Can a free AI tutor close Singapore's tuition gap?

Nothing in Singapore schooling is measured as carefully as the part the Ministry of Education does not run. Households here spent $1.8 billion on private tuition in 2023, according to the Department of Statistics, and the shape of that number matters more than its size. Households in the top fifth by income spend more than four times what households in the bottom fifth spend on tuition.

Two MPs put that to Parliament in February 2025. Dr Wan Rizal asked about rising tuition expenditure and what it means for educational equity. Mr Yip Hon Weng asked whether social inequality might widen. MOE’s reply was calm and largely unmovable: spending growth had broadly tracked household income and inflation, excessive tuition can take the joy out of learning, and the ministry cannot regulate what a family chooses to buy.

Which leaves the question sitting there. If what money buys is an extra hour of explanation, and a machine can produce that hour for close to nothing, does the gap just close by itself?

The serious version of the argument

It is not a marketing line. Two NIE academics, Emeritus Professor Looi Chee Kit and Dr Wong Lung Hsiang, made the case in The Straits Times in April 2025. Their argument was not that a chatbot matches a good tutor. It was that AI which can plan, track and adapt could extend the reach of human mentors, and could offer, in their words, “a practical and inclusive alternative to costly tuition as the sole route to academic success”. They also made a point that gets lost in the efficacy debates: this kind of tool gives non-judgmental feedback, which matters to a child who has learnt to dread being marked.

The comparison that counts is not chatbot versus star tutor. It is chatbot versus nothing at all, because nothing at all is what a lot of children get after 6pm.

What the trials actually found

The strongest evidence comes from a Stanford randomised trial of a system called Tutor CoPilot, run with 900 tutors and 1,800 K-12 students from historically under-served communities. Students whose tutors had the tool were four percentage points more likely to master topics. The number worth pausing on is the next one: for students of lower-rated tutors, mastery improved by nine points. It cost about US$20 per tutor per year.

Read the design carefully, though, because it is the whole story. The AI was not tutoring the child. It was coaching the tutor, live, mid-session, nudging them towards guiding questions instead of straight answers. The gain landed where the existing support was weakest. That is a real equity result, and it is not the same result as handing a child an app. It is closer to what we described in the AI labs want your child’s teacher: the useful intervention keeps arriving through an adult.

The other serious trial ran in Nigeria. A World Bank team (De Simone, Tiberti, Barron Rodriguez, Manolio, Mosuro and Dikoru) tested six weeks of extra English sessions with Microsoft Copilot for first-year secondary students, and reported gains of 0.31 standard deviations overall and 0.23 on English, the main outcome. They ranked it among the most cost-effective ways to improve learning that anyone has measured, worth something like one and a half to two years of ordinary schooling.

Then comes the sentence the excited coverage tends to skip. The programme helped students across the ability range, but the largest effects went to female students and to those with higher initial academic performance. In a setting where the alternative was genuinely nothing, the free tool still paid out most to the children who were already doing better. Availability is not the same thing as equality.

Singapore already gives every child one

This is the part that tends to get missed in the argument. There is an adaptive AI practice system sitting inside the Student Learning Space, free, in the account your child already has.

MOE said the quiet part out loud when it announced the work in July 2021: the aim was to improve learning outcomes “especially among the lower-progress learners”. The Adaptive Learning System recommends resources and practice questions, gives immediate feedback, and paces itself to what a child is ready for. GovTech’s January 2025 write-up put its coverage at upper primary Mathematics and upper secondary Geography, so it is narrow, and we could not find published outcome data on how much it has moved results. Worth knowing both things.

Still: free, national, already paid for, and very easy to never open.

The ministry is buying adults, not software

Look at what MOE actually spent money on at this year’s Committee of Supply, on 3 March 2026. It folded UPLIFT and GEAR-UP into a single package and extended it from 100 schools to 157, with additional manpower and funding, benefiting around 20,000 students a year. Alongside that sit the Learning Support Programme, Learning Support for Mathematics, and the Collaborative Tuition Programme’s subsidised tuition.

Manpower. People, in other words, which is the expensive option, chosen while free software was available.

Our view, clearly marked as ours: that ordering is right, and the trials support it. Free general-purpose AI has been sitting on every phone in Singapore for years. If mere availability closed gaps, this one would already be closing. The scarce resource was never the explanation. It is the adult who sits the child down and stays until the question is finished. (We build AI mentoring, so we have an obvious commercial interest in believing AI tutoring works. Take the caveat with the claim.)

What to do with that this week

Ask your child’s form teacher one question: is the Adaptive Learning System switched on for my child’s subject, and has anything been assigned through it? The answer tells you whether the free option is actually running for your child, or just exists.

If you are already using a general chatbot for homework at home, watch one thing only: does it ask or does it tell. A tool that answers leaves less behind than a tool that questions, which is the finding underneath when AI helps kids learn.

And judge it over a fortnight, not an evening. The signal is not whether the worksheet got finished. It is whether your child tried the next question before reaching for the machine.