ChatGPT for studying: where it works, and where a board does better

Most students now use ChatGPT, or something like it, to study. It is quick, it is patient, and it will explain the same thing six ways without sighing. It is also the wrong shape for a lot of studying, and the research on how students use it is less flattering than the experience of using it. This post sets out where a chatbot is good, where its shape fails, what the evidence says about studying with one, how to use it well anyway, and what changes when the same AI writes on a board instead.

Where a chatbot is good

A chatbot suits one-answer tasks. Ask what a word means, why a step in a derivation is allowed, what the difference is between two terms you keep confusing, and the answer is a short paragraph you read once. It will reword a definition until it clicks. It gives a decent first explanation of something you have never seen. It can check one line of a calculation or one sentence of an argument. It is also a good place to think out loud at two in the morning when nobody else is awake.

Each of those is a question with one answer, read once, and the chat shape handles one answer at a time well. Studying is mostly not that.

A worked example: one question, two shapes

Take a question every biology student meets: compare mitosis and meiosis. Here is what the two shapes give you.

In chat. You get an opening sentence, a paragraph on mitosis, a paragraph on meiosis, and then, if you asked nicely, a table with six or seven rows: purpose, number of divisions, daughter cells, chromosome number, genetic variation, where it happens. It is correct. It is also four screens tall on a laptop, and to compare "one division" with "two divisions" you scroll between the paragraph and the table. Ask a follow-up about crossing over and the table is gone. Tomorrow, the whole thing is one entry in a sidebar list of forty conversations.

On a board. The same question in Ask produces two columns side by side, mitosis on the left and meiosis on the right, with the matching rows aligned: one division beside two divisions, two identical cells beside four different ones. Under each column, a chain of the phases with a small drawing at each. Between the columns, an arrow at meiosis I marked "crossing over happens here", because that is the row where the two diverge. Ask the follow-up about crossing over and a new box appears next to that arrow, not at the bottom of a scroll. Tomorrow the board is in your library under its own name, and zooming out shows the whole comparison on one screen.

Nothing in the second version is information the chatbot lacked. The difference is entirely in shape: placement, persistence, and the ability to point at one part and ask about it.

Where the shape fails

Each of these is a shape problem, not a model problem. The same model that wrote the chat answer can write the board. What differs is whether the output lands as a column of text or as objects in space.

What the evidence says

The largest controlled study so far of students using a chatbot to study is from Bastani and colleagues at Wharton, run with nearly a thousand high-school maths students in Turkey over the 2023–24 year [1]. Students who had access to a standard GPT-4 chatbot during practice sessions did much better on the practice problems, around half again as many correct. On the closed-book exam afterwards, the same students did worse than students who had no AI at all, by about 17 per cent. A second group used a version of the chatbot tuned to give hints rather than answers. Their practice scores went up even more, and their exam scores were no worse than the no-AI group, but not better either.

The authors' explanation is that students used the chatbot as a crutch: they asked for the answer, copied it, and did not learn the method. The hint-only version prevented the harm but did not add learning on top. The finding has been widely repeated since and fits older work on how people judge their own learning.

That older work is on illusions of competence. Koriat and Bjork showed that when the answer is in front of you, you reliably overestimate how well you would recall it later [2]. Reading a fluent explanation feels like knowing it. Dunlosky and colleagues' review of ten study techniques rated rereading and highlighting, the two closest to reading chat output, as low-utility, and rated practice testing and spaced practice as the two highest [3]. A chatbot's default mode, a clean explanation you read once, is the low-utility end.

None of this says do not use a chatbot. It says the default way of using one, asking and reading, is the way that does not work, and that any tool you study with has to push you towards recall.

How to use ChatGPT well for studying

If a chatbot is what you have, these habits move it from the harmful end of the Wharton study to the useful end.

• Attempt first, ask second. Write your answer, then paste it and ask what is wrong with it. You get feedback on your reasoning instead of an answer to copy.
• Ask for hints, not solutions. "Give me one hint towards the next step, not the answer" works, and matches the version of the tool that did no harm in the study.
• Make it quiz you. "Ask me five short-answer questions on this chapter, one at a time, and grade each before the next." Then answer without looking anything up.
• Explain it back. Type your own explanation of the topic from memory and ask it to find the gaps. Every vague word it flags is something you do not know yet.
• Move the result out. Whatever you learned, write it into your own notes in your own words. A chat log is not a place you will revise from.

The pattern in all five is that the recall stays yours and the chatbot does the labour: writing questions, finding gaps, marking.

What changes on a board

Put the same question to Ask, which runs on an AI whiteboard, and the explanation is written onto an infinite board: headings, notes, arrows, diagrams, placed in space. A comparison becomes two columns you read across. A process becomes a chain you follow. A hierarchy becomes a tree. Tomorrow the board is still there under its own name, and you can zoom out and see the whole topic at once.

It is also yours. Everything the AI writes is the same text and shapes you make yourself, so you fix the one wrong sentence rather than re-asking. Select a part of the board and ask about only that part. Write your own attempt in an empty area beside the AI's version and compare them.

Add your chapter or slides as a source and the board works from your material, and each claim links to the page it came from. Ask Quizbook for a quiz and it writes the questions onto the board, grades written answers on the reasoning rather than the wording, grades diagrams you label or draw, and gives you a report card linking every miss to the passage to reread. Ask Tutor and the topic is taught out loud while the AI draws, with checkpoint questions before it moves on, which is the hint-and-check pattern rather than the answer-and-copy one.

The five habits above still apply. The board makes them the default rather than something you have to remember to do.

A fair way to split it

Keep the chatbot for quick questions, definitions, and a first explanation of something new. Use a board for anything you need to see, anything you need to keep, anything from your own files, and anything you need to be tested on. Bord is free to try as a guest, with no account.

Questions