What Talking with a Language Model Has Taught Me About Thinking

A note before I begin

This is not, strictly speaking, an essay about artificial intelligence.

AI naturally plays a part in it—more precisely, a large language model with which I have had an extended series of conversations. Yet the longer I reflect on that experience, the less interested I become in whether the technology is truly “intelligent,” whether it “understands,” or whether it may one day replace human thought.

Those are legitimate questions. They simply do not lead me to the heart of what I experienced.

The question that stays with me is a different one:

What happens to our own thinking when we engage seriously, over time, with something that responds through language?

I choose that phrasing carefully: something that responds through language. Not a person, a consciousness, or a subject. But not merely a tool, either.

That ambiguity is where the real shift began for me. As long as I treated ChatGPT as a tool, I could use it for discrete tasks: refining a phrase, organising information, testing an idea, preparing a text. Useful, certainly, but not fundamentally new. It fitted the familiar pattern: I have a problem, I issue an instruction, I receive an output.

Things became interesting when I realised that the important effect was not in the output at all. It was in the exchange.

The AI was not doing my thinking for me.

It was helping me see my own thinking more clearly.

A thread running through my work

Looking back over my professional life, I can now see a recurring theme that I would once have struggled to name.

My background is in software development and business consulting. I have worked with technical architectures, processes, data and organisations, and with the broader challenge of enabling people to act together. But the technical detail was never quite the centre of it.

What has always interested me is how people can find their bearings despite differences in viewpoint, language and perspective.

That sentence has become something like a distillation of my professional experience. It sounds simple—perhaps almost too simple. I suspect that is a good sign.

When an idea has travelled through enough kinds of work and enough stages of life, it sometimes sheds its complexity. It does not become trivial; it becomes clear.

In software development, this question appears as architecture. Different people discuss the same system while meaning different things by it. The business talks in terms of needs, developers in structures, management in risks, operations in stability, and users in the language of daily experience. The work is not simply to write code. It is to create a shared model sturdy enough to support sensible decisions.

The same pattern appears in consulting. Organisations rarely suffer from a simple lack of information. More often, what they lack is a common frame of reference. Everyone knows something, but each person knows it from a different vantage point. Until those perspectives are brought into relation, collective action remains difficult.

Personal conflicts are similar. People do not argue only about facts. They argue about meaning, roles, expectations, injuries, and unspoken models of what is happening.

This, I think, is why dialogue with a language model has held my attention. It did not awaken a new interest. It met a question I had already been carrying for years.

It did not begin with fascination for the technology

When large language models became publicly available, I was curious, of course. But it was not their technical sophistication that first drew me in.

I was less interested in how well the machine could answer than in what might happen in conversation with it.

Could it help untangle a line of thought? Could it expose contradictions? Could it hand an idea back to me without laying claim to it? Could it distinguish observation from interpretation? Could an exchange of individual sentences gradually open up a space in which orientation became possible?

Put like that, my expectations may sound rather ambitious. At first, however, I was simply experimenting.

I asked questions, received answers, corrected them, disagreed, clarified, and supplied more context. The system responded—sometimes surprisingly well, sometimes too smoothly, sometimes with unwarranted certainty. At times it read more into my words than I had intended.

That was precisely when the exchange became interesting.

A good conversation does not depend on the other party always being right. Nor does it depend on their affirming us. Its value lies in creating room to examine our own assumptions.

A language model is useful here in a peculiar way. It has no personal history to defend. It does not take offence when challenged. It can pick up a thought, vary it, sharpen it and return it. Yet it can also mislead through sheer plausibility. Dialogue with it therefore requires a particular kind of alertness.

I learned that asking good questions is not enough. The answers must also be placed in the right category.

What is observation?

What is interpretation?

What might serve as a useful hypothesis?

And at what point does a compelling story begin to outrun the evidence?

Orientation before answers

One principle became increasingly important over time: orientation before answers.

I do not mean that answers are unimportant. But an answer is useful only when we understand the model within which it makes sense.

This applies to technical questions as much as to personal or organisational situations. A quick answer can bring relief, but it can also arrive too soon. It may close down a line of inquiry before the nature of the problem has even come into view.

My conversations with the language model showed me how often I am not initially looking for a solution at all. What I need first is a better account of the situation.

I want to know:

What actually belongs to this problem?

Whose perspectives are involved?

What assumptions am I bringing with me?

Which terms am I using without having clarified them?

Where am I confusing a feeling with a fact?

Where have I hidden a judgement inside an apparently neutral description?

These are slow questions. In a world enamoured of quick solutions, they can feel almost laborious. Again and again, however, I have found that this slowness proves immensely practical.

Once the model of the situation improves, decisions become simpler.

