AI tutors can help you learn more in less time, or they can quietly do the work for you. Three studies show where the line is, and what to look for before you pick one.
Most students who try an AI tutor want one thing: to learn faster, for real, so the material is still there on test day. The good news is that AI tutors can do that. The catch is that the same technology can also make you worse at a subject while your homework grades go up. Which one you get depends less on the AI and more on how the tutor is built. This guide walks through the best studies so far, what separates a real tutor from a chatbot, and a checklist you can use before you commit to any app.
Yes, a well-designed AI tutor works. In a 2025 Harvard trial, students using a carefully built AI tutor learned more than twice as much as students in an active-learning class, in less time. But a plain chatbot that hands out answers can backfire. In a large high school study, students who practiced with one scored 17% lower on the exam than students with no AI at all.
So the question to ask is not "does AI tutoring work?" It is "does this tutor make me think?"
In 1984, the education researcher Benjamin Bloom published a short paper with a big claim. Students taught one to one, with regular checks for understanding, did about two standard deviations better than students taught in a normal class. In his words, "the average tutored student was above 98% of the students in the control class" (Bloom, 1984).
Bloom called it a problem because nobody can afford a personal tutor for every student. He challenged researchers to find group methods that came close.
Two sigma turned out to be optimistic. Bloom's numbers came from small studies over a few weeks. When later researchers pooled dozens of randomized tutoring programs for K-12 students, the average effect was about 0.37 standard deviations (Nickow, Oreopoulos and Quan, NBER, 2020). That is far smaller than two sigma. It is still one of the largest effects in education research. Tutoring works. It just does not work like magic.
That is the bar an AI tutor has to clear: a patient, one-to-one teacher that is available at 11 p.m. the night before a test, at a price most families can pay.
There are now a few serious studies, not just surveys of how students feel. Three are worth knowing.
Greg Kestin and colleagues ran a randomized trial with 194 students in a large undergraduate physics course at Harvard (Kestin et al., 2025). Half the time, students learned a topic from an AI tutor at home. The other half, they learned it in class through active learning, which is already a strong way to teach. The groups then swapped.
The results:
The design is the part most headlines skipped. The team did not just point students at ChatGPT. The tutor walked students through each problem one part at a time. Its instructions told it to keep students active, avoid overloading them, and encourage a growth mindset. The researchers also gave it complete step-by-step solutions written in advance, so its feedback would be accurate. The authors are careful to say AI tutoring will not always beat class, for example on tasks that need complex synthesis of several ideas.
Rori is an AI math tutor that runs over WhatsApp on basic phones. A study followed about 1,000 students in grades 3 to 9 across 11 schools in Ghana. Students used Rori for two 30-minute sessions a week for eight months, on top of normal math class (Henkel et al., 2024). Their math growth scores were higher than the control group's, with an effect size of 0.37.
That is close to the pooled effect of human tutoring programs, from a cheap tool on a low-bandwidth network. The authors ask readers to interpret it with care: it covers only the first year, and schools, not individual students, were assigned to groups.
The most useful study is the one where AI hurt. Hamsa Bastani and colleagues gave nearly 1,000 high school math students in Turkey access to GPT-4 during practice sessions (Bastani et al., 2025). One group got a plain ChatGPT-style chat ("GPT Base"). Another got "GPT Tutor," which was told to give hints instead of answers and was loaded with teacher-written solutions and common mistakes. A third group had no AI.
During practice, both AI groups did better: grades rose 48% with GPT Base and 127% with GPT Tutor. Then came the exam, with no AI allowed. The GPT Base group scored 17% lower than students who never had access. The GPT Tutor group largely avoided that harm.
Three details from the paper stay with me as a teacher:
That last point is the real risk. Answer-first AI feels like progress. The bill arrives on test day.
Put the three studies side by side and the pattern is plain. The tools that helped all made students do the thinking. The tool that hurt did the thinking for them. A study of university students learning Python found the same split. Using AI to generate solutions let students cover more topics but understand each one less. Asking it for explanations deepened understanding (Lehmann, Cornelius and Sting, 2024).
In practice, four behaviors separate a tutor from a chatbot.
It asks before it tells. A tutor starts with "What do you think the first step is?" A chatbot starts with the solution. Socratic questioning is slower per problem and much faster per skill.
It remembers your gaps. If you mixed up sine and cosine last week, a good tutor brings it back this week. A chatbot starts every chat from zero.
It checks your work, not just its own. A tutor wants to see your attempt and finds the exact step where it went wrong. That is the feedback Kestin's team built in.
It plans ahead. A tutor knows your exam is in three weeks and spaces your review. Research shows that spreading practice over several sessions leads to better long-term memory than cramming the same time into one (Cepeda et al., 2006).
If you want the full method for studying this way, see our guide on how to study with AI. For the case against answer-first use, read is ChatGPT bad for learning?
Neither one wins on every count. Here is how they compare.
Many families end up using both: a human tutor once a week, and an AI tutor for the daily practice in between.
Use this list whatever app you are looking at, including ours.
How the main options compare. Khanmigo, from the nonprofit Khan Academy, is built on a large free library of lessons and is free for teachers. Parents and adult learners in the US pay a monthly or annual fee (Khanmigo). ChatGPT now has a study mode that uses Socratic questions and hints, available on its free plan. The trade-off is that you can switch it off mid-chat. Photo solvers such as Photomath are fast at showing answers and steps, which is useful for checking work and risky as a first move.
How Kuest does it. Kuest is an AI tutor built around the habits above. Its Guide me mode gives hints one step at a time, and Show me walks through a full worked solution when you ask for one; parents can lock the hint-first mode on. Quiz and flashcard answers are graded on the server, so they never reach your browser before you submit. Exam prep builds a study plan to your test date with spaced review, and you can see and delete what the tutor remembers about you. It runs code in 10 languages and draws graphs for math. Our line is that you can study anything 10x faster with the right tutor. That is a goal, not a lab result: the research above is about tutors with this kind of design, not about Kuest itself.
Parents can read more on our page for parents, and teachers on the page for teachers. Plans are on the pricing page.
Some are, with limits. ChatGPT's study mode is on its free plan. Khanmigo is free for teachers but paid for learners and parents. Kuest has a free plan with a daily allowance, and paid plans add more usage and features such as mock exams.
No, and the best studies do not claim it can. In the Harvard trial, the AI tutor taught material from a course designed and run by instructors. AI tutors are strongest as daily practice and one-to-one help between classes.
It can be, if it makes you practice real question types under time and reviews your mistakes. Look for timed practice, explanations of why each wrong option is wrong, and a plan that runs to your test date. See our SAT page.
The evidence says yes when the tutor gives hints and checks your steps, and no when it hands out answers. The Bastani study of high school math students showed both results in one trial.
Yes. In one study, a plain chatbot was right only about half the time on high school math problems. Good tutors reduce this with worked solutions, math tools and code runners, and they ask you to check the result yourself.
Kuest is the AI tutor built to help you study anything 10x faster, by teaching you instead of answering for you. Start free.