TL;DR: In 1997, a machine beat the greatest chess player alive, and the world asked the wrong question about it. Last issue we learned intelligence isn’t one single ability. This issue asks the next question: of all those abilities, which ones have machines actually acquired, and which remain uniquely human?
The Question Everyone Got Wrong
Something happened in 1997 that a lot of people genuinely believed they’d never live to see.
The reigning World Chess Champion, Garry Kasparov, sat down across from an opponent unlike anything he’d ever faced.
It didn’t breathe. It didn’t blink. It never smiled after finding a brilliant move. It didn’t celebrate when it won.
It was IBM’s Deep Blue, and after six games, it beat him.
The headlines came fast. “The computer defeated the world’s greatest chess player.” To a lot of people, this felt like the beginning of something new.
But underneath the excitement, the world quietly asked the wrong question.
People asked, “Has the machine become intelligent?”
Almost nobody asked, “What kind of intelligence did the machine actually possess?”
Those two questions sound similar. They’re not the same question at all.
Deep Blue could defeat the greatest chess player in history. It couldn’t recognize a cat. It couldn’t understand a joke. It couldn’t learn a new game without engineers going back in and rewriting its software. It couldn’t tell you what chess actually was.
And yet, it was unquestionably brilliant at one specific thing.
So what had humanity actually built? An intelligent machine? Or something that only looked intelligent from one very specific angle?
Last issue asked what intelligence actually is. This issue asks something different: if intelligence has many ingredients, which of those ingredients have machines actually acquired, and which remain uniquely human?
Today’s AI writes poetry, generates images, explains physics, discovers mathematical proofs, diagnoses diseases, drives cars. And a five-year-old child still effortlessly does things that remain astonishingly difficult for our most advanced models to pull off.
This isn’t a contradiction. It’s a clue. And it’s the clue this entire issue is built around.
Comparing human intelligence with machine intelligence is like comparing birds and airplanes. Both fly. They don’t fly in remotely the same way. If you misunderstand that difference, you’ll misunderstand almost everything happening in AI right now.
Why This Isn’t a Competition
Whenever people compare humans with AI, they picture a competition. Human versus machine. Who’s smarter? Who wins? Who replaces who?
That’s an understandable instinct. We compare everything. But intelligence refuses to behave like a race.
Human intelligence and machine intelligence often produce similar-looking outcomes while getting there through entirely different processes. A human recognizes a friend’s face. A neural network recognizes a friend’s face. From the outside, the behavior looks almost identical. Inside, almost everything about it is different.
That’s what this issue is actually about. Not who wins. What’s actually different underneath the surface.
Human Intelligence Was Never One Scale
In the last issue, we saw that intelligence isn’t a single ability. Memory, learning, reasoning, creativity, adaptation, they’re different capacities working together, not one dial turned up or down. That changes the question we should be asking about AI. Instead of “is the machine as intelligent as a human,” the better question is: which specific parts of intelligence has it actually learned?
Psychologists themselves have never agreed that intelligence is one thing, which is worth remembering before comparing it to anything artificial. Charles Spearman argued for a single general intelligence factor underlying many cognitive abilities. Howard Gardner pushed back hard, proposing a theory of multiple intelligences, eight distinct types including linguistic, spatial, musical, and interpersonal, arguing that being gifted in one says almost nothing about your ability in another. Robert Sternberg proposed a third view: a triarchic theory built from analytical, creative, and practical intelligence, the last of which formal IQ tests tend to miss entirely.
The point isn’t deciding who was right. It’s recognizing that even human intelligence was never one scale to begin with. Once you accept that, the whole “is AI as smart as us” framing starts to fall apart. A model can plausibly be developing something like Sternberg’s analytical intelligence at a superhuman level while having nothing resembling Gardner’s interpersonal intelligence at all. The honest question is always: which kind of intelligence, and compared to which specific task?
How Humans Work
Evolution’s Masterpiece, Not Evolution’s Blueprint
Here’s a fact about the human brain that deserves more attention than it usually gets.
It wasn’t designed. Nobody planned it. No engineer optimized it. No committee selected its architecture.
Instead, it emerged over roughly four billion years of evolution. Every neuron, every connection, every instinct, every bias is the accumulated result of countless generations surviving just well enough to reproduce.
The human brain isn’t optimized for elegance. It’s optimized, loosely and imperfectly, for survival. That’s a very different design target than “get the highest score on a benchmark,” and it explains a lot about why our intelligence looks so inconsistent from the outside.
