# How AI Actually Works
Nineteen ninety-seven. Chess hall in New York City.
Garry Kasparov — the greatest chess player alive — stares at a computer screen. Deep Blue just moved. The audience holds its breath.
And Kasparov resigns.
Not because he's tired. Because he can't see a way forward. A machine just beat the human brain at the thing humans thought made us special: strategy.
Except here's what nobody tells you about that moment.
Deep Blue wasn't *thinking.* Not the way Kasparov was. It was calculating — checking two hundred million positions every second. Brute force wrapped in silicon.
Kasparov was visualizing future games... feeling patterns in his gut... drawing on every match he'd ever played.
Deep Blue was doing arithmetic. Really, really fast arithmetic.
That difference? That's the whole story.
When we say "artificial intelligence," we're using a term from nineteen fifty-six. John McCarthy coined it at a summer workshop at Dartmouth. He and Marvin Minsky and a handful of others gathered to explore one question: can machines think?
They believed they could crack it in a summer. Two months. They submitted a proposal asking for thirteen thousand dollars and the time of ten men.
They were optimistic. Also... cosmically wrong about the timeline.
But here's what they got right. They understood that intelligence might not *require* consciousness. It might just require the right kind of computation.
That insight — that thinking might be separable from feeling — that's either liberating or terrifying, depending on the day.
Fast forward to nineteen eighty-six.
Geoffrey Hinton and two colleagues publish a paper on something called backpropagation. It's a method for training neural networks — these mathematical structures loosely inspired by how neurons in your brain connect and fire.
Backpropagation lets these networks learn from their mistakes. You show the network a picture of a cat, it guesses "dog," and the algorithm adjusts thousands of tiny numerical weights so next time it gets closer to "cat."
Think of it like this. You're a kid learning to shoot free throws. You miss left. Your brain — without you consciously deciding anything — adjusts the angle and force for next time.
Backpropagation is that... but instead of muscle memory, it's adjusting numbers. Thousands of them. Simultaneously.
This wasn't new science. The math had been around since the seventies. But Hinton made it work at scale.
And nothing happened.
For twenty years... almost nothing happened.
Neural networks were interesting. They were also slow and expensive and couldn't compete with simpler methods. Hinton kept working on them anyway. His colleagues thought he was wasting his career.
There's a particular kind of loneliness in being right too early.
Until twenty twelve.
Alex Krizhevsky, a grad student working with Hinton, enters a competition called ImageNet. The challenge: look at photos and identify what's in them. Dogs, cars, bridges, mushrooms. Fourteen million images, twenty thousand categories.
The best systems at the time — built by teams at Microsoft, Oxford, the University of Tokyo — were getting about seventy-five percent accuracy.
Krizhevsky's neural network hits eighty-five percent.
That ten-point gap? That's not incremental improvement. That's the sound of a field waking up. It's the moment when everyone realizes the old rules just stopped applying.
Suddenly everyone sees it. If you give a neural network enough data and enough computing power, it doesn't just get a *little* better. It gets qualitatively different. It starts seeing patterns no human programmer explicitly told it to look for.
Patterns we didn't even know existed.
So here's how it actually works.
Imagine you're trying to teach a computer to recognize faces. The old way — symbolic AI from the seventies and eighties — you'd write rules. "A face has two eyes above a nose above a mouth." You'd code in proportions and distances. Expert systems. Explicit logic. Hundreds of if-then statements that some poor programmer had to think up and debug.
The new way?
You show the network ten thousand faces. You don't explain what a face *is.* The network builds its own internal representation across layers of artificial neurons.
The first layer might detect edges — just lines at different angles. The second layer combines edges into shapes. Curves. Corners. The third layer recognizes patterns like "two dark circles near each other."
And by the tenth layer, it's got something we might call the concept of "face" — except it's not a concept in any human sense. It's a mathematical transformation. A point in ten-thousand-dimensional space that clusters faces together and keeps non-faces far away.
Numbers in... numbers out.
