Blog / Sep 05, 2026

Why Computers Don't Understand Anything

And why we're still not entirely sure humans do.

There's a weird thing we do whenever a computer does something impressive.

It recognizes a face, so we say it sees.

It translates Japanese into English, so we say it understands Japanese.

It writes code, so we say it knows how to program.

It answers a hard question, so we say it knows the answer.

And when an LLM writes something that sounds thoughtful, we start arguing about whether it's conscious.

But there's a more basic question underneath all of this, one we usually skip past on our way to the fun argument:

What the fuck does it actually mean to understand something?

Because before asking whether a computer understands anything, we probably need to figure out what "understanding" even is. Nobody agrees. That's where this gets weird.

1. Start with something stupidly simple

Take a calculator.

Type:

2 + 2

You get:

Obviously, the calculator knows two plus two is four.

Right?

No. Not in the way we normally mean know.

There's no mathematical soul in there going, ah yes, two apples plus two apples, four apples, beautiful. It doesn't know what an apple is. It doesn't know what "two" is. It doesn't know that 4 is a number, or that a number is a thing that exists.

It's flipping electrical states according to a rule.

And the answer is still correct.

That's the first crack in the whole thing: correctness and comprehension just came apart, and nobody noticed.

2. Symbols are not what they represent

Computers are obsessed with symbols. Everything eventually becomes a representation of a state — numbers, letters, images, audio, video, programs, files. A photograph of a dog becomes millions of numerical values. The word:

DOG

might be stored as something with zero resemblance to an actual dog. The machine never gets the animal — just the representation.

Now compare:

CAT

A human looks at that and instantly drags in a mess of associations: an animal, fur, whiskers, claws, a particular yowl, your childhood cat, the general concept of "pet," the smell of one, what happens when you step on its tail.

None of that lives in the three letters. C-A-T is just ink. The interesting part is everything the symbol drags in behind it — and that's exactly where computers start getting interesting too.

3. The Chinese Room problem

Philosopher John Searle proposed a thought experiment now infamous enough to have its own Wikipedia arguments section.

You don't speak Chinese. You're locked in a room. Someone slides Chinese characters under the door. You have a rulebook: if you see symbol A followed by symbol B, output symbol C. You follow it perfectly. People outside send increasingly complex sentences. You keep producing flawless, appropriate replies.

Eventually someone outside concludes: the person in there understands Chinese.

But you don't. You have no idea what any of it means. You're pushing symbols around a table.

From the outside, though, your behavior is indistinguishable from someone who actually understands Chinese. Which raises an ugly question:

Is producing the right behavior enough to count as understanding?

If yes, the room understands Chinese. If no — what exactly is missing, and how would you ever detect its absence?

4. Syntax vs. semantics

Two useful words.

Syntax is structure and rules. Semantics is meaning.

The cat chased the mouse.

has grammar. Scramble it —

Mouse the cat chased the.

— and the syntax collapses. But syntax was never the same thing as meaning. A machine can manipulate symbols by rule without possessing what those symbols mean, and that's basically the entire premise classical computing was built on.

Input, rules, output. Computers are ferociously good at this and need to understand none of it to do it. Your CPU doesn't understand your program; it executes instructions. Your hard drive doesn't understand your photos; it stores bits. Your router doesn't understand the website; it forwards packets.

And yet the whole stack, end to end, produces something that feels drenched in meaning — to us.

5. So where does meaning come from?

Here's the uncomfortable part: maybe meaning was never in the symbol. Maybe meaning is the relationships between things.

TREE

You don't understand it because T-R-E-E secretly contains the essence of a tree. You understand it because that symbol sits at the center of a web:

tree → plant → grows → roots
tree → wood → furniture, fire, paper
tree → forest → ecosystem, rain, deforestation

Pull the symbol out of the web and it's dead. Nothing in isolation means anything. Meaning is the wiring, not the node — and that reframing is exactly what makes modern machine learning worth taking seriously.

6. Then machine learning showed up

Traditional software:

Rules + Input → Output

Machine learning inverts part of that. Instead of hand-writing every rule, you hand the system examples:

image → cat
image → dog
image → cat
...

and it adjusts its own internal parameters until it gets good at producing the right outputs. Nobody tells it cats have two triangular ears, whiskers, and this facial structure. It finds the statistical regularity on its own.

Show it something new and it says:

CAT: 97.3%

Did it understand the cat, though? That's the harder question, and confidence is not the same thing as being right.

