The Primary Source

What the research and the regulators actually say about technology in primary schools.

Issue 01Term 4 · 2026about 20 min read

Written for Australian primary school leaders.

Published quarterly by Code Camp, part of Cluey Learning.

To reach the editors email: editor@codecamp.com.au

From the editors

Those of us who have worked around education and technology for many years cannot remember a technology arriving in schools with this much noise and this little useful guidance.

Every week brings another claim. AI will transform learning. AI will hollow out thinking. Children must start now. Children should be nowhere near it. If you run a primary school, none of that helps you decide anything on Monday morning.

So we have started The Primary Source, which reviews the original research, government reports and regulatory documents, and asks what we actually know. And what we do not. AI is the loudest question this term. It will not always be, and this brief follows what is changing in technology for primary schools.

We come at this from inside the sector. Code Camp runs after-school and holiday camp coding and technology programs in Australian primary schools, which means the questions in this brief are ones we often get asked, rather than ones we invented. It also means we may have a commercial interest in some of the answers, and we will say so when we do.

One thing in this first issue surprised us. Most of the research quoted in the current public debate, and every cognitive-offloading study we examined for this issue, was conducted on older adolescents and adults rather than primary-aged children. That does not mean there is no risk. It means we should stop talking as though we already know what the risk is.

The more immediate problem for your school is probably not AI at all. It is privacy. Australia's Children's Online Privacy Code is coming, and this will likely force a lot of schools to look hard at something that has nothing to do with AI: how many separate companies now hold information about their students, and who agreed to that.

That is where we have started.

The brief

What changed this quarter

Children's Online Privacy Code - statutory deadline 10 December 2026

government/regulator

The OAIC released the exposure draft on 31 March 2026; consultation closed 5 June 2026. The Code must be registered by 10 December 2026, the statutory deadline set under the Privacy and Other Legislation Amendment Act 2024. The commencement date and any transition period are not yet settled - the OAIC invited submissions proposing them. You will see this reported as "in force from December". That is not yet accurate.

→ OAIC: Children's Online Privacy Code · see In Depth, below

eSafety: 64% of 10–12 year olds have used an AI assistant

government/regulator
survey

Talking to machines, published August 2026, surveyed 1,950 Australian children aged 10–17 between 2 February and 4 March 2026. 78% had used an AI assistant, most commonly ChatGPT - 87% of teenagers, and 64% of children aged 10–12. Of those who had used an assistant or companion, 54% had done so for personal or social reasons, and 20% were using them daily or more often. 8% had used an AI companion.

→ eSafety Commissioner: Talking to machines

Government response to Study Buddy or Influencer tabled April 2026

government/regulator

The House Standing Committee on Employment, Education and Training made 25 recommendations on generative AI in Australian education, including integrating GenAI into the national curriculum and establishing product standards for education AI tools. The Government's response was tabled in April 2026 (response document).

→ Department of Education

Social media minimum age: nine months on

government/regulator

Since 10 December 2025, age-restricted platforms must take reasonable steps to prevent Australians under 16 holding accounts. Services primarily designed for education or health are exempt, as are messaging-only apps and online games; YouTube content remains viewable without an account. Schools are not responsible for enforcement, but eSafety's guidance for educators is worth reading before Term 1. Whether under-16s migrate to less-regulated platforms is an open question rather than a measured outcome, and one worth watching.

→ eSafety: what the restrictions mean for educators

National product expectations for AI edtech move into implementation

government/regulator

The Australian Framework for Generative AI in Schools has applied since Term 1 2024, built on six principles: teaching and learning; human and social wellbeing; transparency; fairness; accountability; and privacy, security and safety. Education Ministers endorsed the 2024 Framework Review in June 2025. National product expectations for AI edtech have since been developed across three workstreams: privacy and information security, human rights and wellbeing, and education impact. Individual components are now at different stages of rollout, piloting and further development. If you are signing multi-year AI contracts now, build in a review point.

→ Department of Education

In depth

The privacy question coming for your edtech stack

Sometime before 10 December 2026, Australia will register a Children's Online Privacy Code. It is the most consequential thing happening to primary school technology this year, and it is almost entirely absent from the conversation schools are having about AI.

Here is what this means for primary school principals and what they should do about it this term.

