The Cognitive Debt of AI: Why I'm Restricting My Kids' Access to 'Smart' Assistants
An engineering-based perspective on why AI in education often creates 'cognitive debt' and how to use prompt engineering to force AI to act as a Socratic tutor rather than an answer-generator.
The best first AI workflow is usually a repeated task with a clear input, clear output, and a human approval step.
Choose one weekly process, draft the output with AI, and track what a human changes before approval.
Keep human judgment in the loop for production code, customer-facing content, and anything that represents the portfolio voice.
The Cognitive Debt of AI: Why I'm Restricting My Kids' Access to 'Smart' Assistants
As a software engineer, I live by the principle of abstraction. If I can use a library to handle authentication or a managed service to handle database scaling, I do it. It saves time and allows me to focus on the business logic that actually differentiates my product.
But there is a dangerous crossover when we apply this engineering mindset to child development. In software, we call it technical debt. In education, it’s cognitive debt.
The Problem: From Tool to Oracle
When a student uses an LLM to "solve" a math problem or write an essay, they aren't using a tool—they are using an oracle.
In engineering, a tool provides leverage for a task you already understand how to perform. An oracle, however, provides the result without requiring the underlying logic. When a child uses AI to bypass the struggle of learning, they are essentially running a black-box script. The output is there, but the "source code"—the mental model required to solve the problem—is never compiled in their brain.
What I Noticed
I see a pattern in how younger students interact with AI. It’s a closed loop: Prompt -> Answer -> Copy/Paste -> Done.
According to the American Psychological Association, the risk isn't just that they get the wrong answer; it's that they bypass the "desirable difficulty" required for long-term memory formation. If the brain doesn't struggle with the syntax of a sentence or the steps of an equation, it doesn't build the neural pathways necessary to master those concepts. We are effectively training kids to be prompt engineers before they’ve learned to be thinkers.
Where AI Fits
AI shouldn't be an answer-generator for students; it should be a Socratic tutor.
If we want to use these tools in a way that respects cognitive development, we need to shift the interaction model from Product-Oriented (give me the solution) to Process-Oriented (help me understand the steps). AI is excellent at scaffolding, but only if the constraints are set correctly.
My Working Approach
I treat my children's access to AI like I treat a production environment: no root access.
I don't give them a raw ChatGPT interface. Instead, I use local, constrained environments (like Ollama) that act as a "tutor-only" wrapper. If the AI is allowed to provide the answer, it will. Therefore, I build the friction directly into the system instructions.
Practical Example: The Socratic Guardrail
If you want to turn an LLM from an oracle into a tutor, you have to force it to refuse the easy path. Here is the system prompt I use:
"You are a Socratic tutor for a student. Your goal is to help the user learn, not to provide answers.
- If the user asks for a solution, refuse to provide it.
- Ask the user to explain their current understanding or their current attempt at the problem.
- Provide hints that lead them to the answer, but never the answer itself.
- If the user is stuck, ask them to identify which specific part of the process is confusing."
What I Would Avoid
Don't fall into the trap of "AI is evil" rhetoric. It’s not about banning the technology; it’s about understanding the abstraction layer.
Avoid letting children use AI for "drafting" or "ideation" until they have proven they can do the work manually. If they can’t write a paragraph without an LLM, they don't need an LLM to "help" them write—they need to learn the fundamentals of composition. Using AI to skip the foundational phase is how you create permanent mental atrophy.
Try This Next
If you want to manage this at home, move the interaction off the cloud-based web interfaces.
- Use Local Models: Run a model like Llama 3 or Mistral locally using Ollama. This removes the "search-and-copy" temptation because the model isn't connected to the live web.
- Audit the History: Review their chat logs. If you see "Can you write this essay?", you know the system prompt failed.
- Build Friction: Force the AI to require a 3-step explanation from the child before it provides any feedback.
We don't need to fear AI, but we must protect the struggle. The struggle is the learning.
