Addiction: When the Learning System That Builds Habits Turns Against You

Introduction

In 2007 a team in Bordeaux ran an experiment that should have ended a very popular story about addiction.

They gave rats a choice. One lever delivered intravenous cocaine. The other delivered water sweetened with saccharin, which has no calories and no drug in it at all. Nothing but a taste. Most of the rats chose the sweet water. Not occasionally. Overwhelmingly, and they kept choosing it even when the cocaine dose was raised, and even in animals that had already been taking cocaine for a long time [1].

Read that again. Not a stronger drug. Not a bigger dose. Sugar water.

Hold that result against the story you have probably heard. Drugs flood the brain with dopamine. Dopamine is pleasure. The pleasure is so enormous that nothing natural can compete, so the person is captured. If that story were right, no rat with a cocaine lever would ever bother with sugar water.

Something in the explanation is wrong.

This article is about what is actually wrong with it, and it turns out to be the most basic thing. The brain system that addiction damages is not a pleasure system. It is a learning system. It is the same machinery that turns a route to work into something you drive without thinking, that makes you reach for your phone before you have decided to, that lets a smell pull up a memory you had not thought about in twenty years. That machinery is one of the best things about being a brain. In addiction it keeps working exactly as designed, on the wrong material, and the result is a person who wants something they no longer like.

That is not a metaphor. Wanting and liking are separable in the brain, they run on different chemistry, and only one of them grows with repeated drug use. Once you know that, most of the confusing parts of addiction stop being confusing.

The other thing you should know before we start is that the science here is not finished. Read almost any hospital or government page about addiction and you get a single clean account delivered with total confidence. Read the journals and you find researchers arguing hard, in print, right now, about whether drug use is habitual or deliberate, whether compulsion is even the right word, and whether calling addiction a brain disease has helped anyone. That argument is more interesting than the fake consensus, and this article will not hide it from you.

The Electrode That Started the Confusion

The pleasure story has a birthday. In 1954, at McGill University, James Olds and Peter Milner implanted electrodes into the brains of rats and let the animals press a lever to stimulate themselves [2].

The rats pressed. Then they pressed again. With electrodes in certain sites they pressed until they collapsed from exhaustion, ignoring food, ignoring water, ignoring everything else on offer. Olds and Milner had found what the newspapers immediately called the pleasure centre.

Every popular account of addiction you have ever read descends from that lever.

It was a genuine discovery and it reshaped the field. It also carried a hidden assumption that took forty years to dig out. The assumption is buried in the word "pleasure". The rats worked for the stimulation, so the stimulation must have felt good. That inference seems so obvious that it barely registers as an inference at all.

But working for something and enjoying something are not the same fact. One is behaviour you can see. The other is an experience you cannot. Olds and Milner measured the first and named it after the second.

By the 1980s the pleasure centre had been mapped onto a chemical. Roy Wise and Michael Bozarth argued that dopamine in the mesolimbic pathway was the common currency of reward, and that drugs of abuse all worked by pushing on it [3]. Every addictive drug, from cocaine to nicotine to heroin, does raise dopamine in the nucleus accumbens, though by different routes. The theory was elegant, it explained a lot, and it hardened into the version you now find on hospital websites.

Then people started testing the pleasure part directly, and it fell apart.

The Signal That Was Not Pleasure

In the early 1990s Wolfram Schultz was recording from single dopamine neurons in the midbrain of monkeys while they learned simple tasks. What he saw did not look like pleasure.

When an unexpected drop of juice arrived, the dopamine neurons fired a burst. Fine. That fits pleasure. But once the animal had learned that a light predicted the juice, the burst moved. It stopped firing at the juice and started firing at the light. The juice still tasted the same. The animal still drank it. The dopamine response to it was gone.

And when the light came on and the juice failed to arrive, the neurons went quiet. Below baseline. At exactly the moment the juice was due.

Schultz, working with Peter Dayan and P. Read Montague, showed in 1997 that this pattern matched something specific: a reward prediction error, the quantity at the centre of a class of learning algorithms in computer science [4]. The neurons were not reporting how good things were. They were reporting how much better or worse things were than expected. Better than expected, fire more. As expected, say nothing. Worse than expected, fire less.

That is a teaching signal. It is the correction term that tells the rest of the brain to update its predictions. If your expectations are already accurate, there is nothing to teach, so the signal falls silent even though the reward is still there and still enjoyable.

