AI in schools: what the Australian evidence says.

The rule in Australia

Education Ministers approved the Australian Framework for Generative AI in Schools on 5 October 2023, implemented from Term 1 2024. It sets six principles and 25 guiding statements, and it replaced the earlier state-level bans — New South Wales had announced one in January 2023 AIE-10. The national setting is enabling rather than prohibitive. Individual schools still set their own rules underneath it, so the policy that governs your child is the school’s, not the framework’s.

How much students actually use it

In a 2025 survey of more than 3,000 Australian high-school students run by Elevate Education — a tutoring company reporting on its own survey, so weaker evidence than the government and peer-reviewed sources elsewhere on this page — three-quarters reported using AI at least a few times a week and almost a quarter reported daily use; ChatGPT was the most-used tool at 34% AIE-13. Teachers are using it too: 66% of Australian lower-secondary teachers reported using AI in the past year — the fourth-highest rate in the OECD, against an average of 36% AIE-14.

Take the student figure as indicative rather than precise, given who ran it. The direction is hardly in doubt, though, and the teacher figure is firmer: AI use is ordinary behaviour rather than an edge case to be prevented, in a system whose teachers are among the world’s heaviest AI users.

What it does to learning

The evidence points both ways, and the direction depends on how the tool is used rather than whether it is.

That trial is the clearest thing in the evidence: the pedagogically safeguarded tutor produced a 127% in-session improvement against 48% for the plain interface. The difference was not the model. It was whether the tool was built to make the student do the thinking. Both trials were run overseas — Turkey and the United States — so they describe the tools rather than Australian classrooms; what makes them worth reading here is that the tools are the same ones.

Usage data points the same way. Anthropic’s analysis of over half a million higher-education conversations found nearly 47% were direct request-and-receive exchanges rather than collaborative ones AIE-12. That figure is a usage category, not a measure of learning — but read beside the trial above, it describes the mode of use that trial found costly.

The largest test so far, and the thing it found

Both trials above are small and short. The biggest randomised test of the personal-tutor claim ran for two years across 18 middle schools in the United States, giving randomly assigned students an AI tutor deliberately configured to coach rather than hand over answers — the favourable case, not a straw man. It raised maths achievement by 1.3 national percentile ranks per term, roughly 0.06 to 0.08 standard deviations across a school year, and the researchers say plainly that those gains resemble what the same practice platform delivers with no AI at all AIE-16.

The reason is not that the tutor was bad. It is that it went largely unused. In that United States trial 96% of students opened it at least once, but the median student messaged it on only a third of the days they practised — and, the number worth sitting with, in only 17% of the sessions where they made a mistake. Five times out of six, at the exact moment a tutor earns its keep, the student did not turn to it. What they did send was mostly bare answers or clicks on suggested prompts.

A human tutor does not wait to be asked. That is the honest difference, and it is a more useful one than any argument about which model is cleverest: the research question has moved from whether the tool can teach to whether a fourteen-year-old, unsupervised, will let it. On the current evidence, mostly not.

The thing you can see at the kitchen table

One more study is worth knowing about, with its limits stated first: it follows 26,811 Chinese secondary students over 30 months, it is not Australian, and it compares students who chose to adopt AI against those who did not, which is weaker than a trial. What it found is still the most useful pattern in this evidence. Homework marks went up 18% and homework time went down 30% — while closed-book exam marks fell 20% within six months, and entrance-exam results fell 18 and 24% for those Chinese students, with the full effect only appearing after about two years AIE-17.

Read the two halves together. Work done at home got better; work done under supervision got worse. A parent watching homework improve has, on this evidence, very little information about whether their child is learning more — and possibly the opposite.

But the study does not say AI use causes this. It says outsourcing does. The losses in that Chinese sample concentrated among the roughly 80% of AI users whose homework got markedly faster and markedly better at the same time, and the authors report that the Chinese students who kept taking about as long as everyone else showed only small losses AIE-17. The tell is not that a child uses AI. It is that the work stopped costing them anything.

If the school says AI wrote it

Treat a detector result as a reason to ask questions, never as proof. A peer-reviewed study found GPT detectors misclassified over half of non-native-English essays as AI-generated — an average false positive rate of 61.22% — while being near-perfect on native-English US 8th-grade essays AIE-05. The bias falls hardest on students who learned English as an additional language and on anyone whose writing is plainer or more formulaic than the average.

OpenAI withdrew its own AI-text classifier six months after launch, citing low accuracy AIE-04. Turnitin’s published guidance says its detection is tuned to hold false positives below 1%, explicitly trading away some sensitivity to do so AIE-06. Even taken at face value, a sub-1% false-positive rate applied across every assignment in a school still produces wrongly flagged students.

An Australian assessment authority has now said the same thing in writing. In an official notice effective Term 4 2026, NSW’s Education Standards Authority tells schools they should not rely on AI detection tools as the main safeguard against malpractice and that such software, if used at all, is “only one input for teacher professional review, not as sole proof of AI use or malpractice”. It goes past a caution: schools are directed to check that their malpractice policy does not rely on detection tools at all AIE-18.

Read that with its boundary attached. NESA governs the HSC in New South Wales and has no authority in Victoria, and this page makes no claim about what VCAA has or has not said on detectors. What the notice tells a Victorian family is not a rule that binds their school — it is that when an Australian assessment authority examined the same detectors this page describes, it reached the same conclusion, and put it in a directive rather than a suggestion.

The practical answer is not a better detector. It is a record of the work: drafts, notes, a version history, and a student who can talk through why the piece says what it says. That is also, not coincidentally, what writing taught properly produces.

Where these claims come from

Every figure on this page is drawn from the AI in education domain of the Evidence Register, where each claim appears with its verbatim quote and full citation. Newly published research is read on the frontier page as it lands.

How this changes the teaching

Argo Academics is a one-to-one English tutoring practice in Melbourne, where Callum Hay teaches every lesson personally. In a weekly lesson the work is visible as it is made — the thinking happens in front of someone, which is the part no tool can stand in for and the part a finished essay never shows.

If what you are weighing is whether your child’s writing has become something they assemble rather than something they think through, the two-minute diagnostic asks about exactly that.