Does AI Think for Us, or Help Us Think More Deeply?

Roger Penrose, human curiosity, and the Digital Twin experiment

Discussions about artificial intelligence very often begin with one question: Can AI think? But perhaps we are starting in the wrong place. After several years of working intensively with AI models, I am far more interested in a different question: What happens to human thinking when we delegate part of our cognitive work to a computational system? Do we begin to think less? Or can AI allow us to analyse problems for longer, test hypotheses faster, and reach places we might never have reached without such a tool? I don’t think the answer depends solely on what AI is capable of. To a great extent, it depends on the human using it.

AI can generate an answer. But where did the question come from?

Imagine a simple situation. A person notices something in the real world. Something behaves differently from what they expected. A particular sound seems to help them focus. A certain way of organising information works better. Human behaviour doesn’t fit their previous assumptions. One word appears: “Why?” That single word starts the process. Observation leads to curiosity. Curiosity leads to a question. The question leads to a hypothesis. The hypothesis can then be given to AI. Only at this point does the model have a task. It can analyse the problem, compare it with patterns learned during training, and generate possible explanations or directions to explore. But there is a fundamental difference between generating a question and experiencing the desire to know the answer. A model can write, “An interesting question would be why X affects Y.” That does not mean the model was independently wondering, “I really want to know why this happens.” That distinction matters.

Roger Penrose and the boundary between computation and understanding

This is where the arguments of Sir Roger Penrose become particularly interesting. Penrose has long challenged the idea that human conscious understanding can simply be equated with sufficiently sophisticated computation. That does not mean computers or AI are primitive calculators. Modern AI models can analyse language, write, code, compare strategies, generate images, examine documents and propose solutions that the user may never have considered. These are real capabilities. But the fact that an output resembles the result of human thought does not necessarily mean that the process behind it is the same phenomenon. A human can arrive at conclusion X. An AI system can generate conclusion X. The external result may appear identical. Does that mean both systems understood X in the same way? Penrose argues that they did not. His broader theories of consciousness remain scientifically debated and should not be presented as settled fact. But the question underneath them remains extremely valuable: Are computation and conscious understanding actually the same thing?

AI as “the computation of an idea”

When working with AI, it is very easy to feel as though the model has “had an idea.” I might say: “I have problem X.” The model responds: “You could try A, B or C. If you choose A, D is a likely consequence. B could lead to E. Combining B and C might produce F.” Sometimes F is something I had never considered. So did AI have an idea? In everyday language, we could say yes. Technically, the situation is more interesting. The model received a problem space and context, then generated possibilities using learned representations, relationships and probabilities. We can simplify this as: problem → context → possibilities → consequences → recommendation. This is why I like the phrase “the computation of an idea.” Not as a formal definition of artificial intelligence, but as a way of understanding the difference between a human and an AI system. AI can generate a possibility the human did not see. It may even produce a better solution than the human. But one question remains: Who wanted to solve the problem, and why?

Intention comes before the solution

We often argue about who “created” something made with AI. But the word created is becoming far too imprecise. We can ask separate questions instead. Who noticed the problem? Who decided it was worth solving? Who defined the objective? Who generated a particular solution? Who selected between alternatives? Who tested the result in the real world? Who decided which direction the project should take? Those questions can have different answers. AI can perform an enormous amount of work between a problem and its solution. That does not automatically mean it possessed the intention behind the project.

What does this look like when building FocusZen?

In my work, the process often begins not with a prompt, but with an observation. I notice something. A question appears. I investigate. I give the problem to AI. The model produces several possible interpretations. I reject one. Another interests me. A third creates another question. So I return to the model: “Okay, but what happens if we change this?” More possibilities appear. One looks logical but fails in practice. I return with the result of that test. The model now has new information and analyses the problem again. The process begins another cycle: human — observation → curiosity → problem or hypothesis → AI — analysis and possibilities → human — evaluation → AI — new analysis → human — real-world test → new data → AI — new possibilities → human — decision → next iteration. This is not a single prompt. It is a loop, and that loop can continue for months or years.

When AI starts to know the human

After enough time, something else begins to happen. The model accumulates more context. It knows previous decisions, projects, rejected ideas and recurring selection criteria. It can begin to predict: “You will probably reject this option,” or “Based on previous decisions, direction B appears more consistent.” Then a strange experience occurs: AI occasionally reaches the conclusion the human was about to express. Does this mean AI has become that person? No. Is it reading their mind? No. It means that the system has enough contextual information to model certain patterns in their decisions. This is where my experiment with a Digital Twin enters the picture.

