It's just probability.

There is a confident objection, and you have heard it: we cannot trust generative language models, because they only produce the statistically likely next token. They do not know; they guess with high resolution. A human, the argument implies, is a different kind of thing.

The objection is half right, and the half it gets wrong is the half that matters. A language model is, mechanically, a probability distribution over what comes next. But the second clause, that this is what separates it from a human mind, does not survive contact with the last forty years of cognitive neuroscience. The brain is also, mechanically, a probabilistic prediction engine. The senses in which the model is "just probability" are, one by one, senses in which you are too.

This is not the claim that models are as trustworthy as people. It is a narrower and harder claim: "it's just probability" is not a difference between them. It is the thing they share. If probabilistic computation disqualified a system from being trusted, it would disqualify human cognition first, because we have far better instruments for reading the machine's uncertainty than for reading yours.

01The brain is a prediction machine

The dominant framework in computational neuroscience is not that the brain records the world and reacts to it. It is that the brain continuously predicts its next input and corrects itself by the error. This is predictive coding¹, generalised as the free-energy principle² and the Bayesian-brain program⁴, and synthesised for a wide audience by Clark³. On this account perception is a controlled hallucination held in check by prediction error: the brain runs a generative model and reconciles it against the signal.

That already rhymes with a language model. But the rhyme is now measured, not asserted. When researchers record human cortex during natural language and compare it to a transformer, they find the brain performing the same operations a large language model performs: predicting the next word before it arrives, and registering surprise proportional to how unexpected it was⁵. Model internal representations predict human neural responses to language better than any prior theory, and the models that predict the brain best are the ones best at next-word prediction⁶. Brains and algorithms, in the careful phrasing of one study, partially converge⁷.

02The mechanical correspondence, made explicit

For the reader who builds these systems, here are the two substrates side by side, mechanism by mechanism, so the analogy is not hand-waving.

• The distribution is literal. A transformer's final layer emits a vector of logits, one number per token; softmax turns them into a probability distribution; sampling draws from it. "It only predicts the likely next token" is a plain description of softmax over logits. The brain's homolog is population coding: a belief is a pattern of firing rates across many neurons, read out downstream. A distribution, not a stored symbol.
• Attention is retrieval. Self-attention is soft, content-addressable lookup: each position forms a query, matches it against keys, and reads back a weighted blend of values. That is associative memory, the same operation the hippocampus performs when a partial cue completes a pattern.
• The decision is a race with a temperature. Sampling temperature scales the logits before softmax: low temperature sharpens toward one winner, high temperature keeps competitors alive. That is the threshold and noise of a sequential-sampling decision. Lower either and the race commits sooner. The number of tokens still carrying real weight at a given step is a live measure of how open the race is, readable directly from the model's output.
• Learning is error times activity. Backpropagation adjusts each weight in proportion to how much it contributed to the error. The brain's plasticity rules combine pre-synaptic activity, post-synaptic activity, and a global error or reward signal, and they compute the same shape of update by other means.

None of this is a metaphor laid over the model. It is what the model is. And each line has a biological counterpart that is not a metaphor either.

03Your certainty is a readout, not a fact

The intuition behind the objection is that a human belief is a discrete, held thing, true or false, present or absent, while a model only ever has probabilities. But this is not how human decisions are made either. In the sequential-sampling tradition that has dominated mathematical psychology since Ratcliff⁸, and in its leaky, competing-accumulator refinement⁹, a decision, including a judgement of true or false, is the outcome of noisy evidence accumulating toward a threshold. The verdict you experience as crisp is the readout of a continuous, probabilistic race between competing options. Change the noise, the starting point, or the threshold, and the same evidence yields a different verdict.

There is no hard binary underneath. What feels like knowing is a competition that happened to finish where it did, at the confidence it happened to reach. A model's logits are not a lesser substitute for a human's certainty. They are the same kind of object, exposed. The difference is that we can read the model's distribution directly and can only infer yours.

04You confabulate too

The failure people fear most in language models is the confident fabrication, the "hallucination," the fluent detail that was never true. It is presented as the model's signature defect. It is also, precisely, how human memory works.

You do not retrieve a stored record. You reconstruct a plausible one, and the reconstruction is editable by the question you were asked: change the verb from "hit" to "smashed" and people remember faster speeds and broken glass that was never there¹⁰. Worse, the act of remembering rewrites the memory. A recalled trace becomes labile and has to be re-stored, so each recall is a chance to overwrite it¹¹. What you keep is not the experience but a summary skewed toward its emotional peak and its end¹², with emotionally arousing moments written deeper than the rest¹³. A human memory is a probable reconstruction presented with confidence. The word for that, when a model does it, is "hallucination."

