Mention "AI Tutor" in a room full of educators and you'll see the full spectrum: the quizzical look, the eye-roll, the deadpan indifference, even the excited entrepreneurial dollar-sign.
Never mind that it should theoretically work very well. Does it? In decades of experience in English language learning, machine-mediated skills acquisition works extremely well. It puts the language lab in individual learners' hands. Before today's AI, that was a godsend. Now, with personalised and accelerated learning solutions, language learners do even better.
One caveat: AI isn't the all-encompassing tool. It cannot replace the human connection that is the reason for language learning in the first place.
The hype is easy enough to find: tutors that are available at any hour, never lose patience and adapt to each student. The research is read far less often, and it is more interesting than either the enthusiasts or the sceptics tend to admit, because the same kind of tool can help a great deal or quietly do harm, depending on how it is set up and whether a teacher stays involved.
What the better trials show
Two recent randomised trials give the optimistic case its firmest footing. In Edo State, Nigeria, a World Bank team ran a six-week after-school programme in which public secondary school students used ChatGPT twice a week, with teachers trained beforehand to introduce each topic, keep students on task and help them make sense of the AI's feedback (including spotting its occasional hallucinations). Students in the programme gained about 0.3 standard deviations, a result the researchers say surpasses 80% of the rigorously evaluated education interventions they compared it with, and English, the programme's main focus, improved along with the rest (De Simone and colleagues, World Bank, 2025).
At Harvard, Gregory Kestin and colleagues compared an AI tutor used at home with an active-learning lesson in class, in an introductory physics course of 194 students who each experienced both. Learning gains with the AI tutor were more than double those in class, and the typical student finished in 49 minutes against a 60-minute lesson (Scientific Reports, 2025). The caveat matters as much as the result, of course: this was one course at one university, and the tutor was built by the course's own instructors with expert-written prompts and worked solutions, which is a long way from an app bought off the shelf.
What happens when the tutor simply gives answers
Hamsa Bastani and colleagues gave nearly 1,000 high school maths students in Turkey either a plain GPT-4 chat interface, a version designed to give hints rather than answers, or no AI at all. Students with the plain version improved their practice grades by 48% but then scored 17% lower in the exam, taken without AI, than students who had never had access. The hint-giving version improved practice grades by 127% and largely avoided the exam penalty (PNAS, 2025).
A smaller and more recent study points the same way. Researchers from Carnegie Mellon, Oxford, MIT and UCLA gave 1,222 adults roughly ten to fifteen minutes of help from an AI assistant and then removed it without warning. Afterwards they solved fewer problems than people who had worked without AI throughout (57% against 73% in the first fractions experiment), and in two of the three experiments they skipped more problems rather than persevere (Liu and colleagues, 2026 preprint). It is a preprint, and its participants were adults online rather than pupils in a classroom, so I would not lean on it alone, but it describes much the same thing teachers mean when they complain that students reach for the answer before they have tried the problem.
Where the teacher fits
The most useful evidence on that question comes from Stanford's Tutor CoPilot trial, in which an AI suggested teaching moves to human tutors as they worked, rather than tutoring the students itself. Across 900 tutors and 1,800 students from under-served communities, students whose tutors had the tool were 4 percentage points more likely to master a topic, and 9 points more likely where the tutor had been rated less effective beforehand, at a cost of about US$20 per tutor a year (Wang and colleagues, 2024).
Read together, the pattern fits my own experience in language teaching more or less exactly. The gains in Nigeria came with a teacher in the room, the Harvard tutor was designed by teachers, Tutor CoPilot made human tutors better rather than redundant, and the harm in Turkey came from a tool that did the work for the student.
What this means for Malaysian schools
If your school is weighing an AI tutoring platform, three questions seem to me worth asking before the price:
- Does it give answers or hints? The Turkish trial suggests this one design choice can decide whether a tool helps or harms, so ask the vendor to show you, with a real exercise, what happens when a student asks it for the answer, and whether your teachers can change that behaviour.
- What is the teacher's role? None of the positive results above came from leaving students alone with a chatbot, so budget for the teacher time to plan, supervise and follow up, which makes the platform an addition to teaching rather than a cheaper substitute for it. Any cost case built on replacing teachers or tutors deserves a good deal of scepticism.
- How will you know whether it worked? Practice scores with the tool switched on are the figure most likely to flatter it. Test what students can do without it, as both the Turkish trial and the adult study did, and do so some weeks later rather than the following day.
The questions still open
Most of this research is short-term, so we do not yet know whether the gains hold a year later. We know little about who actually gets access when schools adopt these tools at their own pace, and whether the students who most need help are the ones who receive it. And the studies on persistence raise a harder question than effectiveness: if students come to see learning as something an AI does for them rather than something they do, what is lost that a test score will not show?
The takeaway
AI tutors aren't magic, and they aren't toys. They are specialised tools that work best when they do what they do well (immediate feedback on practice, endless patience with repetition, availability at any hour) and when teachers do what they do well: believing in students, explaining the why, and knowing when to push and when to step back. The evidence so far favours schools that use both, with the teacher deciding what the tool is for.
Revised 24 October 2026: two research citations in the first version could not be verified and have been replaced with the published and preprint studies linked above.