The evidence for this held up under every method the field could throw at it. Schultz mapped the physiology in detail [5]. Pascale Waelti and colleagues ran a blocking experiment, a classic test from animal learning theory, and found the dopamine response followed formal learning theory rather than reward magnitude [6]. Hannah Bayer and Paul Glimcher showed the signal was quantitative, scaling with the size of the error and not just its sign [7]. Mathias Pessiglione and colleagues moved it into humans with drugs that raised or lowered dopamine and watched reward learning shift accordingly [8].

Notice what none of those experiments measured. Nobody asked the monkey whether the juice was nice. The signal was defined arithmetically, as the difference between what was predicted and what arrived, and it behaved arithmetically. You can build a machine that does this. People had already built several, which is why the match mattered so much.

The last gap was causation. Correlations are cheap. In 2013 Elizabeth Steinberg and colleagues used optogenetics to fire dopamine neurons artificially at the moment when a real prediction error would have occurred, and produced learning that should not have happened [9]. The signal does not merely accompany learning. It causes it.

That is the moment the idea stopped being an interpretation and became a mechanism. If you can create learning by injecting a signal at the right instant, the signal is doing the teaching.

Three decades on, the reward prediction error account is one of the better established ideas in systems neuroscience, refined rather than replaced [10][11]. If you want the version of this story told properly on its own terms, we have written about dopamine as the molecule that teaches your brain separately. Recent reviews still frame dopamine primarily as a teaching signal rather than a hedonic one [12], and work by John Salamone and Mercè Correa adds a second job to the list, which is willingness to spend effort [13].

None of those jobs is pleasure.

How a Drug Breaks a Teaching Signal

Now put a drug into that system and watch what happens.

A natural reward produces a prediction error once. You eat the unexpected cake, dopamine spikes, your brain updates. Next time you see cake you predict it correctly, and the spike moves to the sight of the cake. The error shrinks toward zero, because your predictions got better. That is the system working.

A drug does not play by those rules. Cocaine blocks dopamine reuptake. Amphetamine forces release. Nicotine and opioids act on the neurons that gate dopamine cells. However they do it, the drug produces a dopamine surge pharmacologically, downstream of any prediction. Your brain can predict the drug perfectly and the surge still arrives at full size.

So the correction term never corrects. Every single use is treated by the learning system as an outcome better than expected, forever.

Nothing about this requires the drug to feel wonderful. It only requires it to arrive unpredicted, chemically, every single time.

Think about what that means for everything the signal touches. The cues that preceded the drug get their predictive value strengthened again and again, without limit. The actions that led to it get reinforced again and again. Steven Hyman, Robert Malenka and Eric Nestler laid this out as the core of addiction as a disorder of reward-related learning and memory [14]. The drug is not just rewarding. It is a permanent teaching signal that keeps writing the same lesson into a system built to stop learning once it has got the answer.

Underneath, real synaptic change follows. The mechanisms that write ordinary memories, including long-term potentiation at glutamate synapses, are recruited by drug experience, and Peter Kalivas has argued that a lasting disruption of glutamate balance in the accumbens is what leaves the system unable to let go [15]. Structural work continues to find drug-related remodelling in these circuits, including changes to the perineuronal nets that stabilise striatal connections [16][17].

Here is the flow of that failure, in one picture.

The important box is the one on the right. The value assigned to the cue keeps climbing, because nothing ever tells it to stop. And the thing that climbs is not enjoyment.

Wanting Without Liking

This is the part that explains the experience, and almost nobody outside the field is told about it.

In 1993 Terry Robinson and Kent Berridge published a theory that split reward into pieces that had always been treated as one thing [18]. Their claim was that "wanting" something and "liking" something are produced by different brain systems, that only wanting depends heavily on dopamine, and that only wanting gets amplified by repeated drug exposure.

The evidence for the split is odd and specific. Berridge had spent years measuring liking in rats by the facial reactions they make to taste, a set of movements so conserved that human newborns make the same ones. Sweet produces rhythmic tongue protrusions. Bitter produces gapes. These reactions can be scored without asking anyone anything.

Take dopamine away from a rat, almost entirely, and it stops working for food. It will starve in front of a full bowl. That looks like a total loss of pleasure. Put the food in its mouth and it makes every normal liking reaction, in normal amounts. It still likes the food. It just does not go and get it [19].

Read that again, because it is the whole argument compressed into one animal. Motivation collapsed. Enjoyment did not. If wanting and liking were the same thing, you could not separate them with a lesion, and yet you can.

Run the experiment the other way and you get the mirror image. Sensitise the dopamine system and the animal works harder for reward without showing any increase in liking reactions at all [20].

Liking, it turns out, lives in small fragile patches. Berridge and Morten Kringelbach mapped hedonic hotspots in the nucleus accumbens shell and the ventral pallidum where opioid or endocannabinoid signalling amplifies pleasure reactions, and these sites are tiny, a cubic millimetre or so in a rat [21]. Wanting runs on a much larger and more durable system, and that system is the one drugs sensitise.