7. The interesting thing about neural networks

A neural network doesn't file knowledge in neat boxes. There's no address in memory that reads:

CAT = furry animal, whiskers, purrs

Information is smeared across enormous numbers of parameters. Patterns lock onto other patterns. Features tangle together. Representations become useful for predicting other things — and something genuinely strange falls out of that: the system starts capturing structure, not just examples.

The famous case: models learn that

Paris : France
Tokyo : Japan
Berlin : Germany

hold the same relationship to each other, without anyone ever writing that rule down. That's not a lookup table. That's a network finding shape in the data and reusing it — which is a big part of why these systems feel like they're "understanding" rather than memorizing.

But feeling like understanding and being understanding are not obviously the same claim. Same wall, again.

8. What about LLMs?

Here's where the argument gets spicy.

An LLM takes in tokens, runs them through a gigantic network, and predicts what token is likely to come next. That's a simplification, but it's a true one.

So someone says: it's just predicting the next word.

Technically correct. Also a little misleading, because predicting the next token well can demand modeling an absurd amount of structure.

The glass fell from the table and shattered because it was...

The likely completion is fragile. Why? Because the model has absorbed relationships between glass, falling, impact, breakage, and causality itself. Nobody hand-coded:

IF glass falls
THEN glass may shatter

The network found that structure statistically. Which is exactly why "just autocomplete" doesn't settle anything — autocomplete gets genuinely dangerous to dismiss once the thing doing the completing has built a working model of the world underneath it.

9. But does an LLM actually know what a glass is?

Here's where I think you have to be careful.

Ask an LLM what happens if you drop a glass on concrete, and it'll give you a good answer. But it has never held a glass. It's never felt the weight shift in its hand right before the drop, never heard the specific crack of glass on stone, never had to sweep up the pieces and find one under the fridge three days later. It doesn't have a hand. It doesn't have a kitchen.

Its "glass" is a shape in a statistical field, assembled from other people's descriptions of glasses.

Humans build concepts from language, perception, physical interaction, memory, emotion, consequence, and social experience, all layered on top of each other. A kid doesn't learn fire is hot by reading the sentence fire is hot. Eventually the world reaches out and corrects you. That's a fundamentally different channel of information — and it's not obviously replaceable with more text.

10. Information isn't knowledge

These words get used interchangeably. They're not the same thing.

37.7749, -122.4194

That's information. Useless on its own. Tell you it's San Francisco, and it's more useful. Show you a map, and now you have a representation. Explain where it sits relative to LA, and you have a relationship. Tell you to leave at 6 AM tomorrow because traffic is brutal, and the information just became actionable — you can plan around it.

A rough hierarchy:

DATA → INFORMATION → REPRESENTATION → KNOWLEDGE → UNDERSTANDING

Don't mistake this for a law of nature. It's a ladder someone built to think with, and the rungs are fuzzy as hell. But it's useful fuzz.

11. So what is understanding?

My favorite working definition:

Understanding is the ability to build and use a useful internal model of something.

Not knowing facts about it. Not being able to describe it. A model you can run.

If you understand gravity, you can predict what happens when something falls. If you understand programming, a brand-new bug doesn't stop you cold. If you understand a person, you can guess how they'll react to something you've never seen them face. If you understand a joke, you're not just holding the words — you know why it's funny.

The load-bearing part is generalization. Understanding is what lets you go past the examples you've actually seen.

12. And here's where things get uncomfortable

Build a system that's never seen a particular problem. Hand it the problem. It produces a correct solution. Hand it a totally different problem. Solved. Again. Again.

Eventually its internal representations get rich enough that it starts making correct predictions about situations it has never once encountered.

At what point do you stop saying it's manipulating symbols and start saying it's understanding?

There's no obvious line — which is the actual problem, not a rhetorical flourish. We don't have an understanding detector. We can watch behavior, poke at internal representations, and test generalization, but we can't crack a system open and read off its subjective experience.

Inconveniently, we can't do that with other humans either.

13. How do you know I understand anything?

You don't. Not directly. You infer it.

If I say the sky is blue because shorter wavelengths scatter more strongly off atmospheric molecules, you might conclude I get it. Or maybe I memorized that sentence yesterday and have no idea what a wavelength is.

So you push:

Why does the sky turn red at sunset?

I answer.

What if the atmosphere were different?

I answer.

What if there were no atmosphere at all?

I answer, and now you're actually convinced — not because any single answer was impressive, but because I kept generalizing to cases I couldn't have simply memorized. You're inferring an internal model from a pattern of outputs.