And this is where it gets strange.
We don't fully know what happens in those middle layers. We can see the inputs. We can see the outputs. But the internal logic? Black box. Billions of parameters adjusting in ways we can measure but not always interpret.
It's like asking someone why they find a song beautiful and they say "I don't know, I just do."
Except the person is made of math.
Timnit Gebru — one of the sharpest voices in AI ethics — calls this the interpretability problem. And she's shown it's not just academic.
When facial recognition systems are trained mostly on lighter-skinned faces, they can be thirty times less accurate on darker-skinned individuals. Point-eight percent error rate for light-skinned men. Nearly thirty-five percent error rate for dark-skinned women.
The bias isn't in the algorithm. It's in the data. The AI learned our prejudices because we fed them in.
It's like training a chef entirely on Italian food and then being shocked when they can't make proper Thai curry.
Except the stakes are whether someone gets falsely arrested.
But let's go back to that question. Can machines think?
Twenty sixteen. Seoul, South Korea.
Lee Sedol — one of the greatest Go players in history — sits across from AlphaGo, a program built by DeepMind.
Go is harder than chess. *Way* harder. More possible board positions than atoms in the observable universe. Chess computers could win by brute force — check every possibility far enough ahead.
Go requires something else. Intuition, maybe. Creativity. The ability to see the whole board as a living system.
Game two, move thirty-seven.
AlphaGo places a stone on the fifth line.
Commentators call it a mistake. A move so strange it has to be an error in the code. Lee Sedol leaves the room. He needs fifteen minutes to process what he's seeing.
Because it's not a mistake. It's brilliant. A move no human would play because it violates three thousand years of accumulated intuition.
And it works.
Fan Hui — the European Go champion who'd lost to AlphaGo months earlier — was watching. He said later it was like seeing a new color. Something that shouldn't exist... but does.
AlphaGo wins four games to one.
Here's the thing — it didn't learn by memorizing human games. It learned by playing *itself.* Millions of times. Reinforcement learning. Try something, get a reward or penalty, adjust.
It invented its own strategies. Strategies that are now changing how humans play a game we've been playing since five forty-eight BC.
So is that thinking?
Or is it just really, really sophisticated pattern matching?
David Chalmers — philosopher of consciousness — argues we might never know. Because we don't even agree on what thinking *is.*
When you recognize your friend's face in a crowd, are you computing? When you feel sad after a breakup, is that information processing?
Maybe consciousness and computation are different things. Maybe they're not. Maybe asking "do machines think?" is like asking "do submarines swim?"
Depends what you mean by swim.
What we *do* know is this. Modern AI is powerful... and narrow.
GPT-3 — released by OpenAI in twenty twenty — has a hundred and seventy-five billion parameters. It can write poetry, answer questions, generate code, explain quantum mechanics in the style of a pirate.
Feed it text, it predicts what comes next. That's it. That's the whole trick. Predict the next word. Do it so well, across so much training data, that the predictions *look* like understanding.
But it doesn't know what a word means. Not the way you do.
It's never tasted coffee or felt rain or had someone break a promise. It's never wanted anything. It's a statistical model of language, trained on hundreds of gigabytes of text scraped from the internet. Books, Wikipedia, Reddit arguments, fanfiction, medical journals, conspiracy theories. Everything.
Which means it's learned our conspiracy theories, our bad grammar, our cruelty, our brilliance, our contradictions.
It's a mirror made of probability distributions.
And here's the cost nobody talks about enough. Training GPT-3 produced as much carbon as five cars over their entire lifetimes. More than five hundred metric tons of CO2.
AI is energy-hungry. The bigger the model, the bigger the environmental bill. We're burning fossil fuels to teach computers to write sonnets.
So why build these things?
Because sometimes they work in ways nothing else can.
Twenty twenty. DeepMind releases AlphaFold, a system that predicts how proteins fold. This is a problem that's stumped biologists for fifty years.
Proteins are chains of amino acids that twist into 3D shapes, and the shape determines what the protein does. A protein for digesting lactose looks completely different from a protein that fights infections.
