How to study with AI without cheating yourself

AI is too helpful to study with carelessly. It will hand you a perfect summary, you will read it, and you will feel like you have learned something. Then the exam asks you to produce it and nothing comes. This is not a hunch; it has now been measured. This post goes through what the research found, the one rule that avoids it, a five-step plan that follows the rule, the prompts to use at each step, and the signs that you have slipped.

The problem, measured

In 2024 a team at Wharton ran a controlled trial with nearly a thousand high-school maths students in Turkey [1]. One group practised with a standard GPT-4 chatbot. A second group practised with a version tuned to give hints and never the answer. A third practised with no AI. On the practice problems, the chatbot group did about half again as well as the no-AI group, and the hint group did better still. On the closed-book exam that followed, the chatbot group scored about 17 per cent lower than the students who had used no AI at all. The hint group scored the same as the no-AI group.

The authors' reading is straightforward: students with the unrestricted chatbot asked it for answers, copied them, and never learned the method. The hints version removed the harm but did not add learning on top. In other words, the AI did the one thing a student cannot outsource, and the student let it.

The mechanism is older than chatbots. Koriat and Bjork called it the illusion of competence: when the answer is in front of you, you reliably overestimate how well you would recall it on your own [2]. A fluent explanation feels like knowledge. Chatbots produce fluent explanations on demand, so the illusion is now available at any hour.

The rule

Let the AI do the work that is not learning, and keep the work that is.

Building the questions, finding the page, drawing the diagram you already understand, taking the notes while you listen, checking your working: that is labour, and AI should do it. Recalling the answer, explaining the idea in your own words, deciding what you got wrong: that is learning, and it has to be yours. Every use of AI for studying can be sorted into one of those two bins. If you are about to ask it for something in the second bin, stop and attempt it first.

A plan that follows the rule

1. Get it taught oncePut the chapter or slides on a board and have Tutor teach it out loud while it draws, or have Ask lay it out for you to read. Interrupt where you get lost. This is input, so keep it to one pass. A second pass through the same explanation is where the illusion starts.
2. Explain it backClose the explanation, or scroll it out of view. On a blank part of the board, write the idea yourself in boxes and arrows, from memory. Then open the AI's version beside yours and compare. Every vague word in yours is a gap, and every box the AI has that you do not is a gap. This step is the one most people skip, and it is the step that does most of the work.
3. Get testedAsk Quizbook for a quiz with written questions, not only multiple choice. It grades the reasoning, so a memorised phrase without the logic behind it scores low. Use the help button when you are close, so you finish the recall yourself. Skip when you have nothing, and let the miss stand; a guessed answer teaches you less than an honest blank.
4. Read only the gapThe report card links each miss to the passage in your file. Reread that passage, fix your own note on the board, and move on. The whole chapter is not the problem; one gap is. Rereading the whole chapter is how a two-hour session produces twenty minutes of learning.
5. Come back in a few daysOpen the board, redo the explain-it-back step on a fresh patch of board, and ask for another round of questions. Forgetting in between is what makes the second recall stick. Three days, then a week, then before the exam.

Why each step is there

Each step maps onto a result from the memory literature, and knowing the result makes the step easier to keep doing when it feels slow.

One pass of input. In Dunlosky and colleagues' review of ten study techniques, rereading rated low-utility: it feels productive and adds little after the first pass [3]. Hearing a topic taught once is the cheapest way to get the structure in your head. Hearing it twice mostly buys the feeling.

Explaining back and getting tested. Roediger and Karpicke's test-enhanced learning experiments found that students who studied a passage once and then tried to recall it remembered more a week later than students who studied it four times, even though the restudy group felt more confident [4]. Karpicke and Blunt extended this: practising retrieval beat spending the same time building a concept map, on both factual and inference questions [5]. Steps 2 and 3 are retrieval practice in two forms, one free and one prompted.

Reading only the gap. This is retrieval practice with feedback targeted at the error, which is the condition under which it works best. It also keeps the session short, which keeps you doing it.

Coming back later. Spaced practice is the other high-utility technique in the Dunlosky review, and Cepeda and colleagues' synthesis of 254 studies found that spreading study over time beats massing it in nearly every condition tested, with the best gap growing as the exam gets further away [6]. The forgetting between sessions is not a cost; it is what makes the second retrieval effortful enough to count.

Prompts that keep the recall yours

The difference between the harmful and harmless groups in the Wharton study was one thing: whether the AI gave answers or hints. You can impose that yourself with how you ask. These work in Bord and in any chatbot.

Before you ask anything: "I am going to write what I think the answer is. Do not correct me yet." Then write it. Then: "Now tell me what is wrong or missing in what I wrote, without giving me the full answer."

When stuck on a problem: "Give me one hint towards the next step, not the solution."

To get tested: "Ask me six questions on this chapter: two multiple choice, two fill in the gap, two that need a written explanation. Grade each on the reasoning."

After a miss: "Point me to the part of the file that covers this. Do not summarise it."

At the end: "List the three things I got wrong today as questions, with no answers, so I can try them again on Thursday."

In Quizbook the third and fourth prompts are the default: written questions are graded on reasoning, and every question in the report links to its passage. In Tutor, checkpoint questions come with help, skip and retry, which is the hint-not-answer pattern built in.

Signs you are cheating yourself

• You have read three AI summaries of the same topic and not written one yourself.
• You asked for the answer before you tried the question.
• Your notes are all in the AI's words. None of the sentences are ones you would say.
• You could recognise the right answer in a list but could not say it to a friend.
• Your practice scores are high and your mock exam scores are not. That gap is the Wharton result, in one person.
• Studying has stopped feeling difficult. Retrieval that works is effortful; if it is easy, it is probably rereading in disguise.

Each of these feels productive and is not. The fix for all of them is the same: close the explanation and produce something from memory.

Why a board helps

A chat log encourages reading. A board encourages doing. Your attempt sits beside the AI's explanation and you can see where they differ, which is the explain-it-back step with the comparison built in. The quiz is on the same board as the notes, so a miss turns into a fix in the note rather than a sigh. Your file is on the board too, so "read only the gap" is a click on the citation, not a search through a PDF. Over a term, the board becomes a map of the subject you made, with the parts you found hard marked on it, which is the thing to open the night before the exam.

Questions