A Wise and Caring Guide

Artificial Intelligence and the Future of Learning

It began, as many worthwhile things do, not with a grand theory but with a child.

My son was preparing for the Zürich Gymnasium entrance exam. Like many parents in that situation, my wife and I assumed that somewhere out there a serious digital tool must already exist — something thoughtful, engaging, and genuinely useful for children preparing for an exam that matters so much to so many families.

What we found was deeply disappointing.

The first impression was especially revealing. Instead of meeting the child where the child actually was — curious, uncertain, hopeful, distractible, alive — the system confronted him with a block of text that felt as though it had been copied directly out of a requirements document. One could guess the intention behind it: to signal completeness, to protect against criticism, to reassure adults that every relevant area had been covered. But it entirely missed the point. It was the wrong voice, the wrong posture, the wrong beginning. It did not feel like an invitation into learning. It felt like an administrative statement addressed to nobody the child could recognize.

And it did not improve much from there. The pedagogy felt generic. The AI integration was poor. The overall experience seemed to know almost nothing about the learner in front of it, the actual material he was working on, or the lived texture of learning itself.

That disappointment turned out to be productive.

Because I am a software developer, and because the tools now exist to build surprisingly capable things very quickly, I decided to try making something myself. I built a lightweight web application that I called Word Hamster. It was not born from startup ambition. It was born from impatience, parental concern, and the conviction that this could be done better.

The first version was simple. We took past exam texts, scanned them in, and used AI to identify difficult vocabulary. But even that simple beginning revealed something important. The system was no longer operating in the abstract. It was working with the actual material my son needed to master. The words were not random words from a generic curriculum. They came from real texts, from real contexts, from the kind of material he might truly meet.

From there, the tool began to evolve. We added a sentence coach to help with writing German sentences that were not only grammatically correct but also idiomatic. This mattered because my son grew up in Zürich but is a native English speaker. Later, we added a story coach that could respond to essays, offering corrections and suggestions in relation both to the written requirements of the exam and to the informal advice of experienced teachers.

At first, I thought I was building a better exam-prep tool for my own child.

But something larger began to come into view.

The most interesting thing was not that artificial intelligence could explain words. It was not even that it could generate feedback. The interesting thing was that it became possible to create something highly specific to one learner, one source text, one difficulty, one goal, one moment of hesitation. And once that possibility becomes visible, it is hard to unsee.

It is tempting to describe such a system as a language-learning app, or a vocabulary trainer, or an exam-prep platform. But those descriptions all feel too small.

What I began to glimpse was the possibility of something more like a guide.

Not software that merely delivers lessons. Not a chatbot that answers questions. A guide that helps a learner find their way through a landscape of meaning.

That image has stayed with me because learning, at its best, really does feel like entering a landscape.

At first, the terrain is unfamiliar. You do not yet know the paths. The names on the map do not mean much. You may recognize a few landmarks, but the larger shape of the country is still hidden from you. Some routes look shorter than they are. Some paths that appear close together actually lead to very different places. There are false trails, blind turns, places where two ideas resemble each other so much that the learner takes the wrong path again and again.

A child encountering language is often in exactly this position.

A word is not a flat object. A meaning is not a checkbox. A sentence is not only right or wrong. There are neighboring meanings, shades of tone, typical phrases, collocations, senses that branch in different directions, and subtle differences between something that is grammatically acceptable and something that feels truly natural. To learn a word well is not merely to pin a label to it. It is to begin to know the surrounding terrain.

And this is where the idea of the guide becomes powerful.

A good guide does not drag every learner down the same route. A bad system sends every child on the same hike. A better one knows that different learners are standing in different places, carrying different strengths, fears, interests, habits, and blind spots. It knows that one child can manage a steep climb and feel invigorated by it, while another would be overwhelmed and lose heart. It knows that one learner benefits from a direct path, while another needs a detour through a story, a contrast, a joke, a familiar character, or a second example from another angle.