Which is precisely the test we run on machines. We just trust it more when the thing answering has a face.

14. The brain isn't a magical exception

This is where the conversation usually gets defensive.

Computers can't understand — they're just machines.

Okay. What are brains?

Physical systems. Neurons, proteins, lipids, electrical spikes, chemical signaling, blood, ions. There's no ghost in there doing the real understanding while the neurons handle the paperwork. Your brain is matter, doing extremely complicated things to itself — and somehow that produces thought, memory, imagination, emotion, language, reasoning, and whatever consciousness actually is.

Nobody has closed that gap. That's the elephant standing in the middle of every one of these arguments, and it doesn't go away just because we're more comfortable staring at the AI instead.

15. The brain doesn't retrieve memories like files

We tend to picture memory like:

MEMORY.DAT → retrieve → play memory

Human memory is much messier than that. Recalling something isn't pulling a clean recording off a shelf — it's reconstruction. Fragments get reassembled. Context bleeds in. Emotion bleeds in. Your current mood bleeds in. And every time you replay a memory, you risk slightly rewriting it.

That's a very different architecture from a hard drive, and it points at something interesting: maybe intelligence was never about storing everything faithfully. Maybe it's about building compressed models good enough to reconstruct what matters, when it matters.

16. Compression might be the secret

You can't store every detail of the world — nobody does. You don't remember every frame of every conversation, every leaf, every sound from yesterday. Your brain throws almost all of it away and keeps the shape.

"I went to that restaurant with my friends."

That's it. Not the molecular state of the room — a compressed pointer you can unpack later. And that compression is exactly what science formalizes: instead of memorizing every falling object's trajectory, we wrote down equations of motion. Instead of memorizing every organism, we built a theory of evolution. Instead of memorizing every sentence in a language, we learned grammar.

Understanding looks a lot like finding the structure that generates the observations, rather than hoarding the observations themselves.

17. But compression alone isn't understanding

Here's the catch: you can compress something with zero understanding.

A ZIP file compresses a book. It doesn't understand a single sentence of it. A neural network compresses statistical structure out of billions of examples — and that alone doesn't prove it grasps what any of them mean.

So maybe understanding needs more layers stacked on top:

Compression
+ Representation
+ Prediction
+ Causal models
+ Interaction
+ Generalization

Or maybe that list is still missing the one ingredient that actually matters. We don't know. Anyone who tells you they've nailed the final recipe is probably selling something.

18. Prediction is a surprisingly powerful test

One clean way to think about intelligence:

Can the system predict what happens next?

A good model of the world buys you prediction almost for free. Understand a ball, and you can predict it falls. Understand a language's grammar, and you can predict a sentence is broken before you can say why. Understand a person, and you can often guess their next sentence before they say it.

Prediction isn't the same thing as understanding, but understanding makes prediction cheap. And that's the quietly wild part of modern machine learning: a system trained hard enough on "predict what comes next" can get forced into building a genuinely sophisticated internal model of the thing it's predicting, as a side effect of chasing accuracy.

19. The strange thing about language

Language doesn't just describe reality — it lets us build worlds that were never real to begin with.

If the king had never entered the room,
the assassination would never have happened.

There's no physical object anywhere that this sentence points to. You're reasoning inside a hypothetical you just invented mid-sentence.

LLMs can do a version of this too — reason across hypotheticals, fictional worlds, mathematical objects, alternate histories, machines that don't exist and never will. That means whatever's happening inside them supports manipulating concepts with zero physical referent.

Does that count as understanding? Maybe. Maybe not. But it's a hell of a lot more interesting than "fancy autocomplete," and it's worth sitting with that discomfort instead of resolving it too fast.

20. Could a sufficiently complex system develop understanding?

I don't know, and "I don't know" is the only scientifically honest answer available right now. Roughly three ways this could go:

One: understanding requires something biological

Maybe consciousness and real understanding are tied to properties specific to living neurons — something silicon simply can't replicate no matter how it's arranged. If true, machines will only ever simulate understanding, forever, no matter how good the simulation gets.

Two: biology isn't special

Maybe substrate is irrelevant and organization is everything. If a system reproduces the right functional processes, understanding might fall out of the architecture regardless of what it's built from — carbon, silicon, whatever.

