What we mean when we say "intuition"

A friend of mine, a brilliant programmer and one of the sharpest thinkers I know, sent me an email last week with an argument I haven't been able to put down.

He said today's AI is, if anything, purely intuitive. Think about how it's trained. It gets a flood of data, looks for patterns, adjusts parameters, predicts an outcome, checks it against a reference, and tunes itself until it's good enough. When you give it a prompt, statistical functions return a result. There's no thinking behind it. Just fast pattern recognition, the way human intuition fires before reasoning does. The model doesn't know why it answered the way it did. It answers instinctively.

I wanted to agree with him. Partly because the analogy is elegant, and partly because I have felt the same thing while using AI. Sometimes it gives you an answer with the speed and confidence of someone who just gets it. It jumps to a framing, a sentence, a structure, or a solution before you have fully explained the problem yourself.

But something about calling that intuition bothered me.

He's right about the mechanics. And he's not alone. This is one of the most common ways people now describe AI in serious discussion. Even researchers sometimes use the word intuitive when describing what large models do.

So if pattern recognition is intuition, then yes, AI has it. But that is the whole question. Is pattern recognition what we actually mean when we say intuition?

Two things wearing the same word

The philosopher Michael Polanyi called human intuition tacit knowledge. The background sense through which deliberate thought takes shape. Not a hunch, not a guess. A whole layer of knowing we can't fully articulate, but that organises everything we do consciously on top of it. The classic example is recognising a face in a crowd. You know who it is instantly, but you couldn't describe exactly what features you used to do it.

Daniel Kahneman later described something close to this as fast thinking. Cognitive scientists Hugo Mercier and Dan Sperber argue that reasoning evolved, in part, to make our intuitive judgments shareable. To test them, refine them, and exchange them inside a community. Intuition and reasoning don't simply compete. In human life, they develop together.

That distinction matters.

When a large model completes a sentence, it performs pattern recognition so dense it can feel intuitive. But there is no tacit dimension underneath it in the human sense. No body that learned by being wrong. No community of practice that slowly shaped which patterns matter. No lived history of consequences. It has learned associations across text, image, and code at a scale no human could match. What it doesn't have is the context that makes those associations meaningful.

This stops being philosophical the moment you open Figma, a strategy deck, or a client presentation.

The body knows before the sentence does

Harvard cognitive scientist Elizabeth Bonawitz puts it well. Human minds are computational. They use Bayesian processes, a way of updating belief about the world by weighing prior experience against new evidence. Each time you see something happen, you nudge your sense of how likely it is to happen again. Most of cognitive science assumes the brain does something like this, quietly, under the hood. But human minds are, in Bonawitz's words, better than Bayesian. We make leaps that probability alone can't explain.

In her lab, kindergarteners playing strategic games made informed moves faster than a purely Bayesian system. Five-year-olds, with a fraction of the data, outperformed the math.

How? One answer is that human judgment is not only computed. It is felt.

The neurologist Antonio Damasio spent decades studying patients whose brain damage had taken away their ability to feel emotions. Their logic was intact, but they struggled to make decisions. His explanation became known as the somatic marker hypothesis. Bodily signals, shaped by emotional experience, help us move before conscious reasoning has finished its report. When you sense something is wrong before you can say why, that isn't magic. It may be information your body is summarising for you, drawn from every time something similar happened and mattered.

This is the part of intuition AI doesn't have, and as far as we know, can't have without a body. A model has no stakes. It doesn't flinch. It can't be wrong in a way that costs it anything. There's no autonomic system tightening, no chemical record laid down for next time. The pattern matching may look similar from the outside. The substrate underneath is not.

The disembodied problem

When I think about this, I keep coming back to the same image. A model produces text. A person produces a sentence. From the outside, the outputs can look similar. But everything that built them is different.

Behind your sentence is a body. A body that has felt cold and warm, sat in uncomfortable meetings, been disappointed by people, fallen in love, been wrong in public, been right when nobody believed it. The sentence carries all of that whether you notice it or not. Behind the model's sentence is a probability distribution over the next word. That isn't an insult. It is just a different thing.

Tony Prescott and Stuart Wilson, who work in cognitive robotics, argue that robotics is a useful test bed for understanding the kind of intelligence that emerges through bodies acting in the world. The broader point is important: real intelligence did not evolve as text prediction. It evolved through organisms that sensed, moved, failed, adapted, and survived. Without that loop, you can get correlation at scale, but not the same kind of understanding.

Melanie Mitchell, one of the clearest voices in AI research, makes a similar point in different words. Today's systems capture correlations at immense scale, but they lack the flexible abstraction that gives human reasoning its context. They know what tends to follow what. They don't know what any of it is for.

This is why human experts still catch things models miss in ways that look almost magical. The surgeon who feels something different in the resistance of the instrument. The negotiator who can tell a client is hiding something from a tiny shift in tone. The designer who knows a layout is wrong before she can say which proportion is off.

I know that last one well.

In digital work, a page can be technically correct and still feel wrong. The sitemap can make sense. The copy can be clear. The components can be polished. The visual system can be consistent. And still, something can feel dead. Or too eager. Or like it is trying to impress instead of helping.

