Artificial Intelligence as a Homework Tutor

What problem were we trying to solve?

Homework is a cornerstone of learning in schools around the world. But for homework to truly help, students need timely feedback — someone to point out mistakes, explain why an answer is wrong, and ask follow-up questions that deepen understanding.

In practice, that rarely happens. Teachers are stretched thin, private tutors are expensive, and most homework ends up being a one-way exercise with no real interaction. For students learning English as a second language (ESL), this gap is especially costly.

What did we do?

We designed a way to use GPT-4 as an interactive homework tutor. Instead of simply answering questions or grading responses, the AI guides students through exercises with hints, asks follow-up questions, and gives personalised feedback — much like a patient human tutor would.

Crucially, setting up new exercises requires very little effort from teachers. This was important because creating content is one of the biggest barriers to adopting AI tools in real classrooms.

How did we test it?

We ran a Randomised Controlled Trial (RCT) — the gold standard for measuring cause and effect — across four high-school classes. This is the kind of rigorous experiment used in medical research, now applied to education.

• ATreatment group — students did their homework through AI tutoring sessions instead of traditional exercises.
• BControl group — students did the same homework the traditional way, with no AI involvement.

Both groups covered the same material and were assessed before and after the study, so we could measure the actual learning gains attributable to the AI tutor.

What did we find?

Students who used the AI tutor showed greater grammar improvement than their peers who did traditional homework. They also reported feeling more engaged, more motivated, and — in many cases — more comfortable making mistakes because they were talking to an AI rather than a person.

Why does this matter for teachers and schools?

• Low setup effort. Teachers don't need to rebuild their curriculum. Existing exercises can be adapted quickly.
• Scalable feedback. Every student gets personalised guidance, not just the ones who raise their hand or can afford a tutor.
• Real-world evidence. This wasn't a lab experiment — it happened in actual classrooms, with real students and teachers.
• A safe space to practise. Many students feel less embarrassed making mistakes with an AI, which encourages them to try more and learn faster.