Does AI Tutoring Work? How to Tell, and What We Have Measured
Short answer: It depends on the tool, and on how the student uses it. An AI tutor can help a student work through a problem when it asks them to try the next step and explain why. A tool that hands over a finished solution can finish the worksheet without teaching the step. Aristotle's team has measured whether the tutor makes a sound teaching move at hard moments in real sessions. That is a measure of pedagogy in the next turn, not a claim about grades.
When does AI tutoring help, and when does it not?
It helps when the student is doing the work: saying what they tried, explaining a step, and staying with the part that broke. It does not help when the tool does the assignment and the student copies it down.
That is the same split as between a tutor and a homework machine. A chatbot is trained to answer. A tutor is trying to leave the problem open long enough for the student to work.
Aristotle is a voice AI tutor. The student talks through the problem on a shared whiteboard, and Aristotle asks the next question instead of giving the answer. Parents get a summary after each session.
How can you tell if an AI tutor is working?
You do not need a study to check a single session. After the student works with the tool, ask them to do three things without it:
- Explain the step. If they can say why the step is allowed, in their own words, they understood that move. If they can only repeat what appeared on screen, they may have copied it.
- Change one thing. Swap a number, flip a sign, or turn the problem into a word problem. Memorized steps break. Understood steps bend.
- Find a mistake. Show a worked solution with one wrong line and ask them to find it.
During the session itself, listen for questions. A tutor that is teaching asks what the student has tried and why. A tutor that is finishing the work keeps talking until the solution is complete.
If you want to test a tool before a real assignment, ask it for the answer to a problem the student already finished, then stay quiet.
What has Aristotle actually measured?
In November 2026, Aristotle published What would a great tutor do?, a research post by Jaiden Reddy, Vivek Vajipey, Sophie Liu, and Hunter McNichols.
The team sampled 500 live Aristotle sessions (staff, test, and scripted sessions excluded) and isolated 791 pivotal problem-solving turns. Those are the moments where a student is attempting a step, has just done one correctly, has given a wrong answer, or is stuck. They then compared how Aristotle, OpenAI's Study Mode, and Google's Guided Learning responded at each of those turns.
The question they scored was not "did the student learn over a semester." It was whether the tutor's next move was pedagogically sound: focused, restrained, and aimed at leaving the work with the student. The authors say the scores measure next-turn pedagogical decisions, not longitudinal learning outcomes.
Across all 791 turns, Aristotle's pass rate was 54%, compared with 34% for Study Mode, 33% for Guided Learning, and 35% for Claude Sonnet 5.5 on the Guided Learning prompt. The authors also write that 54% is a starting point, and that AI tutors still fall short on harder moments.
That is useful evidence about how a tutor behaves when a student is stuck or wrong. It is not evidence that an AI tutor raises a test score, and this post does not claim that.
What can an AI tutor not do?
An AI tutor is not a teacher who knows the student's classroom. It is not a person who can notice when a student needs a break, or who will be there in the same chair every Thursday. It cannot cover a topic it does not know, and it cannot replace the practice of explaining a step to a parent or a teacher.
It also cannot help if the student only wants the answer. Any tool can be used that way if the student pastes the problem and accepts the first solution. The parent summary after an Aristotle session is one way to see whether the session was teaching or answer-giving.
For a longer look at how Aristotle is built to teach, see the AI tutor page. The Intro plan is free with 3 one-hour sessions, so a family can try it on real work. Current plans are on pricing.
Sources
- Reddy, Vajipey, Liu, and McNichols, What would a great tutor do?, 11 November 2026. 791 pivotal turns from 500 live sessions; pass rates 54% (Aristotle), 34% (Study Mode), 33% (Guided Learning), 35% (Claude Sonnet 5.5 on the Guided Learning prompt). The post states that scores measure next-turn pedagogical decisions, not longitudinal learning outcomes.
- Aristotle, AI tutor, accessed 18 November 2026.
- Aristotle, Why AI chatbots make bad tutors, accessed 18 November 2026.
- Aristotle, Pricing, accessed 18 November 2026.
Questions parents ask
Does AI tutoring work?
It can, when the tool keeps the student thinking and the student uses it that way. A tool that finishes the problem for them is doing a different job. Aristotle's published measurement is about the quality of the next teaching move in real sessions, not about grades or test scores.
How do I know if it is working for my child?
After a session, ask them to explain the step, try a slightly changed problem, and find a mistake in a worked solution, without the tool. If they can only repeat what was on screen, the session probably handed them the work.
Has Aristotle published any research on this?
Yes. What would a great tutor do? reports a 54% pedagogical pass rate for Aristotle on 791 pivotal turns from 500 real sessions, compared with 33–35% for the frontier tutoring modes they tested. The authors are clear that this is not a study of learning outcomes over time.
Is an AI tutor a replacement for a human tutor?
No. It is available when the student sits down, and it can work across more than one subject in the same plan. A person is still the right fit when the student needs a consistent relationship, or help the software does not cover. Plenty of families use both.