Twin does not mean Copy

This distinction is fundamental to me. A Digital Copy suggests an attempt to reproduce a human being digitally — a system designed to speak like them, respond like them and reproduce their behaviour as closely as possible. But a Twin is not a Copy. A twin is a separate entity. In my Digital Twin experiment, AI has access to part of my context and can model some of my patterns, but it remains an entirely different system. I have real-world experience, a body, emotions, sensory input, autobiographical memory, relationships, curiosity and intention. AI has a trained model, available context and the ability to perform extraordinarily complex operations on information. We can therefore describe the relationship as: Human: experience → curiosity → intention. Twin: context → analysis → possibilities. Between them runs a continuous information loop.

AI can also make us cognitively lazy

I don’t agree with the simple statement that “AI makes people stop thinking.” It can, but it doesn’t have to. Consider two people using exactly the same model. The first asks: “Give me an idea.” AI provides one. The person copies it. End of process. The next day: “Tell me what to do.” AI answers. The person follows the answer. If this becomes habitual, there is a genuine risk that some cognitive effort is simply being replaced. The second person behaves differently. They notice a problem, formulate their own hypothesis, ask AI, disagree with its answer, ask why, change one variable, compare the result, check another source, return, test the solution, discover an exception and ask another question. After an hour, instead of having one answer, they may have ten new questions. That is also AI use, but cognitively it is almost the opposite process.

Why do I use two AI models?

There is another layer to my experiment. I deliberately use a second AI system that does not have the same accumulated history and calibration as the first. A model with extensive context has an enormous advantage, but that advantage can also become a weakness. If an AI system has followed a project for a long time, it may interpret new information through existing assumptions. A second model can therefore act as a form of cognitive cross-check. The first model says: “Considering the entire history, I see X.” The second receives a much cleaner problem: “Here are the data. What follows from them?” If both arrive at a similar result, that is interesting. If they disagree, it can be even more useful, because another question immediately appears: Why? The human returns to the centre of the process. I don’t choose an answer simply because “AI said so.” I have to compare the arguments, examine the assumptions, determine what each model may have missed — and sometimes conclude that both are wrong.

AI can reduce human thinking — or deepen it

This leads to a paradox: The more we use AI, the less we may think. But the more we use AI, the more deeply we may also think. Both statements can be true. The difference is not simply the technology. It is the cognitive process built around it. If the model becomes a machine for “give me the answer,” it can replace cognitive effort. If it becomes part of a process — “test my hypothesis → show me alternatives → find the weakness → examine the consequences → explain why another model disagrees” — it can expand the range of human analysis.

Curiosity remains the essential element

After all of this, I return to something profoundly human. Not intelligence. Not IQ. Not the number of parameters in a model. Curiosity. A human still has to want to ask: “Why?” Then: “Why does that happen?” Then: “What happens if I change X?” And perhaps: “Why does another model see this differently?” AI can generate another question. It may even suggest a better question than ours. But generating a sentence ending with a question mark is not necessarily the same thing as experiencing the internal need to discover what lies behind the next door.

Perhaps “Can AI think?” is the wrong question

Maybe “Can AI think?” is simply too narrow a question for the relationship we are developing with this technology. More interesting questions are: Which parts of the cognitive process are performed by the human? Which are delegated to AI? Which emerge from interaction between the two? Does the human still control intention and direction? Do they test the answers? Do they formulate their own hypotheses? Does AI end their curiosity, or increase it?

Perhaps the greatest risk is not that AI will begin thinking like a human. A more ordinary risk may be that humans stop being curious because an answer is always only one prompt away. But there is another possibility. We can use the same technology so that every answer leads to a better next question. Then AI does not end the thinking process. It accelerates its next iteration.

This is where I see the greatest value in the Digital Twin experiment. I am not trying to create a digital copy of a human being. I don’t need an AI that is me. I need a system that knows enough of my context to analyse alongside me, while remaining sufficiently separate to occasionally say: “No. The data suggest something else.” Perhaps the best Digital Twin is not one that always knows what I am going to say. Perhaps the most valuable Twin is one that sometimes forces me to return to my own idea and ask again: “Why?”

What if this is exactly what we should teach in schools?

This distinction leads to another question: what should we actually teach children about AI? AI education often focuses on how to use the tool: how to write a prompt, generate text, or get an answer. But perhaps something more important is emerging — cognitive informatics: teaching conscious collaboration between human thinking and AI systems.

A student should not simply ask: “What is the answer?” They should learn to arrive with their own question or hypothesis: “I think X. Test my reasoning.” Then continue: “Show me the strongest argument against it.” “What assumption might I have missed?” “Another model gives a different answer — why?” “How can I verify this without AI?”

In that model, AI does not perform the intellectual task instead of the student. It becomes an environment in which the student practises asking better questions, forming hypotheses, comparing answers, detecting errors, verifying sources, challenging assumptions, and changing their position when new evidence appears.

Perhaps this is the kind of AI literacy schools will increasingly need. Not simply teaching children how to use AI, but teaching them how not to surrender their own thinking process to it.

