Key facts
- Where it comes from
- A small number of midbrain neurons: the ventral tegmental area and substantia nigra pars compacta
- What it signals
- Reward prediction error: the gap between what was expected and what occurred
- What it does not signal
- Pleasure. A fully predicted reward produces no dopamine response at all
- Wanting versus liking
- Dopamine drives wanting; liking depends more on opioid and endocannabinoid systems
- Three pathways
- Mesolimbic (motivation), mesocortical (cognition), nigrostriatal (movement)
Where dopamine comes from, and where it goes
Before the argument, the anatomy, because half the confusion about dopamine dissolves the moment you know where it is made and where it is delivered.
Dopamine in the brain comes from a remarkably small number of cells. They sit in the midbrain, in two adjacent regions: the ventral tegmental area (VTA) and the substantia nigra pars compacta, the latter named for its dark pigmentation, visible to the naked eye in a dissected brainstem. In the human brain these populations number in the hundreds of thousands, against a total of tens of billions of neurons. It is a rounding error of a system, and it exerts an influence out of all proportion to its size, because each of its neurons sends an enormously branched axon to a vast number of targets. Dopamine is a broadcast system, not a point-to-point one.
Where those axons go determines what dopamine does. There are three pathways worth knowing, and they do genuinely different jobs. Conflating them is the second most common dopamine error, after the pleasure story.
Mesolimbic
From the ventral tegmental area to the nucleus accumbens and surrounding ventral striatum. This is the pathway of reward learning and of motivation: the drive to pursue, the willingness to expend effort, the pull that a cue exerts once it has come to predict something good. It is also the pathway most heavily implicated in addiction, because every drug of abuse, by one route or another, raises dopamine here.
Mesocortical
From the ventral tegmental area to the prefrontal cortex. Here dopamine supports working memory, cognitive flexibility and the maintenance of goals. Its relationship to performance is famously an inverted U: too little dopamine and prefrontal representations are unstable, too much and they become rigid. This is one of the reasons the naive idea of "more dopamine is better" is wrong even on its own terms.
Nigrostriatal
From the substantia nigra pars compacta to the dorsal striatum, part of the basal ganglia. This pathway is essential to the initiation and vigour of movement and to the learning of habits. Its progressive degeneration is what produces the motor signs of Parkinson's disease: the slowness, the rigidity, the tremor, the difficulty starting a movement at all.
D1-like and D2-like
Dopamine acts on G-protein-coupled receptors sorted into two families with opposite intracellular effects: D1-like receptors, which raise cyclic AMP, and D2-like receptors, which lower it. The same molecule therefore does opposite things depending on which receptor it lands on. In the striatum, D1 and D2 receptors are expressed on largely separate populations of output neurons, which is why dopamine does not simply turn the striatum up or down but rebalances it.
The nigrostriatal pathway alone should give the pleasure theory pause. If dopamine were the pleasure chemical, why is the great majority of it devoted to a circuit whose principal job is the control of movement, and whose failure produces a disorder of walking?
How the pleasure story got started
Bad ideas rarely come from nowhere. The pleasure theory of dopamine came from two respectable places, and understanding why they were misread is instructive.
The self-stimulation experiments. In 1954, James Olds and Peter Milner implanted electrodes in the brains of rats and gave the animals a lever that delivered a brief electrical stimulation to the site. With electrodes in certain regions, the rats pressed the lever with extraordinary persistence, returning to it again and again, in some conditions to the neglect of other activities. The regions in question turned out to overlap with dopaminergic pathways. The interpretation offered, and swiftly popularised, was irresistible: the electrode had been placed in the pleasure centre, and the rats were pressing for pleasure.
Notice what the experiment actually shows, and what it does not. It shows that stimulation of these regions is powerfully reinforcing: it makes the animal do the thing again. It does not show that the animal is experiencing pleasure, because a rat cannot tell you, and "it kept doing it" is compatible with the animal wanting something intensely without enjoying it at all. That distinction sounds like philosophy. It is not. It is the distinction that the next fifty years of research turned out to hinge on.
The drug findings. The second source was pharmacology. Cocaine blocks the dopamine transporter, so dopamine that is released is not cleared and accumulates. Amphetamine forces dopamine out of the terminal and reverses its transporter. Nicotine, opioids and alcohol raise dopamine in the nucleus accumbens by various indirect means. The pattern is consistent, and it looked like a proof: addictive drugs feel good, addictive drugs raise dopamine, therefore dopamine is what feeling good is made of.
