Aiodic

Back to the table

83 Advanced Ed Tutoring Applications

Tutoring

Personalized teaching and explanation.

What it is

AI tutoring means using an AI model as a personal teacher: something that explains an idea at your level, asks you questions to see what you've understood, gives you practice, and changes its approach when you get stuck. The same chatbot you use to draft emails can do it, if you ask it to behave like a tutor rather than an answer machine. A growing number of products are built specifically for it, from Khan Academy's Khanmigo to the "study" and "learning" modes now built into the big chatbots.

The difference between tutoring and just asking questions is worth spelling out, because it's the whole point of this element. When you ask a chatbot "What is compound interest?", it gives you a tidy explanation and stops. You read it, nod, and feel you understand. A tutor does something different. It asks what you already know. It explains one piece, then asks you to use it. When you get it wrong, it doesn't just correct you; it works out which part of your thinking went astray and goes back there. And a day or a week later, it asks you again.

Good AI tutoring usually combines five moves:

None of this needs special software. It's mostly a matter of the prompt you write, and in particular of giving the AI a tutor's role with clear rules, which is role prompting put to practical use. The built-in study modes are, underneath, a carefully written set of instructions of exactly this kind.

This element is marked Advanced not because the technique is hard, but because doing it well means resisting the easiest path. A chatbot's default is to be helpful by giving you the answer. A tutor's job is to be helpful by making you find it. Getting the second behaviour, and noticing when you've slipped back into the first, takes some deliberate effort from you.

Why it matters

One-to-one teaching works, and most people have never had it. Education researchers have known for decades that a student with a personal tutor usually learns far more than one in a class of thirty. The catch has always been cost: a good human tutor is expensive, so personal tutoring has mostly been for people who could pay. An AI tutor is available at two in the morning, doesn't lose patience on the fifth attempt, and costs little or nothing. For an adult teaching themselves bookkeeping, a new language, video editing or basic coding, that changes what's possible.

It meets you where you are. A course or a book is written for an imagined average reader, so it's too slow in some places and too fast in others. A tutor can skip what you know and slow down on what you don't. If you're a novelist trying to understand how book royalties are calculated, it can use your own book's numbers. If an explanation doesn't land, you can ask for it again with a different example, as many times as you like, without embarrassment.

The research shows it can help, and also that it can hurt. This is the part that makes tutoring an Advanced element. The studies in the next section show AI tutors producing large learning gains when they're designed to teach. They also show that students using a plain chatbot to get through practice problems did better in practice and worse on the test afterwards, because the chatbot did the thinking for them. The same technology produces both results. What decides which one you get is how it's set up and how you use it.

Feeling that you've learned isn't the same as learning. Reading a clear explanation produces a strong sense of understanding, and that sense is often wrong. In a well-known 2006 experiment, Henry Roediger and Jeffrey Karpicke had students either reread a passage or practise recalling it. Rereading felt better and even did better on a test five minutes later, but on tests two days and a week later the students who had practised recalling remembered substantially more, and the rereaders were more confident despite remembering less. An AI that explains beautifully can leave you in the same trap. A tutor that makes you answer questions gets you out of it.

The explanations can be wrong. AI models can state false things fluently and with complete confidence, which is called hallucination. In a classroom that's a real hazard, because a learner, by definition, can't yet tell a correct explanation from a plausible wrong one. Any honest account of AI tutoring has to include this, and the practices below are partly about catching it.

Real-world examples

Bloom's "2 sigma problem" (1984)

The modern interest in tutoring starts with a short paper by the educational psychologist Benjamin Bloom, published in the journal Educational Researcher in 1984 under the title "The 2 Sigma Problem: The Search for Methods of Group Instruction as Effective as One-to-One Tutoring". Drawing on studies by his doctoral students at the University of Chicago, Bloom reported that the average student taught one-to-one by a tutor performed better than about 98 percent of students taught in a conventional class. In statistical terms that is two standard deviations, or "two sigma", above the class average.