We’re brilliant at recognizing faces. We’re consistently bad at intuitive probability and statistics. These aren’t random flaws that got accidentally left in. Evolutionary theorists frame them as trade-offs, systems optimized for fast, “good enough” decisions under real-world energy and time constraints rather than perfect logical accuracy.
Imagine designing an organism from scratch, with limited energy, limited time, limited biological resources. Would you build a perfectly rational creature? Probably not. Perfect reasoning burns enormous energy. Instead you’d build something that’s good enough. Fast pattern recognition. Quick decision-making. Social cooperation. Danger detection.
The result wouldn’t be perfectly logical. It would survive. That’s us.
Twenty Watts
One of my favorite facts in all of neuroscience sounds almost made up.
The adult human brain runs on roughly 20 watts of power. Roughly the same energy draw as a dim household light bulb.
Sit with that for a second. Around eighty-six billion neurons continuously communicating. Vision updating dozens of times a second. Language coming out effortlessly. Balance adjusting without a single conscious thought.
You recognize your mother’s face instantly. Catch a falling glass before it hits the floor. Notice when someone sounds anxious despite smiling right at you. Dream. Imagine. Plan tomorrow.
All of it powered by about the same electricity it takes to light a desk lamp.
Modern AI couldn’t be more different. Training a frontier language model requires enormous clusters of specialized processors consuming megawatts of electricity, and even running a single large trained model can consume computational resources that dwarf the brain’s energy budget for comparable tasks.
Why biological computation is this efficient relative to current artificial systems remains a genuinely active area of research. The brain doesn’t do precise floating-point arithmetic. It communicates through noisy electrical spikes. Its memory is imperfect. Its computations are massively parallel. Its architecture keeps rewiring itself over time. It shouldn’t be this efficient. It is anyway.
Nature appears to have found remarkably efficient solutions that researchers are still trying to understand, and some believe cracking that efficiency might matter just as much as improving today’s neural networks.
Learning From Almost Nothing
Imagine walking into your kitchen with a two-year-old. On the table sits a fruit neither of you has seen before. Someone says, “This is a rambutan.” The child touches it, maybe tastes it. The next week, out shopping, they point across the aisle: “Rambutan!”
One example was enough.
Now imagine introducing that fruit to a large AI system. It can identify it, provided its training data already contained many prior examples pulled from books, websites, and photographs.
As we saw last issue, children often learn astonishingly quickly from just one or two experiences. Modern AI has narrowed that gap considerably in recent years through techniques specifically designed to reduce the number of examples needed, but it hasn’t closed it.
François Chollet, the creator of the ARC benchmark, argues this distinction sits at the actual heart of intelligence itself. His argument is that intelligence shouldn’t be measured by the total skill a system displays, but by how efficiently it acquires new skills when it runs into unfamiliar problems. A system that needs vast amounts of prior experience just to solve every task might display genuinely impressive performance, and still generalize poorly the moment it hits a truly novel situation.
Maybe intelligence isn’t measured by what you already know. Maybe it’s measured by how quickly you can learn what you’ve never seen before.
Self-Knowledge Machines Don’t Have
There’s a second piece to learning that rarely gets mentioned alongside it: knowing what you don’t yet know.
In 1982, the psychologist Albert Bandura introduced the concept of perceived self-efficacy: our judgment of how well we can execute the actions needed to handle a situation we haven’t faced yet. Bandura found something that sounds almost too simple to matter, except it turns out to explain a great deal. People who accurately believed they were capable of a task performed better at it. People who avoided tasks because they doubted their own capability paid a real cost for that inaccuracy, in wasted effort and missed opportunity.
Later research digging into international student assessment data found that a learner’s self-efficacy, their belief in their own capability, was a stronger predictor of achievement than their self-concept or their anxiety about a subject.
No AI system currently has anything resembling this. A model doesn’t hold a genuine, evidence-based belief about its own capability that it then checks against reality and revises. It doesn’t know what it doesn’t know in the way a student knows they’re shaky on fractions. It can output a confidence score, but that number isn’t self-knowledge in Bandura’s sense; it’s a statistical byproduct, not a judgment the system has formed about itself through experience.
Humans build accurate self-efficacy over years of trying things, failing at some of them, and updating an internal model of their own capability accordingly. That loop, of acting, noticing the gap between expectation and result, and quietly revising your sense of yourself, doesn’t have a clean machine equivalent yet.
Memory: Reconstruction, Not Recording
In Issue #3 we saw that human memory isn’t a storage device. Modern neuroscience goes even further.
Every time you remember your tenth birthday, you’re not replaying a perfect video hidden somewhere inside your skull. You’re rebuilding that moment from scratch. Your brain stitches together sights, emotions, sounds, smells, expectations, and fragments of other memories into something that feels complete, even though it isn’t.
