Should children use AI?

The evidence does not support a single answer, and the honest response is age-banded and conditional. What it does support is a sharper claim than most guidance offers: how a child uses AI matters far more than whether they do, and the difference between the two designs is the difference between help and harm.

This page separates what the developmental literature establishes from what technology commentary asserts, because the two are routinely blended and only one of them has evidence behind it.

The finding that should shape every decision

Bastani and colleagues gave nearly a thousand high-school mathematics students access to a GPT-4 tutor. Grades rose while the tool was available: 48 per cent with a plain chat interface, 127 per cent with a guardrailed tutor designed to make the student do the work.

Then access was withdrawn. The plain-interface group scored 17 per cent lower than students who never had access at all. The guardrailed group largely did not.

Same model. Opposite outcomes. The variable was interface design, not exposure. Which means the question "should children use AI" is badly posed, and the answerable question is "under what design, doing what."

Why this happens

Learning is produced by effortful retrieval, not by receiving correct answers. Conditions that make study feel harder improve long-term retention, and conditions that make it feel fluent improve immediate performance while worsening retention. A system that supplies the answer removes exactly the effort that would have built the capability, and it does so while the homework still gets done, so nothing signals the loss to the child, the parent or the teacher.

See desirable difficulty and cognitive load.

By age, with the confidence stated

Under roughly 8. The evidence base is thin and this research declines to give confident guidance. UNICEF has set out requirements for AI systems affecting children, and the sensible default is caution, not because harm is demonstrated but because it is unstudied and the developmental stakes are high. Anyone offering confident advice here is not working from evidence.

Roughly 8 to 13. The foundational skills being built, reading fluency, arithmetic, writing, are exactly the ones most easily short-circuited. If AI is used, it should be doing something other than producing the output: explaining a concept after an attempt, generating practice problems, checking work the child has already done. Never producing the thing being assessed.

Roughly 14 to 18. Prohibition is not workable, and the serious institutions have stopped pretending otherwise. The useful shift is from banning to teaching a rule they can apply themselves: form your own view first, use the tool second, and be able to explain the answer without it. That is Human at the Start in a form a teenager can actually use.

The thing parents ask that nobody answers

"How do I know if it is harming them?" The test is the same one adults should apply to themselves: can they still do it unaided? Not whether the homework is good, which is the only signal most households have and the one signal that cannot detect the problem. Ask them to explain a piece of finished work without the device. That takes five minutes and tells you more than any monitoring software.

Where the evidence is genuinely absent

The Bastani study is one subject, one age group, one country, unreplicated. There is no longitudinal evidence on children's cognitive development under sustained AI use, because the technology is too new for it to exist. Claims about a generation being made incapable, and claims that this is just another moral panic about a new tool, are both currently unsupported.

A separate and less-studied concern deserves naming. Bill Gates, writing in August 2026, raised AI companions and children's relationships, noting he doubted he would have put in the same work as a young man had an endlessly available and agreeable companion existed. That is a hypothesis from someone with no incentive to raise it, not a finding, and this research treats it as a question worth studying rather than a conclusion.

Change the design, not the ban

• Change the design, not the ban. Guardrailed use helped; unrestricted use harmed. That is the whole practical finding.
• Attempt first, always. The child produces something before the tool is opened. This single habit preserves most of the value.
• Test unaided, occasionally and without drama. Explain this to me without the screen.
• Separate tedium from difficulty. Formatting a bibliography is tedium. Constructing the argument is the point. They look identical to a tired teenager at 10pm.
• Do not rely on detection. It does not work reliably and its errors fall hardest on second-language writers. See does AI detection work.

Related SuperSkills research

On learning, how humans learn with AI and desirable difficulty. On school assessment, assessing students when AI can do the assignment. On the adult version, am I becoming dependent on AI. On what is and is not established, what we actually know.

Key sources

• Bastani, H. et al. (2025). Generative AI can harm learning. PNAS.
• UNICEF (2025). AI and children: requirements and recommendations.
• Bjork, R. A. and Bjork, E. L. Desirable difficulties in theory and practice.
• Liang, W. et al. (2023). GPT detectors are biased against non-native English writers. Patterns, 4(7).

About this research

Rahim Hirji is the author of SuperSkills (Kogan Page, 2026) and founder of The SuperSkills Intelligence Company. This page declines to give confident guidance for under-8s because the evidence does not exist. It is not clinical or educational advice. Reviewed quarterly.