The Classroom AI Evidence Is In, and It Is Not What Parents Were Promised
Early data on children using AI in school points to worse outcomes, not better ones, and it is time to take the numbers seriously.
Sign up for my Substack daily AI newsletter here.
See my AI Training course portfolio for corporate Business Leaders here.
Follow me on LinkedIn: https://www.linkedin.com/in/johanosteyn/
Every week brings new advice on how parents should prepare children for a future shaped by artificial intelligence. The tone is confident, the tips practical, and the underlying assumption largely unchallenged: give children the right skills, expose them early to the technology, and they will thrive. Meanwhile, the first serious international evidence on what happens when children actually use AI in class is arriving, and it is more sobering than the advice columns suggest. Adoption is racing ahead of pedagogy. Rules are being written after children have already logged in. Learning outcomes, where they can be measured, are not moving in the direction parents were led to expect. This is not a case for banning AI from schools. It is a case for reading the evidence before we scale it further.
CONTEXT AND BACKGROUND
The first thing to note is how quickly children have adopted these tools. In a recent survey of American nine-to-seventeen-year-olds, Common Sense Media found that AI is now routinely part of children’s daily lives, yet only just over half of them have been taught how to use it safely, and nearly half have never discussed it with a parent. Adoption is high. Guidance is thin.
The classroom evidence is beginning to catch up. The OECD’s Digital Education Outlook 2026 examines what happens when students, teachers, and pairs of the two actually use generative AI in real learning environments, and its findings should temper the enthusiasm on both sides of the debate. International bodies are only now converging on what children should learn about AI at all, with UNESCO’s competency framework listing twelve capabilities across four dimensions and placing critical evaluation at the centre.
INSIGHT AND ANALYSIS
The uncomfortable finding across this early evidence is that using AI is not the same as learning with it. A child who finishes an essay with a chatbot has produced an essay. Whether they have learned to write is a different question, and the answers now emerging suggest it is often no. In one Wharton-led controlled study of high-school pupils, those given unfettered access to GPT-4 performed strongly while it was in front of them, then scored about seventeen per cent worse than a control group when the assistance was taken away. Fluency with the tool did not become mastery of the subject.
This matters because most of the advice being offered to parents assumes the opposite. If AI in class were reliably making children learn faster and think better, the correct response would be to lean in. If it is not, then racing to adopt it before pedagogy catches up is not neutral. It is the education equivalent of prescribing a drug before the trial has finished.
I have previously written about this, arguing that no matter how capable the technology becomes, educational accountability cannot be delegated to a machine, and that teachers, schools and parents remain answerable for whether a child has learned.
IMPLICATIONS
The implications for South Africa are sharper still. More than eight in ten Grade 4 learners cannot read for meaning in any language, and the country’s foundation-phase textbook procurement is in disarray. Layering AI tools onto classrooms whose learners cannot yet decode a sentence is not innovation. It is a distraction, and often a costly one.
Schools do need a serious policy for how AI is used in classrooms, but a policy is not the same as adoption for its own sake. The right sequence is to fix the foundations, teach children to interrogate what a model produces, and only then scale AI-assisted learning where evidence shows it helps. This is cognitive governance applied to education. Humans, in this case teachers and parents, must remain net contributors to the child’s learning, and accountability for whether they learn cannot be handed to a chatbot.
CLOSING TAKEAWAY
The temptation, faced with a technology this rapid, is to assume that anyone urging caution is a Luddite. That is a mistake. The most honest reading of the early evidence is not that AI has no place in schools, but that the confident advice columns are running ahead of what we actually know. Our children will live longer with this technology than any of us, which is precisely why we owe them classrooms built on measured learning, not on parental hope, and tools that build independent minds rather than dependent users. The question is not whether they can use AI. It is whether we can prove they are learning when they do.
Author Bio: Johan Steyn is a prominent AI thought leader, speaker, and author with a deep understanding of artificial intelligence’s impact on business and society. He is passionate about ethical AI development and its role in shaping a better future. Find out more about Johan’s work at https://www.aiforbusiness.net
Comments