The logical hole: the inference from "drugs of abuse raise dopamine" to "dopamine is pleasure" is only valid if the reason drugs are taken compulsively is that they are pleasurable. That premise is exactly what the evidence has since undermined. Long-term addicts routinely report that the drug has stopped being enjoyable while the craving for it has grown, which is an incoherent report if wanting and liking are the same thing. The pharmacology did not prove that dopamine is pleasure. It proved that dopamine is deeply involved in compulsive pursuit, which is a different claim and, as it turns out, the true one.
So the pleasure story was a reasonable first guess from two suggestive lines of evidence. The reason to abandon it is not that its founders were foolish. It is that a better experiment came along, and the pleasure story failed it outright.
The experiment that killed it
Wolfram Schultz and colleagues recorded, in awake behaving monkeys, from single dopamine neurons in the midbrain, while the animals learned a simple task involving a cue and a reward such as a drop of juice. This is the finding, and it is worth reading slowly, because everything else on this page follows from it. It comes in three stages.
Stage one: an unexpected reward
Give the monkey a drop of juice out of the blue, with no warning. The dopamine neuron fires, a clear burst above its baseline rate, at the moment the juice arrives. So far, so consistent with the pleasure story. Juice is nice, the neuron fired, the neuron must be signalling niceness.
Stage two: a cue that predicts the reward
Now introduce a cue, a light or a tone, half a second before the juice, and repeat it many times until the animal has learned the association. Something remarkable happens. The burst migrates backwards in time. The neuron now fires to the cue, and it no longer fires when the juice arrives. The juice is unchanged. It is the same volume, the same sweetness, delivered to the same thirsty animal, and the animal drinks it with every sign of enjoyment. And the dopamine neuron is silent.
Stage three: the cue, but no reward
Now give the cue and then, at the moment the juice is due, withhold it. The neuron does not merely fail to fire. At precisely the time the reward was expected, its firing drops below baseline: a pause, a dip, a negative signal. The neuron marks the absence, and it marks it at the exact moment the absence occurs.
Sit with stage two for a moment, because it is the fatal one. A fully predicted reward, however pleasant, produces no dopamine response at all. If dopamine were a pleasure signal, this could not happen. The monkey is enjoying the juice. The pleasure is there. The dopamine is not. A signal that disappears when the pleasure is expected is not tracking pleasure.
And stage three tells you what it is tracking, because the neuron reports something even when nothing happened. Nothing is not an event. But nothing when something was expected is very much an event, and it is exactly the event a learning system needs to hear about. The dopamine neuron is not reporting the world. It is reporting the world's departure from prediction.
Reward prediction error: the quantity (reward received) minus (reward predicted). Better than expected gives a positive error and a burst of firing. Exactly as expected gives an error of zero and no change in firing. Worse than expected, including an expected reward that fails to arrive, gives a negative error and a dip below baseline. Every one of the three stages above is a direct read-out of this single quantity, and no other simple quantity accounts for all three.
Schultz, Dayan and Montague set this out in Science in 1997, connecting the physiological finding to the temporal difference learning algorithms already developed in machine learning, in which an agent learns precisely by computing the discrepancy between predicted and obtained value and using it to update its predictions. That the brain had independently arrived at the same error term was one of the more satisfying convergences in the history of the field, and it turned a curious recording into a theory.
What a prediction error actually is
It is easy to nod along to the phrase "prediction error" without grasping why it is the right thing for a brain to compute. So consider the problem from the learner's point of view.
You want to learn which things in the world lead to good outcomes. Suppose you simply strengthened your association with every good outcome as it arrived. Then a reward you already fully expect would keep strengthening its own prediction indefinitely, without limit and without purpose, because you already knew it was coming. You would be spending learning capacity on information you already possess. Worse, you would have no way to notice when something you predicted stopped happening.
Now suppose instead that you learn only from surprise. A reward that arrives unexpectedly is news, and you update towards it. A reward that arrives exactly as predicted is not news, and you do nothing, because your model is already correct. A predicted reward that fails to arrive is bad news, and you update away from it. Learning stops automatically when prediction becomes accurate, which is exactly when learning should stop. The error term is self-extinguishing.
