Use AI to Learn to Code (Without Letting It Learn for You)

Last month, a student of mine spent four weeks using ChatGPT to build projects she was proud of. She showed me a travel planner app, a quiz game, and a portfolio site. Then I asked her one question: "Can you write a loop that filters a list without asking the AI first?" She froze. She had copied the code, but the code had not copied into her brain.

She had learned how to ask for code. She had not learned how to code.

Here is the direct answer: You can use AI to learn to code, but only if you treat it like a tutor, not a vending machine. A vending machine gives you a cold drink when you put in money.

• Most people use AI that way: they type a request and get finished code. That gives you finished work with no learning.

• A tutor makes you think, then checks your thinking.

• The difference is not the tool. It is how you use it.

A real experiment with nearly 1,000 students proved this. Unrestricted AI made practice scores jump 48% but exam scores drop 17%. Tutor-style AI produced learning that stuck. I will explain the numbers, then give you five rules to follow so AI helps you grow instead of replacing your thinking.

What did the experiment actually prove?

Unrestricted AI helps you practice but hurts what you remember.

Researchers at a university in Turkey ran a randomized experiment with nearly 1,000 high school math students. They split the students into three groups: one with unrestricted access to GPT, one with a guarded "GPT Tutor" that gave hints instead of answers, and one with no AI at all.

The unrestricted group scored 48% better during AI-assisted practice. That sounds great. But when the AI was removed for the real exam, the same group scored 17% worse than the control group that never used AI. The students had practiced more and learned less.

The GPT Tutor group scored 127% better in practice and scored about the same as the control on the exam. Their learning stuck. The study also found AI-assisted students were overly optimistic about what they had learned. They felt productive, but the knowledge was not in their heads.

The paper is published on arXiv and comes from researchers at Corvinus University of Budapest. It concludes that uncontrolled AI use "leads to disengaged students and low understanding of material."

Why does unrestricted AI wreck learning?

When the AI gives a complete answer, your brain never does the work of forming it.

Learning happens when you struggle just enough to build a mental model. Think of it like lifting weights. If a machine lifts the bar for you, your muscles never grow. The vending machine gives you the product without the process.

You feel productive because you are producing output. But producing output and building understanding are different things. The experiment showed this clearly: the unrestricted group scored better in practice and worse on the exam. They had been productive, not educated.

I saw this with another student last term. He used Copilot to complete every assignment. His code worked, his grades in practice were high, and he felt confident. Then the final exam required writing code in a blank editor. He failed the practical section. The same pattern as the experiment, in my own classroom.

How do you ask for a concept instead of code?

Ask "how does this work?" not "what is the code for this?"

The way you phrase your question decides how much you learn. "Write a function that reverses a string" gives you output. "Explain how string reversal works step by step, then I'll try it myself" gives you understanding.

Make the AI explain the idea, not produce the solution. Ask for analogies. Ask for the mistake you would likely make. Ask it to check your reasoning instead of your code.

Do not ask for the full solution to your exercise. The GPT Tutor in the experiment gave hints and explanations, not complete answers. That tutoring approach produced real learning, with no drop in exam scores. The students in that group learned as well as the control, and practiced far more.

Why must you type the code yourself?

Copying code by hand is not wasted time. It is the learning time.

When you type code, your fingers learn the syntax. Your eyes learn the structure. Your brain connects the logic to the movement. When you copy-paste, only your clipboard learns.

The experiment's 17% drop came from students who relied on answers, not those who typed practice themselves. The difference was not the AI. It was who did the typing.

A student I mentor switched from copy-paste to typing everything manually. For more on how to study programming effectively, see this guide. He still used AI for explanations, but he always typed the solution himself. After three weeks, he told me he could finally read code without translating it in his head. That is the feeling of real progress.

If the code is long, type it in chunks. Then close the AI and rewrite it from memory. That act of recall is what moves knowledge from short-term to long-term memory.

How do you use AI to generate practice problems, not answers?

Ask the AI to create new exercises for you, then solve them without help.

The AI becomes a tutor that provides practice, not a vending machine that provides solutions. Instead of "solve this for me," say "create five similar problems at this difficulty level, and I will solve them before you check."

Here are some sample prompt templates you can copy and try right now. These are for beginners, but you can change the topic to match your lesson.

• Practice problems:

  Create five beginner-level JavaScript exercises about arrays and loops. Do not give me the answers. I will try to solve them first, then you check my work.

• Concept explanation:

  Explain how a for loop works in JavaScript using an analogy. Do not write any code. Just explain the idea step by step.

• Debugging hint:

  Here is my code. Do not fix it. Give me one hint about where the problem is, and let me try to fix it myself.

