Coding Park
No, AI Shouldn’t Code for Your Students
enAISTEMPython19 novembre 2026

No, AI Shouldn’t Code for Your Students

A
Amine Lajmi
Coding Park

Philosophers are warning about AI’s effects on learning. Mathematicians describe an unprecedented disruption of their discipline. In tech, some already argue that writing code will soon become unnecessary: understanding architecture, systems, and the big picture will supposedly be enough.

I do not share the alarmism of the first group. But I believe both sides are missing the point.

Some look at the issue from television studios. Others look at it through their tools and promises of productivity. Few look at it from a classroom where twelve eleven-year-olds are learning Python, or from a training session where data engineers are learning to design pipelines with Kafka or Flink.

That is where I am speaking from.

What senior engineers know, and what money cannot buy

When a software architect tells you that a design will not hold up in production, it is not simply because they read it in a book. It is because they have dealt with the incident at three in the morning. They have watched a poorly partitioned Kafka pipeline collapse under load. They have spent years doing the work.

That perspective, that expert judgment, cannot be bought. It is acquired through repeated practice, mistakes, and a gradual understanding of what works and what breaks.

Senior engineers who delegate part of their coding to an LLM can generally evaluate what the machine produces, spot a hallucination, and choose between two architectures. That discernment comes from ten, fifteen, or twenty years of writing code.

How will the next generation acquire that judgment if they never practise?

What we see in the classroom

After a few sessions, an eleven-year-old student asked us directly: “So I just copy the solution, right?”

At the other end of the spectrum, a Master’s student discovers the AI assistant in their Databricks environment. Barely two minutes after the instructor shares the notebook, the student uses it to solve every cell.

This is not an intelligence problem. It is an instructional design problem. If a shortcut exists, learners will take it. That is normal. It is deeply human.

During a workshop at the European School, I was looking at the teaching dashboard with Ayman, one of our instructors. We could see a student’s engagement sequence in real time: away, idle, reading the tutorial, writing code, then stuck.

“What does stuck mean?” Ayman asked me.

Good question. Intuitively, the student writes, deletes, returns to the editor, and tries again. These are not certainties produced by fixed rules, but behavioural heuristics. The instructor sees the signals and knows when to step in, and who needs help.

The next step we are exploring is to detect that moment of struggle and bring in AI at the right time. Not to provide the answer, but to ask the right question.

What Luna does, and what it does not do

Luna, our AI, is not designed to code on the learner’s behalf. It is not a general-purpose assistant. It is a tutor grounded in the notebook, the instructions, and the criteria defined by the instructor.

In a conventional notebook, a general-purpose assistant can explain an error or generate a complete function in seconds. That is useful for moving fast and effective for driving product adoption. But it is often counterproductive for learning: students reach the correct result without necessarily understanding the reasoning, the constraints of the exercise, or the decisions they were expected to make.

The instructor remains the source of pedagogical truth. They define the objectives, validations, expected answers, and assessment criteria. Luna operates within that framework. It first checks what the learner already understands, then offers an explanation or a progressive hint rather than a ready-made answer.

In this example, Luna understands the notebook context and questions the learner before helping them move forward.

Two forms of guidance in the same notebook

Luna beside the notebook

Learners can ask Luna about the notebook, an instruction, a cell, or an error message. In the demonstration above, Luna begins by asking what the learner understands about the event contract. When the learner says they are stuck, Luna explains the concept from the beginning instead of completing the requested function.

Checkpoints built into the learning path

The notebook can also include open-ended questions assessed by Luna against criteria prepared by the instructor. We do not only check whether the program returns the correct result: learners can be asked to explain why they chose a particular method, how they interpret a chart, or what limitations they see in their analysis.

Coding exercises remain deterministically checked through assertions defined by the instructor. AI is used where a simple value comparison is no longer enough: understanding, justification, and critical thinking.

A native Jupyter notebook, not a closed format

The learning content, starter code, and validation information are carried in the metadata of the .ipynb file. The notebook remains readable in editors that support the Jupyter format. In Coding Park, this metadata enables the enhanced experience: integrated validation, checkpoints, and contextual guidance.

Instructors therefore work with a format they already know. They prepare an instructor notebook containing their instructions, solutions, and validations, then generate the version intended for learners. Coding Park adds the execution environment and the learning layer without locking the content into a proprietary format.

The same problem at a different scale

What we have started building for teenagers learning Python, including isolated workspaces, engagement signals, comprehension checkpoints, and an AI tutor bounded by the instructor’s objectives, also addresses the needs of an engineering school, a bootcamp, or a team training its data engineers.

The context and complexity change. The principle remains the same: learners need to practise, make mistakes, and understand, while instructors need the means to intervene at the right time.

AI as a tutor, not a shortcut

The useful debate is no longer whether AI should be allowed or banned. Learners will use it. The real question is the role we want it to play.

Learning means accepting that you will not understand immediately. It means staying with a problem, trying again, and eventually grasping something you could not even articulate before. That friction is at the very heart of the learning process.

A future in which AI solves everything on our behalf is not a future of liberation. It is a future in which we gradually lose the ability to think for ourselves and, with it, our sense of what we are doing.

I do not subscribe to a vision of humanity kept on a universal-income drip, reduced to watching its own tools. Intelligence is not centralised in data centres. To us, it is distributed, decentralised, embodied, and acquired through practice and effort.

That is the vision we want to bring to life with Luna: using AI to preserve the effort through which learning happens, not to remove it.

In Coding Park Academy, Luna adds contextual explanation and feedback to a native Jupyter notebook within the framework set by the instructor. The experience is already available in our free trials. Automatically detecting when a learner is stuck and having Luna intervene proactively is the next step, not a feature we claim to have completed already.

You can watch Luna guide a learner and try the experience in our online Jupyter notebook, or explore the Coding Park Academy teaching modules.

Try the AI-guided Jupyter workspace

Open a native Jupyter notebook in an isolated cloud workspace and see how Luna guides learners without taking over.