Bloom's "problem" was that society couldn't afford a tutor for every student, so he challenged researchers to find group methods that came close. For forty years the paper has been the reason people get excited about tutoring machines, and AI companies quote it often.

It's worth knowing that later research found a smaller effect. In 2011 Kurt VanLehn reviewed many studies comparing human tutoring with no tutoring and put the average effect at about 0.79 standard deviations, not 2, and found that well-designed computer tutoring systems came surprisingly close to that. That's still a large effect by the standards of education research. The honest summary is that personal tutoring helps a lot, not that it performs miracles, and that the two-sigma figure should be treated as the high end of a real benefit.

The Harvard physics tutor (autumn 2023, published June 2025)

In the autumn 2023 semester, Gregory Kestin and Kelly Miller, who teach physics at Harvard, ran a randomised controlled trial in Physical Sciences 2, an introductory physics course for life-science students and Harvard's largest physics class. About 194 students took part. For two lessons, on surface tension and fluid flow, half the class learned in a normal active-learning class session (already a well-regarded, interactive way to teach) while the other half learned the same material at home from an AI tutor built on GPT-4. The next week the groups swapped, so every student tried both.

The tutor wasn't plain ChatGPT. Kestin built it with detailed prompts that included step-by-step solutions to the problems, so the model had correct answers to work from rather than having to generate them, and with instructions to guide students through one step at a time instead of giving answers away. The results, published in Scientific Reports on 3 June 2025, found that the median learning gains of students using the AI tutor were more than double those of the students in class, and that they got there in less time: the median time on the AI lesson was 49 minutes, against a 75-minute class. Students also reported feeling more engaged and motivated.

Two details matter for anyone copying the idea. First, the tutor was carefully built by teachers who knew the material, and supplied with correct solutions, which is a form of grounding. Second, the researchers present it as a complement to teaching, not a replacement; in the Harvard Gazette's report, Kestin stressed that an AI tutor shouldn't do students' thinking for them. The study covered two lessons in one course, so it shows what a well-designed tutor can do, not that any chatbot will do the same.

When the chatbot becomes a crutch: high school maths in Turkey (autumn 2023)

A team of researchers from the Wharton School and the University of Pennsylvania, led by Hamsa Bastani, Osbert Bastani and Alp Sungu, ran a field experiment with nearly 1,000 students in grades 9 to 11 at a high school in Turkey during the autumn term of the 2023 to 2024 school year. Across four 90-minute practice sessions, some students had a plain GPT-4 chat interface ("GPT Base"), some had a version prompted to act as a tutor ("GPT Tutor"), and some had only their books and notes. The tutor version was told to give hints without handing over the answer, and its prompt included the teacher's correct solutions and common student mistakes.

During practice, with the AI available, both groups did better than the students without it: 48 percent better with GPT Base and 127 percent better with GPT Tutor. Then came an exam with no AI. The GPT Base students scored 17 percent lower than the students who never had AI at all. The GPT Tutor students scored about the same as the control group, so the hint-giving design removed the harm, though in this study it didn't produce a gain on the exam either. The researchers' explanation was that many students used the plain chatbot to get answers, using it as a crutch, and so never built the skill themselves. Worse, those students didn't think they had learned less.

The study also measured how often plain GPT-4 got the practice problems right. Asked for the answer, it was correct only about half the time, with logic errors in 42 percent of cases and arithmetic errors in 8 percent. The paper first appeared online in July 2024 as "Generative AI Can Harm Learning", and was published in the Proceedings of the National Academy of Sciences (PNAS) in 2025 as "Generative AI without guardrails can harm learning". It's the single best piece of evidence for why an AI tutor should make you work.

