Our view on AI
AI is a great tool. It isn't a tutor.
I run X Tutoring and still tutor students myself, so I think about this a lot. AI tutors are getting good, and I'm not going to pretend that's nothing. But the research so far says the same thing I see in sessions: AI helps most when a person is still in charge of the learning. Here's where we draw the line, and why.
Jin, founder
Last updated November 2026 · More research we follow · Suggest a study
01 · What is AI good at?
Practice, patience and memory.
An AI will explain a concept as many times as a student needs, at 11pm the night before a test, and it can remember every question a student has ever gotten wrong. No human tutor keeps that record in their head.
When it's built carefully, it can teach well. In a randomized trial in a Harvard intro physics course, students using an AI tutor designed around good teaching practice learned more than twice as much (median test gain) as in an active-learning class, in a bit less time. In Nigeria, a six-week after-school program where students worked with an AI tutor, with teachers in the room, raised test scores by about 0.3 standard deviations.
Where we stand: AI is a real tool for practice between sessions. Note the fine print, though: the Harvard tutor took months to build, was given worked solutions to every problem, and was tested on two short topics. That isn't the same thing as a student opening a chatbot.
Kestin et al. (2025), Scientific Reports 15:17458 · paper · De Simone et al. (2025), World Bank Policy Research Working Paper 11125 · summary
02 · Where does AI go wrong?
When it does the thinking for the student.
In a study of nearly 1,000 high school math students, those given plain ChatGPT during practice did 48% better on the practice problems. When the AI was taken away for the exam, they scored 17% lower than students who never had it. A version of the AI that gave hints instead of answers mostly avoided that drop.
A separate pair of experiments found the same pattern: students who used AI to generate solutions covered more topics but understood each one less, while students who used it to ask for explanations understood more.
Where we stand: any student can get a homework answer in seconds now, so a tutor whose value was getting homework done is in trouble. What's left is getting the student to do the thinking and building the confidence to do it alone, which was always the point.
Bastani et al. (2025), "Generative AI without guardrails can harm learning," PNAS 122(26) · paper · Lehmann, Cornelius & Sting (2025), "AI Meets the Classroom," working paper · paper
03 · What does a tutor do that AI can't?
Notice, judge, and care.
An AI can't notice when a student says "yeah, I get it" and doesn't. It can't sense that a kid who was fine last week is quiet today. A good tutor catches that in the first few minutes and changes the plan. They know which student needs a push, and which one needs to hear that everyone can be good at math before they'll try the hard question. And kids work harder for someone they don't want to let down.
The research on why tutoring works points the same way. Researchers name the tutor–student relationship as one of the reasons: a consistent tutor builds a better understanding of the student's needs, and positive relationships are linked to better academic, social and motivational outcomes.
Where we stand: your child keeps the same tutor, session after session. That's on purpose.
Robinson, Kraft, Loeb & Schueler (2021), Design Principles for Accelerating Student Learning with High-Impact Tutoring, EdResearch for Action · brief · Nickow, Oreopoulos & Quan (2024), AERJ · our research page
04 · What happens when AI replaces the teacher?
The biggest test so far hasn't gone to plan.
Alpha School, a U.S. private school chain, has students spend about two hours a day on its AI learning platform and promises they'll learn twice as fast. In October 2026 the New York Times reported on the school's own internal data: fewer than half of its K–8 students hit that "2x" goal in both math and reading last year. Families who left described a model that asks an unusual amount of self-motivation from a child.
Where we stand: we don't think the goal of education is learning as fast as possible. And motivation isn't a side issue you can hand to software. For most kids, it's the whole game.
The New York Times, "Students at an A.I. School Were Supposed to Learn Faster. Many Aren't." (October 23, 2026)
05 · So what does work?
AI that makes the human tutor better.
In a randomized trial with about 900 tutors and 1,800 students, tutors were given an AI assistant that suggested how to respond in the moment. Their students were 4 percentage points more likely to master a topic, and 9 points more likely with tutors who had been rated lower. The tutors asked more guiding questions and gave away fewer answers. It cost about US$20 per tutor per year.
Where we stand: this is the model we believe in. AI behind the tutor, helping them teach, rather than in front of the student, replacing them.
Wang, Ribeiro, Robinson, Loeb & Demszky (2025), "Tutor CoPilot: A Human-AI Approach for Scaling Real-Time Expertise," working paper · paper