same buttons in the same app. The difference is which side of the work it sits on. Everything below is about putting it on the right side.

Should you use AI to study at all?

Yes, with one condition: use it for the part of studying that still makes you do the work, not the part that replaces it.

This is not a "AI good, humans bad" argument. A large body of work on how people learn keeps coming back to a small set of techniques that move the needle — practice testing, working material more than once over several sessions, and explaining or generating the idea in your own words — and a widely cited survey of learning techniques (Dunlosky and colleagues, 2013) rates those as the highest-yield general-purpose ones, with passive re-reading and highlighting among the lowest. None of that changes because a fast model can type at you. The learning still happens in the same place it always has: in the act of pulling the answer out of your own head.

So the question was never "AI or no AI." It is "will I still be the one doing the recalling?"

The one split that decides everything

There are really two ways to use AI on a topic, and they feel about the same inside the app.

The version that does the work for you. You paste in a page and ask it to explain, summarise, or answer. You read the output. It is clear, well-organised, and you feel like you understand it. This is smooth. It is also, in the parts that matter, almost exactly as good as re-reading your notes — because the retrieval never happens, and retrieval is the part that builds the memory. The finding behind this is the testing effect: Roediger and Karpicke showed in 2006 that being tested on material improves long-term retention compared with re-studying it, and the effect is one of the most robust in learning research. If you skip the test, you skip the gain.

The version that makes the work harder. You still do the recalling; the AI just changes what you are recalling, or checks you. You ask it to turn a dense page into practice questions, then you answer them yourself. You ask it to explain a concept a different way when the textbook did not land. You write your own answer and ask it to poke at where it is wrong. This feels worse — slower, more effortful — and that is the point. Deliberately making a task feel a little harder is a well-known way to make the learning stick harder, what researchers call desirable difficulties.

The two versions produce a similar amount of text. The first one produces understanding on paper and a gap in your head. The second one produces a gap on paper and learning in your head. Same app, opposite direction.

How to use AI so it builds memory instead of skipping it

Concretely, the useful uses are the ones that keep the answer on your side of the conversation:

• Turn dense material into practice questions. Give it a page — a mechanism, a set of terms, a historical argument — and ask for five or ten questions of increasing difficulty, then hide the answers and write them from memory. This is the highest-leverage use, and it works for exactly the subjects that hurt most: anatomy, law, drug names, vocabularies in another language. If you are drowning in lecture slides or a chapter you cannot read all of, the same move applies — see how to work from bullet-point slides and how to memorise a large amount of material.
• Explain the thing back to it, the wrong way on purpose. When you are stuck, tell the AI how you understand it, including the part you think is shaky, and ask it to find the flaw in your version. This is the generation and testing idea pointed at a hard concept.
• Check, not produce. Write your own answer first. Then ask the AI where you are right, where you are wrong, and what you missed. You did the recall; it did the grading.
• Spaced it out. Do the questions today, then the same five again in a day or two, then a few days after that. The spacing is something you do — spread the same retrieval across more than one day rather than cramming it into one sitting — and it is one of the few things that reliably beats a single long session.

A worked example. Say you have a page on a class of drug: its mechanism, its main side effects, what to monitor. The lazy move is "summarise this for revision," then read the summary and feel like you know it. The useful move is "turn this page into eight practice questions, hardest last," then answer each one from memory before you look. The first move reads smoother tonight. The second one is the one still there when the question shows up on the day.

If you are staring at a dense textbook page or a photographed whiteboard, one honest way to get those questions without typing them out by hand is a photo-to-flashcards app — you photograph the page and it turns it into question-and-answer cards you then have to answer yourself. SnapDeck AI, for example, does exactly that: photograph a page, notes or whiteboard, import a PDF or paste text, and it makes the cards in seconds, with a study mode where you flip the card and mark what you know versus what to review. Handwriting the cards yourself is completely fine and free, and any flashcard tool that forces you to write the answer does the same job; the photo version just saves the typing so the time goes into the recall instead.

The common thread: in every one of these, you produced the answer. The AI did the asking, the checking, or the card-making. That is the side of the work you want it on.

When AI quietly works against you

The other side of the split is easy to slip into because it feels productive:

• Writing the deliverable. The essay, the problem solution, the report you submit. If the AI wrote it, you did not learn it, and it is the part your course rules will treat as yours to own. This is the thing behind "am I an idiot for not using ChatGPT to write papers?" — and the answer is no, because having it write the paper does not make you the person who knows the material.
• Skipping the struggle. The first time a question is hard, that difficulty is not an obstacle to be removed; it is the load you are training under. If the moment you feel stuck is the moment you reach for the answer, you have moved from the "harder" version to the "saves you" version without noticing.
• Trusting it when it should not be trusted. It is fluent and it does not look unsure, so it can hand you a confident wrong answer — a name, a date, a mechanism that is slightly off. In a field where precision is the point, that is not a small thing. Verify anything you will actually be graded on against your own source.

How do you know which kind of AI study you are doing?

Run the one-line test: close the app and try to do the task alone.

If you can do it — write the answer, work the problem, state the mechanism — the studying did something, and the AI helped you do the work. If you cannot, and it only made sense while the app was showing you the answer, then the app did the work and you got a good read. Both take the same twenty minutes; they are not the same twenty minutes.

This test is also a good place to notice when you have drifted. You meant to use AI to quiz yourself, and somewhere you started reading what it produced instead. The test catches it in one line, and it costs you nothing to run.

A few honest limits

• It is a tool, not a study system. It does not replace the schedule, the sleep or the fact that spaced retrieval over days beats one cramming session. It can make a session more useful; it cannot replace the session. For the schedule side, a revision timetable that survives real life is where most of the actual gains are.
• Course rules vary a lot. What is fine as practice and what counts as AI-assisted submitted work is set by your course, not by the tool. Check the specific rules you are under before you do anything you plan to hand in.
• Keep the answer on your side. Every time, the habit is the same: you produce, the AI checks. The moment that habit flips is the moment the tool stops helping.

The short version

Use AI to study, but only on the side that makes you do the work. Generate the practice questions, explain a stuck concept, check your own answer, space the retrieval over days — and keep the producing of the answer on your end. Let it write the essay or the answer and you will feel productive and learn almost nothing. The whole thing fits on one line: close the app, try it alone, and if you can do it, the studying did something.