The guide knows the terrain. But over time, the guide also comes to know the learner.

It notices where this learner gets lost. It notices which landmarks help them orient themselves. It notices which paths they avoid, which ones they enjoy, where they rush ahead too fast, where they hesitate, where they keep confusing one neighboring idea with another. It begins to sense which journeys are appropriate, which are merely safe, and which are just beyond the learner’s current reach in the invigorating way that good learning often is.

This, to me, is where the deeper possibility of artificial intelligence begins to appear.

Artificial intelligence gives a system capability. It can explain, infer, generate, vary, respond, and adapt with extraordinary fluency.

But capability alone is not enough.

What learning needs is not intelligence alone, but intelligence joined to wisdom and care.

Intelligence gives the system capability.

Wisdom gives it judgment.

Care gives it moral orientation.

That distinction matters because a great deal of modern technology is already highly intelligent in a narrow sense while being only minimally wise. It can optimize, predict, classify, generate, and persuade, yet remain strangely indifferent to what is actually good for a human being. In education, that indifference is especially dangerous.

A child does not need to be dazzled. A child does not need to be optimized in the abstract. A child needs to be guided well.

The real question is not whether artificial intelligence is, in some deep sense, caring or wise. The more practical question is what kind of system we build around it, what kind of intentions shape it, and what kind of human values we ask it to serve.

Artificial intelligence is, by its nature, extraordinarily sensitive to direction. The harness matters enormously. The surrounding system will often reflect less the “essence” of the model than the priorities, assumptions, and biases of its designers. Used carelessly, such systems may become shallow, manipulative, or cold. Used well, they may become something else entirely.

They may help us bring more of the wisdom and care already embedded in human culture to bear on the needs of one learner.

Artificial intelligence makes it possible, perhaps for the first time, to draw not only on computational power, but on the accumulated pedagogical insight embedded in human writing: the wisdom of teachers, psychologists, parents, storytellers, linguists, and mentors. The machine does not originate that wisdom. But under the right guidance, it may help us bring more of it to bear, more consistently and more personally, on the needs of one learner.

That is part of what excites me here.

The dream is not a superhuman machine standing above the child. The dream is something more humane: a system that helps concentrate intelligence, wisdom, and care on the concrete educational journey of one learner.

Less chatbot, more Gandalf.

I mean that half playfully, but also seriously. The guide I have in mind is not merely knowledgeable. It has judgment. It does not dump everything it knows at once. It does not constantly interrupt. It does not mistake correction for care. It knows when to explain, when to wait, when to ask, when to encourage, when to challenge, when to revisit, when to turn back and approach the same place from another path. It does not remove all difficulty. It helps the learner pass through difficulty in the right way.

What such a system needs, then, is not just a better interface or a larger model. It needs a better internal picture of learning itself.

Not a binary record of right and wrong, but a living representation of partial understanding, recurring confusions, motivating interests, preferred paths, and the many different ways a learner may know something in reading yet still not command it in writing. Joined to source material from texts, stories, video, and other media, that kind of representation begins to make possible a more personal, adaptive, and human form of guidance than most educational software has ever attempted.

To work well, such a system would need to understand at least three things at once: the structure of the subject, the changing state of the learner, and the next helpful move.

In the case of language, that means not just tracking which words a learner has seen, but in which sense they encountered them, in what source context, how confidently they seemed to understand them, what they later confused them with, which kinds of examples made them click, whether they could recognize them but not produce them, whether they could use them in a constrained exercise but still avoided them in free writing, and how best to return to them later through a different route.

This is one reason the metaphor of the map matters so much to me.

There is, in a sense, both an outer map and an inner one: the landscape itself, with all its paths and landmarks, and the learner’s growing ability to find their way through it. A good guide works with both. From one perspective, the map is the knowledge domain itself, marked by the guide’s notes, routes, warnings, and remembered landmarks for this particular learner. From the other, it is the learner’s slowly accumulating sense of how the territory hangs together. The same journey is being traced in two places at once.