This is why "teaching signal" is the right description and "pleasure signal" is not. Dopamine is not telling the rest of the brain how good things are. It is telling the rest of the brain by how much it was wrong. And a signal of that kind has a natural job to do at the synapse: dopamine released onto striatal and cortical synapses modulates their plasticity, biasing which recently active connections are strengthened. It arrives as a global broadcast saying "whatever you just did, do more of it" or "do less of it", and the circuits that were active a moment before are the ones that take the instruction. Compare long-term potentiation, where the strengthening rule is local; the dopamine error term supplies the missing third factor, an evaluation of whether the local activity was worth keeping.
There is one more elegant detail. In stage two, the burst does not disappear; it moves. It transfers to the earliest reliable predictor of the reward. If you then put a second cue before the first, and train that, the burst moves again, to the new earliest predictor. Value propagates backwards in time along the chain of predictors, which is exactly what temporal difference learning does, and it is why a long sequence of actions leading to a distant reward can be learned at all. The pull you feel when you see the first step of a familiar path to something good is, in this account, a dopamine burst that has walked backwards from the reward to the doorway.
Wanting and liking come apart
Schultz's work shows that dopamine is not a hedonic signal. Kent Berridge and Terry Robinson's work shows what it is instead, from a different direction, and their result is the one that most directly answers the question a reader is left with: if dopamine is not pleasure, why does it feel so much like it matters?
Their answer is that reward is not one thing. It has at least two dissociable components, and ordinary language collapses them because in everyday life they usually travel together.
Wanting
The pull towards a thing: the drive to seek it, approach it, and expend effort for it. Berridge and Robinson call the underlying process incentive salience, the attribution of motivational magnetism to a stimulus or its cues. Wanting need not be conscious desire, and it need not be pleasant. It is a force acting on behaviour.
Liking
The actual pleasure of the thing when you have it. The sweetness of the sugar in the mouth. This is a distinct process with its own neural basis, concentrated in small hedonic hotspots in the nucleus accumbens and ventral pallidum, and it depends on opioid and endocannabinoid signalling rather than on dopamine.
How you ask a rat
You cannot ask a rat whether it enjoyed something, but you can watch its face. Sweet tastes placed in the mouth elicit a stereotyped pattern of rhythmic tongue protrusions; bitter tastes elicit gapes and headshakes. These orofacial reactions are conserved across species, appearing in rats, in monkeys, and in human infants, which is what makes them usable as an objective index of hedonic impact rather than of motivation.
The result
Deplete a rat's dopamine and it stops working for food. It will not press a lever, it will not cross a barrier, and if the depletion is severe it may fail to feed itself even with food present. Yet place a sweet solution directly into its mouth and the liking reactions are essentially intact. The pleasure is still there. The pursuit is gone.
Read that last card again, because it is a genuinely astonishing fact. An animal that still visibly enjoys food, whose hedonic machinery is working normally, may starve amid food because it lacks the dopamine to want it. Enjoyment and pursuit, which every intuition insists are the same phenomenon seen from two sides, are running on different hardware, and one can be removed while the other remains.
The dissociation runs in the other direction too. Sensitising the dopamine system in animals increases wanting, the effort they will expend and the strength with which cues pull them, without a corresponding increase in liking reactions. Wanting up, liking flat. If a single quantity called reward were doing all this work, neither result could occur. Two quantities can come apart. One cannot.
Why the two normally look like one: in a well-functioning organism in a natural environment, the things you like are also the things you want, because the wanting was learned from the liking in the first place. The prediction error that trained the cue was generated by an outcome that was, in fact, good. So wanting tracks liking closely enough that language never needed separate words for them. The systems only reveal themselves as separate when something drives them apart: a lesion, a drug, a pathology, or an experimenter with a syringe. That is precisely why the distinction was not obvious for so long, and precisely why it explains so much once you have it.
Why this explains addiction
Now put the two halves together, because the payoff is large.
Dopamine carries a prediction error, a teaching signal that says "that was better than you expected, update towards it". In ordinary life this signal is honest: it fires when the world genuinely surprises you with something good, and it fades as your predictions improve, which is why the tenth good meal at a familiar restaurant is pleasant but no longer thrilling. The signal is supposed to extinguish as learning completes.
