GLP-1s and Mental Health: What 10 Studies Actually Found (2026)

A cohort of 240,618 patients drawn from a large electronic-records network, compared against matched patients on other diabetes and weight drugs. New suicidal ideation was substantially lower on semaglutide in both the obesity and diabetes groups. ⚠ Observational, so it cannot establish cause — but note the design choice that makes it worth reading: comparing against another drug rather than against nobody, which removes much of the difference between people who seek treatment and people who do not.

PMID 38182782 ↗DOI 10.1038/s41591-023-02672-2 ↗

The strongest observational evidence here, and it is a study design problem solved rather than a large number. Sweden's and Denmark's national registers followed 124,517 people starting a GLP-1 against 174,036 starting an SGLT2 inhibitor — a comparator chosen because it treats the same patients. Over a median 2.5 years, suicide deaths were 0.13 against 0.18 per 1,000 person-years, a difference that did not reach significance. Self-harm was null too. National registers capture everyone, which removes the reporting bias that cripples the database studies below.

PMID 39226030 ↗DOI 10.1001/jamainternmed.2024.4369 ↗

The best randomized psychiatric-safety data available at the obesity dose: four STEP trials pooled, 3,377 participants, with suicidality measured prospectively on a validated scale rather than reconstructed afterwards. Reports came in at 1.0% against 0.7% on placebo — a difference well inside chance — and depression scores improved slightly more on the drug. ⛔ One limitation shapes everything: these trials excluded people with serious existing mental illness, so this says little about the population most at risk.

PMID 39226070 ↗DOI 10.1001/jamainternmed.2024.4346 ↗

The largest randomized-only synthesis: 27 trials, roughly 78,432 participants, across six drugs and both indications. Nothing reached significance — not suicide, not self-harm, not suicidal ideation — and depression-related reports were numerically lower on the drugs. ⚠ The honest caveat is statistical rather than clinical: none of these trials was powered to detect rare psychiatric events, so pooling underpowered trials produces a wide confidence interval rather than a clean absence.

PMID 40105856 ↗DOI 10.1001/jamapsychiatry.2025.0091 ↗

The pharmacovigilance analysis that documented the formal signal, using the FDA's spontaneous-reporting system after the 2023 European alert. Reporting rates for suicidal ideation ran elevated for both semaglutide and liraglutide, and the signal persisted after excluding reports involving antidepressants. ⛔ What this design cannot do is tell you the rate of anything: nobody knows how many people took the drug without reporting an event, and a drug in the news generates reports for that reason alone. It records that a signal existed, not that a risk does.

PMID 38087976 ↗DOI 10.1080/14740338.2023.2295397 ↗

The same team repeating the analysis in the World Health Organization's global database, drawing on reports from more than 130 countries — and the signal held. That rules out one specific explanation, that the original finding was an artifact of how Americans report adverse events. ⚠ It shares every other limitation: no denominator, no timing, and heavy confounding by why people were prescribed the drug in the first place.

PMID 39433133 ↗DOI 10.1016/j.jad.2024.10.062 ↗

The first randomized test of any drug in this class against alcohol use disorder: 48 adults, none of them seeking treatment, on low-dose semaglutide or placebo for nine weeks. It reduced how much people drank in a laboratory setting, how much they drank on the days they drank, and how much they craved it, and cut cigarettes among smokers. ★ Note precisely what it did not move: how many days people drank, and drinks per calendar day. Small, short, and genuinely important as proof that the cohort signals reflect real pharmacology.

PMID 39937469 ↗DOI 10.1001/jamapsychiatry.2024.4789 ↗

A target-trial emulation in 222,942 patients, comparing semaglutide against seven different diabetes drugs among people carrying a tobacco use disorder. Starting it was associated with fewer tobacco-related medical encounters and more cessation prescriptions and counseling visits, and the pattern held against all seven comparators. ⚠ Consistency across comparators is a real strength. The unfixable problem is that wanting to quit is unmeasured, and people who pursue a weight drug may differ on exactly that.

PMID 39074369 ↗DOI 10.7326/M23-2718 ↗

The same approach applied to cannabis use disorder, across cohorts of 85,223 with obesity and over a million with diabetes. New diagnoses were roughly 44% and 60% lower on semaglutide, with recurrence down too. It fits the pattern the alcohol, tobacco and opioid work shows, which is what makes a shared reward-pathway mechanism plausible. ⚠ Cannabis use is poorly captured in medical records, so what is being measured is diagnosis rather than use.

PMID 38486046 ↗DOI 10.1038/s41380-024-02498-5 ↗

The substance-use series taken to its most serious endpoint: 33,006 patients with diabetes and existing opioid use disorder, where twelve-month overdose risk ran 42 to 68% lower on semaglutide depending on the comparator, holding across sex, race and severity. ⚠ People who stay engaged with medical care both take their prescriptions and overdose less often, and no adjustment fully separates those. No randomized trial in opioid use disorder has reported.

PMID 39320894 ↗DOI 10.1001/jamanetworkopen.2024.35247 ↗