And here’s the strange part. Each act of remembering actually changes the memory itself. Psychologists call this reconsolidation. Memory isn’t read-only. It rewrites itself every single time it gets opened.
That sounds like a flaw. It might actually be one of the greatest strengths human intelligence has. Imagine if every experience stayed perfectly frozen forever. Your understanding of the world could never improve. Instead, each new experience quietly reshapes the old ones. Your understanding evolves. The facts stay connected. The meaning changes.
Machines remember very differently. A trained neural network distributes what it learns across billions, sometimes trillions, of numerical parameters called weights. No single weight stores the concept of “cat.” No single neuron contains the entire idea of “gravity.” Knowledge is spread across the entire network, what researchers call distributed representation.
This makes neural networks surprisingly powerful. It also makes them strangely fragile in ways that trip people up. Unlike humans, who typically add new knowledge without erasing old knowledge, neural networks trained sequentially on new tasks can suffer something called catastrophic forgetting. Train a model on one task, then train it on another, and sometimes the second task completely overwrites the first. It’s almost as if someone learned Spanish and immediately forgot English as a result.
Humans and many animals perform lifelong, cumulative learning naturally, layering new knowledge on old without erasing it. Artificial systems are still working on matching that flexibility.
Embodiment: The Body Problem
Close your eyes for a second. Imagine biting into a lemon. Even without a lemon nearby, you can almost taste it. Your mouth might even start producing saliva.
Humans don’t merely process information. We develop intelligence through constant interaction with a physical world. Gravity teaches us balance. Pain teaches us caution. Touch teaches us texture. Long before children understand physics, they possess an intuitive grasp of weight, motion, and force, built through direct bodily experience like falling. Thousands of times, over and over.
This idea has a name in cognitive science: embodied cognition. A meaningful body of researchers argue that intelligence can’t be fully separated from the body experiencing the world. Your brain didn’t evolve in isolation. It evolved alongside hands, eyes, muscles, skin, balance, hunger, pain, movement.
Today’s large language models have never fallen off a bicycle, never burned a hand, never felt cold, never smelled rain before a storm. They learn almost entirely through symbols, words, images, tokens, and the statistical relationships between them. When a language model explains what an apple tastes like, it isn’t remembering the taste. It’s modeling how humans typically describe the taste. Those aren’t the same thing at all.
Whether genuine understanding requires physical embodiment, or whether it can emerge purely from symbolic and linguistic relationships at sufficient scale, is one of the most actively contested open questions in current AI research. Some researchers believe today’s language models are already showing forms of abstract understanding that surprise even the people who built them. Others argue that without embodiment, something fundamental will always be missing, no matter how large the model gets.
The debate remains genuinely unresolved. But everyone agrees on one thing at least. Humans and machines arrive at knowledge through profoundly different journeys, even when they land on similar answers.
How Machines Work
Seeing Isn’t the Same as Understanding
Take a quick look around the room you’re in right now. Without much effort, you’ve probably identified dozens of objects. A chair. A table. A window. Maybe a phone. It feels instantaneous. Almost effortless.
For decades, researchers genuinely believed computer vision would be relatively easy. Cameras already existed. Computers could already process images. How hard could recognition really be? The answer turned out to be astonishingly hard.
A child recognizes a cat after seeing only a handful of them. Early computer vision systems needed carefully hand-engineered rules. Detect edges. Detect corners. Measure texture. Every improvement required a fresh burst of human ingenuity.
Then deep learning arrived. Instead of writing rules for vision by hand, researchers started training neural networks to discover their own rules. Convolutional Neural Networks revolutionized image recognition.
And yet these systems still make mistakes a human never would. Researchers discovered that changing just a handful of pixels, sometimes imperceptible to the human eye, could completely fool a neural network. A panda becomes a gibbon. A stop sign becomes a speed limit sign. Nothing visible changes for us. Everything changes for the machine.
These are called adversarial examples, and they reveal something genuinely fascinating. Machines often see patterns humans simply cannot. Humans see meaning machines still struggle to grasp. We’re not looking at the world the same way at all. We’re arriving at similar answers through completely different representations underneath.
Do Machines Actually Reason?
Last issue we defined reasoning as constructing knowledge that wasn’t explicitly given, connecting separate facts into something new the way you looked at a wet kitchen floor and an open window and instantly knew it had rained, without anyone telling you so directly. The question now is whether today’s AI systems are actually doing that, or merely producing outputs that resemble it.