• Revision from memory:

  I just wrote a function to reverse a string. Close this chat. I will rewrite it from memory. When I come back, check my code and tell me what I missed — but do not fix it. Give me one hint about where the problem is, and let me try to fix it myself.

You can use these as they are, or change the topic to what you are learning this week. The key is to make the AI give you work, not answers.

This forces you to apply what you learned to new situations. That is where real understanding lives. Solving the same problem twice teaches far less than solving ten variations of it.

You can also ask the AI to give you a skeleton with blanks to fill, or to act as a reviewer after you finish. Reviewing your work is a different skill than writing it, and both matter for exams and job interviews.

Pair this with a real tutorial from MDN's JavaScript guide. After a week, you will see the difference in what you actually remember when the AI is off.

Why should you set a no-AI checkpoint every day?

Write at least one piece of code completely without AI, every single day.

This is the rule that most students skip, and it is the most important one. The experiment showed exactly what happens when AI is always available: your brain outsources the work and forgets how to do it. A checkpoint forces your brain back on duty.

Pick a small task. A function, a loop, a CSS layout. Something you can finish in 10 minutes. Write it in a blank editor with no AI open. If you get stuck for more than 5 minutes, look up the concept, not the full answer.

Think of it like learning to drive. You can watch someone drive for weeks, but until you sit behind the wheel with no instructor's hand on the wheel, you have not learned to drive. The no-AI checkpoint is your solo drive.

One student I know called this his "brain gym session." He did 15 minutes of no-AI coding every evening. If you are still deciding whether AI is worth using at all while you are a beginner, read my earlier post on whether it is worth learning to code with AI.

That will help you decide. When his course ended, he was the only person in his study group who could write a complete function in an interview-style setting. Everyone else had relied on Copilot and froze.

How do you track learning, not just output?

Keep a log of what you can do without AI, not just what you built with it.

Most beginners measure progress by projects completed. That metric lies when AI does half the work. Instead, track what you can reproduce from memory. If you cannot rebuild a project without AI, you did not learn it.

At the end of each week, ask yourself: what can I write from scratch? List the specific skills. If the list is short, slow down and practice before starting new material. The experiment's overconfidence finding shows that students believed they had learned more than they actually had. A skill log keeps you honest.

This connects to why you forget what you learn to code in the first place. The forgetting curve hits everyone, and escaping tutorial hell is a natural next step after setting your no-AI checkpoint. AI hides the forgetting until the exam reveals it.

Which approach should you pick?

Use AI for explanations, practice generation, and review. Never use it to produce the answer you will be tested on.

The experiment's conclusion is simple: unrestricted AI makes you feel productive but learn less. Guarded AI that acts like a tutor produces learning that sticks. The tool did not change. The same person, using the same AI, went from solving a problem in five minutes to not being able to start it from a blank screen. That is not a tool problem. That is a practice problem. The usage did.

Start with this rule today: for your next coding exercise, ask the AI to explain the concept but not to write the solution. Then type it yourself. Then close the AI and rebuild it from memory. That sequence takes 30 minutes and does more for your learning than an hour of generating code you do not understand.

Keep your no-AI checkpoint daily, and you will avoid the 17% drop. And if you feel stuck in tutorial loops, escaping tutorial hell is the natural next step after setting your no-AI checkpoint.

Frequently asked questions

Is it cheating to use AI to learn to code?

No, if you use it like a tutor. Start with this prompt: 'Explain this error in one sentence, then give me an exercise to find the bug on my own.' It becomes cheating when the AI produces the answer that you submit or present as your own understanding without doing the work yourself.

Can I use ChatGPT to debug my code?

Yes, but ask for a hint first. Say "where should I look?" instead of "fix this." You learn more from finding the bug yourself with a pointer.

Do I need to avoid AI completely as a beginner?

No. The experiment showed that tutor-style AI improved practice scores by 127% with no harm to real learning. The problem is unrestricted use, not AI itself.

How much time should I spend with AI vs without it?

Use this rough split: explain with AI, practice without it. For every 10 minutes of AI explanation, spend another 10 typing code yourself and 5 minutes rebuilding from memory.

Will using AI make me dependent on it?

It will if you never practice without it. Daily no-AI checkpoints prevent dependence. The experiment's 17% exam drop shows what happens when you skip them.

What should I do when I am stuck and AI is not helping?

Write a simpler version of the problem. Break it into smaller pieces. If you are still stuck after 10 minutes, look up the specific concept on MDN before asking the AI for anything more than an explanation.

Is the 48% and 17% result from a real study I can read?

Yes. The paper is published on arXiv and describes a randomized experiment with nearly 1,000 high school math students in Turkey, with three comparison groups.

How is this different from the debate about whether to learn to code with AI at all?

That question is about whether to use AI. This post is about how to use it so you still learn. Both matter, but the method is something you control completely.