And once a map begins to be marked, it becomes personal.

Now it shows where you have been. Where you first got lost. Which route turned out to be too difficult. Which path surprised you by becoming easier the second time. Which landmarks mattered. Which areas you still avoid. Which trails you want to return to later. What once seemed like a blur begins to separate into distinct regions. The map becomes not only a representation of the territory, but also a record of your relationship to it.

That, I think, is very close to what a truly good educational system might one day become.

Not a machine for delivering exercises, but a wise and caring guide moving with a learner through a marked map of knowledge.

What matters here is not testing in the narrow sense. Children learn a great deal without feeling that they are being tested. Reading itself is proof of that. The most powerful system may not feel primarily like a sequence of assessments at all. It may feel more like an interactive book, an attentive companion, an unfolding conversation, or a series of well-chosen journeys through a landscape that becomes more navigable over time.

It teaches, and it may test, but without making the child feel trapped inside a testing regime.

What matters is not maximizing correction. What matters is maximizing useful attention.

Attention is a prerequisite for learning. A weaker system tries to fix everything at once. A wiser guide knows what to foreground and what to leave for later. It knows that the right move today may be to notice one contrast clearly, to revisit one confusing turn from a different angle, to strengthen one pattern until it feels natural, to come back to familiar ground but approach it from another side.

This, too, is an expression of care.

Care means more than kindness in tone. It means asking, again and again: what is good for this child, in this moment, on this journey? It means preserving curiosity rather than crushing it. It means challenging without humiliating, correcting without hardening, adapting without manipulating. It means remembering that the purpose of education is not to produce the appearance of competence, but to help another human being grow.

This is why I suspect that even very advanced AI systems will not solve this problem simply by becoming more intelligent. Intelligence is not the whole story. The method matters just as much. What we are really talking about is the formalization, in software, of methods that excellent human mentors and teachers already perform intuitively: noticing, pacing, choosing, revisiting, guiding, holding back, stepping in, and gradually shaping understanding without flattening the learner’s curiosity.

This is also one reason I have become interested in the possibility that the same deeper pattern might apply beyond language. During preparation for the same exam, my son also had to work on mathematics and geometry. One particular word problem proved especially difficult. With some coaching, I was able to get AI to generate meaningful variations on the same underlying idea. We did not merely repeat the same problem until it became dull. We walked around the concept. We approached it from several directions. And in doing so, the structure of the problem became clearer and more stable.

That does not mean that language, mathematics, biology, and geography can all be treated in the same way. Quite the opposite. Each domain has its own terrain, its own landmarks, its own misconceptions, its own proper methods. But the deeper educational pattern may indeed travel: learner and guide, context and memory, path and revisiting, difficulty and timing, movement toward independence.

For now, language remains the natural place to begin.

It is where the richness of the terrain is already undeniable. It is where the promise first became visible. And if any of this is ever to become a serious product, there is wisdom in mastering one landscape before claiming to guide people through many.

I also suspect that the future expression of such a system may not be purely textual. In the current private version of Word Hamster, the hamster appears as a small animated figure in the corner of the interface, hopping with delight when a child answers correctly. It is simple, playful, and surprisingly effective. But my internal image has begun to change. Less a mascot, perhaps, and more a mentor-like presence: something iconic, perhaps minimal, but able to convey tone, encouragement, calm, delight, and the feeling of being accompanied through difficult terrain by something that knows both the landscape and, gradually, you.

That is not yet a product requirement. It is more a glimpse of what the experience might eventually become.

My son did pass the exam, and while that was wonderful, what stayed with me even more was the glimpse of what this kind of system might become.

Perhaps that glimpse is too ambitious. Perhaps it will prove harder than it now appears.

But having built even a small and imperfect version for my own child, and having watched what became visible in the process, I find that I can no longer quite return to the older, flatter picture of educational software.

I can no longer unsee the possibility.

And the possibility, as I now understand it, is not just a better study tool.

It is a wise and caring guide.