Can AI Personalise Learning Better?

The problem, in numbers

Personalised learning is one of the oldest promises in education technology: give each student practice pitched at what they already know, and move them on once they have mastered it. In our reading, the problem in Australia is not a shortage of products but a shortage of evidence about them. The Australian Education Research Organisation (AERO) said in its 2024 research agenda that more rigorous and relevant evidence is needed for Australian students to realise the benefits of education technology.^{Source 1} Its October 2026 explainers for teachers and policymakers say there is still limited research on the impact of AI use in schools.^{Source 1}

That leaves the international research on intelligent tutoring systems: software that estimates what a student knows and adapts the next problem or hint. In our reading, this body of work, built up over decades and mostly in mathematics, gives a more useful answer than the generative AI studies of the past two years.

This review asks whether adaptive software teaches better. Our review of tutoring costs deals with price per student, and our grading review with marking.

What improves, and why

The mechanism is simple. A tutoring system breaks a subject into skills, infers from each answer which skills a student has mastered, and chooses the next practice and hints accordingly. In our reading, it works best where three things hold: the skill can be broken into steps, an answer can be checked as right or wrong, and progress can be measured on a test aligned to the practice. Algebra fits. An essay argument fits less well.

The meta-analyses support the general case. Kulik and Fletcher reviewed 50 controlled evaluations and found a median gain of 0.66 standard deviations over conventional teaching, enough to move a median student from the 50th to the 75th percentile.^{Source 2} They also found the size of the gain depended heavily on whether it was measured on locally developed or standardised tests, and evaluations with unconventional control groups or flawed implementations showed small effects.^{Source 2}

Ma and colleagues, in a second meta-analysis, pooled 107 effect sizes from 14,321 participants. Intelligent tutoring beat teacher-led, large-group instruction (g = 0.42) and textbooks or workbooks (g = 0.35), but showed no significant difference from individualised human tutoring or small-group instruction.^{Source 3} In our reading, that is the clearest summary of what personalisation buys: something close to individual attention that a full class cannot otherwise get, not something better than a good tutor.

What the evidence actually says

The effect shrinks as studies move closer to ordinary schools. Steenbergen-Hu and Cooper reviewed 34 samples on K-12 mathematics and found average effects of g = 0.01 to 0.09, which they describe as no negative and perhaps a small positive effect.^{Source 4} Effects were larger for interventions shorter than a school year, and larger for general students than for low achievers, which the authors say raises the question of whether such software could widen achievement gaps.^{Source 4}

A 2025 systematic review of AI-driven tutoring systems in K-12 schools found 28 studies with 4,597 students. Seven of eight studies comparing them with teacher-led instruction reported significant positive effects, but only one of four comparing them with non-intelligent computer tutoring found an advantage.^{Source 5} Most studies were quasi-experimental and many interventions were short, which the authors warn may inflate effects through novelty.^{Source 5}

Large randomised trials give the most sober picture. In a RAND trial across seven US states, schools randomly assigned to Cognitive Tutor Algebra I saw no effect in the first year. In the second year the effect was positive, about eight percentile points for the median student, and statistically significant only in high schools.^{Source 6} In a trial of 2,850 Year 7 students in 43 Maine schools, randomised by school, the ASSISTments online homework platform raised end-of-year maths scores, with a larger effect for lower-performing students.^{Source 6} An evidence registry records that effect as g = 0.18, or 0.29 for low achievers, and notes that students averaged 14 hours of use against about 24 planned.^{Source 6} In our reading, these trials show adaptive practice can work at scale, with modest gains that take time to appear.

Chatbot tutors built on large language models are a newer and separate question; our tutoring costs review covers the trial evidence on them.

Where it falls short

• Motivation is not measured. The 2025 review did not consider interest, attitude or motivation,^{Source 5} and in our reading the trials measure test scores, not whether students want to keep going. Lower-than-planned use in the Maine trial is a reminder that the software only helps when students open it.^{Source 6}

• Gains depend on the test. In one meta-analysis, the size of the gain depended heavily on whether outcomes were measured on locally developed or standardised tests.^{Source 2}

• Equity cuts both ways. One trial found larger gains for low achievers;^{Source 6} one meta-analysis found smaller ones.^{Source 4}

Where governance decides the outcome

In our reading, the main governance choice is design: whether a tool hands over answers or makes students work for them. NSW has made that choice for its public schools. The department says NSWEduChat's student mode, open to students in Years 5 to 12, does not just give direct answers but provides guidance and asks open-ended questions.^{Source 7} It reports teacher efficiencies and high engagement from 50 trial schools, but publishes no student learning outcomes.^{Source 7} AERO asks school leaders and teachers to monitor for risks to learning and to set clear expectations for AI use.^{Source 1} Given the conflicting equity evidence, schools should check whether a product's gains hold for their lowest achievers, and treat any efficacy claim without a comparison group as untested.

How to think about adoption

Use adaptive practice where the skill is stepwise and mastery can be tested; in our reading, mathematics is where most of the evidence sits. Plan to run it for at least two years before judging, since a large randomised algebra trial found nothing in year one. Measure results on standardised tests as well as the software's own mastery scores, and track how much students actually use it. Prefer tools that withhold answers and prompt reasoning. Keep the teacher responsible for the part no model supplies: the reason to keep practising.