A drug that raises dopamine pharmacologically breaks that rule. It does not produce a prediction error because the world exceeded expectations. It produces one directly, by chemical means, bypassing the computation entirely. And crucially, it produces one every single time, no matter how well predicted the drug is, because the drug does not know it was predicted. Where a natural reward's teaching signal shrinks to nothing as the reward becomes expected, the drug's teaching signal does not shrink. It says "better than expected, update towards it" on the thousandth occasion exactly as loudly as on the first.
The consequence is a learning system that can never finish learning. The cues that predict the drug, the place, the ritual, the people, the time of day, accumulate incentive salience without limit, because the error term keeps arriving to reinforce them. Wanting is built, and built, and built, and nothing in the mechanism ever tells it to stop.
Meanwhile liking, which runs on a different system with no such exploit, does what pleasure normally does: it habituates. Tolerance develops. The drug becomes less enjoyable.
The incentive-sensitisation account: Robinson and Berridge's proposal that addiction is characterised by a progressive divergence of wanting from liking. The dopamine-dependent wanting system becomes sensitised, so cues exert an ever stronger pull, while the hedonic impact of the drug itself flattens or declines. The addict is left craving intensely something that no longer gives much pleasure. This is not a paradox to be explained away. It is the direct prediction of the model, and it matches what people with addiction consistently report.
That last point deserves emphasis. The pleasure theory of dopamine has to treat the addict's report, that they no longer enjoy the drug but cannot stop taking it, as some kind of confusion or bad faith, because on that theory craving and enjoyment are the same object. The wanting-liking account takes the report at face value and explains it. When a theory makes a strange first-person report come out true and mechanically expected, that is strong evidence for the theory.
Two caveats, because this is a domain where confident stories do harm. Addiction is not reducible to dopamine. It involves stress and withdrawal circuitry, changes in prefrontal control, learning and habit systems in the dorsal striatum, and a great deal of social and economic context that no neurotransmitter accounts for. The incentive-sensitisation model is an influential and well-supported framework, not a complete theory of a complex human condition, and it should not be used to make anyone's recovery sound like a matter of chemistry alone.
Why this explains motivation, apathy and Parkinson's
The same framework explains something far more ordinary than addiction: why some things feel worth doing and others do not.
Effort is costly. Any organism deciding whether to act must weigh what it will get against what it will cost to get it. Dopamine, particularly in the nucleus accumbens, appears to be central to that computation, and specifically to the willingness to pay a cost for a reward. In a widely used task, an animal chooses between a small reward it can have for nothing and a larger reward that requires climbing a barrier or pressing a lever many times. Animals with intact accumbens dopamine choose the effortful, larger option. Animals with accumbens dopamine depleted shift to the easy, smaller one. They have not stopped preferring the larger reward when both are freely available; what has changed is what they will pay for it.
Read that as a definition and it is quite close to what we mean, in everyday speech, by motivation. Motivation is not liking the outcome. It is being willing to bear the cost of getting it. The apathy seen in some clinical populations, an inability to initiate and sustain goal-directed effort that patients themselves often describe as caring but being unable to get started, becomes intelligible on this account. Nothing has gone wrong with their capacity to enjoy. Something has gone wrong with their capacity to want strongly enough to pay.
Which brings us to the disease that is the strongest anatomical argument against the pleasure theory. Parkinson's disease is caused by the progressive death of dopamine neurons in the substantia nigra pars compacta, and it is classically described as a movement disorder: bradykinesia, rigidity, resting tremor, difficulty initiating movement. On the pleasure theory of dopamine, this is inexplicable. Why would losing the pleasure chemical make it hard to walk?
On the account developed here, it is not merely explicable but expected. Dopamine in the nigrostriatal pathway sets the vigour and the readiness of the motor system, energising movement much as it energises pursuit; it is a signal about whether acting is worth it, applied to the motor apparatus. Remove it and movements do not become unpleasant; they become effortful, slow and hard to start, exactly as reported. And the same account predicts something that the movement-disorder framing does not, which is that many people with Parkinson's also experience apathy and motivational impairment that are not simply a psychological reaction to being ill, and that can be present early. A disorder of a single transmitter that impairs both movement and motivation looks strange only if you thought that transmitter was for pleasure. It looks entirely coherent if you think it is for the vigorous pursuit of what is worth pursuing, in the muscles and in the mind alike.
