On this page
Any student with a phone now has a chatbot that can solve their homework in seconds. The useful question is no longer whether AI can solve the problem. It is whether the student learns anything when it does.
Two studies published in 2025 give a clear answer: it depends almost entirely on how the AI is designed and used.
When AI made students worse
In a field experiment published in PNAS, researchers gave nearly a thousand high school maths students access to one of two AI tools during practice sessions. One worked like a standard chatbot ("GPT Base"). The other ("GPT Tutor") was set up with safeguards designed to protect learning, such as giving hints instead of full answers.
During practice, both groups did much better: grades rose by 48% with the standard chatbot and 127% with the tutor. The real test came when the AI was taken away. Students who had used the standard chatbot then scored 17% lower than students who had never had AI at all.
They had used it as a crutch. The tutor's safeguards largely removed that harm.
When AI helped students learn more
A randomised trial at Harvard, published in Scientific Reports, compared an AI tutor with an in-class active learning lesson in an introductory physics course. Both were built on the same research-based teaching practices. Students using the AI tutor learned significantly more, in less time, and said they felt more engaged and motivated.
The AI tutor walked students through problems one step at a time and prompted them to think instead of handing over complete solutions.
How to use AI without it becoming a crutch
Whatever tool you use, these habits keep the thinking with you:
Try first. Spend at least ten minutes on a problem before asking for help.
Ask for a hint, not the answer. "What should I look at first?" teaches more than "solve this".
Redo it alone. After an explanation, close the chat and solve the problem again from a blank page.
Let it quiz you. Ask the AI to give you questions on a topic and check your answers.
Explain it back. Describe the solution in your own words and ask the AI to point out gaps.
How Fermi's AI tutor teaches
Fermi is an AI-powered STEM learning platform built on this idea. Its one-to-one AI tutor holds a line of reasoning across a whole session, notices where a student is stuck and decides whether to explain, ask a question or show. It writes on a board, draws the figure, runs a simulation and checks understanding as it goes.
Problem Library: worked problems with a step-by-step solver that explains the reasoning rather than printing an answer.
Concept Library: a lesson for every concept in physics, chemistry, biology, mathematics and computing.
Simulation Lab: interactive simulations for ideas that are easier to see than to read.
Courses: university-level courses taught by working scientists, with the AI tutor on every lecture.
The standard to hold AI tutors to
The research points the same way as decades of work on teaching: students learn by doing the thinking themselves. A good AI tutor makes that easier, and a bad one quietly does the thinking for you. That is the standard we think every AI tutor, including ours, should be held to.
