What is AI tutor? What the evidence says in 2026
An AI tutor is a chatbot that teaches with hints and questions instead of answers. Here's how it works, real examples, and what two studies found.

An AI tutor is a chatbot set up to teach you rather than hand you answers. It asks you questions, checks your reasoning, gives hints in small steps and adjusts to what you already know, a bit like a patient human tutor who's available at 2am. The best evidence so far says a well-designed one can help you learn faster, but a plain chatbot used as an answer machine can leave you worse off once it's gone.
That second part is the bit most marketing skips, and I think it's the most useful thing to know before you pick one.
Short answer
An AI tutor is a large language model (the same kind of model behind ChatGPT or Gemini) wrapped in instructions that make it behave like a teacher: it withholds the final answer, asks you to try the next step, and explains where you went wrong. Khan Academy's Khanmigo, ChatGPT's study mode and Gemini's Guided Learning are the best-known examples. Whether it helps you depends far more on that teaching layer, and on how you use it, than on which model sits underneath.
The mechanics behind an AI tutor
Here's the short version. A general chatbot is trained to be helpful, and "helpful" usually means giving you the answer as fast as it can. A tutor flips that default.
The change mostly happens in the instructions, not the model. When OpenAI launched study mode in ChatGPT on July 29, 2025, it described the feature as "custom system instructions" written with teachers and pedagogy experts, aimed at things like encouraging active participation and managing cognitive load. Google went a step further: its Guided Learning mode in Gemini runs on LearnLM, a family of models fine-tuned for learning, and it leans on open-ended questions instead of finished answers.
So a typical tutoring loop looks like this:
- You share a problem or topic.
- The tutor asks what you've tried or what you already know.
- It gives the smallest useful hint.
- You attempt the next step, and it checks your work.
- If you're wrong, it points at the specific mistake; if you're right, it moves on or raises the difficulty.
Each step is a chance for the model to slip, too. That matters because the tutor is grading your work with the same model that can misread a fraction or invent a formula.
A worked example with numbers
Say you're stuck on: "A shirt costs $40 after a 20% discount. What was the original price?"
A plain chatbot replies "$50" with a one-line explanation. Done in five seconds, and you've learned roughly nothing.
A tutor would ask something like "If the discount is 20%, what percentage of the original price are you paying?" You say 80%. It then asks you to write that as an equation, and you get 0.8 × P = 40. You divide, you get P = 50, and it asks you to check: 20% of $50 is $10, and $50 minus $10 is $40. Same answer, maybe three minutes instead of five seconds, but you've built the step (turning "20% off" into "pay 80%") that the next ten problems need. The common wrong answer, $48 (adding 20% back onto $40), is exactly the kind of slip a good tutor catches by making you do the check.
Real examples
These are the names you'll run into most. I've kept the table to facts I could confirm on each vendor's own pages.
| Tool | Who makes it | Price | Tutoring approach |
|---|---|---|---|
| Khanmigo | Khan Academy | Free for teachers; $4/month or $44/year for learners and families | Personalised tutoring tied to Khan Academy's course library, plus writing and coding feedback |
| ChatGPT study mode | OpenAI | Launched for logged-in Free, Plus, Pro and Team users | Guiding questions and step-by-step help instead of quick answers |
| Gemini Guided Learning | Inside the Gemini app | Built on LearnLM; probing questions, step-by-step breakdowns, images and diagrams |
Khanmigo's the one I'd look at first for a younger student. On the Khanmigo plans page, the family plan lets a parent add children, see the history of their chats and get moderation alerts, which a general chatbot doesn't really offer. At $4 a month it's also cheap (cheaper than one session with most human tutors, for what it's worth).
Then there are homework apps that scan a photo and explain the solution. They sit closer to answer engines, and our Gauth AI review covers where that style works and where it doesn't. If maths is your main subject, our list of math AI tools compares solvers and tutors side by side.
Benefits and limitations
Two studies are worth knowing, because they point in different directions and both are right.
The case for: a Harvard physics trial
Gregory Kestin and Kelly Miller ran a study with 194 students in Harvard's Physical Sciences 2 course in fall 2023. Each group had one week with an active-learning class led by an instructor and one week with an AI tutor at home, then the groups swapped. In the preliminary analysis, learning gains in the AI-tutored lessons were about double those of the in-class lessons, and students also reported feeling more engaged. The peer-reviewed paper came out in 2025.
