Here is a question that sounds simple but leads somewhere surprisingly deep: why do you reach for your phone the moment you wake up, even when you know you’d be better off not doing it?
The answer involves a lever, a sick rat, a Cambridge psychologist, and one of the most important ideas in both neuroscience and artificial intelligence. Let’s explore.
The Experiment That Changed How We Think About Behaviour
In the early 1980s, a psychologist named Tony Dickinson at Cambridge University was running a deceptively straightforward experiment. He trained rats to push levers to release food pellets. The rats learned quickly — push the lever, get food, feel happy. A classic reward loop.
Then Dickinson introduced a twist. He laced the food pellets with a chemical that made the rats feel mildly ill. Not dangerously sick — just unpleasant enough to make the food distinctly unappealing. After recovery, the rats were given a pile of the same pellets and, as expected, they ate far fewer than before. The food had lost its appeal.
But here came the fascinating question: what would happen the next time the rats saw the lever?
Two predictions were possible. If rat behaviour is purely mechanical — governed by what psychologists call Thorndike’s law of effect — the rats should still rush to the lever. After all, pushing the lever had been rewarded hundreds of times, and nothing had yet “unrewarded” the act of pushing itself. The lever, in that mechanical view, is still a good lever.
But if rats are capable of something more sophisticated — if they can mentally simulate the consequence of pushing the lever (imagining the food that now makes them sick) — they would hesitate, or stop pushing altogether.
Dickinson’s finding was clear: rats who had experienced the food paired with illness pushed the lever almost 50% less than rats who had not. The rats were not just responding to a stimulus. They were imagining the outcome — and choosing accordingly.
This was powerful evidence that even simple mammals can vicariously simulate future choices and adjust their behaviour based on imagined consequences, not just past rewards.
When Thinking Stops: The Discovery of Habits
Dickinson pushed the experiment further — and made an even more surprising discovery.
He varied how many times the rats had performed the lever-pressing task before the food was devalued. Rats that had pushed the lever around 100 times behaved smartly: once the food was made unpleasant, they stopped pushing. They had simulated the outcome, recognised the problem, and updated their behaviour.
But rats that had pushed the lever 500 times? They kept pushing — with the same energy as before — even though they knew the food would make them sick.
What had happened? The behaviour had become a habit.
After hundreds of repetitions, something fundamental had shifted in the brain. The action was no longer connected to its goal. It had become an automated motor response — triggered directly by the sight of the lever, bypassing the part of the brain responsible for planning, simulation, and conscious decision-making.
Dickinson called this phenomenon insensitivity to outcome devaluation: the behaviour had been repeated so many times that the brain no longer bothered to check whether the outcome was still desirable. It just… acted.
Two Systems Living in One Brain
This discovery revealed something profound about how mammalian brains — including human ones — are organised. There are not one but two distinct modes of decision-making operating inside us at all times:
Mode 1 — Goal-Directed (Slow, Thoughtful)
The brain pauses, simulates options, considers consequences, and makes a deliberate choice. This is driven by the prefrontal cortex — the brain’s planning centre. It is flexible, intelligent, and accurate — but costly in time and energy.
Mode 2 — Habitual (Fast, Automatic)
The brain recognises a familiar trigger and fires an automatic response — no simulation, no deliberation, no goal-checking required. This is driven by the basal ganglia — an ancient, efficient part of the brain. It is fast and effortless — but inflexible and sometimes wrong.
You will recognise this duality immediately. It appears across many fields under different names:
| Field | Slow, Thoughtful Mode | Fast, Automatic Mode |
| Neuroscience | Goal-directed behaviour | Habitual behaviour |
| AI | Model-based learning | Model-free learning |
| Behavioural Economics | System 2 (Daniel Kahneman) | System 1 (Daniel Kahneman) |
If you have read Daniel Kahneman’s landmark book Thinking, Fast and Slow, you already know System 1 and System 2. Now you know the biological machinery behind them.
Neither Mode Is Better — Both Can Fail
Here is the nuance that matters most: neither system is superior. Both exist for good reasons.
Habits are not flaws. They are one of evolution’s most elegant solutions to the problem of limited cognitive energy. Imagine if you had to consciously deliberate every single action — tying your shoes, making coffee, driving a familiar route. You would be exhausted before lunchtime. Habits free your brain’s expensive planning resources for the decisions that actually matter.
But habits become dangerous when they fire in the wrong context — when the world has changed but the brain has not updated. The rat pushing the lever for food that now makes it sick is a perfect metaphor for every human behaviour we continue out of pure automation, long after it has stopped serving us.
The brain attempts to manage this balance intelligently — deploying habits when a situation is familiar and stable, and triggering deliberate simulation when something is uncertain, new, or important. But it does not always get the balance right. And when it gets it wrong, the result is the full spectrum of irrational human behaviour: from reaching for the phone on autopilot to repeating patterns in relationships, business, or policy that everyone can see are no longer working.
What This Means for Artificial Intelligence
This is where the neuroscience connects directly to AI literacy — and why it matters for all of us.
Today’s dominant AI systems — large language models and most machine learning tools — operate almost entirely in model-free mode. They learn statistical patterns from enormous amounts of data and respond to inputs automatically, without pausing to simulate consequences, check goals, or imagine alternative futures.
In other words, they are extraordinarily capable habit machines. Fast, fluent, efficient — and, like the 500-push rat, sometimes confidently wrong in ways they cannot self-correct.
Building AI systems that can also operate in model-based mode — that can pause, simulate, reason about consequences, and update goals in light of new information — is one of the defining challenges of the next generation of AI research. It is the difference between a system that can perform a task and one that can understand why it is doing it.
The rat at the lever, in its tiny moment of hesitation, was doing something that our most powerful AI systems still cannot reliably do. It was asking: “If I do this, what will happen — and do I still want that?”
The Takeaway
Our brains are not single engines. They are a collaboration between an ancient automatic system that acts on instinct and a more recent, deliberate system that pauses to imagine and plan. The interplay between these two modes — habits and goals, fast and slow — shapes every choice we make.
Understanding this is not just fascinating neuroscience. It is a window into why AI systems behave the way they do, what they are currently missing, and what the next breakthroughs in machine intelligence might look like.
The most important question in AI today is not just “how do we make models bigger?” It is “how do we teach machines to pause — and wonder?”