Imagine a rat standing at a fork in a maze. To the left: water. To the right: food. The rat pauses. It looks left, glances right, hesitates — then makes its choice.
In that tiny moment of hesitation lives one of the most profound secrets of biological intelligence. And understanding it may hold the key to building AI that truly thinks, rather than just predicts.
Let’s explore what neuroscience tells us about how mammals make decisions — and why it matters enormously for the future of artificial intelligence.
The Pause That Changes Everything
When a mammal faces a decision that isn’t straightforward — when the choice is genuinely close, or something unexpected occurs — something remarkable happens in the brain. The animal pauses.
This pause is not confusion or failure. It is a deliberate, elegant biological mechanism called vicarious trial and error. The brain essentially runs a private simulation before committing to real-world action.
Think of it like this: instead of just acting and hoping for the best, the brain briefly “plays the movie” of each possible choice — all without the animal taking a single step.
This three-step process unfolds in the agranular prefrontal cortex (aPFC) — the front region of the brain most associated with planning and complex decision-making.
Step 1: Uncertainty Triggers the Simulation
The aPFC is constantly monitoring the world. In normal circumstances, when everything is unfolding as expected, it stays quiet. The brain runs on autopilot — what scientists call model-free learning — making quick decisions based on habit and past reward.
But the moment something unexpected happens, or a decision becomes genuinely difficult, the aPFC springs into action. It detects a critical signal: disagreement.
Different clusters of neurons — called columns — in the aPFC are each making their own predictions about what should happen next. When those predictions align, the brain proceeds confidently. When they disagree — “turn left for water” versus “turn right for food” — the level of disagreement is read as a measure of uncertainty.
Uncertainty is the brain’s trigger for simulation. The aPFC sends a signal that effectively pauses ongoing behaviour, buying time for something more sophisticated to happen. The rat freezes at the fork — and begins to imagine.
Step 2: The Brain Simulates Its Options
Now the real magic begins. Rather than running through every conceivable action — the rat could, theoretically, do a million things — the brain is efficient. It only simulates the options it was already considering.
This is elegantly economical. The aPFC doesn’t start from scratch. It takes the predictions its own columns are already generating and plays them out, one by one, like short mental films.
One group of neural columns plays out the scenario: turn left, follow the corridor, reach the water. Another group plays out: turn right, follow the corridor, reach the food.
Fascinatingly, experiments with rats in mazes have shown that during this process, the aPFC and the sensory cortex become synchronised in a unique way. One scientific interpretation is that the aPFC is essentially asking the sensory cortex to render each simulation — to create a vivid mental experience of what each path would feel like — and the sensory cortex responds, step by step, all the way to the imagined goal.
It is, in every meaningful sense, the brain running its own internal virtual reality.
This process closely mirrors how AlphaZero — the AI system built by DeepMind to master chess and Go — approaches decision-making. AlphaZero does not evaluate every possible move on a board. It plays out the moves it already predicts are most promising, exploring those branches deeply before choosing. The biological brain and the AI system arrived at strikingly similar solutions to the same problem: how to make a good decision without exhausting infinite possibilities.
Step 3: The Brain Votes — and Chooses
Once the simulations are complete, the decision still needs to be made. So who casts the final vote?
That role belongs to the basal ganglia — an ancient part of the brain, present even in early vertebrates, whose core function is to accumulate signals for competing actions until one crosses a threshold and wins.
Picture it like an election happening in real-time inside the brain. Multiple populations of neurons are each advocating for a different action, growing more and more excited, until one side surpasses the choice threshold — and that action is selected.
Here is the beautiful part. During simulation, the basal ganglia does not know whether it is experiencing a real event or an imagined one. It simply responds to the signals it receives. If imagining going left to drink water releases more dopamine — the brain’s reward signal — than imagining going right to eat food, then the “vote for left” accumulates faster. The threshold is crossed. The rat turns left and heads for water.
The net result is that the simulation has vicariously trained the basal ganglia. The brain rehearsed the decision in imagination, reached a confident conclusion, and then acted — as if the choice had already been made.
Why Does This Matter for AI?
Here is the insight that connects this beautiful neuroscience to the world of artificial intelligence.
Today’s most powerful AI systems — large language models like ChatGPT, Claude, or Gemini — are extraordinary at pattern recognition. They predict the next most likely word, image, or action based on statistical patterns absorbed from vast amounts of training data. They are, as I often describe them, very sophisticated autocomplete machines.
But they do not pause at the fork in the maze. They do not experience genuine uncertainty, trigger simulations, and run mental experiments before committing to an answer. They do not accumulate votes through imagined experience. They do not have a basal ganglia checking whether the dopamine of one imagined path outweighs another.
In other words: they lack the biological architecture of decision-making under uncertainty.
This is precisely why many researchers — myself included — believe that the path to more general, more trustworthy artificial intelligence does not run through simply scaling up existing models with more data and more compute. It runs through understanding mechanisms like causality, simulation, and model-based reasoning — the same mechanisms that allow a rat to make a wise choice at a fork in a maze.
The brain is not just a pattern-matcher. It is a world model — a system that can imagine futures it has never seen, test them in a safe internal space, and use those imagined experiences to guide action in the real world.
Building AI systems that do the same is one of the most important open challenges in the field today.
The Takeaway for Curious Minds
You do not need to be a neuroscientist to appreciate what this tells us. Every time you hesitate before making a difficult decision, every time you mentally “play out” a scenario before acting — you are doing exactly what that rat does in the maze. You are running your aPFC, triggering simulations, and letting your basal ganglia vote. The difference between biological intelligence and today’s AI is not just about size or speed. It is about this capacity for purposeful imagination — the ability to pause, simulate, and choose.