What it is

The Code is a mandatory instrument made under the Privacy Act, created by the Privacy and Other Legislation Amendment Act 2024. That Act required the Information Commissioner to develop an Australian Privacy Principles code about children's online privacy, and the statutory deadline for registering it is 10 December 2026. The OAIC published an exposure draft on 31 March 2026 and closed consultation on 5 June 2026.

The 10 December 2026 date is the deadline for registration, not a confirmed commencement date. The OAIC explicitly invited submissions proposing an appropriate commencement date and transition period, and at the time of writing neither has been settled. You will encounter confident claims that the Code "comes into effect on 10 December". Treat them carefully. What is settled is that the exposure draft has been published and the final Code has not yet been registered, and that it may apply to many of the online services your school uses.

Who counts as a covered service. Narrower than it first sounds. The draft applies to entities already covered by the Privacy Act that provide a social media service, a relevant electronic service or a designated internet service, where that service is either likely to be accessed by children or primarily concerns children's activities. Online health services and carriage service providers such as internet providers are excluded. Whether a particular supplier of yours is covered needs its own assessment. The next section sets out what that does and does not include for your school.

Based on the exposure draft, the Code would require those services to, among other things: design the service in children's best interests before collecting, using or disclosing their personal information; obtain consent for targeted advertising to children; give age-appropriate privacy notices in clear, accessible language; enable children to request deletion of their personal information; notify children when a parent has consented on their behalf; notify children about geolocation tracking by other users; and restrict direct marketing.

Who it actually binds - and why that isn't the reassuring news it sounds like

The Code applies to online services covered by the Privacy Act where children are likely to access the service, or where the service is primarily concerned with children's activities. That second limb goes further than the Privacy Act itself, which refers only to services likely to be accessed by children. The OAIC's own worked examples of it include online school management systems that monitor student performance.

Whether the Privacy Act binds your school depends on your sector. Government schools are not APP entities - they are not covered by the Privacy Act, and are instead subject to state and territory privacy legislation. Non-government schools are generally covered where annual turnover exceeds $3 million. A smaller non-government school may also be covered where it provides a health service and holds health information, even if providing that service is not its primary activity. Whether a particular school qualifies is fact-specific and worth advice. Note too that Privacy Act coverage generally and the narrower scope of this Code are two different questions.

So for many principals, the Code does not regulate you directly. It may regulate some of the companies you buy from, depending on each supplier's own position under the Privacy Act.

That is not, however, where your exposure sits. Your exposure sits in the trust relationship. The Code is going to raise the floor on what children's online services must do, generate press coverage, and give parents a vocabulary - best interests, age-appropriate, deletion - they did not previously have. They will not direct those questions at a Californian vendor. They will direct them at you, at the school gate, in Term 1.

The consent-stacking problem

The OAIC's public consultation report is worth reading in full, and not for the legal analysis. Its consultation, run in May and June 2025 ahead of the exposure draft, received 337 submissions - including 161 from primary school students, 74 from secondary students and 102 from parents and carers - and it quotes them directly.

One parent's account of a primary school onboarding will be recognisable to almost every reader of this brief:

"At my youngest child's school, she was connected to 16 edtech apps via one platform consent form. The privacy risks in many of the added layers in the schools tech stack were very high. The school had no answer for me when I asked why they didn't explain the different apps on top of the school portal - this eroded the trust I had in the school to look out for the best interests of my child."

That quotation is reproduced from the OAIC report with punctuation lightly tidied. Others in the same report describe feeling that their children's information is "floating around in a lot of small company start-up like databases", and raise concerns about early-childhood and school platforms that "store detailed records of our kids".

This is the real issue for primary schools, and it long predates AI. Over roughly a decade, primary schools have accumulated technology stacks - portal, LMS, reading app, maths app, wellbeing tracker, communication platform, photo-sharing tool, assessment tool - usually procured individually, often by different staff, frequently consented to once in a bundle at enrolment, and rarely audited since. Each one holds data about children who cannot meaningfully consent for themselves. AI tools are simply the newest layer on a structure that was already under-governed.

What we don't know

Three things we do not know. The final Code text is not registered, so the obligations above are drawn from the exposure draft and may shift. The commencement and transition timing is unresolved, which makes precise compliance planning impossible this term. And how the Code will interact with state and territory privacy regimes covering government schools is not yet clear - which means government-sector principals should expect their own jurisdiction's guidance to be the operative document, whenever it arrives.

What this means for your school

Four things you could start this term. None of them require the final Code text, and all of them are useful regardless of what it says.