That row about sensitisation is the one to hold on to. It is the only row where the two columns behave differently over time, and time is what addiction is made of.

Now read the table again as a description of a person.

Someone with a long addiction reports that the drug stopped being fun years ago. They are not lying and they are not in denial. Liking really did fade. Wanting did not, because wanting is the part that sensitises, and sensitisation can persist for a very long time after use stops [22]. The result is a pull toward something without the payoff that would once have justified it. Berridge and Robinson call the intensified pull incentive salience, and it attaches to cues at least as strongly as to the drug itself [23].

Individual animals differ in how much they do this, which turns out to matter enormously. Shelly Flagel and colleagues studied rats that learn a cue predicts food. Some rats, the goal-trackers, go to the food tray when the cue appears. Others, the sign-trackers, approach and interact with the cue itself, treating a metal lever as if it were the reward. Both groups learned the prediction equally well. Only the sign-trackers needed dopamine for the cue to acquire its magnetic quality [24]. Dopamine was required for the cue to become wanted, not for it to become informative [25].

That is the cleanest available demonstration that predicting and wanting are different operations, and that dopamine does the second one.

Thirty years after the original paper, Robinson and Berridge published a full reassessment of the theory, holding it up against everything that had accumulated since [26]. It is worth noting what they did next, because it is rare. In 2026 they published a paper asking whether their own theory can actually account for opioid addiction, which has features that fit it awkwardly [27]. Aldo Badiani and colleagues had already argued that opiate and stimulant addiction differ in ways the field keeps glossing over [28]. Good theories get tested by the people who built them.

The Cue That Outlives the Drug

If cues are where the value accumulates, cues are where the trouble lives.

Anna Rose Childress and colleagues showed people with cocaine addiction a video of drug-related material while scanning their brains. The drug users reported craving and showed increased blood flow in the amygdala and anterior cingulate. The cocaine-naive comparison group watched the same video and showed nothing of the kind [29]. This was a small study, fourteen detoxified cocaine users and six comparison subjects, and all of the drug group were men, which is worth holding in mind. The effect has since been replicated many times and across substances, and modern work is trying to turn cue reactivity into something measurable enough to guide treatment [30][31].

A cue is not a memory of the drug. It is a prediction that the drug is coming, and predictions are things your brain acts on before you have finished thinking.

What makes cues genuinely strange is what they do with time.

You would expect a cue's power to fade during abstinence. Time away, less craving. Jeffrey Grimm and colleagues tested that in rats and found the opposite. Cue-induced drug seeking got stronger over the first sixty days of withdrawal, not weaker, tracking a rise in brain-derived neurotrophic factor in the mesolimbic system [32]. The field calls it incubation of craving.

Sit with that for a second. The dangerous moment may not be day three. It may be week six, when everyone including the person themselves has decided the hard part is over.

The memory side of this is being worked out now. Drug-associated memories appear to be stored in identifiable cell populations, and the hippocampus is more involved than the older accumbens-centred story allowed [33]. That connects addiction to the way emotional arousal shapes which memories survive, and to the role of the amygdala in attaching significance to a stimulus.

You might think the fix is obvious. If a cue was learned, unlearn it. Present the cue, withhold the drug, let extinction do its work. Cue exposure therapy has been tried seriously for decades and the results have been persistently disappointing in the clinic, largely because extinction learning is context-bound and does not travel home with the patient [34]. Extinction does not erase the original memory. It writes a second, more fragile one on top.

That distinction matters more than it sounds. A memory you have learned to override is still there, waiting for a place where the override was never practised. Which is more or less a description of going home.

Newer approaches try to get at the original trace instead, by reactivating a memory and interfering with it during the window when it is briefly unstable. A 2026 trial using a memory retrieval and aversive conditioning procedure reported durable reductions in gaming craving alongside changes in fronto-insular circuits [35]. That is one study. It is not a treatment you can go and get. It is an indication of where the mechanism is being attacked.

From Choice to Habit

Here is where the field gets its most famous story, and also its biggest current argument. Take the story first.

Behaviour comes in at least two flavours. Goal-directed action is sensitive to what you expect to get. If the outcome stops being worth having, you stop doing the thing. Habitual action is sensitive to the situation instead. The cue appears, the response follows, and it keeps following even when the outcome has lost its value.

You have both systems and you use both constantly. Deciding where to eat tonight is goal-directed. Putting your seatbelt on is not.