After-school English in Nigeria (June to July 2024)

In June and July 2024, a World Bank team ran a six-week after-school programme in Benin City, in Edo State, Nigeria, for first-year senior secondary students. Teachers guided students through sessions using Microsoft Copilot, powered by GPT-4, to practise English grammar and writing. Students were randomly assigned to take part or not, so the researchers could compare the two groups fairly.

The World Bank's working paper, published in May 2025, reported an improvement of 0.31 standard deviations on an assessment covering English, AI knowledge and digital skills, and 0.23 on English alone. The team estimated the gain as equivalent to 1.5 to 2 years of ordinary schooling, and said it outperformed about 80 percent of the education programmes in developing countries that have been tested this way. Students who attended more sessions gained more, and the benefits showed up in their end-of-year exams too.

Two points keep this in proportion. The AI wasn't working alone: teachers designed the sessions and guided the students, which is the human in the loop doing real work. And the paper found the largest effects for girls and for students who were already doing well, which is a reminder that AI tutoring can widen gaps as well as close them if the students who struggle most don't get extra support.

The chatbots add a tutor switch (2025)

In 2025 the major chatbot makers built tutoring behaviour into their own apps. Anthropic introduced a "learning mode" for university users of Claude in April 2025, designed to guide students with questions rather than hand them answers, and made a Learning style available to all Claude users in August 2025. OpenAI launched "study mode" in ChatGPT on 29 July 2025, describing it as step-by-step guidance with questions to check understanding, available on free and paid plans.

Both are, at heart, carefully written instructions, the same idea as a system prompt. OpenAI said study mode was powered by custom system instructions and acknowledged this could mean inconsistent behaviour and mistakes across conversations. And the learner stays in control: as Education Week reported at launch, a student can simply switch study mode off and get the answer. That's the honest limit of every tutor mode. It can offer to teach you, but it can't make you want to learn. Names, availability and features change often, so check what your chatbot offers at the time you read this. (Earlier examples of the same idea, Khanmigo and Duolingo Max, both launched on 14 March 2023; see role prompting and API.)

An everyday scenario: the author and the spreadsheet

Imagine Dana, a self-published author, who wants to understand her own royalty reports and build a spreadsheet to track them. She's avoided spreadsheets for years. She asks a chatbot to act as a patient tutor, tells it she's comfortable with arithmetic but has never written a formula, and asks it not to give her answers she could work out herself.

The tutor starts by asking what she wants the sheet to tell her. It explains one idea, cell references, using her own sales figures, then asks her to write the formula for one month's royalties. She gets it wrong, and instead of correcting it the tutor asks what she expected the formula to do and where the number came from. She spots the mistake herself. Twenty minutes later she explains back, in her own words, how the sheet works; the tutor points out that she's mixed up gross and net, which was the misunderstanding underneath her original confusion. Three days later she opens a new chat, pastes her notes, and asks to be quizzed. She remembers most of it, and the parts she's forgotten are exactly the parts she now practises.

Try it yourself

This takes about 20 minutes, plus five minutes a few days later. You need any chatbot and something you genuinely want to learn: a concept from your work, a skill for your business, a topic you've always felt shaky on. Keep it small enough for one sitting, such as "how email open rates are calculated" rather than "marketing".

  1. Choose your topic and write down what you already know. One or two sentences. This is for the tutor, and it's also a baseline you'll compare against later.
  2. Turn on a tutor mode if your chatbot has one, or set up the tutor yourself. Paste this at the start of a fresh chat:
I want you to be my tutor. The topic is: [your topic].
About me: [one or two sentences about your background and what you already know].
My goal: [what you want to be able to do or explain at the end].

