How do I raise a child who thinks for themselves?

By protecting the part of their day where they are stuck. Independent thinking is built by struggling with a problem before help arrives, and the tool now available to every child removes that struggle faster and more pleasantly than anything before it. The single most useful finding for a parent comes from a field experiment that gave teenagers an AI tutor, watched their marks rise, then took it away: the group with unrestricted access scored below students who had never had it at all, while the group given a version that withheld answers was largely spared. The design of the help decided the outcome. At home, you are the one designing the help.

The experiment that should set your house rules#

Bastani and colleagues ran a field experiment in Turkish high schools. Students with unrestricted access to a GPT-4 assistant improved their practice performance by 48 per cent. Students given a guardrailed tutor, built to prompt rather than to answer, improved by 127 per cent. Then the tools were withdrawn for the exam. The unrestricted group scored 17 per cent below students who had never used AI at all. The guardrailed group largely avoided that penalty.

Nothing about the amount of use explains that difference. Both groups used it heavily and both improved while they had it. What separated them was whether the tool made the child do the thinking. That is the whole question. Configuration and habit decide it, and neither of those is what a screen-time rule regulates.

It also explains why the usual parental instruments fail. A time limit does not distinguish between forty minutes spent arguing with a model about a history source and forty minutes spent pasting in questions. A ban produces neither. The variable that mattered in the only experiment to test it is one a rule about hours cannot reach.

Difficulty is the mechanism, not the obstacle#

The learning science here predates AI by decades and is unusually settled. Bjork and Bjork's work on desirable difficulties found that conditions which slow performance during practice, spacing, interleaving, testing yourself, tend to improve long-term retention, while conditions that make practice feel fluent tend to worsen it. Kapur's productive failure studies found students who attempted a problem before instruction outperforming those taught the method first, despite failing more during the attempt. Roediger and Karpicke found retrieving information from memory beating restudying it, again with the harder condition winning.

The common thread is that the feeling of learning and the fact of learning run in opposite directions. A child who found the homework easy has a weaker signal than a child who found it hard. An assistant that removes friction is optimising precisely the variable that predicts the loss.

Two more findings are worth a parent's attention because they describe how the illusion works. Sparrow and colleagues found people remembering where to find information rather than the information itself when they expected it to remain available, which is the Google effect. Fisher and colleagues found that searching the internet for explanations inflated people's estimates of their own internal knowledge, even on unrelated questions afterwards. Access to an answer feels like possession of it. Children are not unusually susceptible to that; adults are just as bad, so this one is difficult to model at home.

What a summary costs that reading three sources does not#

Melumad and Yun ran seven experiments comparing what happens when people learn a topic from an AI summary against learning it from the underlying sources. One experiment generated a single synthesis and then rewrote the identical facts as six articles presented as links. Participants given the summary spent less time engaging, reported learning less, produced shorter advice with fewer named facts, and their advice was three times more similar to each other's. Rated comprehensiveness did not move at all, which is the uncomfortable part: the summarised version did not feel worse while it was being used.

Adding real-time source links to the summary did not fix it, because only about a quarter of participants clicked any link. That is the design every homework tool has converged on, tested, and failing. The full argument is at should I let AI summarise everything I read.

For a child, the practical translation is narrow and useful. Reading three sources and disagreeing with one of them is a different cognitive act from reading a synthesis of the three. The synthesis is faster, feels complete, and removes the moment where a young person notices that two adults they respect say incompatible things. That moment is most of what independent judgement is made of.

Boredom goes first, and nothing replaces it by accident#

In What I Tell Parents About AI, published on 3 May 2026, Rahim Hirji sets out the rules he actually runs at home rather than the ones that sound good in a talk. Phones out of bedrooms. The laptop treated as a tool and the phone treated as a relationship. A life that is not on a screen, with the observation that boredom is a developmental requirement and that something has to be put in its place once a device removes it. And the rule underneath the others:

I let her struggle before I help. The struggle is the lesson. Removing the struggle removes the lesson.

He is candid that he does not enforce any of it perfectly, and the candour is the point: the aim is knowing which two or three battles matter and not folding on those. The same essay gives the sequence he wants a child to learn, which is build first, augment second: the model is something you argue with about a character you have already drawn, a source you have already read, code you have already written and cannot fix. The work is the child's first. That formulation is the parental version of the augmented mindset and of human at the start.