The standard laboratory test for the difference is outcome devaluation. Train an animal to press a lever for a food. Then make that food worthless, either by feeding the animal to satiety on it or by pairing it with mild illness. Then offer the lever again. A goal-directed animal presses less. A habitual animal presses the same as ever, because it is no longer consulting the outcome at all.

Henry Yin, Barbara Knowlton and Bernard Balleine located the anatomy. Lesions of the dorsolateral striatum left animals able to represent what the outcome was worth but unable to form the habit, keeping their behaviour goal-directed for longer [36]. Damage a different striatal region and you get the reverse. The basal ganglia are not one system for movement. They contain at least two learning systems that compete for control of the same action [37]. We have written elsewhere about how ordinary habits become automatic, and that is the process being described here, in its healthy form.

Barry Everitt and Trevor Robbins built the addiction version. Their account has drug seeking starting out goal-directed, controlled by the ventral striatum, and gradually migrating to habitual control in the dorsal striatum with extended use, ending in something they called compulsion [38]. The anatomical bridge is a set of spiralling connections from the ventral striatum up through the midbrain to progressively more dorsal striatal territory, and David Belin and Everitt showed that cocaine seeking habits actually depend on that serial link [39]. Blocking dorsal striatal dopamine reduces cue-controlled cocaine seeking [40]. Recent work is still probing what changes at those synapses [41][42].

The strongest single piece of evidence for the end state came in 2004. Louk Vanderschuren and Everitt trained rats to seek cocaine, then paired the seeking with an aversive signal. Early in training the signal suppressed drug seeking, exactly as punishment should. After extended cocaine self-administration it stopped working. Sucrose seeking in the same design stayed suppressible [43]. The animals were not afraid of nothing, and they had not simply come to value cocaine more. Something about the control of the behaviour had changed.

Punishment is a blunt instrument, and you should be suspicious when an animal stops responding to one. But the sucrose control is what makes this result hard to wave away. The same rats, the same aversive signal, a different reward, and the suppression still worked. Whatever changed was specific to the cocaine.

The same year, Véronique Deroche-Gamonet, David Belin and Pier Vincenzo Piazza reported that rats given long access to cocaine developed a cluster of behaviours resembling human diagnostic criteria, including seeking during signalled unavailability and resistance to punishment, and that these behaviours predicted relapse [44]. The detail that matters most is easy to skim past. Only a minority of the animals did this. The rest took plenty of cocaine and never became addiction-like. Later work found that patterns of intake and craving early on predicted which animals would go that way [45].

Keep that minority finding in view for the rest of this article. It reappears every time anyone counts. Whatever addiction is, it is not the automatic consequence of putting a drug into a body.

Human evidence exists too, though there is much less of it than the confidence of the popular account suggests. Sjoerds and colleagues scanned 31 abstinent alcohol-dependent patients and 19 matched healthy controls on an instrumental learning task built to separate goal-directed from habitual control. The patients leaned more on habit, and the neural pattern shifted accordingly [46]. That is one study with 31 patients. It is the good one.

Everitt and Robbins updated their own account a decade on and were careful about what remained open [47]. David Belin and colleagues reframed it as a failure of control over maladaptive incentive habits, which merges the habit and incentive accounts rather than choosing between them [48].

The Argument Nobody Puts on the Page

Now the part that every page-one search result for this topic leaves out.

A serious body of work says the habit account is wrong, or at least badly overextended. Lee Hogarth argued in 2020 that drug use in humans remains goal-directed throughout, and that what escalates is not automaticity but the value of the drug relative to everything else, driven by negative emotional states [49]. On that reading a person is not a machine executing a stimulus-response loop. They are making a choice, under conditions that have made the drug the best available option for how they feel right now. Hogarth and colleagues had earlier mapped out the associative mechanisms in the transition from recreational use to addiction without needing habit to carry the weight [50].

It is worth being clear about how different that picture is. In the habit account the person has become a mechanism. In this one they are still deciding, and the decision has been distorted by how they feel and by how little else is on offer. Those two descriptions imply completely different things about what would help.

Vandaele and Ahmed went after the construct itself the same year, arguing that the behavioural signatures used to identify habit in the laboratory do not survive contact with the complex choice situations that real drug use involves [51]. Devaluation insensitivity in a rat pressing one lever in a box may simply not be the same phenomenon as a person continuing to drink.

Then the numbers arrived. Giannone and colleagues published a meta-analysis of the published animal experiments on habit and alcohol in 2024, combined with a review of the human studies. What they found was not a switch. It was a graded, probabilistic shift in decision bias, with wide variation between studies and between animals. They also argued for keeping habit and compulsion apart as concepts, because collapsing them has caused a great deal of confusion [52].

