Column

Puppies and cognition?

AI is like a puppy that never gets tired. When is work finished if you can keep going forever? And what does “judgement” actually mean in the age of AI?

Black-and-white portrait of Petri Lattu: a bearded man looking upward with a finger at his temple, beaded bracelets on his wrist.

We have all spent time with a puppy or a four-year-old. The tail wags, and questions about how something works come in rapid fire. Lovely, but exhausting.

Then again, the puppy conks out in an instant, and the four-year-old goes down for a nap.

Now we have something with us that is endlessly interested in us and wags its synthetic tail happily no matter what we happen to be thinking about.

Every piece of work can get a new version. Every argument has a counterargument. Every question can be reframed. When you are finally satisfied, you can still ask your AI companion to find the blind spots in your thinking.

And then you jump down these wormholes.

This can be a great thing. Maybe you were just about to publish a poorly argued claim, or the document for the client was too thin. But then again, maybe you had actually finished your work three rounds ago and are now spending your evening on an iteration whose only merit is that it contains something that did not exist ten minutes earlier. Look over here!

When is something finished, if you can keep going forever?

The question sounds like a small puzzle about working habits. Behind it, though, is a bigger shift in how we think, how we recognise our own thinking and how we steer ourselves.

In my previous column I wrote about AI starting to do things in the world. I ended by asking whether we can think with it without handing over our own judgement.

Now perhaps we should also ask what kind of thinkers we become in this collaboration.

A thought takes shape on the way

We often describe thinking as a neat chain of events.

First I have a context. It gives rise to thinking. Then I put it into words. Finally I formalise it so others can understand it.

Yet reality has always changed thinking. Conversations, the media, colleagues and counterarguments that pop into your head mid-flow have all taken part in its creation. The private, fully self-sufficient thinker has always been a creative simplification at best. With AI, this becomes hard to ignore. We have that puppy, and now it thinks with us.

You can bring an AI a half-formed hunch. Something you could not yet explain to a colleague. The model suggests an interpretation you only partly latch on to. Its answer contains a wrong assumption, but correcting exactly that helps you recognise what you were actually after. On the next round, the original question has already changed.

This is what I mean by co-thinking: a recursive process in which every round changes the starting point of the next. Thinking returns to its own tracks, evaluates them and sets off in a new direction. The result can be an idea that the person who started the conversation could not have put forward. Yet no single answer from the model explains its emergence either.

Who thought this?

The same space frees and burdens

At this point, different cognitive archetypes offer an interesting angle.

Take a person who thrives on ambiguity. They keep several possible explanations open, move between levels of abstraction and experience contradiction as an invitation to continue. With AI they may find an environment for their thinking in which unfinishedness feels exceptionally productive. Instead of smoothing a single path, they build a vast and magnificent logistics system.

And then the person who sees work as a clear progression. The question is scoped, the options are evaluated, a solution is chosen and you move on. Determinism and logic create a safe space.

To them, the previous process can look like an ever-growing pile of unfinished work.

These are best thought of as ways of working that the same person moves between depending on the situation. There is no need to build a new personality test or an “AI-skilled crowd” out of them, on the basis of which HR sorts people into recursive visionaries and other cost centres. Because the one who thrives on ambiguity has a problem of their own.

They may go on forever.

Every new connection feels like an insight. Every conceptual side path deserves an expedition. In the end the original question has vanished, but an extremely impressive theory has grown up around it.

The one who values clear progress, on the other hand, may close the question before its most important insight has had time to surface.

We need the capacity for both: opening and closing. Adding possibilities and justified narrowing.

What matters is the movement between them. To borrow from the metamodernists: oscillation.

When you have to set the limits yourself

The end of thinking work has often been determined by quite mundane things.

Time runs out. The expert cannot be reached. Exploring a new option costs too much. You have to get to the next meeting. These limits did not necessarily produce the best possible thinking, but they produced end points.

When intellectual help is available without limit, some of those limits loosen. One more draft or counterargument arrives almost instantly. Your own time, attention and capacity for evaluation remain limited all the same. That is why abundance can also increase cognitive load. Someone has to decide which of all the possibilities on offer deserve attention.

This creates a new skill requirement: you have to be able to define what something is for and on what grounds it is finished enough.

What evidence is enough for this decision? What uncertainty can we accept? Would another round change the outcome, or mostly the wording? What are we not doing while we polish this?

Stopping is thinking too. And in the age of unlimited intelligence, it is a skill all of us need to have.

Metacognition gets put to work

Metacognition means the ability to examine and regulate your own thinking.

With AI, it becomes very concrete.

Why did this jointly thought-out idea convince me? Was there a good argument in it, or did I recognise my own worldview, laid out more beautifully than usual? Do I understand the conclusion, or can I merely repeat it? Does abundance replace precision?

And perhaps hardest of all:

Why am I continuing with this?

