What the evidence says about AI tutors

What the research found

When the AI is built to teach

At Harvard, Gregory Kestin and colleagues compared an AI tutor, designed with the same teaching practices as the course, against an in-class active learning lesson. Students learned significantly more, in less time, with the AI tutor (the line quoted at the top of this page).

In Ghana, about 500 students used Rori, a math tutor on WhatsApp, for one hour a week during study hall. Their growth scores were substantially higher, an effect size of 0.36 (Henkel et al., 2024). In Nigeria, a six-week after-school program with Microsoft Copilot (GPT-4) raised English results:

The effect on English, the main outcome of interest, was of 0.23 standard deviations
(De Simone et al., 2025)

AI can also help human tutors. In Tutor CoPilot, students whose tutors had AI suggestions were 4 percentage points more likely to master topics, and the tutors were

less likely to give away the answer to the student
(Wang et al., 2024b)

ChatGPT-written hints produced learning gains like human-written hints in Pardos and Bhandari (2024).

When the AI just answers

unfettered access to GPT-4 can harm educational outcomes
(Bastani et al., 2025)

In that study the plain chatbot raised practice grades and cut exam grades by 17%, while the version with learning safeguards mostly avoided the harm. See Guide, don't hand over answers.

Measuring teaching, not just answering

Research groups are now testing whether models teach well, not only whether they are right. Google's LearnLM work found pedagogy-tuned models preferred by educators, and an evaluation taxonomy for AI tutors separates the two jobs:

highlighting which LLMs are good tutors and which ones are more suitable as question-answering systems
(Maurya et al., 2025)

subject expertise, indicated by solving ability, does not immediately translate to good teaching
(Macina et al., 2025)

Older research on computer tutors is the long view: reviews found intelligent tutoring systems raised test scores (Kulik and Fletcher, 2016; Ma et al., 2014), nearly as much as human tutors in VanLehn (2011).