Craving and decision-making in addiction are more tightly bound than scientists once thought. A landmark study published in Nature Mental Health now shows that craving does not simply follow addictive behaviour. It actively reshapes how the brain learns, and it does so differently depending on the substance involved.

Researchers at Yale School of Medicine and the Icahn School of Medicine at Mount Sinai led the work. They found that craving alters the brain’s core learning mechanisms in real time, creating a cycle that can make addictive choices feel almost automatic.

How Craving Distorts Decision-Making in Addiction

To examine how craving shapes decision-making in addiction, researchers recruited 132 participants reporting moderate to heavy use of either alcohol or cannabis. Each person completed a computerised task with two virtual slot machines. One machine paid out 80 per cent of the time, the other only 20 per cent. The winning machine switched positions without warning several times throughout the game.

Rewards varied by block. In one block, a correct choice earned a coin image. In the other, participants received a substance-related image, such as a beer photograph for the alcohol group or a bong image for the cannabis group. Researchers recorded craving and mood at regular points throughout.

When money was the reward, participants quickly identified the better machine. They adjusted sensibly after a loss. Their choices followed a clean, rational pattern. Once substance images replaced the coins, craving began pulling decisions off course.

Alcohol and Cannabis Push Learning in Opposite Directions

Computational modelling revealed that craving biased the brain’s learning rate, the speed at which new information updates future choices. The direction of that bias, however, split sharply between the two groups.

In alcohol drinkers, higher craving accelerated learning from alcohol-related outcomes. The brain grew more sensitive to reward signals tied to drinking, locking in choices faster. Researchers describe this as a feedback loop. Craving sharpens attention on substance-related wins, which then reinforces the pull back toward drinking. That loop helps explain why decision-making in addiction to alcohol can feel so hard to interrupt.

Cannabis users showed the opposite pattern. Higher craving slowed learning from cannabis-related prediction errors. The brain became less responsive to new information, not more. Earlier research has linked long-term cannabis use to diminished learning and memory, and this study adds a computational dimension to that picture.

“There is something that distinguishes the mechanism of alcohol use versus cannabis use,” said Dr Kaustubh Kulkarni, first author and a psychiatry resident at Yale. “It could be the actual pharmacologic effect of alcohol on the brain, which is pretty different from cannabis, or it could be something else entirely.”

The pharmacological gap between the two substances is significant. Alcohol typically hits in concentrated episodes and takes longer to produce effects. Cannabis takes effect within minutes and users tend to consume it more gradually. Alcohol works across multiple receptor systems, producing euphoria and disinhibition. Cannabis targets cannabinoid receptors and produces relaxation. These biological differences likely underpin the diverging computational patterns the study uncovered.

How Craving and Decision-Making Fuel Each Other

Perhaps the most striking finding was the bidirectional relationship between craving and decision-making in addiction. Craving shaped choices, but choices also shaped craving.

Both anticipated outcomes and actual results drove shifts in momentary craving across both groups. This fits a Bayesian model of craving, the idea that it rises from a combination of prior expectations and real-time evidence. Seeing a cannabis or alcohol image after a win raised craving through two routes: what the brain predicted would happen, and what it actually observed.

Craving also proved dynamic rather than fixed. It shifted trial to trial across the experiment. That variability matters clinically. A person’s craving at any moment reflects an ongoing negotiation between memory, expectation, and immediate experience, not a static background urge waiting to be triggered.

Computational Scores Can Predict Addiction Risk

The team tested whether the computational measures from the task could predict clinical addiction risk, over and above demographics and standard behavioural metrics.

For alcohol users, the combined model incorporating computational, demographic, and behavioural variables performed best. It explained a meaningful share of variance in AUDIT and ASSIST risk scores, with a correlation of r = 0.545 (P less than 0.001). Higher learning rates, elevated baseline craving, and greater sensitivity to outcomes each linked positively to higher alcohol addiction risk.

For cannabis users, demographics alone proved sufficient. Adding computational parameters did not improve predictions. That gap suggests craving and learning interact differently in cannabis use disorder, or that a different set of computational measures may be needed to capture its risk profile.

Not all addictions follow the same internal logic. Tools that map one substance use disorder well may not translate to another.

What Craving Science Means for Addiction Treatment

Professor Xiaosi Gu, director of the Computational Psychiatry Unit at Yale and senior author of the study, sees the findings as pointing toward more precise interventions. “This could explain why breaking the addictive cycle feels so difficult, as the brain is adapting constantly,” she said.

Craving and decision-making in addiction reinforce each other in a loop. Managing craving may therefore weaken the loop itself, not just reduce discomfort. The researchers argue their models provide the first mechanistic explanation for why craving-focused interventions sometimes succeed.

“My hope is that these gamified behavioural assays will be used to assess patients in the clinic,” said Professor Gu. “Unlike brain imaging, which requires expensive equipment, behavioural tasks can be administered anywhere, even remotely during telemedicine visits.”

The team now plans neuroimaging studies to identify which brain circuits carry out these computational processes. Once researchers identify those circuits, pharmacological treatments, brain stimulation, or targeted psychotherapy could address them directly.

The framework may also reach beyond addiction. Professor Gu suggested it could shed light on craving-like states in eating disorders and in conditions involving social isolation, situations where a persistent internal urge overrides deliberate judgement.

Limitations Worth Noting

The sample included people at moderate to high risk, not those with a confirmed substance use disorder diagnosis. The task also required participants to actively seek out substance cues, which may not reflect the experience of someone in recovery or trying to abstain. Without neuroimaging data, the specific brain regions driving these effects remain unknown.

Even so, this study offers the first empirically tested computational framework linking momentary craving with reinforcement learning across two separate substance-using populations. That foundation now supports a clearer research path toward understanding, and eventually disrupting, how craving drives harmful choices.

The study, “A computational mechanism linking momentary craving and decision-making in alcohol drinkers and cannabis users,” was published in Nature Mental Health (2026).

Source: dbrecoveryresources