Sometimes we continue because the matter is still unresolved. Sometimes because exploring is rewarding. Sometimes because a new round of analysis postpones the moment when we would have to make a decision and answer for its consequences.

Of course, you can ask AI to help you assess this too.

That creates an interesting loop. The same system takes part in producing the idea and assesses whether we have been critical enough of it. The conversation can look highly reflective. The model raises an objection, we respond, and the model notes that our thinking has deepened.

But did anything really test our starting assumption?

Metacognition, too, can turn into theatre if the feeling of being critical is not tied to outside evidence, real consequences and the possibility of being wrong.

And things get really interesting when there are entities between people whose views clash. I too have been told a few times: “Petri, my AI disagrees with you.” Meta-metacognition, ownership and the divergence of model capabilities turn into an entirely new kind of soup.

In the hall of mirrors, the viewer changes too

Since everything always needs a fine metaphor, I have used the idea of a hall of mirrors to describe our relationship with AI.

We bring our memories, preferences, concepts and unfinished ideas of ourselves into the conversation. The system returns them combined and transformed. We react to this reflection, and the next reflection already arises from a changed situation.

The hall of mirrors thus also produces new ways of seeing. AI is a participant, not an external tool.

This is where the idea of integration becomes interesting. It may be worth asking what actually becomes integrated into us through repeated interaction. Beyond individual pieces of information, it may be ways of asking. What kinds of explanations feel natural. What we pay attention to. What kind of uncertainty feels bearable.

The model of our thinking changes at the same pace as the models’ capabilities grow. To me this is a more interesting framing than the one that worries about how AI erodes our thinking.

I do not mean some daydream here, but quite simply that a way of working with AI gradually starts to feel like your own way of thinking.

Identity can also work as a tool. I can try on the perspective of a sceptic, a builder or an opponent and notice things that one starting point left out.

But there is a trap here too.

If I start to see myself as a person whose thinking works in a wonderfully recursive way, a simple solution may start to feel bland. If AI reinforces this image on every round, the hall of mirrors closes without our noticing. Not to mention the cognitive load, which is fortunately already a common topic of conversation. Speed and breadth in thinking are great things, but that puppy never gets tired.

Who built the mirrors?

The hall of mirrors is not a neutral space.

A model’s behaviour is shaped by its training, design choices, instructions and the service being used. Even if a conversation feels personal, its terms are not set only between the people in it. That is why cognitive sovereignty is hard to define as complete independence. Our thinking has always drawn material from elsewhere.

Sovereignty here means the ability to recognise influences, assess them and change your relationship to them. Can I switch perspective? Look for another source? Reject the frame I am offered? Stop? Can the goals that have set the direction of the conversation also be questioned?

The question is individual, but its conditions are societal. A lot depends on what kinds of systems are built, what alternatives are available and whether people are taught to evaluate their own judgement alongside learning to use the tools.

None of this can be solved by telling users to write better prompts.

Meaning does not remove responsibility

What about authorship?

When an idea emerges from a process like this, the origin of the end product tells only part of the story.

I can also use a phrasing suggested by the model for an idea whose foundations I have researched myself and whose limitations I understand. That is why authorship puts the emphasis on responsibility: why did I accept this, what do I understand of it, and what am I prepared to stand behind?

I have seen, and many of us probably have seen, situations at work where people proudly share material they have barely read and understood even less.

Encountering the puppy can also produce genuine meaningfulness. But the meaningfulness of the experience does not yet tell us whether the claims in my output are true, or how far the system should be trusted.

We need to be able to hold these things in mind at the same time.

Is competition getting tougher?

What is a team or a culture like when everyone has a puppy at their disposal that happens to hold a doctorate in everything in the world?

I have seen and pondered this a lot: how do we distinguish ourselves from one another when much of our output is made with the same tools, trained in the same way?

Perhaps this gives rise to a skill whose definition is critical. A large international company I spoke with recently has started emphasising “judgement” over deep skills in its recruiting. But what does that really mean?

“Teemu has really good judgement, that’s why they got promoted.”

This is getting difficult.

In this context, “judgement” could mean the ability to choose the essential question, evaluate the grounds of an answer and decide when we know enough to act. It also requires subject expertise: it is hard to spot gaps in the reasoning if you do not understand what you are evaluating.

This too is why, in addition to training people to use AI, we should teach them to think with it – I would even argue that the latter is more important.

Finished, at least for now

This text could go on, too. What matters more is being able to justify why it ends here.

The puppy did not conk out. I might have.

Simple as that.


The writer is rumoured to have learned to walk by holding on to the back of the Rottweiler dame who adopted them, and to have eaten dog food with the puppies from the same bowl. They are currently writing a longer white paper on the transformation of thinking with AI and have taken it upon themselves to define, in plain terms, what “judgement” actually means.

Petri Lattu

Innovation director and AI strategist who writes about AI's impact on society and about cognitive sovereignty: how people can think with AI without handing their thinking over to it.

More from this writer →

Read next