At first, I often can't explain it. I just know we are not there yet. Then, after a few hours or days, the reason becomes visible: the hierarchy is saying the wrong thing, the rhythm is off, the concept is too decorative, the interaction is clever but not useful, or the page is answering a question nobody actually asked.

That first discomfort is not random. It is years of looking at work that almost works.

And this is the difference I keep coming back to. AI can produce options. Sometimes very good ones. It can help me name a pattern faster, sharpen an argument, or see a blind spot. But it doesn't carry the memory of the last time a beautiful idea failed in a client room. It doesn't remember what it feels like when a page is impressive but not trusted. It doesn't have a nervous system trained by consequences.

Human experts are not just better pattern matchers. They are pattern matchers anchored in years of consequence.

What about giving AI a body?

The researchers know all this. The field is moving fast.

A team led by researchers at the University of Edinburgh published work in Nature Machine Intelligence on ELLMER, an embodied large language model connected to a robot arm. The robot makes coffee and decorates plates. The interesting part isn't the task itself. It is that the system uses visual and force feedback, the pressure the robot senses when it touches something, as part of how it decides what to do next. A wave of labs is now working on vision-language-action models, where AI doesn't just talk about the world, but learns to act in it.

The bet is simple. If embodied cognition is what makes human intuition possible, give the machine a body and see what happens.

It might work. I genuinely don't know. But I notice that even the best of this research is still building intuition from the outside in. A sensor reports pressure. A model interprets the reading. An action follows. What's missing is what a body actually has, which is something at stake. The robot arm doesn't care if it drops the cup. It doesn't feel embarrassed. It doesn't lie awake remembering it. The sensorimotor loop is there, but the loop that connects experience to consequence to memory to identity isn't, at least not yet.

Maybe one day it will be. Until then, the gap is real, and it is worth naming.

Why this matters

You might think this is a definitional argument, and definitional arguments are usually boring. But this one changes how we work.

I use AI almost every day. I use it to think, write, structure, compare, challenge, and get unstuck. I don't want to pretend it is just a toy. It is already part of my work. Sometimes it helps me move faster. Sometimes it helps me see a better angle. Sometimes it gives me a sentence I wish I had written myself.

But I don't want to outsource the part of the work where judgment lives.

AI can generate options. It can make the page fuller, faster. It can produce arguments, variations, headlines, concepts, and structures. But it cannot yet tell me which option carries the right weight for this client, this audience, this moment. It doesn't know what should be left unsaid. It doesn't know when a clever idea is too clever. It doesn't know when a design is solving the brief but missing the emotion.

If we accept that AI is intuitive in the same way humans are intuitive, we'll start treating its outputs as if they carry the same epistemic weight as the gut feeling of an experienced expert. We'll say the model has a sense for this or it just knows. And then we'll be surprised when it confidently produces something subtly wrong in a way an experienced human might have caught.

The human's wrongness is constrained by having lived in the world. The model's isn't.

Calling it pattern recognition instead of intuition isn't a downgrade. It's a more honest description. Pattern recognition is genuinely powerful. It is most of what we do too. But human intuition adds something the model can't, at least not yet: a body that has been wrong in real situations, with real consequences, and adjusted accordingly.

My friend is right that AI looks intuitive. He's right that the mechanics can seem similar. Where we differ is on what the word covers.

For now, intuition still belongs to those of us with skin in the game.

Although, if a robot ever pulls into a parking spot and just knows it'll fit, I'll reconsider.

Sources

If you want to read further into the ideas behind this piece:

• Michael Polanyi, The Tacit Dimension (1966). The book that gave us "we know more than we can tell." The foundation of the tacit knowledge argument.
• Daniel Kahneman, Thinking, Fast and Slow (2011). The fullest treatment of System 1 thinking and the conditions under which intuition can be trusted.
• Hugo Mercier and Dan Sperber, The Enigma of Reason (2017). The argument that reasoning evolved to make intuition shareable inside a community.
• Antonio Damasio, Descartes' Error: Emotion, Reason, and the Human Brain (1994). Where the somatic marker hypothesis is laid out for a general reader.
• Melanie Mitchell, Artificial Intelligence: A Guide for Thinking Humans (2019). A clear-eyed account of what current AI systems can and cannot do, and why flexible abstraction is the missing piece.
• Tony J. Prescott and Stuart P. Wilson, Understanding brain functional architecture through robotics. Science Robotics, 2023. A robotics-based argument for understanding intelligence through bodies acting in the world.
• Ruaridh Mon-Williams et al., Embodied large language models enable robots to complete complex tasks in unpredictable environments. Nature Machine Intelligence, 2025. The ELLMER framework connecting large language models with visual and force feedback in robotic tasks.
• Harvard T.H. Chan School of Public Health, Intuition in the Age of AI essay series (2025). Where the Polanyi, Kahneman, Mercier and Sperber, and Mitchell threads are pulled together.
• Harvard Gazette, Is AI dulling our minds? (2025). The interview with Elizabeth Bonawitz on why human minds are better than Bayesian.