In learning experience design, we often pursue efficiency.
If a tool speeds up a process, reduces friction, or removes an obstacle, our first reaction is usually positive. And that makes sense. In L&D, we often work with limited time, limited resources, multiple audiences, and constant pressure to deliver better content faster.
But after several months of integrating language models, agents, and AI-assisted workflows into learning processes, I started noticing something that concerns me.
Sometimes, the path of least resistance is also the path where learning stops.
AI can help us think better, write better, analyze better, and design better. But it can also do something much more silent: it can complete the process for us before we have the chance to develop the skill that process was supposed to build.
I call this phenomenon the Zone of No Development.
I am not presenting it as a closed academic concept, but as a design lens. A simple way to detect when a tool stops working as support and starts working as a replacement.
The Zone of No Development appears when a technology reduces cognitive effort so much that the person completes the task, but does not develop the capability the task was meant to strengthen.
The result may be correct. The document may look good. The answer may sound solid. But the person has not necessarily learned. They only reached the end with the help of a machine that did too much of the journey for them.
And that is the risk: not failing, but moving forward without growing.
From the Zone of Proximal Development to the Zone of No Development
To explain this idea, it helps to return to a classic concept: Lev Vygotsky’s Zone of Proximal Development.
The ZPD describes the distance between what a person can do on their own and what they can achieve with help. In education, that help is often understood as scaffolding: guidance, a question, a hint, a structure, or an intervention that allows the learner to go further without removing their active role in the learning process.
For years, many educational technologies were designed around that logic. The tool helped learners organize, practice, visualize, remember, or access information. The goal was to expand the learner’s capacity, not replace it.
But generative AI changes the scale of the problem.
A language model does not only offer a hint. It can write the full answer. It does not only help organize ideas. It can produce the entire analysis. It does not only suggest a structure. It can draft the conclusion, solve the dilemma, build the presentation, and justify the decisions.
This is not necessarily bad. In fact, it can be extremely useful in many professional contexts. The problem appears when that capability enters a learning experience without clear limits.
That is when we cross an invisible boundary.
We leave a zone where support helps us move forward and enter a zone where support removes development.
That is the Zone of No Development.
It does not happen because AI is bad. It happens because AI is too good at solving things the learner still needed to process.
The danger is not using AI. The danger is delegating thought.
I want to be clear: I am not writing this from a place of rejecting AI.
Quite the opposite.
I use LLMs every day. I work with AI-assisted workflows. I experiment with agents. I maintain my own homelab. I design knowledge systems, test local models, and constantly think about how to integrate these tools into real learning, productivity, and work processes.
That is exactly why the Zone of No Development concerns me.
Because the more powerful the tool becomes, the more important it becomes to decide which part of the process must remain human.
AI can be an extraordinary ally when it helps us explore, compare, organize, practice, review, or improve. But it becomes dangerous for learning when it turns the user into a simple approver of outputs.
That shift may seem small, but it is enormous.
One thing is using AI to think more clearly. Another thing is using AI to avoid thinking.
One thing is asking a model to help us compare arguments. Another thing is asking it to make the decision for us.
One thing is using AI to receive feedback on our own conclusion. Another thing is asking it to write the conclusion before we have built a position.
The difference is not only in the final result. It is in who did the most important cognitive work.
And in learning, that matters a lot.
Productivity without process
One of the major problems with AI in education and training is that it can create an illusion of competence.
The person submits something correct, clear, and well structured. From the outside, it looks like they learned. But when we ask them to explain the reasoning, justify a decision, or transfer that skill to a new context, the cracks begin to show.
The product is there. The process is not.
This can happen in a university, in corporate training, in an internal course, or in any activity where AI can complete the task without requiring real participation from the user.
A person can finish an analysis without having analyzed.
They can submit a reflection without having reflected.
They can produce a strategy without having made strategic decisions.
They can complete a learning activity without having learned much.
That is the heart of the Zone of No Development: an experience where the task moves forward, but the person does not.
And for those of us working in L&D, this is a critical problem. Because our job is not simply to get people to complete activities. Our job is to design the conditions for people to develop judgment, capability, confidence, autonomy, and transfer.
If AI improves delivery but weakens development, we are not designing better learning. We are designing a faster version of compliance.
Friction can also be part of the design
In many digital environments, friction is treated as the enemy.
Fewer clicks. Fewer steps. Less effort. Less time. Fewer obstacles.