Large Language Models appear to do something similar. Ask ChatGPT a multi-step logic problem and it will often break the task into steps and arrive at a correct answer. At first glance, this resembles human reasoning closely.
Whether this constitutes genuine reasoning, in the sense of internally modeling and testing hypotheses, or extraordinarily sophisticated pattern completion that mimics the output of reasoning without the underlying mechanism, is a genuinely open and actively contested research question. Some research demonstrates that language models can solve novel problems that plausibly weren’t directly present in training data, suggesting real generalizable capability. Other research demonstrates that model performance degrades sharply when superficial surface features of a problem are altered while the underlying logical structure stays identical, suggesting heavier reliance on pattern-matching than robust reasoning.
The honest answer is that we don’t fully know yet. For decades, researchers believed reasoning had to be explicitly programmed through symbolic logic, rule by rule. Today’s language models weren’t built that way at all. Instead, they learned statistical relationships across an enormous amount of text. And somewhere inside those relationships, behaviors that resemble reasoning started to emerge on their own, which surprised almost everyone, including a lot of the researchers who built the models in the first place.
We already explored Turing’s 1950 reframing of “can machines think” in Issue #2, so I won’t revisit the Turing Test here. The only idea we need is this: we never observe intelligence directly, in humans or machines. We only ever observe behavior and infer the rest. That limitation applies equally in both directions, which is exactly why debates about machine understanding remain so difficult to settle.
The Transfer Problem
One of the biggest differences between humans and machines isn’t how well they solve problems. It’s how broadly they can transfer what they’ve already learned to something new.
Teach a child to stack blocks, and a week later that same child starts stacking books, then cups, then stones. Nobody taught these as separate tasks. The child recognized the underlying idea sitting underneath all of them: balance, support, gravity. The concept transferred on its own, the same way learning to ride a bicycle makes balancing on a motorcycle easier, or learning one programming language makes the second one come faster.
Machine learning has historically struggled here. A chess engine can’t suddenly play poker. A language model doesn’t automatically become an excellent robotic surgeon. Each new capability has typically required substantial additional training, though this specific gap is an area of active research effort, not a permanently fixed limitation. Meaningful progress has been made, even if fully general, human-like transfer remains unachieved.
Chollet argues that this specific capacity for efficient generalization to unfamiliar problems is the more meaningful marker of intelligence. Not solving today’s problems. Learning tomorrow’s. Maybe that’s exactly why children remain one of the greatest mysteries in artificial intelligence right now. Their knowledge is tiny. Their flexibility is enormous.
Creativity: Combination Versus Compulsion
For a long time, creativity was considered uniquely human. Then generative AI arrived. Today, machines create astonishing images, write novels, compose music, design products, and generate output many people find genuinely creative by ordinary standards.
Does that mean machines are creative? The answer genuinely depends on what creativity actually is. A lot of psychologists describe creativity not as producing something from nothing, but as combining existing ideas in genuinely new and useful ways. Under that specific definition, modern AI certainly demonstrates remarkable creative ability.
But human creativity often starts somewhere completely different. Curiosity. Emotion. Frustration. Wonder. A child builds imaginary worlds not because someone asked them to. Imagination itself is rewarding to a child, full stop.
Machines rarely possess anything resembling intrinsic motivation. They create because someone asks them to. Humans often create because they genuinely cannot help themselves. Maybe that’s one of the deepest remaining differences left standing. Not the ability to generate. The desire to generate in the first place.
What Still Separates Them
Intrinsic Motivation
Humans generate their own goals from internal states: hunger, curiosity, social bonds, awareness of our own mortality. Current AI systems pursue objectives specified externally, by human designers, training processes, or prompts.
When a child learns to walk, nobody hands them a reward function. Children stand up because exploration itself is rewarding, on its own, no external prize needed. When scientists spend decades studying black holes, they aren’t maximizing benchmark accuracy. They’re driven by pure curiosity. When an artist paints through the night, there may be no audience waiting, no prize, no external objective sitting at the end of it. The act itself is the point.
Current AI systems don’t wake up wanting to understand the universe. They don’t become fascinated by mathematics on their own. They respond. They optimize. They generate. Those are extraordinary achievements in their own right. They’re still different from intrinsic motivation.
This isn’t a claim about some permanent, unbridgeable gap. Some researchers are actively working on systems with something functionally resembling curiosity built directly into how they learn. Whether that ever becomes anything like human intrinsic motivation in a deeper sense, versus a clever engineering approximation of its outward behavior, remains genuinely unknown.
But today, right now, the difference is unmistakable. Humans pursue meaning. Machines pursue objectives. And humans, uniquely so far, know something real about their own capacity to get there.