Three: we're asking the wrong question

Maybe "understanding" was never binary to begin with. Maybe it's a spectrum:

UNDERSTANDING
      ↑
human
      │
advanced AI
      │
animal
      │
simple learning system
      │
thermostat
      │
static program

A thermostat has some representation of temperature. A calculator has some representation of numbers. An LLM carries around a staggeringly complex representation of language and concept. A human has an embodied, persistent, self-referential model of the whole show. Maybe these aren't different answers to yes/no — they're different points on the same axis, and we've just never had to grade on a curve before.

21. There's another problem: consciousness

Understanding and consciousness might not even be the same phenomenon. A system could plausibly understand something with zero subjective experience of understanding it. Or maybe real understanding requires consciousness as a precondition. Nobody knows, because consciousness is still one of the least-cracked problems in science or philosophy.

You can map every neuron firing in someone's brain in exquisite detail. None of that data contains, anywhere, the actual experience of I am me, sitting here, reading this sentence.

There's a gap between what a system does and what it's like to be that system. And that gap is enormous — arguably the real subject this entire essay has been circling the whole time.

22. Maybe intelligence isn't what we thought it was

For a long time, intelligence meant "things humans are good at." Then machines started beating us at chess. Then Go. Then image recognition. Then translation. Then writing code. Then broad reasoning across domains nobody trained them on directly.

Every single time, the goalposts moved.

That's not intelligence, it's just calculation.

That's not intelligence, it's just pattern recognition.

That's not intelligence, it's just language prediction.

Maybe the uncomfortable discovery underneath all this goalpost-shifting is that intelligence is less magical than we assumed — that a lot of what we've been calling intelligence just falls out of systems that build rich enough models of their environment.

If that's right, the interesting question stops being can a machine think like a human? and becomes:

What kinds of thinking become possible on a completely different substrate?

Machines don't have to think like us. They don't forget the way we forget. They don't perceive the way we perceive. They don't experience time, or embodiment, or mortality, the way we do. They might eventually land on forms of cognition we don't have words for yet — because we invented all our words for minds inside a single, narrow, biological case study.

23. So, do computers understand?

Here's my answer, such as it is.

Some computers clearly represent, predict, transform, and generalize information. Whether that earns the word understanding depends entirely on which definition you're holding.

If understanding means can you manipulate representations in ways that support accurate prediction and generalization — then yes, modern AI systems do something that looks a lot like understanding, full stop.

If understanding means do you possess a grounded internal model of the world — the answer gets a lot murkier, a lot faster.

If understanding means do you subjectively experience what you understand — we currently have no method for checking that, in an AI or, arguably, in anyone but ourselves.

And if understanding means do you understand it exactly the way a human does — then no, probably not. But that's not automatically a failure. A bat doesn't experience the world like a human does. Neither does a bee. Neither does a dog. It would be pretty arrogant to conclude from that alone that they understand nothing.

Different architectures can build different, valid models of the same reality.

24. The question I actually care about

I don't think "does AI understand?" is the interesting question anymore. The more interesting one is:

"What kind of understanding can exist in a system fundamentally unlike us?"

We've spent thousands of years studying human cognition almost exclusively, because it was the only cognition we had access to. Now we're building systems with an architecture nothing like the wet machinery inside our skulls, and watching them develop capabilities that would've sounded absurd twenty years ago.

Maybe they don't understand. Maybe they do. Maybe understanding is a word we're going to have to retire and replace, the same way we retired "computer" as a job title once machines took it over.

That's a far more interesting fight than another round of "it's just autocomplete."

25. And maybe we're asking the wrong question about ourselves too

One last uncomfortable thought.

We started by asking: do computers understand anything?

We end up forced to ask: how do we know that we understand anything?

You don't have direct access to my mind. I don't have direct access to yours. We infer understanding in each other from behavior, communication, prediction, and consistency — from the model of "what's going on inside" that we build off someone's outputs over time. It's the exact same test we're running on the machines. We just don't notice, because we've been running it on each other since before language had a word for it.

Maybe the line between

machine

and

mind

was never as clean as it felt. Maybe understanding isn't some special ingredient that shows up only when biology is involved. Maybe it's just what it looks like when a system gets good enough at modeling the world, itself, and the relationship between the two.

Or maybe that's completely wrong.

We don't know yet. That's the fun part.

One last thought

A calculator doesn't know what 2 + 2 = 4 means. A human does.

But somewhere in the enormous, mostly unmapped space between a calculator and a human, we're building new things — right now, on purpose, at scale — and we still don't have good instruments for telling what's actually happening inside them.

Maybe the machine doesn't understand.

Maybe we just haven't figured out what "understanding" looks like when the mind isn't made of meat.