And in many cases, that is correct. Nobody wants a confusing platform, unnecessarily complex navigation, or an experience that wastes time.
But in learning, not all friction is bad.
Sometimes, a certain level of difficulty is exactly what allows learning to happen.
Robert Bjork works with the concept of desirable difficulties: challenges that may make learning slower in the moment, but deeper and more durable in the long term. This is not about making things complicated for the sake of it. It is about designing forms of effort that force people to retrieve, connect, reorganize, and apply knowledge.
Manu Kapur, through the idea of productive failure, shows something similar from another angle: facing a problem first, even imperfectly, can better prepare someone to understand a later solution.
In other words: not every mistake is a waste of time. Not every doubt is inefficiency. Not every difficulty should be removed.
Sometimes, difficulty is the space where learning is built.
The Zone of No Development appears when AI removes that difficulty before it has served its purpose.
When it answers too quickly.
When it solves too completely.
When it reduces effort at the exact point where the learner needed to practice.
When it turns a learning activity into a transaction of results.
That is why the challenge is not to make AI less useful. The challenge is to design when it should help, how it should help, and when it should stop.
A simple example: the conclusion AI should not write
Imagine an activity where a student must analyze a case, compare evidence, and build their own conclusion.
The student opens a chatbot and writes:
“Write the conclusion for me.”
An unrestricted model will probably do it. It may even produce a conclusion that is clear, coherent, and convincing.
But from a learning perspective, that answer may be a problem.
Because the conclusion was not just a final product. It was the place where the student had to demonstrate that they could connect ideas, prioritize evidence, and take a position.
In that case, a better response would not be to provide the complete conclusion, but to redirect the request:
“I will not write the conclusion for you, but I can help you build it. First, let’s identify your three main pieces of evidence. Then, we can look at the pattern between them and decide what position you can defend.”
That small refusal is not a technical limitation.
It is pedagogical design.
The tool is still helping, but it is not taking over the moment where development was supposed to happen.
And that is where Mini Brains come in.
Mini Brains: a simple answer to a complex problem
Mini Brains come from a fairly simple idea: if we are going to work with language models, we need a clear way to define how we want them to behave in a specific situation.
We do not always need to create an application, train a model, or build complex infrastructure. Sometimes we need something much more accessible: a well-designed unit of context.
That is what a Mini Brain is.
A structured Markdown file that contains the identity, purpose, rules, boundaries, reference knowledge, and interaction patterns we want a model to follow.
I like to think of it as a video game cartridge.
You load it into the model and it says: this is the world, these are the rules, this is your role, this is what you can do, and this is what you cannot do.
It is not “creating an AI” from scratch. It is guiding the interaction with an existing AI.
That difference matters.
Mini Brains are a low-tech solution to a high-tech challenge. They do not require someone to be a software engineer. They do not depend on a specific platform. They do not force us to build a closed product. They can be used in a local model, in a free version of an LLM, or in a corporate environment, as long as the system allows context to be loaded or pasted.
Their value is not in technical complexity. Their value is in instructional clarity.
And to avoid the Zone of No Development, that clarity is essential.
How a Mini Brain can prevent the Zone of No Development
A Mini Brain can work as a kind of pedagogical firewall.
Not because it blocks the use of AI, but because it defines what type of help is valid and what type of help starts replacing learning.
For example, it can establish that AI may explain concepts, ask Socratic questions, help compare evidence, provide feedback on a student’s answer, suggest improvements, identify contradictions, help plan the work, offer partial examples, or guide reflection.
But it can also establish that AI should not write the final answer for the student, invent a position the student has not built, solve an entire assessment activity, replace human judgment, skip reasoning steps, or deliver closed conclusions when the objective is for the person to think.
This is not a minor detail.
In a regular chatbot, behavior often depends on how the user phrases the request. In a Mini Brain, behavior is governed by a prior structure.
That allows AI to help without taking control.
And that is the key: AI does not disappear. It becomes more intentional.
Portability: pedagogical design should not depend on a platform
One of the aspects I find most interesting about Mini Brains is their portability.
In education and L&D, we often get trapped inside platforms. Each tool has its own logic, permissions, limitations, costs, and product changes. What works in one environment today may not work in another tomorrow.
Mini Brains suggest a different logic.
The pedagogical design lives in a simple, readable, editable file.
That means the value is not locked inside an interface. It lives in the structure: in the rules, the context, the instructions, the boundaries, and the curated knowledge.