Consciousness, Briefly
Consciousness deserves its own issue, so I’ll only make one distinction here. Intelligence concerns the capacity to solve problems and adapt behavior. Consciousness concerns subjective experience, whether there’s genuinely “something it is like” to be a given system.
Those aren’t the same thing. A chess engine displays remarkable intelligence in one narrow domain without any credible claim to consciousness. A thermostat makes decisions without experiencing temperature. Likewise, consciousness wouldn’t automatically imply extraordinary intelligence. They’re separate questions, and current evidence doesn’t let us confidently declare either one present or absent in machines. We’ll return to consciousness properly once we’ve built the technical foundation to engage with it seriously.
Are Machines Becoming Human?
This question comes up everywhere. It might also be the least interesting question we can possibly ask.
Airplanes never became birds. Submarines never became whales. Calculators never became mathematicians. Yet all of them surpassed biology in their specific domain anyway.
Maybe artificial intelligence will follow the exact same path. Not imitation. Innovation. Machines may eventually solve problems in ways evolution never once discovered. Just as airplanes completely ignore flapping wings, future AI may end up ignoring many features of human cognition entirely, while still surpassing us in completely different directions we haven’t even thought of yet.
The future of AI was never really a story about replacing the human mind. It’s a story about expanding the entire space of possible minds that could exist.
A New Way to Think About Intelligence
When this issue started, comparing humans and machines as competitors felt natural. Now that comparison feels incomplete.
Human intelligence is biological. Embodied. Emotional. Adaptive. Curious. Continuous. And, as Gardner and Sternberg both argued in their own ways, plural, made of several loosely related capacities rather than one single dial.
Machine intelligence is mathematical. Scalable. Precise. Distributed. Relentless.
Neither is inherently superior to the other. Each one reveals strengths the other one genuinely lacks.
Humans create meaning. Machines reveal patterns hidden inside unimaginable amounts of data.
Humans pursue meaning. Machines pursue objectives.
Maybe intelligence was never a single ladder leading straight upward. Maybe it’s a vast landscape with many different peaks scattered across it.
For thousands of years, humanity explored only one of them. Artificial intelligence has just begun climbing another.
Key Takeaways
1. Human and machine intelligence often produce similar results through fundamentally different underlying processes.
2. Human intelligence itself isn’t one thing. Spearman, Gardner, and Sternberg each proposed different, sometimes conflicting, structures for it, from a single general factor to eight independent types to a triarchic model of analytical, creative, and practical intelligence.
3. Human intelligence is shaped by evolution, embodiment, emotion, and lifelong adaptation. Machine intelligence emerges from mathematics, optimization, data, and raw computation.
4. Humans remain dramatically more sample-efficient learners in most everyday domains, though this specific gap is actively narrowing, not fixed permanently.
5. Whether current AI systems perform genuine reasoning or extremely sophisticated pattern completion that mimics reasoning’s output remains a real, unresolved question.
6. Embodied cognition, the idea that intelligence requires physical experience of the world, is a serious and evidence-backed position within cognitive science. It’s also still genuinely contested, not a settled fact.
7. Accurate perceived self-efficacy, our evidence-based judgment of our own capability, is strongly linked to human learning and performance, and has no real equivalent in AI systems today.
8. Intelligence and consciousness are separate concepts entirely, and current evidence doesn’t justify treating them as the same thing, or confidently declaring either one present or absent in machines.
Think Like a Researcher
Imagine two systems.
One can memorize every book ever written but cannot ask a single original question. The other begins knowing almost nothing, yet spends a lifetime asking why.
Which one would you call more intelligent?
There may not be a single correct answer. But notice something about how you just answered anyway. You almost certainly reached for a definition of intelligence, whether or not you stated it out loud, that quietly determined which system won.
That’s the actual lesson underneath this entire issue. Most debates about “is AI intelligent” aren’t really debates about AI at all. They’re debates about which definition of intelligence each person is silently using, and most people never stop to check whether they’re even arguing about the same thing.
Looking Ahead
Next Issue
Why Logic Came Before AI
Long before anyone had ever imagined neural networks or machine learning, philosophers asked a deceptively simple question.
Can thinking itself be written down?
If reasoning could be reduced to a set of rules, then maybe those rules could be followed by anyone. Or anything.
That single idea would become one of the greatest intellectual breakthroughs in human history. It would reshape philosophy, lay the foundations of mathematics, inspire modern computing, and centuries later, become the first serious blueprint anyone ever drew up for artificial intelligence.
Our journey now leaves psychology behind and enters the exact language in which intelligence was first formalized.
Logic.