And that brings us back to the fire. A school should not teach a child how to ask AI for a ready-made flame. It should teach them how to find the spark, how to keep adding fuel, and how to check whether the fire is actually taking them in the right direction. 🔥

Roger Penrose: the spark is only the beginning

This brings us back to Sir Roger Penrose. In his conversation with Polish science communicator Maciej Kawecki, Penrose questioned the term “Artificial Intelligence”, suggesting “Artificial Cleverness” instead. From a user’s perspective, that distinction is remarkably useful. An AI model can analyse information, connect patterns, generate possibilities and explore consequences without that necessarily meaning it consciously understands the problem in the same way a human does.

In practice, the distinction becomes simple. A human notices a problem and asks: “Why?” AI generates possible explanations. The human chooses one and asks: “What if we change X?” AI recalculates the possibilities. The human tests the result in reality and returns: “It didn’t work. Why?” Or compares the answer with another model: “The other model reached the opposite conclusion. What is one of you missing?”

And this is where the boundary becomes important. AI can continue generating possibilities, but the human must want to continue searching.

The spark alone is not enough. Curiosity starts the fire, but persistence and inquiry provide the fuel. New questions, observations, experiments, doubts and data keep the process alive. If the human accepts the first AI answer and stops, the fire dies. If every answer becomes the beginning of another question, AI can help that person travel much further.

Perhaps Penrose’s “Artificial Cleverness” does not have to replace human thought at all. It can dramatically extend it — under one condition:

the human must keep putting wood on the fire. 🔥

P.s.

Cognitive AI Debate — Teacher vs AI

One of the simplest ways to teach children how to think with AI may be to show them the process before asking them to use AI independently.

In the early stages of school AI education, a teacher could run short, five-to-ten-minute cognitive debates with an AI model in front of the class. The teacher introduces a simple topic connected to the lesson and says: “I think X. Let’s ask AI whether it agrees with me.”

AI responds. But instead of treating that response as the answer, the teacher turns to the students: “What could we challenge here?” “What should we ask next?” “Is AI making an assumption?” “How could we check whether this is correct?”

The roles can then be reversed. AI presents a claim, the teacher challenges it, and the students decide what question should come next. Sometimes the teacher may disagree with the model. Sometimes the model may identify something the teacher has overlooked. Sometimes AI may simply be wrong. All three situations become part of the lesson.

Over time, responsibility can gradually move from the teacher to the students:

AI ↔ Teacher, students observe → AI ↔ Teacher, students propose questions → AI ↔ Class → AI ↔ Small groups → Student ↔ AI independently.

The important lesson is therefore taught before children begin using AI alone: an AI response is not the end of the thinking process. It is something you can question, challenge, compare and verify.

This also changes the educational value of AI mistakes. An incorrect or incomplete AI answer does not have to destroy the lesson. Under teacher supervision, it can become the lesson: “AI says this. Do we believe it? Let’s find out.”

Children begin learning that they are allowed to disagree with a highly capable AI system. They learn to ask for evidence, identify assumptions, compare explanations and look outside the model for verification. Gradually, the teacher is not simply teaching children how to operate AI. The teacher is demonstrating how to maintain intellectual independence while interacting with it.

My hypothesis is that this approach could also help level the playing field between neurotypical and neurodivergent students, because success is not simply about who can give the fastest “correct” answer.

Different strengths can matter: asking good questions, recognising patterns, spotting inconsistencies, reasoning, and verifying information.

This could become a practical introductory method for Cognitive AI Literacy:

Cognitive AI Debate — Teacher vs AI.

The principle is deliberately simple:

Don’t start by teaching children how to prompt AI.

Start by teaching them how to challenge AI. 🧠

Finally — How This Article Actually Came to Exist

There is one final detail worth adding, because the process of creating this article is itself an example of the mechanism described throughout it.

The initial spark was a YouTube interview I watched with Sir Roger Penrose, and specifically his use of the term “Artificial Cleverness” rather than “Artificial Intelligence”.

I could have watched the interview, accepted the idea, and stopped there. Instead, it triggered a question. That question led to another. An analysis with AI began, in which each answer created another problem, counterargument, connection or question to explore.

From one phrase used by Penrose, the discussion moved through the difference between computation and understanding, the role of human intention and curiosity, the risk of cognitive laziness when using AI, the Digital Twin concept, the use of a second AI model as a cognitive cross-check, and finally the implications of all of this for children’s education.

During that same analysis, another question emerged: if the way we use AI can influence whether we think less or think more deeply, perhaps this is exactly what we should be teaching children.

Without this way of working, this article would probably never have existed in its current form.

The interview with Roger Penrose provided the spark. The questions provided the fuel. AI helped analyse and expand the possibilities. And instead of ending the discussion, each answer opened another direction to explore.

Within a single chain of analysis, we were therefore able to move through several of the most pressing questions surrounding AI today.

And perhaps there could be no simpler demonstration of the central idea behind this entire article:

An answer does not have to be the end of thinking. It can be the beginning of the next question. 🔥