The detail people tend to miss: this wasn't raw ChatGPT. The team wrote lesson-specific instructions so the tutor "would behave like a seasoned instructor", and Miller stressed that it was customised with research-based scaffolding. I'd read the result as evidence for careful design, not for chatbots in general.
The case against: a high school maths experiment
Hamsa Bastani and colleagues gave nearly a thousand high school maths students access to GPT-4 through two tutors. One, called GPT Base, behaved like standard ChatGPT. The other, GPT Tutor, had prompts designed to protect learning. During practice, scores went up 48% with GPT Base and 127% with GPT Tutor. Then access was removed. Students who'd used GPT Base did 17% worse on their own than students who never had AI at all, while the safeguards in GPT Tutor largely cancelled that harm. The authors' explanation is that students used the chatbot as a "crutch".
Put the two side by side and the pattern's fairly clear, I think. Practice scores aren't learning. A tutor that gives answers on request raises the first and can lower the second.
Failure modes
- Wrong maths stated confidently. Language models still make arithmetic and algebra slips, and a student often can't tell.
- Easy escape hatches. In most general chatbots you can switch tutoring mode off or just ask for the answer, so the safeguard depends on your discipline.
- Thin context. A tutor doesn't know your syllabus, your teacher's marking scheme or what you covered last week unless you tell it.
- Privacy. Children's chats are stored, which is why parent visibility (as in Khanmigo's family plan) is worth checking.
- Results don't transfer automatically. The Harvard study used custom prompts for one course; you shouldn't assume a generic app gets the same gains.
Using an AI tutor well
Most of the benefit comes from how you set it up. A short checklist:
- Turn on the tutoring mode (study mode in ChatGPT, Guided Learning in Gemini) before you paste the problem.
- Tell it your level and the exact topic, for example "Year 10, simultaneous equations, I keep messing up the elimination step".
- Ask for hints, not answers, and write each step yourself on paper.
- Check every final answer by substituting it back, like the $50 shirt above.
- Finish with two or three new problems done with the AI closed. That's the part the Bastani study suggests actually counts.
Here's a rough decision guide:
| Your situation | Reasonable choice |
|---|---|
| Parent of a primary or middle school child | Khanmigo family plan, mainly for the oversight and moderation |
| High school or university student who already pays for a chatbot | Its built-in study or guided learning mode |
| Stuck on one homework question tonight | Any tutoring mode, but do the final steps yourself |
| Preparing for a high-stakes exam | An AI tutor for practice, plus a human teacher or tutor to check your weak spots |
Honestly, I'd treat an AI tutor as the first pass. A human's the second. That also matches how Kestin framed it: tutors shouldn't replace in-person teaching, they should help students arrive better prepared. If you want more options, we've ranked the best AI for studying and a set of free AI tools for students.
Next for AI tutors
OpenAI, Google and Khan Academy all now ship a dedicated learning mode or product aimed squarely at students, which is why I'd guess the "chatbot vs tutor" split will probably fade.
The more likely shift is that tutoring becomes the default behaviour whenever a model thinks you're a student.
I'd also expect schools to push for versions they can control. Harvard's Derek Bok Center was already piloting tutor bots in several large intro courses and building resources so any instructor could add one. District tools like Khanmigo's, with rostering and single sign-on, point the same way.
The open question is long-term evidence.
Both studies above are short. Nobody yet knows what a full year of daily AI tutoring does to grades or retention, and I'd be wary of any vendor that claims otherwise. For the wider picture, our explainer on AI in education covers the classroom side.
FAQ
What is an AI tutor?
Short version: it's a chatbot that's been told to teach you, step by step, instead of simply handing over the answer the moment you ask for it. It asks what you know, gives small hints and checks each step you write. Khanmigo, ChatGPT's study mode and Gemini's Guided Learning are the best-known examples.
What is the best AI tutor?
It depends on who's learning. For younger children, Khanmigo is probably the safest pick: it costs $4 a month and gives parents chat history and moderation alerts. For older students who already use ChatGPT or Gemini, the built-in study or guided learning mode is likely enough. Pick one that withholds answers by default, because that's what protects real learning.
Can an AI tutor replace a human tutor?
Not fully, in my view. The Harvard trial found strong gains for introductory material, but it used a custom-built tutor and short lessons. AI tutors still make confident mistakes and don't know your syllabus. For exam strategy, a human's still better.
Are AI tutors bad for learning?
They can be, if they just give answers. In the high school maths experiment, students who used a standard ChatGPT-style tutor scored 17% worse once access was removed. Students using a version with learning safeguards didn't show that drop. So the risk is real, but design and habits largely control it.



