AI Can Code My Son's Game. What Should He Still Learn?

One evening, I was using AI for my business while my son was coding a new game.

There was nothing unusual about either activity. AI had become part of my work, and it was becoming part of the way he made things. If he wanted to add a feature, understand an error or turn an idea into code, help was only a prompt away.

That should have made me optimistic. In many ways, it did.

My son could attempt something that would once have required much more instruction before he could even begin. The distance between imagination and a working result had become shorter.

But I was also aware of an uncomfortable difference between us.

I had spent years making business decisions before asking AI to assist me. My son could receive a convincing answer before he had enough experience to know whether the answer was good.

At first, I thought the question was whether children were losing their basic skills.

Then another question bothered me more:

When the tools change this quickly, what should count as a basic skill in the first place?

We do not teach driving by preserving the horse and carriage

Someone entering driving school today is not expected to learn how to control a horse and carriage first. The carriage was once a mainstream form of transport. It is not a necessary stage on the way to driving a modern car.

Education has always made choices like this. Some techniques remain useful; others become historical knowledge or specialist skills. Asking children to repeat every method used by the previous generation would not prepare them for the future. It would use their limited learning time to preserve our familiarity.

Yet moving from a carriage to a car did not eliminate the need for fundamentals. It changed them.

A driver no longer needs to control reins, but still needs direction, coordination, risk judgment, knowledge of the rules and the ability to decide when not to drive. A navigation system can recommend a route. It cannot make every decision safe.

AI may create a similar change in learning.

My son may not need to memorise every piece of syntax or manually write every line of routine code. But if he does not understand state, conditions, events, data and cause and effect, he may not know why his game works, why it fails or what an apparently small change will break.

The old operation can become unnecessary while the judgment underneath it becomes more important.

Different AI tools are different vehicles

It is tempting to discuss "AI" as though every tool creates the same opportunity and the same risk.

They do not.

An AI tutor that asks a student questions and gives limited hints resembles a learner car with an instructor. A general chatbot that can produce a complete answer is closer to an easy-to-drive private car. An AI agent that can search, edit files, run code and take actions can be more like a large truck: useful, powerful and capable of affecting people well beyond the driver.

All of these vehicles may take a student somewhere. That does not tell us whether the journey was safe or whether the destination was educational.

A working game is one destination. Understanding how the game works is another.

AI may take a child to the first while bypassing the second.

This is why I no longer think the important question is simply, "Did he use AI?" I want to know what work the tool performed and what work remained with him.

Did he decide what the problem was? Could he explain the suggested code? Did he compare alternatives? Could he change the rules of the game when the first version no longer matched his idea? When something failed, did he investigate or simply ask for another complete answer?

The same tool can be a scaffold in one learning process and a substitute in another.

Better performance is not always retained learning

The clearest warning I found came from a large randomised study involving nearly a thousand high-school mathematics students.

Students with access to a standard GPT-4 interface performed much better during assisted practice. Students using a safeguarded tutor designed to give teacher-written hints performed better still. But when the AI was removed for an exam, the students who had used the unrestricted version scored 17% lower than students who had never received AI access. The negative effect was largely removed in the safeguarded tutor group, although that group did not show a clear improvement over the control group on the unassisted exam.

The study was about mathematics in a particular school setting, not my son's game. It does not prove that ordinary AI use makes children worse learners. It demonstrates a narrower and more useful point: a tool can improve performance while it is present without leaving the same improvement in the learner.

There is also evidence pointing away from a simple anti-AI conclusion.

A 2023 CHI study followed 69 learners aged 10 to 17 with no previous text-based programming experience. Half used an OpenAI Codex-based code generator for code-authoring tasks. They completed more tasks, produced higher-scoring code and did not perform worse on manual code-modification tasks. One week later, the AI group scored slightly higher, but the retention difference was not statistically significant.

That study is especially relevant to my question because it concerned young people learning to code. It is also a historical snapshot. The sessions took place in 2022 using Codex, a much more limited system than current AI coding tools. It cannot tell us what today's models—or the tools available when my son becomes an adult—will do.

What the two studies show together is not that AI is good or bad for learning. They show why tool design, task design and what happens after the tool is removed matter.

Schools are already struggling with this distinction

The policy responses developing around the world are not converging on a complete ban or unrestricted use.

Australia's national framework says AI should enhance critical thinking rather than restrict human thought. It also says assignments should clearly state how AI may or may not be used, so that schools can still evaluate a student's ability.

Japan's guidance similarly treats generative AI as a possible reference tool while leaving final judgment and responsibility with people. It advises schools to consider a student's developmental stage and rejects both a uniform ban and a uniform requirement to use AI.

