What Is an AI Tutor? How AI Tutoring Works, With an Example

What is an AI tutor?

An AI tutor is software that works through a problem with you instead of for you. It reads your attempt, finds the step or idea that went wrong, and answers with a hint, a short explanation or a follow-up question. Then it picks the next task based on what you have shown you can and can't do yet.

That last part is the difference from an answer engine. An answer engine gives you a finished solution. An AI tutor treats your mistake as information and uses it to decide what to teach next.

AI tutoring is the use of that loop, attempt → diagnosis → hint → new attempt, for homework, revision and exam prep. The idea is older than chatbots. Researchers have built intelligent tutoring systems for decades, mostly for math and science, with hand-coded rules for each skill. Newer AI tutors add large language models, so you can type or photograph a problem and ask questions in plain language.

What makes something a tutor is how it behaves: it asks for your thinking, responds to the specific step you got wrong, and checks that you can do the next problem on your own.

How AI tutoring works

Most AI tutors, old and new, combine four parts:

• Subject knowledge. The skills, rules and common mistakes for a topic, for example that a number in front of brackets multiplies every term inside.
• A model of you. A running estimate of what you know, updated with every answer, hint request and error.
• A tutoring strategy. Rules for the next move: give a nudge, explain a concept, show an example, or move on to new material.
• An interface. Chat, a step-by-step workspace or a camera scan, on a phone or laptop.

In a 2011 review of tutoring studies, Kurt VanLehn found that tutoring systems giving feedback on each step came close to human tutors: an average effect size of 0.76 for step-based systems against 0.79 for human tutoring, both compared with no tutoring. Systems that only checked the final answer did much worse (0.31). What separated the two groups was how often the system responded: after every step, not only at the end. One limit: the systems in that review were built well before today's chatbot-based tutors, so the numbers don't automatically carry over to them.

A worked example: the same mistake, two kinds of help

A student solves this equation:

$$3(x - 4) = 2x + 5$$

Their work:

$$3x - 4 = 2x + 5 \quad\Rightarrow\quad x = 9$$

The answer is wrong. The 3 was multiplied by $x$ but not by $-4$.

What an answer engine does. It returns the full solution: $3x - 12 = 2x + 5$, so $x = 17$. The student copies $x = 17$ and moves on. Nothing in that exchange tells them which of their own steps was wrong, so the same distribution error turns up on the next problem and on the test.

What an AI tutor does.

1. Hint, not answer. "Before you go on, check your result. If $x = 9$, does the left side equal the right side?" The student works it out: $3(9 - 4) = 15$ but $2 \cdot 9 + 5 = 23$. Something is off.
2. Error diagnosis. "Your first line is where it changed. The 3 multiplies everything inside the brackets. What is $3 \times (-4)$?" The student writes $-12$, gets $3x - 12 = 2x + 5$, and solves $x = 17$. Check: $3(17 - 4) = 39$ and $2 \cdot 17 + 5 = 39$.
3. Follow-up on the same skill. "Try one more: $-2(x + 3) = 4$." This one tests the same rule with a negative factor. The student writes $-2x - 6 = 4$, so $x = -5$, and the check works: $-2(-5 + 3) = 4$.
4. Update the model. The tutor notes "distributing over brackets" as a weak spot and brings it back in later practice.

Both tools reached $x = 17$. Only the second one left the student able to solve the next problem alone. A 2025 field experiment with nearly 1,000 high school math students (Bastani et al., PNAS) found a similar pattern at scale: students given a plain ChatGPT-style assistant did better on practice problems but 17% worse on a later exam without AI than students who never had it. A version built with teacher-designed hints and safeguards largely removed that drop.

AI tutor vs human tutor vs general chatbot

	AI tutor	Human tutor	General chatbot
Availability	Any time, instantly	Scheduled sessions	Any time, instantly
When you make a mistake	Points to the step and asks you to fix it	Points to the step, reads your reaction	Usually rewrites the full solution
Adapts over time	Tracks your errors across many problems	Remembers you and your goals	Little or no memory of your learning
Motivation and accountability	Streaks and reminders at most	Strong: a person expects you to show up	None
Accuracy	Can make mistakes; better tools show every step to check	Can also slip up, but can reason it through with you	Can make confident mistakes
Cost	Often a free tier plus a subscription	Paid per session	Often free or a subscription

The three are not mutually exclusive. One setup that plays to each one's strengths: a weekly session with a person for motivation and big-picture planning, and an AI tutor for daily practice in between.

What AI tutors are good at

• You get feedback the moment you're stuck, instead of finding out at the next lesson that step two was wrong.
• A good tutor tells a sign error apart from a misunderstood rule and responds differently to each.
• It can generate another problem on exactly the skill you just missed, as many times as you need.
• You can ask the same question three times. It costs nothing and nobody is watching.
• It's there at 11 pm before a test, when no one else is.

The limits of AI tutors

• They can be wrong. Language-model tutors can drop a minus sign or state a rule incorrectly. If you're learning a topic for the first time, you may not catch it. Check the result, as in step 1 of the example above.
• They can become a crutch. If the tool will give the full answer whenever you ask, it is easy to stop thinking. In the Bastani study, students using a plain ChatGPT-style assistant did 17% worse on the later exam.
• They only see what you show them. An AI tutor doesn't know you slept four hours or panic during tests. A teacher or parent might.
• They need your data. A model of your learning is personal information. Read the privacy policy, especially for younger students.

A 5-point checklist for choosing an AI tutor

1. Does it ask for your attempt first? Give it a problem and see whether it asks what you tried or dumps the solution.
2. Does it find the exact step? Make a deliberate mistake, like the distribution error above, and see whether it names that step.
3. Does it give a follow-up? After you fix a mistake, a tutor should offer a similar problem to confirm you've got it.
4. Does it fit your course? Test it with problems from your own textbook or past papers, not the demo examples.
5. Can you check it, and trust it with your data? Every step should be visible, and the privacy policy should say what is stored and who sees it.

If you're comparing math tools specifically, see our comparison of AI math tutors.

Where Astra AI fits

Astra AI is an AI tutor that works step by step: it asks questions back and checks your understanding before it gets to the final answer. You can type a question or snap a photo of a problem. Its Exam Prep feature turns notes, PDFs or past tests you upload into a study plan that ends on your test date, with quizzes, flashcards and the tutor to explain what you're stuck on. You can start free without a card.

The same limits apply to Astra AI: it can make mistakes, and it works only from what you type, scan or upload. Use the check-your-answer habit from the example above, and bring anything that still doesn't make sense to your teacher.

Try Astra AI free