01
Build the list. Produce a single register of every digital service that holds student data, including the ones a single teacher signed up for. Owner, purpose, data collected, where it is stored, and the date consent was last obtained. In our experience schools rarely hold this list in one place, and it is usually longer than expected.
02
Separate the stack from the consent. If families consented through a single bundled form, do not assume that gave sufficiently specific and informed consent for every service in the stack. Reviewing the position app by app is a significant job; deciding this year who owns it is not.
03
Put a review point in every new contract. National product expectations for AI edtech have been developed, but important components remain in pilot testing and further development. Do not sign a three-year contract for an AI tool without a clause that lets you revisit it.
04
Tell parents first. Send a short, plain letter this term explaining what your school uses, why, and what you are doing to review it - before the Code generates coverage and the questions arrive unprompted. A school that raises it first is having a different conversation from one that is answering a complaint.
This is a synthesis of published regulatory material, not legal advice. Take advice on your own obligations, which differ by sector and jurisdiction.

The evidence

Read closely

Does using AI make students think less?

What the studies did
A cluster of 2025–2026 studies has examined cognitive offloading: delegating mental work to an external tool. The most widely circulated is Michael Gerlich's 2025 paper in Societies, which surveyed 666 people in the United Kingdom across age bands starting at 17, and interviewed a subset of them. It found a significant negative association between frequent AI tool use and critical thinking scores, statistically mediated by cognitive offloading, with the youngest band showing the highest AI dependence and the lowest critical thinking scores. No primary-aged children were tested. Three other pieces of work are usually cited alongside it, and they are not measuring the same thing. A 2026 experiment in Cognitive Processing put 120 young adults through a logical-fallacy task, with or without ChatGPT. A 2026 cross-sectional survey in Frontiers in Psychology asked 468 Chinese undergraduates about digital-tool offloading generally, not generative AI specifically. And Fan and colleagues, writing in the British Journal of Educational Technology in 2025, ran a randomised experiment with 117 university students split across ChatGPT, human expert, writing analytics and no-support conditions. They coined the warning "beware of metacognitive laziness", in which learners offload not just the task but the monitoring of their own understanding.
What they found
The evidence is mixed and strongly context-dependent, and we want to be careful here, because it is routinely reported as settled. Some studies do associate heavy AI reliance with weaker critical-thinking measures, and EDUCAUSE Review summarised that reading in December 2025 as "better results, worse thinking". Others point the other way. The Cognitive Processing experiment found AI users achieved higher accuracy with less mental effort, and its authors concluded explicitly that AI did not uniformly enhance or reduce critical thinking; how it was used mattered. The Frontiers survey found offloading was positively associated with cognitive self-efficacy, which in turn predicted higher critical thinking, task persistence and learning depth. The most useful result comes from Fan's randomised experiment, because randomisation lets it say something the others cannot: better essay scores did not translate into a demonstrated advantage in knowledge gain or transfer. On present evidence, how the tool is used may matter substantially more than whether it is used.
What they don't show
Three primary limitations. First, these studies are not asking the same question and cannot simply be added together. An experiment on 120 adults doing a fallacy task, a cross-sectional survey of 468 undergraduates about digital tools in general, and a survey-plus-assessment study of AI users are three different designs measuring three different things. Where the design is cross-sectional, causal direction is not established at all: people who already think less critically may use AI more, rather than the other way around. Second, the outcome measures are mostly short-horizon; what happens to a learner's capability over years of AI-assisted schooling is not yet known. Third, and most importantly, none of the cognitive-offloading studies examined for this issue was conducted on primary-aged children. The samples are older adolescents and adults, including university students; the "younger participants" in these studies are typically 17–25, not 7.

