Here is a contradiction worth sitting with. The same small circle of scientists producing the most compelling evidence that GLP-1 drugs might treat addiction are also the people telling you, in print, not to prescribe them for it yet. That is not caution for caution's sake. It is a warning built on the wreckage of their own field's only finished trial.
Start with the story that keeps the enthusiasm alive.
A 53-year-old woman with type 2 diabetes, PTSD, anxiety, depression, and a 36-pack-year smoking history walked into a routine psychiatric follow-up. She smoked within minutes of waking, which placed her in the high-dependence category. She had quit once before, for a year, on varenicline—then relapsed after stopping it over a mood side effect.
In 2023 she started tirzepatide for her diabetes, titrated slowly from 2.5 mg toward 10 mg. Within three weeks, cigarettes suddenly tasted wrong. She described them as "chemicals lit on fire."
The smell disgusted her. Each attempt to smoke brought nausea and vomiting. Two years later, she was still off cigarettes, fighting only rare, tolerable cravings and the occasional vape. Her weight fell from 235 to 156 pounds. In her words, "whatever craving receptor in my brain that wanted the cigarette was lowered after I started this medicine."
This is the kind of anecdote that launches a thousand prescriptions. It has an abrupt onset, a vivid sensory mechanism, a two-year follow-through, and a patient who beat her own poor prognosis. It was published this year as demonstrating "a novel possible mechanism."
Note what it is not. It is a single patient, reporting her own abstinence, with no biochemical verification—no cotinine test, nothing but her account. The authors themselves list weight loss and lifestyle change as possible confounders before concluding that "the temporal relationship, abrupt sensory aversion, and sustained abstinence strongly implicate tirzepatide."
Hold that phrase. "Strongly implicate," from an N of one, is exactly the reasoning the field's leaders are begging clinicians to resist.
"Strongly implicate," from an N of one, is exactly the reasoning the field's leaders are begging clinicians to resist.
Now stack the evidence that makes the leap feel not just reasonable but overdue.
The rodent work is genuinely consistent. Across mice, rats, and non-human primates, independent labs have shown for over a decade that GLP-1 receptor agonists reduce the reinforcing properties of alcohol, nicotine, cocaine, and opioids, and blunt cue-induced relapse. A systematic review of 41 studies (35 preclinical, 6 clinical) found the preclinical signal consistent across drug classes.
And there is a clean circuit to hang it on. A 2025 Science Advances study showed that voluntary cocaine taking lowered plasma GLP-1 in rats, that activating GLP-1-producing neurons projecting to the ventral tegmental area cut cocaine seeking, and that the receptors sit primarily on VTA GABA neurons—which, when engaged, quiet the dopamine neurons driving reward. Mechanism, anatomy, and behavior line up.
Then comes the number that makes headlines. The primary incident-SUD analysis drew on 524,817 initiators (from a base population of 606,434 US veterans with type 2 diabetes), analyzed as emulated target trials, found that starting a GLP-1 drug rather than an SGLT-2 inhibitor was associated with lower risk of incident substance use disorders across the board: alcohol at a hazard ratio of 0.82, cannabis 0.86, cocaine 0.80, nicotine 0.80, opioids 0.75, other 0.87.
Among veterans who already had a substance use disorder, the associations were larger still—fewer emergency visits, fewer hospitalizations, and an SUD-related mortality hazard ratio of reportedly 0.50. A separate VA analysis reportedly compared 12-month all-cause overdose rates in people with opioid use disorder. Two commercial EHR studies of opioid overdose reportedly pointed the same direction.
Six drug classes. Hundreds of thousands of patients. A mechanism you can draw on a whiteboard. It reads like a foregone conclusion.
It is not, and here is the turn.
Not one of those human findings comes from a randomized trial in the indication. The 606,434-veteran study is observational. The overdose study is observational. The smoking case is a case report. Every arrow points the same way, but they are all drawn from the same kind of ink—routinely collected records, matched and weighted, never randomized.
That matters because we have one completed randomized controlled trial in this space, and it tells a more sobering story. Exenatide, a first-generation GLP-1 agonist, was tested against placebo in alcohol use disorder. It failed on the primary alcohol consumption outcomes in the full sample. Reportedly, after 26 weeks, no reduction in heavy drinking days.
The effect appeared only in a secondary analysis—patients who had both alcohol use disorder and comorbid obesity.
Read that subgroup again, because it is the whole problem. The one place the drug "worked" is defined by metabolic phenotype. And metabolic phenotype is precisely the confounder that observational data cannot scrub out.
People prescribed GLP-1 drugs are, by definition, people with diabetes or obesity who fill prescriptions and keep filling them. They differ from comparators in weight, in metabolic health, and in the sheer fact of adherence—the healthy-adherer effect, where people who stick with any medication tend to have better outcomes for reasons the medication had nothing to do with.
A hazard ratio of 0.75 across a cohort cannot tell you whether the opioid benefit was the drug or the kind of person who ends up on the drug and stays on it. Propensity matching helps; it does not settle it. Arguably, the exenatide trial suggests that when you finally randomize, the broad effect can collapse to a subgroup with comorbid obesity.
The clinical case is thinner still on its own terms. In smoking cessation trials, exenatide added to nicotine replacement reportedly helped over six weeks, but dulaglutide reportedly showed no difference in abstinence against placebo in two studies when both arms also got varenicline and counseling. Mixed, not clean.
This is why the sentences from the field's leadership read the way they do.
The teams running the current semaglutide trials for alcohol use disorder reportedly wrote, plainly, to "encourage clinicians and patients to be prudent and avoid placing supposition before science." Their review states the drugs "have the potential to be repurposed" only "pending results from RCTs." The NIDA director's editorial reportedly calls for "fast and rigorous testing," which is not the same as fast prescribing.
They believe the drugs may well work. Several of them are betting their careers on trials designed to find out. That is exactly why they are unwilling to let a case report do a trial's job.
So here is the discipline the evidence actually demands.
"The temporal relationship strongly implicates tirzepatide" is a legitimate line in a case report. It is a hypothesis, offered as one. It becomes dangerous the moment it is read as a prescribing rationale—when N=1 plausibility is treated as N=many proof, and a woman's vivid two-year story is quietly upgraded into a reason to hand the drug to the next person with a craving.
The signals here are stronger than the usual off-label hype. That is precisely what makes the temptation harder to resist, and the confounding-versus-causation gap—the same trap that haunts every observational drug signal—more worth naming out loud.
The mechanism is real. The rodent data are real. The 606,434-veteran association is real. And the only time anyone ran the randomized experiment, the broad effect vanished except in the subgroup that doubles as the confounder.
Wait for the trials. They are running. The honest position is the one the trialists already hold: promising, unproven, and not yet yours to prescribe on the strength of a story about chemicals lit on fire.