A new study has put AI tutors and expert human tutors side by side, with randomly assigned students and a control group. For one hour of GRE preparation, students taught by AI learned about as much as students taught by experts.
That is a promising result. It is also a specific one. This article looks at what the study measured, what it leaves open, and what that means for schools and tuition centres.
The paper is StudentBench, a preprint published on arXiv in September 2026 by researchers at Handshake AI (arXiv:2609.28470). As a preprint, it has not yet been through journal peer review.
The paper also reports a pattern in the Quantitative (maths) sections. Faster AI replies were associated with more messages from students, more messages with more correct practice, and more correct practice with larger gains. These are associations, not a tested cause, but they are worth noticing.
The authors are clear about the limits.
The fourth point is the one this article is about.
The study answers one question well: within a structured session, can AI support short-term exam learning as well as an expert tutor? For one hour of GRE practice, the evidence says it can.
Whether that opportunity becomes part of a student's ordinary week is a separate question, and one the study did not test. As we wrote in Learning to Study Still Takes Support, many students who stop are not missing a better explanation. They are missing a reason to start, a manageable next step, and someone who notices when they have gone quiet.
Short-term learning gains from AI look promising. Turning them into everyday learning still needs to be worked out.
AI can make practice and timely feedback available beyond what one teacher can provide alone. Teachers can use the resulting information alongside their knowledge of each student to decide when to intervene, what to explain differently, and how to support the next attempt. During a lesson that might mean spotting a misunderstanding the practice revealed. Between lessons it might mean checking in with a student who has gone quiet.
Software can help students restart too, with a sensible next question or a reminder of where they left off. But the decision about how to support a particular student, at a particular moment, is still best made by someone who knows them.
The findings are consistent with our emphasis on active practice: students need opportunities to attempt problems, receive useful feedback, and try again. On Axiom, a wrong answer is followed by a one-line reason and a similar question to try, rather than a page to read.
The association between faster replies and greater participation is also familiar to us. We have spent a lot of effort making marking and the next question arrive quickly, because a student waiting for feedback is a student who may close the tab.
The study also leaves open a question that Axiom for Schools is designed to address: how can effective practice become part of a student's regular learning routine? Teachers see per-topic progress for each student and who has been inactive for a week or more, so they can decide when and how to step in. Inactivity is a prompt to check in, not a diagnosis: a student may be studying elsewhere, busy, or unsure how to continue.
In a one-hour GRE study, AI tutoring produced short-term learning gains statistically equivalent to expert human tutors. For one model, inference cost per percentage point gained was far below the corresponding human tutoring cost.
The next question is whether students return, practise regularly, and retain what they learn. Schools and tuition centres can help test that in everyday learning, with teachers who know when a student needs a different explanation, a smaller next step, or a conversation.
If you run a school or tuition centre, see what your teachers would see.