05Even the future is baked into the weights

Consider something that feels like reasoning rather than reflex: valuing a distant outcome less than a near one. It looks like a calculation over time. But a transformer has no clock, only position, and, it turns out, neither does the brain: there is no central timekeeper, only distributed, emergent timing¹⁶¹⁷. A clock-free substrate cannot compute value times elapsed time. It has one option left: bake the discount into the weights.

That is what the evidence shows on both sides. A distant reward is realised less often, so experience depresses the learned weight on distant-reward routes, and histories drawn from an uncertain hazard rate reproduce the classic hyperbolic discount curve¹⁴¹⁵. In humans, uncertainty about whether the future will pay out generates the same hyperbolic discounting¹⁸; in the brain, a single subjective-value signal tracks discounted value with no separate "future" module¹⁹. Two substrates, no clocks, the same solution: probability structure written into the connection strengths.

06Even learning is prediction error, all the way down

The last redoubt is learning: surely a brain learns in some richer way than gradient descent on a prediction loss. But the brain's own teaching signal is a prediction error. Dopamine neurons encode the difference between predicted and received reward and gate plasticity on it²⁰, a biological reward-prediction-error, the same quantity a model minimises. The one clean disanalogy, that the brain cannot run backpropagation as written²¹, is a fact about implementation, not about kind: the brain approximates error-driven credit assignment by other means, and the gap has been narrowing for a decade. Strip the wetware and the functional story is the same on both sides: adjust the weights in proportion to how wrong the prediction was.

Many fear we overvalue language models. The better-supported worry, the one rarely voiced, is that we overvalue humans.

07What actually separates them (it isn't this)

None of this makes a language model trustworthy. It makes the argument untrustworthy. "It's just probability" cannot be the reason to distrust a model, because it is equally true of the mind you are trusting instead. If you want to argue that a given model is less reliable than a given person on a given task, the real grounds are specific and measurable:

• Calibration. Does stated confidence track actual accuracy?
• Error structure. Are the mistakes independent and detectable, or correlated and confident?
• Verifiability. Can the output be checked against a ground truth?
• A persistent, accountable self. Is there a continuous agent that carries consequences across time, or only a fresh instance each call?

These are the differences that bear on trust. Notice that every one of them is measurable, task-relative, and, on the first three, often more favourable to the machine, whose distribution you can read directly and whose calibration you can score, while a human's confidence is a black box you infer from behaviour.

So the objection inverts. "It's just probability" was offered as the property that makes a model untrustworthy and a human safe. It is neither. It is the substrate both of them run on, and of the two, the model is the one whose probabilities you can actually see. The fear that we overvalue language models is reasonable. It should be paired with the harder, better-evidenced fear the objection quietly depends on ignoring: that in the comparison, we systematically overvalue ourselves.

08What is still open

The fourth difference above, a persistent accountable self, is the one that does not dissolve into "you do it too." Here is exactly why. A deployed transformer consolidates only inside its context window; it cannot carry experience from one session to the next, so it has no continuous self that accumulates over time. A brain does. But that is a difference in what has been installed and deployed, not a barrier the substrate cannot cross: give a model persistent memory and online consolidation, and the current gap narrows. Whether that yields a bearing self, one that grounds valence, ownership, and effort, or merely retrieval-augmented continuity, is an open, falsifiable question. It is a prediction this program commits to, not a result it reports.

Three smaller threads stay open, and they deserve to be unfolded, not just named.

The first is the convergence from section 01. The direction is solid: brain and model both predict the next word, and both react to surprise. What is contested is how far the resemblance goes. A model's internal patterns predict neural responses better than any prior theory, but predicting the activity is not the same as proving the brain computes it in exactly the same way. Part of the resemblance may come from both being shaped by the same patterns in language. But even then the parallel holds: the two systems solve the same task the same way. What is open is not whether they run in parallel, but whether the resemblance goes all the way down into the computation itself.

The second is learning, from section 06. The brain cannot run backpropagation exactly as the algorithm is written; it cannot carry the same weight backward through the network the way the method requires. But the space is not empty: there are several active proposals for how the brain reaches the same end by other means (feedback alignment, predictive coding, target propagation), and the gap between them and true gradient learning has narrowed over a decade. What is open is not whether the brain does something backprop-like, but which of these it actually uses.

The third is the smallest of the three: the fast, involuntary advantages a human is born with, the old registers of threat and value that answer before you can think. They may be a timing difference a base model lacks, because it has never had a body or a past that had to survive. Whether that can be installed, or needs a different substrate, we do not yet know.

Unfolding them is the point, not a hedge: an argument that shows exactly where the line falls is harder to dismiss than one that leaves its edges in fog.