Two Ways to Guess the Next Word

Finish this sentence: peanut butter and...

You didn't decide to think "jelly." It was just there, a fraction of a second before you could have stopped it. Your brain had already guessed the end of the sentence before you finished reading it.

That little reflex is the best place I know to start comparing the human brain with a large language model. Both are, at some level, machines that guess what comes next. That shared trick is why the comparison is so tempting, and why it's so easy to get wrong. The overlap is real. So is the gap. And the gap matters more for how we actually work with these systems than the overlap does.

I'm not a neuroscientist, and the science here is still moving. But I spend a lot of time helping people decide how much to trust these tools, and I've found that a rough, honest map of where brains and models rhyme, and where they don't, does more good than either the hype or the dismissal.

Where They Rhyme

Both run on prediction. A language model is trained to do one thing: given what came before, predict what comes next. Everything it can do grows out of that. A growing school of thought in neuroscience, often called predictive processing, argues that the brain works in a surprisingly similar way. It isn't passively receiving the world. It's constantly guessing what it's about to see, hear and feel, then correcting itself when it's wrong. Perception, on this view, is less like a camera and more like a running forecast.

The overlap goes further than metaphor. Researchers have found that the internal activity of language models can predict how people's brains respond to the same sentences, and that the brain, like the model, seems to anticipate upcoming words before they arrive. Nobody is claiming the mechanism is the same. But something about next-word prediction appears to capture a real piece of how we process language.

Both store knowledge as patterns, not files. There's no single neuron that holds your grandmother's face, and no single number in a model that holds the capital of France. In both, knowledge is spread across enormous networks of connections and comes out of how those connections are weighted. That's why both are good at generalizing and bad at telling you exactly where a piece of knowledge came from.

Both reconstruct instead of retrieve. We tend to imagine memory as playback. It isn't. Psychologists have known for nearly a century that people rebuild memories each time they recall them, filling gaps with what seems plausible. That's why eyewitness testimony can be so confident and so wrong. A language model does something similar. It doesn't look up an answer. It generates one, word by word, from patterns. When the pattern is strong, the result is accurate. When it's weak, the result is still fluent.

Both make things up and believe them. In famous studies of split-brain patients, one half of the brain would act on information the other half never saw, and the verbal half would immediately invent a reasonable-sounding explanation for why. Not lying. Just narrating. The brain, it turns out, has a part that is very good at producing a story that fits, whether or not the story is true. If that sounds like what we call hallucination in AI, it should.

That's a lot of rhyme. It's why talking to a good model can feel so uncannily like talking to a person.

Where They Split

Now the differences. These are the ones I think matter most.

The brain has a body. The model has text. Everything you know is tied, eventually, to a world you've touched. You learned "heavy" by lifting things and "hot" by getting burned. Linguists like George Lakoff and Mark Johnson spent careers showing how even our most abstract thinking leans on bodily metaphors: we grasp ideas, we feel close to people, we look forward to things. A language model learns from the record of all that experience, not from the experience itself. It knows how people talk about heat. It has never been warm. Sometimes that difference disappears in the output. Sometimes it's the whole story.

Language is one part of the brain. In a model, it's all there is. This one surprised me. Neuroscientists like Ev Fedorenko have shown that the brain has a fairly specific network for language, and that it's distinct from the systems we use for reasoning, math, planning and understanding other people. People who lose much of their language ability can still solve problems, play chess and get through their day. In the brain, language is a powerful interface sitting on top of other kinds of thinking. A language model is built the other way around. Language is the foundation, and whatever reasoning it can do has to be pulled out of language itself. That helps explain why models can be brilliant with words and still stumble on things a child would find obvious.

Brains learn from a trickle. Models learn from a flood. A child hears somewhere in the tens of millions of words by the time they're fluent. Large models are trained on trillions. That's a difference of many orders of magnitude. Children get there with far less because they aren't just absorbing text. They're pointing, touching, asking, being corrected, and watching faces. The brain is extraordinarily efficient with experience. Models make up for inefficiency with scale.

Brains never stop learning. Models are mostly frozen. Your brain is rewiring itself as you read this sentence. Every conversation changes you a little, and sleep does quiet work consolidating what mattered. A deployed model, by contrast, mostly stops learning when training ends. It can hold information in a conversation, and it can be given documents to read, but the underlying network doesn't change from talking to you. It meets every new day as the version of itself it was when it shipped.

Brains run on a sandwich. Models run on a power plant. The human brain runs on roughly twenty watts, about what a dim light bulb uses. Training and running large models takes data centers. Evolution had a few hundred million years to optimize for efficiency. We've had a few years.

Brains want things. This is the difference I think about most. Your brain predicts in service of something: staying alive, staying connected, getting what you need, avoiding what hurts. It has stakes. It feels the cost of being wrong. A language model predicts because predicting is what it was trained to do. It doesn't care whether its answer helps you, embarrasses you or costs you money. Any care you see in its output came from the data and the training, not from anything it stands to lose. That isn't a flaw to be fixed. It's a fact to design around.

The Wrong Question

People keep asking whether AI thinks like we do. I've come to believe it's the wrong question, or at least not the useful one.

The brain and the model share a strategy. They don't share a situation. One is a predictor inside a body, inside a life, with something to lose. The other is a predictor inside a vast library, with nothing to lose at all. They will often reach the same sentence. They get there for very different reasons, and they fail in very different ways.

The more useful question is: given what each one is, what should each one do?

This is where the mirror and the oracle come back. If you treat a model as an oracle, you're treating it as if it had the things it lacks: grounding in the world, stakes in the outcome, a memory of what happened last time. You hand it judgment it isn't built to carry. If you treat it as a mirror, you use it for what it's actually extraordinary at. Holding a huge amount of language in view at once. Finding patterns. Offering ten framings in ten seconds. Showing you your own reasoning laid out in words, so you can see where it's thin. Then you bring what only you have.

What This Means for How We Work

Once you see the split clearly, the division of labor almost writes itself.

Let the model widen. Let people narrow. Models are excellent at generating options, surfacing patterns and drafting quickly. People are better at choosing, because choosing requires knowing what matters, and knowing what matters requires stakes.

Bring the world to the model. A model only knows the world secondhand. The more real context you give it, your actual data, your actual customers, the thing that happened in the meeting yesterday, the closer it gets to being useful. The quality of the output is capped by the reality of the input.

Check fluency, especially when it sounds right. Both brains and models confabulate. The difference is that we've had thousands of years to build habits for catching each other's stories: evidence, debate, peer review, the coworker who asks "how do you know?" We need the same habits for machine output, and they need to be strongest exactly when the answer is most convincing.

Keep the learning in people. The model you use tomorrow will mostly be the same model you used today. Your team won't be, if you let them learn. Every project should leave people sharper, not just leave a folder of outputs behind. The organization that hands all its thinking to a frozen system stops learning at the same moment the system did.

Same Trick, Different Animal

The brain and the language model both guess the next word. That's a real and genuinely strange discovery, and I don't think we've fully absorbed what it means. It suggests that a surprising amount of what looks like intelligence can come out of prediction alone.

But prediction alone is not a life. The brain predicts from inside a body that can be hurt, a history that can be remembered, and relationships that can be lost. That's where judgment comes from. Not from the guessing, but from what's riding on the guess.

So use the model for what it is: the most powerful pattern engine any of us has ever had access to. And stay clear about what you are: the part of the system that actually has something at stake.

That isn't the weaker role. It's the one that decides what all the guessing is for.