AI in schools: what AERO's new advice means for teachers

Picture a maths class where every student flies through the practice set with a chatbot open. Then the laptops close for the test, and the marks drop below the class that never had AI at all.

That isn't a thought experiment. It's what a large high school trial found, and it sits behind new national advice on AI in schools. Here's what the Australian Education Research Organisation (AERO) now asks of teachers, in plain English, with scripts you can use in class this week.

> Key takeaway: AI should scaffold the thinking, not do it. Ask for a first try before any tool opens, choose tools that give hints before answers, and check learning with the AI switched off.

What does AERO say about AI in schools?

On 14 October 2026, AERO published AI in schools: Considerations for school leaders and teachers. AERO is Australia's national education research body, jointly funded by Commonwealth, state and territory governments.

The explainer is open about the gaps. It says there's limited research on AI and student learning in school settings, and most of it comes from secondary schools and universities.

Its five considerations for teachers and school leaders are:

1. Prioritise evidence-based teaching and learning. AI should help students acquire, retrieve, practise and apply knowledge, not replace the work.
2. Monitor for risks to learning. Watch for over-reliance and "illusions of knowledge".
3. Set clear expectations for AI use. Students should know when AI is allowed and when they work alone.
4. Rethink assessment and feedback. Ask whether take-home tasks still show what a student can do independently.
5. Apply a culturally responsive lens. AI tools are mostly trained on Western data and can misrepresent First Nations knowledges.

Does AI help or harm student learning?

It depends on how the tool is built and used. AERO's own summary is that AI "doesn't have a single, uniform effect on learning".

The clearest example is a field trial published in PNAS with nearly 1,000 high school maths students in Turkey. Classes got either no AI or one of two GPT-4 tutors.

One tutor worked like standard ChatGPT. The other was told to give teacher-designed hints instead of answers.

During practice, both groups did better than students with only their notes and textbook: 48% better with the standard chatbot and 127% better with the hint-giving tutor. But on the closed-book exam, the standard chatbot group scored 17% lower than students who never had AI. The hint-giving tutor group scored about the same as that group.

The researchers found many students simply asked the standard chatbot for the answer and copied it. They also found students didn't notice they had learned less.

What is cognitive offloading, and why does it matter?

Cognitive offloading means handing thinking over to a tool. AERO notes this can be fine, like using a calculator for a big calculation so working memory is free for the real task.

It becomes a problem when the tool does the very thinking the student needs to practise. AERO describes the result as a "performance paradox": better work on the task, less learning underneath.

AERO's chief executive, Dr Jenny Donovan, put it simply when the advice was released. She told the ABC that AI can "disguise the fact that children are using it as a crutch rather than actually a tool to learn".

Should primary school students use AI?

AERO urges more caution before secondary school. Its explainer says younger students are still building the background knowledge needed to spot inaccurate AI outputs, and it points to age restrictions and closer teacher oversight.

Dr Donovan went further in her ABC interview. She said AI should be avoided completely in the first three years of primary school, and used carefully after that alongside practices like explicit instruction.

For anyone building learning tools for younger children, us included, the message is clear. Guardrails and adult oversight are not optional extras.

How can teachers set clear rules for AI use in schools?

Clear lines help students too. At St Mary's Anglican Girls' School in Western Australia, Dean of Teaching and Learning Monica Gordon told Informa Connect that students "wanted to use AI but weren't sure what was acceptable".

The school built a compass that maps six AI functions, and a set of single-purpose AI agents. An ideas agent only generates ideas, and an editor agent gives direction rather than doing the edit.

It's now studying the approach in a two-year study with Murdoch University.

You don't need custom agents to start. You need a few phrases students hear every time.

Scripts for the classroom

• Before AI opens: "Show me your first try. Then you can ask for a hint."
• While they use it: "Ask it for a hint, not the answer. What did the hint make you notice?"
• After they use it: "Close the tab. Can you do the next one on your own?"
• When work looks too polished: "Talk me through how you got here, step by step."
• For a parent email: "At home, try asking: 'Which part did you work out yourself?'"

That last line is the same ask-first idea from our guide to helping kids with homework without doing it for them.

What should teachers check before using an AI tutor?

AERO says school leaders should critically evaluate new tools before students use them, including their limitations, terms and privacy practices. Here are four questions to take into that conversation:

1. Does it give hints before answers? The PNAS trial suggests this one design choice matters a lot.
2. Is it grounded in a real curriculum sequence? Or does it make up a learning path as it goes?
3. Can a teacher or parent see the student's thinking? AERO notes that reviewing AI interactions can reveal misconceptions.
4. What data does it collect, and is it ever sold? Get the answer in writing.

We hold the Archimedes AI tutor to the same questions. Each reply is one short turn with at most one question, and the answer comes only after a real try, when a smaller question still hasn't helped, or when the student asks to be told. Lessons follow a curriculum skill map rather than an invented path, and our privacy page says plainly that we don't sell personal information.

AERO also points out that AI can help students plan, monitor and reflect on their learning. Those planning habits start early, as in our Sunday night routine for kids.

FAQ

What is AERO's advice on AI in schools?

AERO's October 2026 explainer gives teachers and school leaders five considerations: put evidence-based teaching first, monitor for risks to learning, set clear expectations for AI use, rethink assessment, and apply a culturally responsive lens. Its core message is that AI should support the thinking students need to do, not replace it.

What is the performance paradox in AI and learning?

The performance paradox is when students do better on a task while using AI but learn less from it. In a high school maths trial, students using a standard chatbot did better during practice, then scored 17% lower on a closed-book exam than students who never used AI.

Should students use AI for homework?

Only with clear limits. AERO's chief executive told the ABC that parents should avoid giving children unfettered access to general chatbots at home. If AI is used, ask for a first attempt before it opens, use it for hints rather than answers, and check that your child can finish the next problem without it.

Can teachers rely on AI detection tools?

Not on their own. AERO asks schools to weigh the biases and limitations of AI detectors, including their potential to disadvantage students with English as an additional language or dialect. Talking a student through their work, or setting tasks that show thinking in class, can give clearer evidence of what they know.

Want evidence-first notes on AI in the classroom, without the hype? Get the next issue by email.

By Thomas Heal, founder of Archimedes. Thomas is building a patient AI tutor for parents, students and teachers from Perth, Western Australia.

This issue was drafted with an AI assistant, as we describe in building an edtech newsletter, and Thomas reviews and approves every word before it's published. Published 20 October 2026.