How do you study with an AI tutor, and where should you not trust it?

What separates a tutor from a chatbot?

Both give answers, but one puts you to work and the other does the work for you. Ask a chatbot to explain a topic and you get a competent summary, and reading that summary is no different from reading the book: active recall never fires, and the material only appears to have gone in. A tutor runs the other way round, which is what the Socratic method has been doing for two and a half thousand years: asking instead of answering. Teaching through software is not new either; intelligent tutoring systems are a forty year old field. What is new is that the lesson can come out of your own file and the exchange can happen in speech.

Why does explaining beat listening?

Because explaining is a test and listening is not. The moment you try to put a topic in your own words, the gaps you skimmed past while reading become visible. There is a measured name for this: the illusion of explanatory depth. People report knowing how a zip or a bicycle pedal works, then discover otherwise the moment they are asked to describe it. In revision that illusion pairs with the overconfidence effect and produces the night-before feeling that a topic is finished. Roediger and Karpicke 2006 measured it directly: the tested group remembered more than the group that spent the same minutes rereading. The Feynman Technique is the same logic written down as a procedure.

How do you set up twenty minutes?

Opening with a blank page is the worst start. Pick a file before the session, preferably the chapter you are actually on; this narrows what the tool can talk about, and a narrowed tool behaves better. Spend the first five minutes explaining while it listens. Let the place you stumble decide the topic, because that is precisely the zone of proximal development: the band between what you cannot do alone and what you can do with help. For the next ten minutes swap roles, let it explain while you question, and put a why after every answer, which is exactly the chain elaborative interrogation describes. Spend the last five minutes on a quiz from the same section, because once the talking stops only a test shows what stayed.

Why is interrupting a feature?

With a teacher in front of you, you stop them where you lost the thread. With a recording you cannot, which is why listening to a recording is revision rather than tuition. Being able to cut in also manages cognitive load: if three more sentences pile on top of one you did not follow, working memory fills and you lose the rest. The practical rule is to interrupt in the first second you notice you are lost. Waiting is not politeness, it is loss. How you interrupt matters too: having to say a wake word breaks the turn-taking rhythm of ordinary conversation and tends to keep people quiet.

Where should you not trust it?

This is the section that matters most. These tools sit on large language models, which can produce answers with no source behind them; the technical term is hallucination. The known way to reduce it is to tie the answer to a supplied document, which is retrieval-augmented generation, but reducing is not eliminating. Three concrete rules. Verify any date, figure or definition against the file. Take it seriously when the tool says something is not in your document, because a tool that states its limit is better than one that hides it. And do not let a spoken summary stand in for your own notes. Knowing what you do not know is what a student needs most, and a tool that dresses a guess as an answer destroys exactly that.

What does speaking not solve?

Timing. Say you have understood a topic by talking it through; what remains three weeks later is decided by your review schedule, not by the conversation. The forgetting curve has said the same thing for a hundred and forty years, and the spacing effect shows the fix is distributed review. A spoken session is therefore not a study plan on its own but one step inside one: it produces understanding, not retention. Retention comes from the deck and from weak topic tracking. There is a second thing speech does not solve: visual material. A question that turns on reading a diagram cannot be answered by talking, and you have to look at the page.

By Mert Erarslan (Exam prep and subject routes)

Last updated: 2026-08-25 · Written with AI assistance and reviewed before publishing.