Myths about dopamine
Claim: dopamine is the pleasure chemical.
Once an animal has learned that a cue predicts a reward, its dopamine neurons fire to the cue and fall silent at the reward, even though the reward is exactly as pleasant as before. A fully predicted reward produces no dopamine response at all. Meanwhile animals depleted of dopamine still show the normal facial reactions of enjoyment to a sweet taste, yet will not work to obtain it. Pleasure without dopamine, and dopamine without a change in pleasure: the two facts together leave no room for the pleasure theory. Dopamine signals prediction error and drives wanting. Liking is somebody else's job.
Claim: a "dopamine detox" or "dopamine fast" resets your brain's reward system.
The idea is incoherent on its own terms. Sitting in a room avoiding pleasant activities does not deplete dopamine; the dopamine system is not a tank that fills and empties with your entertainments. And you would not want to deplete it if you could: dopamine is required for movement, for effort and for learning, and severe loss of it produces the motor and motivational impairment of Parkinson's disease. What is left after the neuroscience is stripped away is an ordinary behavioural idea, that interrupting compulsive habits can be useful, which may well be true and which needs no reference to any neurotransmitter whatever. The neuroscience is doing no work here except lending an unearned air of mechanism.
Claim: you get a "dopamine hit" every time your phone buzzes.
The phrase is loose in a way that misleads on three counts. It implies a discrete, drug-like dose, when dopamine signalling is a continuous modulation of ongoing circuit activity. It implies pleasure, which is the very error this page exists to correct. And it implies that any unpredictable reward schedule is doing something exotic to your brain, when in fact it is doing exactly what an unpredictable reward schedule has always done, in a pigeon or a slot machine or a letterbox: generating prediction errors and thereby sustaining a checking behaviour. The design of engaging apps is a real phenomenon and worth criticising. It does not become a stronger criticism by being restated in the language of neurotransmitters.
Claim: more dopamine is better; boosting it will make you happier and more motivated.
There is no single dial. Dopamine in the prefrontal cortex has an inverted-U relationship with cognitive performance, so raising it past an optimum degrades working memory rather than improving it. Excess dopaminergic signalling is implicated in psychosis, and dopaminergic medication in Parkinson's disease can produce impulse-control disorders, including compulsive gambling. And because dopamine drives wanting rather than liking, more of it would deliver more craving, not more contentment. If the goal is happiness, an intensified wanting system is close to the opposite of what you would order.
The limits of the story
A page this confident owes the reader its own boundaries, and the prediction-error account, for all its power, is not the last word.
Dopamine neurons are not identical. Recordings show that not all of them behave as canonical prediction-error units: some respond to salient but aversive events, some to novelty, some appear more concerned with alerting than with value, and their responses vary systematically with where in the midbrain they sit and where they project. The tidy picture of a single homogeneous error signal is a simplification of a heterogeneous population, and current work is largely about mapping that heterogeneity.
Nor is the relationship between phasic bursts, measured in milliseconds, and slower changes in the ambient concentration of dopamine, measured over minutes, fully settled. Fast bursts look like teaching signals. Slow, tonic levels look more like a setting of how vigorously to act and how much effort is worth expending. These are probably distinct functional signals carried by the same molecule at different timescales, and much of the apparent conflict between "dopamine is learning" and "dopamine is motivation" dissolves once the timescales are separated.
What is not in doubt is the negative claim, and it is the one that matters most for a reader trying not to be fooled. Whatever dopamine turns out to be in its full complexity, it is not a pleasure signal. That conclusion has survived every attempt to rescue the older view, and it rests on results that anyone can follow: the burst that moves to the cue, the dip when the reward fails to come, the rat that likes its sugar and will not lift a paw for it.
Sources
- Schultz W, Dayan P, Montague PR. A neural substrate of prediction and reward. Science. 1997;275(5306):1593-1599.
- Schultz W. Predictive reward signal of dopamine neurons. Journal of Neurophysiology. 1998;80(1):1-27.
- Berridge KC, Robinson TE. What is the role of dopamine in reward: hedonic impact, reward learning, or incentive salience? Brain Research Reviews. 1998;28(3):309-369.
This page is an educational reference. It is not medical advice, does not concern any medication, and does not diagnose or treat any condition.