Rules for this session:
- Teach in small steps. Explain one idea, then ask me a question about it before moving on.
- Don't give me answers I could work out myself. If I'm stuck, give me a hint first, then a smaller hint.
- When I'm wrong, ask me what I was thinking before you correct me.
- Use examples from my world, not textbook examples.
- If you're not sure something is accurate, say so.
Start by asking me one question to find out what I already understand.
  1. Work through it for about 15 minutes. Answer honestly, including "I don't know". Resist the urge to type "just tell me". If you're stuck, ask for a hint instead. If an explanation doesn't land, say so and ask for a different example.
  2. Ask for a short quiz. When you've covered the main ideas, paste:
Now give me five short questions on what we've covered, one at a time.
Mix easy and hard. Wait for my answer before giving the next one,
and tell me after each one whether I was right and why.
  1. Teach it back. This is the most useful five minutes of the session. Close the explanation, and in your own words write a short explanation of the topic as if to a friend who knows nothing about it. Then paste:
Here is my explanation of [topic], written for a beginner.
Check it like a strict but kind teacher:
1. List anything that is wrong.
2. List anything important that is missing.
3. List anything that is vague or that a beginner would misunderstand.
Don't rewrite it for me. Ask me to fix each problem myself.
  1. Fix it yourself. Rewrite the weak parts and paste the new version. You'll often find that the part you couldn't explain is the part you didn't really understand.
  2. Check one key fact somewhere else. Pick the most important fact or rule from the session and confirm it in a reliable source: official documentation, a textbook, a reputable website. If the tutor was wrong, tell it, and note that the rest may need checking too.
  3. Ask for your review questions and come back later. Before you close the chat, paste:
Write me 8 review questions on this session, with answers listed separately
at the end, so I can test myself in three days without looking.

Save them in a note. In three days, answer them without looking, then check. Whatever you got wrong goes into your next session. That's spaced practice, and it's what turns a pleasant conversation into something you still know next month.

If your chatbot has memory switched on, it may remember what you studied. You don't need it; pasting your notes into a new chat works just as well, and it keeps you in control of what the AI stores about you.

Common mistakes and misunderstandings

Asking for the answer and calling it learning. This is the crutch effect from the Turkish study. If the AI solves the problem, you've learned that the AI can solve it. You'll feel that you understand, and you'll find out otherwise the first time you try alone. Use the AI to check your attempt, not to replace it.

Trusting a confident explanation. A fluent, well-organised explanation can be wrong, and as a beginner you're the person least able to notice. Plain GPT-4 got about half of the Turkish maths problems wrong. Ask the tutor to flag uncertainty, check key facts elsewhere, and be most careful with numbers, dates, laws and anything you'll act on. Asking the AI to review its own explanation can catch some slips (self-critique), but it's no substitute for an outside source.

Letting it be too agreeable. Chatbots tend to tell people what they want to hear. If you say "I think it works like this, right?", many will find a way to agree. Ask it to find what's wrong with your explanation, not whether it's right.

Skipping the quiz and the teach-back. Explanations feel productive, so people spend the whole session on them. The questions are where the learning happens. If you only have ten minutes, spend five of them answering questions.

Cramming it all in one sitting. One long session feels thorough, but you'll forget most of it within days. Two or three short sessions spread over a week will leave you knowing more.

Expecting the study mode to do the work. Tutor modes help, but they can be switched off, they can behave inconsistently, and they can't force you to think. The rules only work if you follow them.

Handing it to a child without checking the rules. AI tutors for children raise questions about what data is collected and kept, who can read it, and whether the product is suitable for the child's age. At the time of writing, OpenAI's terms require users to be at least 13 and to have a parent's or guardian's permission if under 18, and other services have their own rules; check them for any tool you use. In September 2025 the US Federal Trade Commission opened an inquiry into how seven companies, including OpenAI, Meta and Alphabet, test and limit the effects of their chatbots on children and teens. For a child, prefer a tool built for schools with clear data protections, keep conversations free of personal details, and sit in on the first sessions. See privacy.

Thinking it replaces teachers. The strongest results came from tutors designed by teachers, grounded in correct material, and used alongside real teaching. The AI extends a good teacher's reach. It doesn't remove the need for one.

Related elements