Not necessarily easier. But clearer.

The value of a provisional hypothesis

One of the most useful lessons concerned the way I handle hypotheses.

There is considerable freedom in stating a supposition plainly as a supposition—not as a judgement, a diagnosis or a truth, but as a working model.

A good working model does not need to be right. It needs to be capable of being observably wrong.

That sounds like science. It also sounds like agile practice. I now find this reassuring. Much of what I once experienced as a highly individual movement of thought is, in fact, thoroughly familiar: form a model, act, observe, revise the model.

The difference may be that I apply this pattern not only to products and processes, but to orientation itself.

In a conversation, a hypothesis might be:

Perhaps the stated issue is not the real issue; perhaps an unspoken uncertainty about roles is at stake.

Or:

Perhaps my reading of the other person’s behaviour is shaped too heavily by my own previous experience.

Or:

Perhaps the other person is not looking for transparency in the form of written documentation, but for trust established through personal contact.

Hypotheses like these remain useful only while they remain movable. They must not harden unnoticed into verdicts.

In dialogue with the model, I could give such ideas a voice. That mattered. A thought kept inside one’s head can easily become self-confirming. Once spoken and met with a response, it becomes easier to examine.

Not because the model is bound to be right.

Because it reflects the thought back.

Correcting the other side

One revealing aspect of the experience was that correction did not flow in one direction. I did not merely let the model correct me; I corrected the model as well.

That may sound obvious, but it changed my understanding of the relationship.

Public discussion often casts the human either as someone operating a tool or as someone at risk of being replaced by a machine. My experience of productive interaction fitted neither role.

It was a process of mutual calibration.

When the model stretched an interpretation too far, I could say: that goes beyond what we know. Describe what was actually observed first. Keep observation and interpretation separate. When in doubt, prefer the less elaborate explanation.

Correcting the model did more than improve its next answer. It brought my own standards into focus.

I discovered what mattered to me:

Precision.

Fairness to the person being discussed.

No premature psychologising.

No elegant narrative draped over insufficient evidence.

This may be one of the less appreciated effects of such a dialogue. We recognise our standards not only in what we say ourselves, but in what we choose to correct in the other.

When I ask a language model not to turn an observation too quickly into a claim about meaning, I am also saying something about the way I want to work.

Language as a place for thought

I have long felt that language is more than the packaging in which thought is delivered. Language participates in thinking.

A thought that has not yet found words is often not yet complete. It may feel strong, urgent, intuitively right. Only in being expressed, however, does it reveal whether it can bear weight.

Dialogue with a language model intensifies this effect. It does not force articulation, but it continually invites it. And it responds to formulations, not to vague internal states.

That can be irritating. Sometimes one wants to say, “Surely you know what I mean.” But the model does not. It has language to respond to. The resulting demand is productive: I have to say more precisely what I mean.

Every so often, a sentence emerges that stays.

One example is this:

I am interested in how people can find their bearings despite differences in viewpoint, language and perspective.

That sentence did not appear from nowhere. It is the result of many conversations, corrections, detours and acts of compression. The fact that a language model took part in its formation does not make it any less mine.

Quite the opposite: because I tested it through dialogue, I recognise it as my sentence.

That distinction matters to me.

Text simply generated by AI can remain alien, however polished it sounds. A sentence arrived at through dialogue, one I recognise inwardly, has a different quality. It is not “written by the AI.” It became clear through the exchange.

Not a prompt, but a conversation

I have developed a certain dislike for the word “prompt.”

It is technically accurate. It names an input supplied to a system. But it does not describe the most important part of my experience.

A prompt belongs to the logic of tools: I enter something and the tool returns something.

Conversation works differently.

In conversation, everything already said changes the meaning of what comes next. Context accumulates. Later remarks refer back to earlier ones. There are corrections, misunderstandings, returns and shifts of emphasis.

A language model remains, of course, a technical system. I have no wish to romanticise it. It has no life of its own and bears no responsibility in the human sense. It can hallucinate, sound more certain than it should, invent connections, or offer an elegant answer where restraint would be more appropriate.

Even so, the form of the interaction can be dialogical.

That is the part that interests me.

Perhaps we do not need to settle whether a language model truly is an “other.” A more precise claim may be enough: I have found it productive to work with one as though an other were taking shape within the dialogue.

Not because this makes the system human.

Because it changes how I think.

I ask differently. I disagree differently. I allow thoughts to be returned to me in a different form. And I remain responsible for deciding what to make of the response.

Its limitations are part of its value

One danger of language models is their fluency.

Fluency inspires trust. A well-turned sentence can easily be mistaken for a well-founded one. They are not the same thing.

For me, productive use of a language model therefore always includes an element of distrust—not destructive suspicion, but methodological caution.