Hong Kong's 2026 digital-education blueprint uses an unusually clear phrase: schools should guard against students becoming "technologically advanced with lower-order thinking." It places teachers, rather than technology, in charge of judgment and calls for closer supervision when younger students use generative AI.

These policies differ in language and implementation, but they share a concern I recognise at home. Access is not the same as readiness. Producing more is not the same as learning more.

Teachers are also inheriting new work. AI may save time when creating lesson materials or handling administration, but it also creates a need to check generated content, reconsider homework, examine whether submitted work reflects a student's ability and decide what kind of AI assistance is acceptable. The work is not simply disappearing. Some of it is moving from production to verification and governance.

Safe driving includes the other people on the road

The driving analogy also changes how I think about responsibility.

My son is not the only person on the road.

A classmate may not have access to the same paid tools or parental help. A teacher may need to decide whether a polished assignment represents the student's understanding. A creator's writing, images or code may be used without proper recognition. Private information may be entered into a system that stores or reuses it. AI-generated code may later become part of a product used by someone who never knew how it was created.

"Safe use" therefore cannot mean only that the student reaches the desired result without getting hurt.

It also means understanding how the journey affects teachers, classmates, creators, data owners and future users.

The more capable the vehicle becomes, the less reasonable it is to measure safety only by whether it arrived.

Perhaps language is more fundamental than code

If AI can increasingly write code, does that mean speaking, writing and social ability will matter more?

I think they may. I am less certain that they should be dismissed as soft skills.

Language is how a child defines a problem, asks for help, explains a choice, notices ambiguity and coordinates with other people. It is also how the child instructs an AI system. A person who cannot describe the problem clearly may generate a great deal of output without getting closer to the right problem.

Social ability matters for a similar reason. The future will not contain only humans giving instructions to machines. It will still contain disagreement, trust, negotiation, care, responsibility and people who experience the consequences of technical decisions.

But I do not think language and social ability simply replace technical knowledge. A child needs enough knowledge of a field to recognise when an answer is implausible, incomplete or dangerous. An articulate person with no domain understanding can still explain a wrong answer very confidently.

The balance may be different from the one I learned. The need for a balance remains.

The most durable basic skill may be learning itself

Tools will change. Today's mainstream interface may look as old as the horse and carriage by the time my son enters full-time work.

That makes the ability to learn seem like the safest answer. But "learning how to learn" can become another vague phrase unless I can see what it requires.

For me, it includes at least five abilities:

• Remaining curious long enough to investigate rather than accept the first fluent answer.
• Knowing enough about oneself to notice confusion, avoidance, overconfidence and the point at which help is needed.
• Breaking an unfamiliar problem into smaller questions.
• Seeking feedback and changing an approach when the evidence does not support it.
• Carrying useful understanding from one tool or situation into the next.

These abilities are not separate from knowledge. A child becomes better at learning partly by learning real things, making predictions, receiving correction and trying again. We cannot replace every difficult encounter with a lesson about adaptability.

Learning ability is built through learning.

What I now want to see after the game works

I do not want to turn my son's game into a formal examination every time he shows me something he has made. Curiosity can disappear if every project becomes evidence for an anxious parent.

I also do not want a working screen to be the only evidence that learning occurred.

The questions I now find useful are simple:

• Can he explain one important part in his own words?
• Can he change a rule when the requirement changes?
• Can he find or reason about an error instead of replacing everything?
• Can he identify something the AI may have misunderstood?
• Can he tell me what he would need to learn next?

He does not need to complete every project without AI. Adults do not work that way either. What matters is whether the tool is expanding his ability to make decisions or quietly removing every occasion on which that ability would form.

I began with a fear that AI would weaken his basic skills.

I now think that fear was too imprecise.

Some of the skills my generation practised may become less important. Others will remain but appear in a different form. The task is not to preserve every old operation. It is to identify the knowledge and judgment a child needs in order to use new tools without becoming only their passenger.

I do not need my son to walk simply because I learned to walk.

I want him to recognise the destination, understand the vehicle, notice the other people on the road and know when he must take control.

Research and policy sources

• Bastani et al.: Generative AI Without Guardrails Can Harm Learning—Evidence from High School Mathematics
• Kazemitabaar et al.: Studying the Effect of AI Code Generators on Supporting Novice Learners in Introductory Programming
• OECD: PISA 2025 Results, Volume I—Students' Use of Artificial Intelligence for Schoolwork
• Australian Government: Australian Framework for Generative Artificial Intelligence in Schools
• Japan Ministry of Education: Guideline for the Use of Generative AI in Primary and Secondary Education, Version 2.0
• Hong Kong Education Bureau: Blueprint for Digital Education Development in Primary and Secondary Schools
• UK Department for Education: Generative Artificial Intelligence in Education