On this page
You open the chat with one doubt about friction on an incline. Forty minutes later you have read twelve answers, nodded at all of them, and still cannot solve the next question on your own. The fix is a shorter, sharper session.
How many questions should you ask AI in one study session?
Our rule of thumb: about three focused doubts, each followed by one problem you solve without the AI. Then stop, even if the chat is going well. No study has tested the number three exactly; it comes from two findings that point the same way.
First, your mind holds only a little at once. Psychologist Nelson Cowan's review of short-term memory research put the limit at about four chunks of information (Cowan, Behavioral and Brain Sciences, 2001). Each doubt brings a new idea and often a new formula; three leaves room to connect them.
Second, testing yourself beats reading again. In a classic experiment, students who took recall tests on prose passages remembered more two days and a week later than students who restudied them, even though restudying made them feel more confident (Roediger and Karpicke, Psychological Science, 2006). An AI explanation is something you read. The problem you solve afterwards is the test.
Why does a long AI chat feel productive but teach less?
Because reading a clear explanation feels like learning. In a randomised experiment with identical course materials, college students in active classes learned more but felt they learned less than students in passive lectures, partly because active work takes more effort (Deslauriers and colleagues, PNAS, 2019).
A long chat with AI sets the same trap. A field experiment gave nearly 1,000 high school maths students in Turkey access to GPT-4 during practice (Bastani and colleagues, PNAS, 2025). Students using a standard ChatGPT-style interface improved their practice grades by 48%, but on a later exam without AI they scored 17% lower than students who never had access. A version built to give hints rather than answers raised practice grades by 127%, and its students scored about the same as the no-AI group on the exam (Wharton summary of the study). The authors write that, without guardrails, students tried to use GPT-4 as a “crutch”.
There is a machine-side reason too. Language models use information at the start and end of a long input better than information in the middle (Liu and colleagues, TACL, 2024). In a very long chat, details from the middle of the conversation can get overlooked; we explain this in why AI forgets what you said earlier in a long chat.
What does a three-doubt session look like?
Here is a 40-minute session for a Class 11 student revising friction for JEE Main.
Before you open the chat (5 minutes): write three doubts on paper. For example: why is friction μN and not μmg on an incline; when is static friction less than its maximum value; which way does friction act on a box in an accelerating truck.
Doubt one (10 minutes): ask it, read the answer once, then close the chat and solve one incline problem from your textbook alone.
Doubt two (10 minutes): same pattern, with a static friction problem.
Doubt three (10 minutes): same pattern, with a truck problem.
Close (5 minutes): write one line per doubt from memory, then check your notes.
If your doubt is really “I don't know where to start”, that belongs before the session. See how long to try a problem before asking AI for help.
When should you stop an AI study session?
Stop, or start a fresh chat, when you notice any of these:
You have cleared your three doubts.
You are on your third follow-up about the same doubt, which often points to a gap in an earlier concept.
You scroll past an answer without writing anything on paper.
You start asking about things that were not on your list.
You type “just give me the answer”.
You can follow the explanation but cannot restate it.
Three lines that keep the chat active instead of passive:
“Don't explain yet. Give me one question on this and wait for my answer.”
“Check my working and tell me only the first wrong step.”
“Summarise what we covered in three lines so I can start a fresh chat tomorrow.”
Leftover doubts go on a list for the next session or your teacher.
Where Fermi fits in this session
If you are not sure which doubts to bring, Fermi's AI tutor works out what you already know, builds a plan for the session and checks understanding as it goes. For the solve-it-yourself step, the Fermi Problem Library has worked problems with a step-by-step solver that explains the reasoning rather than printing an answer.
Want a full session plan with breaks? Read the 45-minute plan for studying with an AI tutor.