A teacher can adapt a Mini Brain for an activity.
An L&D team can use it to guide a simulation.
A student can load it to practice without letting AI do all the work.
A professional can use it to maintain consistency in a complex task.
That portability matters because it gives part of the control back to the designer, the teacher, or the user.
Instead of depending completely on how a platform decides AI should behave, we can define our own layer of intention.
And in a world where AI changes all the time, that layer of intention becomes a form of continuity.
Mini Brains as an architecture of resistance and collaboration
When I talk about resistance, I do not mean resisting AI.
I mean resisting passivity.
A well-designed Mini Brain does not try to block productivity. It tries to preserve the space where the human still has to participate.
To do that, it can work through three simple layers.
The first is request management. The Mini Brain does not respond only to what the user wants, but to what the activity allows. If the user tries to delegate critical thinking, the system can redirect them toward a more appropriate form of support.
The second is context engineering. Instead of giving the model undefined access to everything, the Mini Brain works with a curated set of information. This helps keep the interaction aligned with the pedagogical objective.
The third is what I often call AI Nutrition. It is not about feeding the model more information, but better information. Analytical frameworks, relevant data, criteria, examples, constraints, and clear rules. Not “empty calories” in the form of ready-to-copy answers, but nutrients that help people think better.
These three layers enable something that, for me, is central: a real human-in-the-loop.
Not as a decorative phrase.
Not as a checkbox at the end.
But as a structure where the person continues to play an active role throughout the process.
The new role of L&D: designing intelligent boundaries
From this perspective, the value of an L&D professional is not only in how quickly they can automate content.
It is in their ability to detect which parts of the process can be assisted by AI and which parts still need to be practiced by the person.
That shift matters.
Designing learning today is not only about designing access to content. It is also not simply about turning materials into faster, shorter, or more automated courses.
Designing learning today means deciding, with strong intention, which moments need support and which moments need effort.
Where it makes sense to give a hint.
Where it makes sense to ask a question.
Where it makes sense to offer feedback.
Where it makes sense to show an example.
And where it makes sense for AI to stop.
Because if we automate everything, we may also automate the part where development was supposed to happen.
That is the risk of total efficiency.
We can create experiences that are faster, cleaner, and more fluid, but less formative.
That is why I believe L&D needs to take on a new role: not only facilitating learning, but designing productive friction.
Not useless friction. Not bureaucracy. Not unnecessary steps.
Friction with purpose.
The kind of friction that forces people to think, decide, justify, review, and improve.
The kind of friction that prevents the user from becoming a spectator of a machine-generated answer.
Automation without design vs. Mini Brains against the ZND
| Feature | Automation without design | Mini Brains against the ZND |
|---|---|---|
| Role of AI | Solves for the user | Supports the process |
| Focus | Finishing the task | Developing judgment |
| Control | Depends on the platform | Lives in the instructional design |
| Interaction | Total fluidity | Productive friction |
| Outcome | Correct product | Defensible learning |
| Human role | Approve or copy | Think, decide, and justify |
The difference is not whether we use AI or not.
The difference is how we design the relationship between the person, the task, and the tool.
This is not about doing less with AI. It is about thinking better with AI.
The conversation about AI in education often gets trapped between two extremes.
On one side, there are those who want to ban it or treat it as an external threat.
On the other side, there are those who want to automate as much as possible because efficiency seems to be the only criterion.
I think we need a more mature conversation.
AI is already part of the learning and work environment. It makes no sense to design as if it does not exist. But it also makes no sense to give it every important moment in the process.
The question should not be: “Should we use AI or not?”
The question should be: what kind of human participation do we want to preserve?
If a person uses AI and ends up thinking better, excellent.
If they use AI and understand more deeply, excellent.
If they use AI to practice, receive feedback, explore paths, compare perspectives, and improve their work, excellent.
But if they use AI to skip all the effort the activity was designed to provoke, then we are not looking at better learning. We are looking at a more efficient but less formative experience.
That is where the Zone of No Development appears.
The Zone of No Development does not appear when we use AI.
It appears when we stop participating in the process.
That is why the answer is not to turn off the technology or return to previous methods. The answer is to design better boundaries, better contexts, and better forms of collaboration.
Mini Brains are my proposal for that: a simple, portable, and open way to remind AI that it is not there to replace human thinking, but to help it grow.
Because true success is not that AI produces a perfect answer.
True success is that, when AI finishes its part, we are still able to think better than before.