A graded shift is a much less satisfying headline than a switch, which may be part of why you have never read about it. It is also the sort of result that tends to survive.

Andreas Heinz and colleagues have pressed the same point from a clinical direction across several papers, asking directly whether compulsion explains addiction and concluding that the answer is at best partial [53][54]. Their more recent work reframes alcohol use disorder around learning mechanisms in the plural rather than around any single one [55]. Meanwhile the animal work continues to test the anatomy at cell-type resolution and does not always come back with the expected answer [56].

None of this means the habit research was wasted. It means the field found something real in the laboratory and then argued about how far it reaches into ordinary life, which is what should happen.

So there are at least four accounts on the table, and they disagree about what the central problem even is.

Notice something about that table. These are not four theories where three must be wrong. They describe different stages, different drugs and different people, and the honest current position is that no single one of them covers the whole thing. Anyone who tells you the question is settled is selling you a simplification.

The Other Engine: Feeling Bad

Everything so far treats addiction as a problem of pursuit. There is a second engine, and it runs on the opposite fuel.

George Koob and Michel Le Moal proposed in 1997 that repeated drug use shifts the brain's set point for reward. The system tries to keep itself in balance against a repeated chemical insult, and the counter-adjustment does not switch off when the drug leaves. What remains is a persistent state in which normal things feel less good than they used to [57]. Koob called the mechanism allostasis, meaning stability achieved by moving the set point rather than defending the old one [58].

You can feel the difference between the two engines in ordinary life, at a much smaller scale. Wanting a coffee at eleven in the morning is one thing. Wanting a coffee at eleven in the morning because without it your head hurts and you cannot think is a different thing, and only one of them is about coffee being nice.

Serge Ahmed and Koob had already shown the behavioural signature in rats. Give animals long daily access to cocaine instead of short access and their intake escalates over weeks, with the escalation tracking a shift in the reward threshold [59]. Short access, stable intake. Long access, a climb.

This is the part that makes early abstinence so much harder than anyone expects. The drug is gone and the counter-adjustment is still running.

On top of that sits recruitment of the brain's stress systems. Corticotropin-releasing factor and dynorphin signalling in the extended amygdala rise with dependence, and Koob has argued that this is the machinery of what he calls the dark side of addiction, where use continues to escape a negative state rather than to reach a positive one [60]. Timothy Baker and colleagues had built a similar case from the human smoking literature, reformulating addiction motivation around negative reinforcement and the avoidance of withdrawal-related affect [61].

Set point is the key phrase there. The system is not simply depleted. It has re-anchored itself around the presence of the drug, so the drug is now required to reach a state that used to arrive for free.

The clinical shape of this is familiar to anyone who has watched a relapse. It rarely looks like someone chasing a good time. It looks like someone who feels terrible and knows one reliable way to stop feeling terrible for an hour. A 2026 meta-analysis of trauma-related craving found negative affect doing exactly this mediating job [62]. Stress and mood are not background details. They are part of the mechanism, which is one reason what stress hormones do to memory and learning matters here.

Koob and Nora Volkow assembled these strands into a three-stage cycle: binge and intoxication, withdrawal and negative affect, preoccupation and anticipation, each mapped onto different circuits [63][64]. It is the most widely used framework in the field and it is genuinely useful, provided you remember it is a framework and not a finding.

The Part of You That Is Supposed to Say No

There is one more piece, and it is the one people intuitively reach for first. What about self-control?

Rita Goldstein and Nora Volkow spent two decades on this and produced the framework known as impaired response inhibition and salience attribution. The claim is that addiction involves both an inflated response to drug-related things and a weakened capacity to override that response, with the second half depending on prefrontal circuitry [65]. Peter Kalivas and Volkow made a related argument that addiction is best read as a pathology of motivation and choice rather than of pleasure [66].

The imaging findings behind this are old and consistent. In 1993 Volkow and colleagues used PET to show reduced dopamine D2 receptor availability in people who used cocaine, persisting three to four months after detoxification, and correlated with reduced metabolism in orbitofrontal cortex and cingulate gyrus [67]. The prefrontal cortex is the part of the brain that holds a plan against a competing impulse, and in long-term substance use it is measurably less active in exactly the situations where you would want it most.

There is something almost cruel in the arrangement. The circuitry that would let you step back and reconsider is the circuitry that has been quietly degraded.

Two cautions belong here, and both matter.

The first is direction. It is tempting to read reduced prefrontal function as caused by drugs. Some of it is. But Jeffrey Dalley, Everitt and Robbins showed that high impulsivity can precede drug exposure and predict which animals go on to escalate, which makes it a risk factor as well as a consequence [68]. Cause and effect are tangled here in a way that cross-sectional brain scans cannot untangle.

