GLP-1 Drugs and Addiction: A Huge Signal, and a Trial That Missed

Two studies landed in 2026 and they point different ways. A cohort of 524,817 US veterans found people starting a GLP-1 developed fewer new substance use disorders of every kind measured — alcohol, cannabis, cocaine, nicotine, opioids — than people starting a comparable drug.^[1] A randomized trial of semaglutide in people who smoke daily then missed both its primary endpoints: craving fell and weight fell, but resistance to smoking and cigarettes per day did not move significantly.^[2] That trial randomized 24 people. Holding both of those in view at once is the honest position.

What the cohort found

The design is worth a sentence, because it is better than most observational work. Drawing on 606,434 US veterans with type 2 diabetes, the researchers emulated eight parallel target trials, comparing people who started a GLP-1 against people who started an SGLT-2 inhibitor — an active comparator rather than nobody, which controls for the fact that people who get prescribed something differ from people who get prescribed nothing.^[1]

Among veterans who already had a substance use disorder, the outcomes that matter clinically moved too: fewer emergency department visits (0.69), fewer hospital admissions (0.74), fewer overdoses (0.61), and substance-related mortality at 0.50 — half.^[1]

What the randomized trial found

Also in 2026, a phase 2a trial gave semaglutide or placebo for nine weeks to adults smoking at least five cigarettes a day, with laboratory sessions measuring how well they resisted smoking and how much they smoked when they could.^[2]

Neither co-primary outcome reached significance. Smoking resistance came in at β 0.16 (95% CI −0.07 to 0.40, P = .16); number of cigarettes at β −0.08 (95% CI −0.25 to 0.08, P = .30).^[2] What did move was craving for cigarettes, which fell significantly, along with body weight.

Wanting a cigarette less is not the same as smoking fewer of them, and this trial separated the two.

Why the two disagree, and which to trust for what

This is the same pattern our alcohol article documents, in a different substance and by the same research group: a large observational signal, and randomized tests that shrink or vanish when they measure behavior rather than desire.

The most likely explanation is not that one study is wrong. Observational data compare people who took a drug against people who did not, and no weighting fully removes the differences between those groups — someone prescribed a GLP-1 is someone engaged with healthcare in ways that themselves predict fewer substance problems. Randomized trials remove that, and pay for it in size and duration. Nine weeks in 24 people is not a fair test of whether someone quits smoking.

So the defensible reading is: the signal is real enough to be worth a properly powered trial, and there is not yet evidence that these drugs help anyone quit anything. Those are compatible statements.

One result that checks out against something else

Among veterans with a pre-existing substance use disorder, suicidal ideation or attempt came in at a hazard ratio of 0.75 (0.67–0.83).^[1] That points the same way as the evidence behind the removal of the suicidality warning from these labels in February 2026, which we cover in the suicidality article — and it comes from an entirely separate dataset and population.

Two independent bodies of evidence agreeing is worth more than either alone. It is still observational, and it is still in a population that is 90%-odd male veterans with type 2 diabetes.

Smoking: five findings from eighty-four people

A pilot trial randomized 84 smokers with prediabetes or excess weight to exenatide or placebo, everyone also receiving a nicotine patch and brief counseling, with biologically confirmed abstinence at six weeks as the outcome. The drug arm did better. A follow-up analysis then asked who it worked for, and reported five moderators.^[4] A larger trial has since tested the question head-on and found nothing — see GLP-1s and quitting smoking for that evidence and its 12-month follow-up.

Then note rows two and three. The trial enrolled smokers with prediabetes and/or excess weight — and reports the drug working best in participants who had neither. A treatment appearing most effective in the subgroup least like the eligibility criteria is a pattern that should prompt suspicion of noise before biology, in five subgroups carved out of 84 people.

The genetic moderator deserves the same caution as every other pharmacogenomic subgroup in this field. Real associations do exist and their effects are small — the genetics are real, and small — and a genotype effect found in a fraction of 84 people is a long way short of that.

What survives is the parent trial’s own result: adding a GLP-1 to a nicotine patch improved short-term abstinence in a pilot. The authors ask for larger prospective studies, which is the right request.

A larger emulation, and the split within the class

A 2026 analysis went further, emulating four separate trials inside a records network covering more than 120 million patients. Adults with type 2 diabetes and no prior alcohol-use disorder starting tirzepatide, semaglutide, liraglutide or dulaglutide were each matched one-to-one against patients starting a DPP-4 inhibitor, and followed 18 months for a first diagnosis.^[3]

Head to head, tirzepatide beat liraglutide directly (0.47, 0.24–0.92). The split is the interesting part: the two newer, more potent agents show the effect and the two older ones do not, in the same drug class.

One more limit worth stating plainly: the outcome is a diagnosis code. Someone has to seek care, and a clinician has to record alcohol-use disorder. That measures diagnosis rather than drinking, and it under-counts in both arms by an unknown amount. The authors’ own conclusion is the right one — that randomized evidence for this indication is now appropriate.