Knowing the shape could change drug design, help us understand diseases, maybe crack Alzheimer's or Parkinson's. But figuring it out experimentally takes months or years per protein. You crystallize the protein, shoot X-rays through it, solve differential equations. It's slow.
AlphaFold does it in hours. For nearly every human protein. Twenty thousand of them. With accuracy that matches experimental results.
John Moult — who's been running protein folding competitions since nineteen ninety-four — said when he saw AlphaFold's results, he nearly fell off his chair. He thought maybe the team had cheated somehow.
They hadn't. They'd just solved a problem everyone thought was decades away.
That's not hype. That's a tool that might help cure diseases. Real ones. That kill real people.
And this is the tension.
AI isn't one thing. It's chess computers and protein folders and facial recognition and chatbots and content moderation and credit scores. Some of it's useful. Some of it's biased. Some of it's actively harmful.
And we're building it faster than we're understanding it.
Richard Sutton — one of the pioneers of reinforcement learning — wrote something in twenty nineteen called "The Bitter Lesson." His point: every time we think AI needs human knowledge, we're wrong. Every time we try to hand-code expertise into systems, we lose.
The systems that win are the ones that scale. More data, more compute, simpler algorithms. Let the machine find the patterns.
Human insight, he argues, doesn't scale. Computation does.
It's bitter because it means decades of work encoding human expertise was kind of a dead end. All those expert systems from the eighties, carefully programmed with rules from doctors and engineers — they lost to pattern matchers with bigger datasets.
But Gebru and others push back hard. Scaling isn't neutral. It amplifies what's in the data. If we don't look carefully at what we're feeding these systems, we're building inequality into the infrastructure.
We're encoding twenty twenty-four's biases into systems that might run for fifty years.
Here's a concrete example.
Amazon built a hiring tool that reviewed resumes. Trained it on ten years of their own hiring data. The AI learned to downgrade resumes that included the word "women's" — as in "women's chess club captain."
Because Amazon's engineering teams were mostly men, so the AI learned that being a woman was negatively correlated with being hired.
They scrapped the tool. But how many companies are using similar systems and haven't noticed?
So here's where we are.
AI works by finding patterns in data. Sometimes those patterns are useful — protein shapes, tumor detection, predicting which bridge needs repairs before it collapses.
Sometimes they're our worst instincts reflected back — who looks "trustworthy," who sounds "professional," who belongs in this neighborhood.
The systems don't understand. They predict. They optimize. They surprise us.
But they don't *want* anything. They're not thinking. Not yet. Maybe not ever.
And maybe that's the thing to hold onto.
When you see an AI do something impressive — when ChatGPT writes you a poem or your phone finishes your sentences — ask: what data trained this? What's it optimizing for? Who benefits? Who gets hurt?
Because the intelligence is artificial... but the choices behind it? Those are deeply, unmistakably human.
There's a researcher named Kate Crawford who studies AI's social impacts. She points out something obvious once you hear it: AI systems aren't artificial and they're not intelligent.
They're made from natural resources — lithium mines, water for cooling data centers, underpaid workers labeling images in Kenya.
And they're not intelligent in any general sense. They're just good at specific tasks we've defined success for.
So maybe we've been asking the wrong question. Not "can machines think?" but "what kind of world are we building with these machines?"
Next time you talk to a chatbot or get a recommendation from an algorithm, try this.
Assume it's not magic. Assume it's math that learned from examples. Then ask: whose examples? What did it learn to ignore? What would it take to teach it something different?
That shift — from "this is so smart" to "what is this actually *doing*" — that's how you stay grounded when the tools are getting more powerful and the explanations are getting harder to see.
The machines aren't thinking. But we still are.
And that matters more than ever.
Because here's the thing about intelligence — artificial or otherwise. It's not just about solving problems. It's about choosing *which* problems to solve.
The AI can optimize. But we decide what optimal means.
And that choice? That's still ours.