Scientists of the late 19th century described animal behaviour in terms of “satisfaction” and “discomfort” – concepts that eventually evolved into our modern understanding of “reward” and “punishment.” These fundamental principles now underpin much of artificial intelligence research, where reward mechanisms drive learning algorithms toward intelligent behaviour.
Consider the humble Nematode C. elegans, whose mere 302 neurons contain sophisticated neuromodulator systems that generate primitive emotional states. Among these chemical messengers, dopamine and serotonin stand as critical regulators of behaviour.
When hunger strikes, dopamine neurons become hypersensitive to food-related stimuli across species, from nematodes to rats, cats or humans. Laboratory experiments demonstrate this dramatically: rats with elevated dopamine transform into impulsive reward-seekers, frantically consuming food and attempting to mate. Conversely, increasing serotonin produces the opposite effect – reducing consumption, tempering impulsivity, and enhancing the ability to postpone immediate gratification.
Despite its popular reputation as the “pleasure chemical,” dopamine’s true role is more nuanced. Rather than signalling enjoyment itself, dopamine mediates anticipation—it’s not about liking, but wanting. This distinction becomes clear when observing neural activity across species: a nematode’s dopamine neurons fire not during feeding, but when the organism detects food nearby. Similarly, a rat’s brain releases dopamine upon encountering cues that predict rewards, not during consumption itself.
This anticipatory mechanism forms the foundation of associative learning, a phenomenon Ivan Pavlov famously demonstrated when he received the Nobel Prize on December 12, 1904. His experiments with dogs revealed how digestive responses could be triggered by stimuli, such as buzzers or lights, that preceded feeding. The dogs’ involuntary salivation in response to these conditional cues suggested something profound: the capacity for learning through association might be far more evolutionarily ancient than scientists had previously recognised.
Bilaterians—animals with bilateral symmetry—possess the ability to learn associations, while radially symmetric creatures like anemones and corals lack this capacity. This evolutionary breakthrough emerged alongside primitive emotions, enabling living organisms to recalibrate their perception of positive and negative stimuli.
Yet associative learning presents a challenge: reality rarely offers clean, singular predictive cues. When a predator appears or food becomes available, countless environmental signals occur simultaneously. How does a brain determine which cue deserves credit for predicting the outcome? Neuroscientists call this the credit assignment problem, and evolution has engineered four elegant solutions.
Eligibility traces create narrow temporal windows—a slug forms associations only when a tap precedes a shock by less than two seconds. This temporal precision ensures that only immediately preceding cues get linked to outcomes. Overshadowing prioritises dominant signals, allowing the most salient cues to form the strongest associations while weaker ones fade.
Latent inhibition prevents familiar background stimuli from forming new associations—essentially filtering out the constant noise of everyday experience. Finally, blocking prevents redundant learning: once a slug associates taps with shocks, it won’t form additional associations between that shock and simultaneous stimuli like temperature or chemical changes.
These four mechanisms constitute the fundamental architecture of associative learning, embedded within the microscopic machinery of neurons and neural circuits across the animal kingdom.
Learning physically reshapes the brain’s architecture. Synapses strengthen or weaken. New neural connections form while obsolete ones disappear.
There is a clever protein machinery in the synapses of bilaterian neurons that detects whether the input neuron fired within a similar time window as the output neuron. This microscopic machinery can sense when an input neuron (perhaps triggered by a gentle tap on your skin) activates in close temporal proximity to an output neuron (response to a mild shock). When these neurons synchronise their activity, the molecular machinery initiates a cascade that reinforces their connection. This remarkable mechanism bears the name of Donald Hebb, the visionary psychologist who eloquently hypothesised its existence in the 1940s, several decades before science confirmed his insight. Today, neuroscientists talk about Hebbian learning: “Neurons that fire together wire together.”
The capacity to learn emerged with the first primitive brains. Early bilaterians such as the nematode C. elegans possessed only rudimentary learning abilities—they could form associations between events occurring within seconds of each other, but lacked the neural machinery for timing prediction, pattern recognition, or spatial navigation.
Yet our human capacity to forge neural connections isn’t some evolutionary innovation unique to our species. Rather, it’s an inheritance from a common ancestor that swam Earth’s ancient seas over half a billion years ago.
The cognitive feats that define us—our mastery of language, our recognition of faces, our appreciation of music, our mental mapping of space—all rest upon these primordial learning mechanisms.
As brains evolved beyond those first bilaterians, nature didn’t reinvent the fundamental synaptic mechanisms of learning; instead, it found increasingly sophisticated applications for them. By the time vertebrates like fish appeared, learning had become the brain’s central purpose.
When animals evolved to seek out beneficial stimuli and avoid harmful ones, their survival hinged on continuously refining these judgments through experience. Modern AI systems lack this fundamental capacity for ongoing adaptation. Consider how self-driving vehicles don’t improve their performance with each mile travelled. The facial recognition software on our devices doesn’t sharpen its accuracy with daily use. Even the conversational prowess of ChatGPT might create an illusion of learning, but the underlying large language model remains static, unable to incorporate feedback from its millions of interactions.
By 2025, the typical AI deployment follows a one-way journey: systems undergo intensive training before being released into operation, after which their knowledge remains frozen in time. This fundamental limitation creates a critical vulnerability. When reality shifts beyond the boundaries of their training data—as it inevitably does—these systems cannot adapt without human intervention and complete retraining. The contrast with biological intelligence couldn’t be starker; while AI systems operate with static knowledge snapshots, brains never stop learning, continuously recalibrating their understanding of the world with each new experience.