This is the trap that catches most popular writing about addiction and the brain. A scan taken after ten years of drug use cannot tell you what the brain looked like before, and the people who end up in these studies are not a random sample of anyone.

The second is that the deficit is not global. People with addiction are not broadly incapable of self-control. Scott Moeller and Goldstein have described a more specific problem with self-awareness and with attributing personal relevance accurately [69]. And the insula, a region involved in reading the body's internal state, keeps appearing in this literature in ways the older models did not predict [70].

Here is the finding that should be on every page about addiction and almost never is. These prefrontal changes recover. Ahmet Ceceli, Charles Bradberry and Goldstein reviewed the cross-species evidence on dysfunction and recovery and found meaningful restoration of prefrontal function with sustained abstinence [71]. Structural measures move too [72]. The brain that changed can change back, which is the same capacity for structural remodelling that got it into trouble in the first place.

Why Only Some People

Most people who drink do not develop alcohol dependence. Most people who are prescribed opioids do not become addicted to them. Most people who gamble do not lose control of it. Any account of addiction that starts and ends with what the drug does to the brain has no way of explaining that, because the drug does roughly the same chemistry in everyone.

So the interesting question is not what the drug does. It is what makes one person's brain respond to it differently from another's.

The rat data make the point first. In Deroche-Gamonet's study, all the animals got cocaine and only a minority became addiction-like [44]. Exposure was constant. Outcome was not.

Genes account for a substantial share of the variation. Kenneth Kendler and colleagues ran a Swedish adoption study of unusual size, following 18,115 adopted children along with 78,079 biological parents and siblings and 51,208 adoptive parents and siblings. Risk of drug abuse was elevated both in the adopted children of biological parents with drug abuse and in relation to disruption in the adoptive home [73]. Genes and environment, both, measured in the same design.

Note what that design rules out. If the effect were purely about growing up in a difficult household, adopted children would not carry their biological parents' risk. If it were purely genetic, the adoptive home would not matter. Both showed up.

Timing matters as much as heredity. Bridget Grant and Deborah Dawson found that age at first use predicted later dependence strongly, with early onset carrying much of the excess risk [74]. Animal work points the same way, with adolescent rats showing greater vulnerability to cocaine at behavioural and electrophysiological levels [75]. The adolescent brain is in the middle of remodelling the exact circuits that addiction recruits.

If you want a single sentence for this section, it is that the drug is one variable among several, and not usually the one with the largest coefficient.

And then there is the part of the picture the biology alone keeps missing. In 1978 Bruce Alexander, Robert Coambs and Patricia Hadaway ran an experiment that later became famous under a name they did not use. They compared rats housed alone in standard laboratory cages with rats living socially in a large open enclosure of nearly nine square metres, giving both groups morphine solution as their only fluid for fifty-seven days and then measuring choice. The isolated rats drank significantly more morphine than the socially housed ones [76].

Be careful with this study, because the internet has not been. It is a rat housing experiment, its replication record is mixed, and it does not prove that human addiction is purely social. What it does show is that identical drug availability produces different drug taking depending on what else is available and what the conditions are like. That is a real finding and it is enough. Work on natural rewards competing with drug reward continues in the same direction [77], and studies keep finding that personality and social context shape whether use escalates [78].

The Lenoir experiment this article opened with belongs in this section too. Rats chose sweetness over cocaine because sweetness was there. Remove the alternative and the choice changes. That is not a small detail about rodents. It is the mechanism by which environments make addiction more or less likely.

Is Addiction a Brain Disease?

You cannot write honestly about this topic without addressing the framing, because the framing is itself under active dispute.

This is not a semantic quarrel. What you call something determines who gets funded to study it, who is expected to treat it, and how a person is spoken to when they ask for help.

The case for is stated most clearly by Nora Volkow, George Koob and A. Thomas McLellan in the New England Journal of Medicine. Addiction involves measurable, persistent changes in brain circuitry; those changes help explain why willpower alone so often fails; and treating addiction as a medical condition supports both research funding and humane treatment [79]. Volkow and Koob had already asked in print why the model was so controversial [80].

That is a serious argument and it has done real good. Framing addiction as a medical condition rather than a moral failure changed how a great many people are treated, and nobody in this debate wants to go back.

The case against is not from cranks. In The Lancet Psychiatry, in the same year as that Volkow and Koob piece, Wayne Hall, Adrian Carter and Cynthia Forlini argued that the brain disease model had absorbed research priorities without delivering the treatment gains it promised, and that it sits awkwardly with the epidemiology of recovery [81].