Five things that survive contact with an actual teenager#

• Ask for the attempt before the answer. Not as a punishment. Because the attempt is what the practice is for, and a child who has tried and failed reads the model's answer completely differently from one who has not.
• Insist on one source they read themselves. One primary text per piece of work, read whole. It costs twenty minutes and it is the only reliable defence against the convergence effect.
• Make them find the mistake. Give the model a question you already know the answer to, together, and look for where it goes confidently wrong. Scepticism about fluent output gets taught by demonstration; being told to be sceptical does almost nothing. See how do I know when AI is wrong.
• Protect something unrecorded. An instrument, a sport, a job, an argument at the dinner table. Independence needs somewhere to be practised that produces no output anyone grades.
• Model it badly, out loud. Say when you did not know something and looked it up, and say what you still are not sure about afterwards. Children calibrate their own confidence against the adults in the room.

Questions for the school, which is improvising too#

Roughly a fifth of schools in England had a policy on safe and appropriate AI use in the Department for Education's 2024 to 2025 survey. Teachers reported using generative AI for lesson planning at 35 per cent and for marking at 5 per cent, and the picture on the pupil side was largely defensive. The full evidence is at how AI will change teaching, and the short version for a parent is that most schools are inventing this in real time without a curriculum or a budget for it.

Four questions worth asking, from the same essay: what is the school's AI policy, and if there is none then that is the policy; how is my child being taught to think when the tool is doing the thinking; how are they being taught to spot when it is wrong; and are they being taught when not to use it. A school whose only answer is a ban is protecting its grading system. A school with no limits is hoping the problem solves itself.

What this page will not tell you#

• Whether AI companions harm adolescent development. Common Sense Media's 2025 survey of American teenagers reported 72 per cent having used an AI companion, on a definition that included general assistants, and the report's own limitations section concedes respondents may have conflated general use. Every available study on the emotional effects is cross-sectional and cannot separate cause from correlation. This research is not publishing an answer until better evidence exists, and it says so at how much should teenagers use AI.
• Whether any of this changes how children develop. The developmental evidence base is thin. Nothing here should be read as a claim about brain development.
• What the right age is. No study establishes one, and anyone offering a number is offering an opinion.
• Whether the Bastani result generalises to your child. It is one field experiment, in one country, in one subject, with one cohort. It is also the only experiment that withdrew the tool to see what remained. That is the reason it carries so much weight on this page, and the reason more of them are needed.
• What the cognitive debt study proves. The MIT Media Lab EEG work on essay writing is a preprint with a small sample, and a published comment has challenged aspects of its analysis. The estate treats it as suggestive rather than settled, at cognitive debt and capability debt.

Key sources

• Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakci, O. and Mariman, R. (2025). Generative AI Without Guardrails Can Harm Learning. Proceedings of the National Academy of Sciences.
• Bjork, R. and Bjork, E. (2011). Making Things Hard on Yourself, But in a Good Way: Creating Desirable Difficulties to Enhance Learning.
• Kapur, M. (2008). Productive Failure. Cognition and Instruction, 26.
• Roediger, H. and Karpicke, J. (2006). Test-Enhanced Learning. Psychological Science, 17.
• Melumad, S. and Yun, J. H. (2025). Experimental evidence of the effects of large language models versus web search on depth of learning. PNAS Nexus.
• Sparrow, B., Liu, J. and Wegner, D. (2011). Google Effects on Memory. Science, 333.
• Fisher, M., Goddu, M. and Keil, F. (2015). Searching for explanations: How the Internet inflates estimates of internal knowledge. Journal of Experimental Psychology: General.
• Hirji, R. (2026). What I Tell Parents About AI. Box of Amazing, 3 May 2026.

Related SuperSkills research#

On the age question, should children use AI and how much should teenagers use AI. On the mechanism, desirable difficulty, productive struggle and how humans learn with AI. On the habits, getting AI to challenge you and using AI without dependency. On what school is doing, teaching and what to tell your children to study.

About this research#

Rahim Hirji is the author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company, and spent two decades in education and education technology. Every study cited here holds a graded entry in the evidence base stating its method, its sample and what it does not prove. Questions about AI companions, loneliness and child development are deliberately unanswered on this page, because the available evidence is cross-sectional and the subject is serious enough that a confident answer would be worse than none.

How this research works  ·  Reviewed quarterly  ·  Found an error? Tell me and it is corrected on the page.