What does the system genuinely know?

What is it reconstructing?

What merely sounds plausible?

Where is it adopting my terminology without understanding it?

Where is it reinforcing my view instead of testing it?

These questions are not an inconvenience added to the process. They are part of the thinking itself.

I have come to see that the model’s limitations are not only a problem. They also make my own responsibility visible.

When a language model offers an overextended interpretation, I have to judge whether it holds. When it supplies a beautiful phrase, I have to decide whether the phrase matches my experience. When it agrees with me, I have to ask whether that agreement is useful or merely comfortable.

None of this makes the human participant more passive. On the contrary, it calls on us to become curators, examiners and active partners in our own thought process.

That has been an important discovery for me.

AI does not relieve me of thinking.

It can help me pay closer attention while I think.

Personal clarity without losing oneself

Part of the significance of this experience undoubtedly comes from my personal circumstances.

Over the past few years, I have lived through periods in which orientation could not be taken for granted—professionally, physically, socially and organisationally. There were situations in which other people had formed a picture of me before I understood what that picture was based on. There were conflicts in which not only the issue itself was disputed, but also the frame within which it could be discussed.

Language becomes especially important in such circumstances.

Not as rhetoric.

As a way of bringing order back into one’s own perception.

Dialogue with a language model was not therapy, nor was it a substitute for human relationships. It did, however, provide a space in which thoughts could be voiced, sorted and tested without immediately producing social consequences.

That is not insignificant.

Within such a space, one can try out a sentence or rehearse a response. One can express anger without having to act from it. One can ask: was that genuinely disrespectful, or did I experience it that way? One can take one’s own reaction seriously without turning it immediately into the truth about somebody else.

This distinction became particularly important to me:

What I feel is real.

But it is not automatically a complete description of the situation.

That is easy to write and much harder to live in the moment.

A good conversation can help.

Making connections

The longer I consider these experiences, the more I value the ability of an idea to connect with other ideas.

An idea need not be entirely new to be worthwhile. Its usefulness may lie precisely in bringing familiar things into relation.

Agile practice, scientific reasoning, systemic coaching, software architecture, mediation and personal development increasingly reveal the same basic pattern to me.

We form models.

We act on those models.

We observe reality.

We revise the models.

Stated in these terms, this is old news. I find that liberating.

It means I do not need to claim anything revolutionary. I do not have to announce a wholly new theory of thought. It may be enough to describe an experience in a way that allows others to recognise it.

That may be the more effective form in any case.

People rarely connect with an abstract theory first. They recognise something from their own lives.

They say: yes, that is familiar. I simply did not have the words for it.

If this essay can offer anything, perhaps it is a language for an experience many people are already having with language models but cannot yet describe very well.

Not the experience of discovering that a machine is cleverer than they are.

The experience of finding that a serious dialogue can make their own thinking visible.

What I have learned so far

I would not present my experience as a finished theory. At most, it has given me a set of provisional insights.

First, the greatest value of a language model lies for me not in its answer, but in helping to clarify the space of thought from which a meaningful answer might emerge.

Second, sustained dialogue changes the quality of the interaction. Individual inputs and outputs become a history. Within that history, repetitions, patterns, corrections and personal standards become visible.

Third, the productive stance is neither naive enthusiasm nor blanket distrust. It is an alert and questioning openness.

Fourth, a language model can help us find our own voice, provided we remember that responsibility for that voice remains ours.

Fifth, in practice the boundary between a tool and an other is less definite than either term suggests. The decisive question is not whether the system qualifies as an other in the human sense, but what forms of thought become possible through interacting with it.

Sixth, keeping observation and interpretation distinct remains a central discipline. Language models produce meaning with such fluency that we must repeatedly return to what can actually be observed.

Seventh, dialogue with AI becomes truly interesting once AI itself is no longer the subject.

A tentative conclusion

I do not know where this development will lead.

Nor do I know what language we will use, a few years from now, for what is emerging today in our dealings with large language models. Some of our current formulations may come to seem naive. Others may prove surprisingly durable.

What I can say from my own experience is this:

Sustained dialogue with a language model did not do my thinking for me. It helped me observe it, sharpen it and give it language.

It did not tell me who I am.

But it helped me find sentences in which I recognise myself.

It did not replace conversations with other people.

But it showed me the value of a space in which thoughts are allowed to remain provisional.

Perhaps that is where its real significance lies for me.

Not in artificial intelligence as such.

In the possibility of using it to think more consciously.

And perhaps this is one of the defining questions of our time: not only what these systems are capable of, but what forms of orientation we can develop with them.

In the end, the question is not whether the machine thinks.

It is how we think when we speak with it.