The argument has sharpened since. In 2025 Chrysanthi Blithikioti and colleagues published a reevaluation in The Lancet Psychiatry, setting out that no specific neural signature of addiction has been identified, that the model handles heterogeneity in recovery poorly, and that the dual use of the model as both a causal theory and an anti-stigma tool has muddled two separate empirical questions [82]. Shane O'Mahony argued the same year that the model can produce a form of epistemic injustice, discounting what people with addiction say about their own experience because the brain scan is treated as more authoritative [83].

You may notice that this is two disagreements wearing one coat. One is factual: does the evidence support calling it a disease of the brain? The other is strategic: does saying so reduce stigma? They can have different answers, and treating them as one question has kept the argument going for thirty years.

Gene Heyman has made the sharpest version of the empirical objection. If addiction is a chronic relapsing brain disease, the recovery statistics should look like those of a chronic relapsing disease. They do not [84].

We are not going to resolve this for you, and anyone who does should be treated with suspicion. What can be said is that both sides agree on more than the argument suggests: the brain changes are real, they are not the whole story, and social and economic conditions do a great deal of work that neuroscience alone cannot account for.

It is also worth knowing that this rewriting is not new. A 2026 paper traced how the addiction chapter changed across six editions of Kandel's Principles of Neural Science, which is as close as neuroscience gets to an official record of what it currently believes [85]. The story has been rewritten repeatedly. It is being rewritten now.

How the Story Changed

Here is the shape of the argument across seventy years.

Read down that column and notice how often the correction came from someone testing the previous generation's favourite idea rather than adding to it. That is what a healthy field looks like from the inside. It is also why any article that gives you one confident paragraph about addiction is giving you a snapshot of one decade.

What Actually Shifts the Behaviour

This is a science article and not a treatment guide, so what follows is about mechanism. It is not advice, and nothing here is a substitute for talking to a clinician.

The most instructive result in the whole behavioural literature is also the simplest. Michael Nader and William Woolverton let monkeys choose between cocaine and food, and then increased the amount of food on offer. Drug choice fell as the alternative got better [86]. Stephen Higgins, Warren Bickel and John Hughes found the same thing in people using cocaine in a laboratory setting [87]. Drug taking behaves like choice behaviour. It responds to what else is available.

It is worth pausing on how unglamorous that finding is. No receptor, no circuit, no scan. Just more of something else on offer, and the behaviour moves.

That principle became a treatment. Contingency management provides tangible reinforcement for verified abstinence, and it has one of the more consistent evidence bases in the field. Higgins, Sarah Heil and Jennifer Lussier laid out the reinforcement logic behind it [88]. Michael Prendergast and colleagues meta-analysed the trials and found reliable effects across substances [89], and Lois Benishek and colleagues confirmed it for the prize-based variant [90].

Sit with the awkwardness of that for a moment. If addiction were pure compulsion, immune to consequences, giving someone a voucher for a clean urine test should do nothing at all. It does something. That is a data point against the strongest version of the compulsion story, and it is why Hogarth and others keep pointing at it.

Craving itself is being tested as a treatment target, with medication effects on cue-induced craving in the laboratory now being linked to real-world drinking outcomes [91]. Not everyone is convinced that craving is one thing or that it is the right target, and the commentary published alongside that work says so directly [92]. Brain stimulation approaches are being reviewed as they accumulate [93][94]. Newer work is also trying to measure addiction through behaviour rather than self-report [95], and stress-cue interactions are turning out to differ by sex in ways the older literature obscured [96].

None of this adds up to a solved problem. It adds up to a set of levers that move the behaviour, which is more than the pure-hijack story predicts should exist.

The Ending That Almost Nobody Prints

Search for addiction and you will read about a hijacked brain, a chronic relapsing condition, a lifelong struggle. All of that appears. What almost never appears is what happens to most people.

Catalina Lopez-Quintero and colleagues analysed the National Epidemiologic Survey on Alcohol and Related Conditions, a nationally representative sample of 43,093 US adults, and calculated the cumulative probability that someone with lifetime dependence would reach remission. For nicotine it was about 84 percent. For alcohol about 91 percent. For cannabis about 97 percent. For cocaine about 99 percent [97]. The dependence subsamples were large: 6,937 for nicotine, 4,781 for alcohol, 530 for cannabis and 408 for cocaine. Remission took years, often many, and most of it happened without formal treatment.

You have almost certainly never seen those figures, and that absence is not accidental. A story about a hijacked brain is more memorable than a story about slow uneven recovery over a decade.