That claim is narrower than it sounds. Primary-aged evidence does exist on adjacent questions. Australia runs a national ICT literacy assessment on Year 6 students, which appears as this issue's Stat of the Term. What is missing is evidence about offloading in young children specifically - whether handing a seven-year-old a tool that completes the task damages the capability the task was meant to build. On that question, we found no study that directly tests this. So when a vendor, a conference keynote or a newspaper column tells you AI harms children's thinking, or transforms children's learning, ask which children. In the studies behind this particular claim, the answer is: not primary-aged ones.
Why it matters anyway
A plausible mechanism is not a finding, and we are not going to tell you this research says something it does not. What survives is narrower and still worth your attention. The cognitive work an AI tool can absorb is exactly the work primary schooling exists to build: decoding, arithmetic fluency, sequencing an argument, tolerating the discomfort of not yet knowing. The studies cited here give no direct evidence, in either direction, about the long-term effects of AI-enabled offloading on primary-aged children. So the decision to put young children in front of tools that complete tasks for them is being made without that evidence, and that is the fact to plan around.
Source: Fan et al., British Journal of Educational Technology, 2025 peer-reviewed · Gerlich, Societies 15(1):6, 2025 (a correction was published in September 2025 fixing a duplicated table; the author states the conclusions are unaffected) peer-reviewed · Cognitive Processing, 2026 peer-reviewed · Frontiers in Psychology, 2026 peer-reviewed · EDUCAUSE Review, Dec 2025 industry · J. Interdisciplinary Studies in Education peer-reviewed

What we're reading

Including what argues against us

Australian teachers are among the world's heaviest users of AI in their work

journalism

TALIS 2024, published 7 October 2025. 66% of Australian lower secondary teachers reported using AI in their work in the past year, against an OECD average of 36%, the fourth highest of 55 systems. Note what that figure is and is not: it covers lower secondary teachers, and it asks about AI use in their work rather than specifically in the classroom. The Australian survey covered 6,040 teachers across 359 schools across school levels, not only the group behind that headline. The number that matters more for a primary school is 47%, the share of Australian primary teachers who reported using AI in the previous year. Of the primary teachers who did not, 84% said they lacked the knowledge or skills to teach using it, which is a professional learning problem rather than a technology one.

→ OECD: Results from TALIS 2024, Australia country note international research organisation ·  commentary by Robin Shields, University of Queensland, in The Conversation

A structured unplugged approach for foundational AI literacy in primary education

preprint

Teaches AI concepts through classification, data representation and evaluation. Mostly unplugged, though it does include a short hands-on machine-learning activity. A small study: 31 fifth-grade students across two classes, of whom 25 were present for the final module and 23 completed both the post-test and the satisfaction survey. The country is not disclosed in the paper, 35% of the students had special educational needs, and it is a preprint rather than peer-reviewed. Directionally interesting, not conclusive.

Evidence for and against banning mobile phones in schools: a scoping review

peer-reviewed

Campbell and colleagues, 2024. Finds limited and inconsistent evidence that phone bans improve academic attainment.

The impact of cellphone bans in schools on student outcomes: evidence from Florida

working paper

Pointing the other way, though more modestly than it is usually reported. The estimated gain was 0.6 percentile points overall and 1.1 points on spring tests in the second year, with 1.3 for middle and high school students. Read the subgroups before you apply this to a primary school: the authors state they find no significant test-score effect for students in elementary school, and no significant effect on elementary absences either. Suspensions rose sharply in the first year, concentrated among Black students, then dissipated, and in elementary schools the ban reduced disciplinary incidents by the second year, though suspensions are far rarer there to begin with. We include both this and the Campbell review deliberately. Two credible bodies of work reach different conclusions, and any vendor or commentator who tells you this question is settled has read one of them.

Stat of the term

50%

of Australian Year 6 students reached the proficient standard in ICT literacy in 2025 - down from 55% in 2022.
From the National Assessment Program – ICT Literacy 2025, a nationally representative sample assessed in May 2025: 5,498 Year 6 students and 4,753 Year 10 students. The Year 10 figure fell further, from 46% to 37%. Conducted by ACER for ACARA. ACARA chief executive Stephen Gniel said the 2025 results "show a decline in student proficiency in ICT literacy and continuing gaps between different groups of students" (ACARA media release). Source: NAP–ICT Literacy 2025 public report.

The assessment does not measure readiness to use AI assistants, and we are not going to claim it does. But set it beside the other number in this issue, 64% of Australian 10–12 year olds having used an AI assistant, and it raises a fair question: what foundational digital knowledge should students have before a school introduces general-purpose AI tools?
Source: Fan et al., British Journal of Educational Technology, 2025 peer-reviewed · Gerlich, Societies 15(1):6, 2025 (a correction was published in September 2025 fixing a duplicated table; the author states the conclusions are unaffected) peer-reviewed · Cognitive Processing, 2026 peer-reviewed · Frontiers in Psychology, 2026 peer-reviewed · EDUCAUSE Review, Dec 2025 industry · J. Interdisciplinary Studies in Education peer-reviewed

The long view

Opinion

The data risk in a primary school is not on the children's phones

This section is opinion. It is the view of this publication, and it is meant to encourage debate and alternative views.