Those numbers are not a promise to any individual person, and they are not an argument that addiction is easy to leave. Look at how long it took. Half of the nicotine cases remitted about 26 years after dependence began, and the figures for alcohol, cannabis and cocaine were roughly 14, 6 and 5 years. That is most of a working life for some of them, and the people who remain dependent are precisely the ones who most need help. But an account of addiction that cannot accommodate those figures is not describing the thing accurately.

This is not a new finding either, which is the uncomfortable part. Lee Robins studied US army enlisted men who had served in Vietnam in 1970 and 1971, following them up after their return and comparing them with a matched group. A large fraction had been addicted to heroin while deployed. Very few remained addicted at home, and readdiction after return was rare. Two decades later Robins wrote a lecture asking why the finding was still not absorbed into either public or scientific thinking about heroin, and examined the defences the field had built against it [98].

The explanation that fits everything in this article is that the men had changed almost every cue, every routine, every social contingency and every source of stress at once. The learning system that had been trained in one environment was returned to a different one.

That does not make addiction a choice in the glib sense, and it does not mean anyone can simply decide to stop. Sensitised wanting is real, incubated craving is real, the dorsal striatum does take over, and the negative affect that drives relapse is not imaginary. What it means is that the system remains a learning system all the way through. Learning systems respond to conditions. That is the entire reason they exist.

If you take one thing from this article, take that. Not optimism, which is cheap. The specific fact that the machinery involved is the machinery of learning, and learning does not have an off switch.

Does Any of This Cover Gambling and Screens?

Gambling disorder was moved into the substance-related and addictive disorders chapter of DSM-5, the first behaviour to be classified alongside drugs, on the argument that it shows similar clinical features and similar circuitry [99]. Gaming, food and short-form video have all since been discussed in the same terms.

Some of that transfer is well founded. Cue reactivity, sensitised wanting and negative-affect-driven use all appear in behavioural addictions. Wei Lei and colleagues reported imbalanced goal-directed and habitual control in people with internet gaming disorder, which is the same signature the alcohol work looks for [100]. Matthias Brand and colleagues have built an integrative model specifically for behavioural addictions [101]. Kyle Burger has applied reinforcement architecture to compulsive overeating [102], and inhibitory control deficits have been reported in problematic short-form video use [103].

There is an obvious risk in all of this, and it is worth saying plainly. Once a word like addiction is available for any behaviour someone does more than they would like, it stops carrying information. Enjoying a game a lot is not the same as a life narrowing around it.

But scepticism is warranted and it exists in print. Tim van Timmeren and Luke Clark asked directly whether habit theory even extends to disordered gambling, and the answer is not a simple yes [104]. Richard Tunney has questioned whether some of the animal paradigms used to justify the extension mean what they are taken to mean [105]. Reviews are increasingly arguing that reward and addiction involve systems well beyond dopamine [106][107].

The honest position is that the learning account transfers better than the pharmacology does. A slot machine cannot block dopamine reuptake. What it can do is deliver unpredictable reward on a schedule that keeps the prediction error large, which is the same trick from a different direction. Whether that is enough to call it the same disorder is exactly what the field is arguing about.

Conclusion

The story you were told is that addiction is about pleasure, that drugs give too much of it, and that the brain is captured by the intensity of the reward.

Nearly every part of that is wrong or misleading. The dopamine signal at the centre of it is a teaching signal, firing to the gap between expectation and outcome, and drugs break it by producing a surge that no amount of prediction can shrink. What grows under that broken signal is wanting rather than liking, which is why a person can be pulled hard toward something that has stopped being enjoyable. What the cues acquire is durable enough to survive years of abstinence, and in animals it grows during the first weeks away from the drug rather than fading. What happens to control is genuinely contested, with serious researchers arguing that it becomes habitual and equally serious researchers arguing that it stays deliberate and driven by the need to feel less bad.

The one thing that unites every account is the word learning. Addiction is what happens when a system built to extract regularities from the world is given an input that never stops producing an error signal.

That framing changes what a reader should take away. Not that the brain is broken beyond repair, because prefrontal function recovers with sustained abstinence and most people with dependence eventually reach remission. Not that it is simply a choice, because sensitised wanting and cue-driven seeking are real mechanisms that operate whether or not anyone consents to them. Something less comfortable than either: a learning system running correctly on material it was never built for, in a person whose circumstances shape the outcome as much as the chemistry does.

If any of this is close to home, the useful next step is a conversation with a clinician or a local health service rather than an article. What this one can offer is a more accurate picture of what is actually happening, which is worth having, because the wrong picture leads to the wrong conclusions about yourself and about other people.