Ask a primary principal about online risk and the conversation usually turns to screens, social platforms and the under-16 restrictions. Those are real concerns, and the restrictions that took effect on 10 December 2025 were a serious piece of public policy. We are not going to argue against them.

But they address a particular set of harms: bullying, unwanted contact, compulsive use. Those harms attach mostly to adolescent social accounts, and your students were largely too young to hold one. For a primary school the restrictions were, in practice, close to a non-event.

The trouble is what they crowded out. For most of a year they absorbed nearly the whole public conversation about children's online safety. And there is a second question in that conversation, one that applies squarely to primary schools, which has gone almost undiscussed. It is not about what children do online. It is about what is done with their data.

For a Year 3 cohort, the largest and least examined data exposure is not a social platform. It is the stack the school itself procured.

Consider what has quietly happened to the role of a primary principal over the last decade. You now sit at the centre of a set of relationships in which a dozen or more commercial entities hold detailed longitudinal records about children - attainment, behaviour, wellbeing, attendance, photographs, sometimes voice recordings - and the main point of visibility your families had over that was a bundled form completed once, at enrolment. You have become a consent intermediary for the children's data industry. Nobody designed this. No one trained you for it. No principal's certification covers it. And in the government sector, where the Privacy Act does not even reach your school, there is often no obvious external party whose job it is to check.

That parent quoted in the OAIC's consultation report - sixteen apps, one form, no explanation available when she asked - was not necessarily describing a badly run school. In our experience many schools would recognise parts of her account. The school may well have had no single register of what it had signed up to. In our experience such registers are far from universal. Once commenced, the Code may bind many of those companies. It will not build the register.

The same blind spot shows up in a second place, and this one is closer to home for us. The reflex response to "we must prepare children for AI" is to procure an AI tool and put children in front of it. For a primary school this is doubly awkward. It is largely blocked. UNESCO recommends 13 as the minimum age for independent classroom use of generative AI, while many general-purpose assistants restrict independent accounts for under-13s and some set higher limits, which excludes most primary-school students from independent use. And it is arguably backwards. The evidence base that exists for young children is not about chatbot fluency; it is about conceptual understanding - classification, data, pattern, the logic of a decision, the limits of a model. Meanwhile the research literature on cognitive offloading, thin as it is on primary-aged children, points the same way: the risk is not that children fail to learn the tool, it is that the tool completes the thinking the schooling was meant to build. There is also a developmental complication, though the research is younger than the confidence with which it is usually quoted. A 2026 systematic review of 35 empirical studies of children interacting with chatbots finds that children readily attribute human-like qualities to them, and that what follows is more tangled than good or bad: children form social ties with these systems, test boundaries with them, and hold a kind of dual consciousness about whether they are real: see the review. Children who overestimate what a conversational system knows, feels or understands need close supervision when they use one.

Declaring our interest: Code Camp offers coding and technology programs for primary school students, so we benefit commercially if you find this argument persuasive. Judge the evidence, not us.

The strongest case against us: primary and secondary are a continuum, not a wall. Year 6 students use their older siblings' accounts, and a policy conversation that treats under-13s as sealed off from the platform economy is fantasy. It is also fair to say that a school stack, however sprawling, is a materially lower-harm environment than an algorithmic feed - the OAIC's parent submissions describe unease and lost trust, not the acute harms that motivated the social media legislation. Proportionality matters, and we may be over-weighting the governance failure because it is the tractable one.

But tractable is the point. You cannot fix the attention economy from a primary school office. You can, this term, find out what your school signed.

We would genuinely value your opinion on this. If you run a primary school and you think our position on this is wrong, particularly if you have built the register and found it was not the problem we think it is, reply to this email. Subject to your prior permission, we may publish some of these alternative views in our next Issue.

The Primary Source is published quarterly by Code Camp, part of Cluey Learning. Code Camp has run after-school and holiday camp coding and technology programs for Australian children aged 5-13 since 2013. We do not sell anything in these pages. If we get something wrong we will correct it in the next issue and add a dated note to the web version. We will not silently edit. For more information take a look at our school programs.

To stop receiving The Primary Source, use the unsubscribe link in any issue or email editor@codecamp.com.au.