Every few months another database study reports that people on these drugs get diagnosed with dementia less often, usually 30 to 50% less. Those studies share one flaw that no amount of statistical adjustment removes: people prescribed a newer, more expensive drug differ from people who are not, in health, in engagement with care, and in ways nobody measured. The randomized picture is thinner and less flattering. The phase 2 liraglutide trial missed its primary endpoint and moved one secondary cognitive measure. Then the two phase 3 trials of oral semaglutide, evoke and evoke+, in 3,808 people with early Alzheimer's, were published in May 2026: over two years, semaglutide did not slow decline on the standard dementia-severity score. That answers the treatment question for early Alzheimer's. Whether these drugs prevent dementia in people who do not have it is a separate question no dedicated trial has answered, and the observational studies below are the case for asking it. Nothing here supports taking a GLP-1 to prevent dementia today.
Ranked papers
#1ELAD
Edison P, Femminella GD, Ritchie C, et al. · Nat Med · 2026
What it measured: How actively the brain used glucose on PET scan, after a year
The phase 2 randomized trial of a GLP-1 in Alzheimer's: 204 participants with mild-to-moderate disease and no diabetes, on daily liraglutide or placebo for a year across UK centers. The primary endpoint, brain glucose metabolism on PET, showed no difference. A secondary cognitive measure, the executive-function domain of the standard Alzheimer's scale, favored liraglutide (0.15, CI 0.03 to 0.28), while daily-function and overall dementia-severity scores did not move. That p-value is unadjusted and the outcome is secondary, which is exactly the shape of result that fails to replicate. It is not a positive trial.
PMID 41326666 ↗DOI 10.1038/s41591-025-04106-7 ↗
#2evoke / evoke+
Scheltens P, Atri A, Feldman HH, et al. · Alzheimers Dement (N Y) · 2026
What it measured: A standard dementia-severity score at two years (no effect, reported May 2026)
The enrollment description for the two phase 3 trials, testing oral semaglutide 14 mg against placebo in people whose early symptomatic Alzheimer's had been confirmed by amyloid testing. It documents who got in: mean age about 72, most at the mildest stage of dementia severity, and cholinesterase-inhibitor use ranging from 42% in North America to 62% in Asia. The results, published in The Lancet in May 2026 for 1,855 people in evoke and 1,953 in evoke+, were negative: semaglutide did not slow decline on the dementia-severity score at two years. This paper shows which patients that result applies to.
PMID 41522368 ↗NCT04777396 ↗DOI 10.1002/trc2.70200 ↗
#3REWIND (cognitive substudy)
Cukierman-Yaffe T, Gerstein HC, Colhoun HM, et al. · Lancet Neurol · 2020
What it measured: How many people crossed into substantial cognitive impairment on standard tests
The first randomized hint, from an exploratory analysis of a cardiovascular trial that followed 9,901 people with type 2 diabetes for a median 5.4 years. Substantial cognitive impairment occurred at 4.05 per 100 patient-years on dulaglutide against 4.35 on placebo, a gap that fell short of statistical significance (hazard ratio 0.93, p = 0.11). It became a 14% relative reduction only after a post hoc adjustment for baseline test scores. Note the absolute difference: three tenths of a case per 100 patient-years. It is the largest randomized cognitive dataset in the class, and the effect it describes is small and fragile.
PMID 32562683 ↗NCT01394952 ↗DOI 10.1016/S1474-4422(20)30173-3 ↗
#4
Wang W, Wang Q, Qi X, et al. · Alzheimers Dement · 2024
What it measured: A first Alzheimer's diagnosis over three years
The study that launched the current wave: more than a million patients with type 2 diabetes in a US health-records network of 116 million people, with semaglutide compared against seven different alternatives one at a time. First-time Alzheimer's diagnoses ran 40 to 70% lower, with the biggest gap against insulin. That last detail is the tell. People started on insulin are sicker and further along, and the largest apparent benefit appearing against the sickest comparator is what confounding looks like. It is hypothesis-generating.
PMID 39445596 ↗DOI 10.1002/alz.14313 ↗
#5
Wang W, Davis PB, Qi X, et al. · J Alzheimers Dis · 2025
What it measured: First-time Alzheimer's disease-related dementia (ADRD) diagnosis
The same group extending the outcome from Alzheimer's to dementia of all kinds, in 1,710,995 patients with type 2 diabetes. Semaglutide was associated with lower overall dementia risk, with hazard ratios from 0.54 against insulin to 0.80 against older GLP-1 drugs. By type, vascular dementia was lower, while frontotemporal and Lewy body dementia showed no association. A vascular signal fits a plausible, unglamorous mechanism (better blood vessels) better than a direct effect on amyloid, and would make it a side effect of cardiovascular benefit rather than a new one.
PMID 40552638 ↗DOI 10.1177/13872877251351329 ↗
#6
Sabbagh M, Boschini C, Cohen S, et al. · Alzheimers Dement (N Y) · 2025
What it measured: Adverse-event incidence in adults aged ≥65
The practical safety question behind the phase 3 trials: is this drug tolerable in the people who would take it for dementia? Pooled across three phase 3a semaglutide programs and 3,529 participants aged 65 and over, adverse events ran at rates similar to the full trial populations, with gastrointestinal effects the most common (44.6% to 73.8% of older participants). Stopping because of side effects was more frequent at 65 and over (9.3% to 12.4%, against 5.7% to 8.7% overall). The authors, several of them Novo Nordisk employees, concluded that age did not change the safety picture, but that higher dropout rate is the result that matters most in this population.
PMID 40337158 ↗DOI 10.1002/trc2.70076 ↗
#7
Younis A, Henney AE, Riley DR, et al. · Diabetes Res Clin Pract · 2026
What it measured: First-time all-cause dementia diagnosis
Three simulated head-to-head comparisons run inside a global records network: tirzepatide and semaglutide each against SGLT2 inhibitors, and tirzepatide against semaglutide. Over two years, tirzepatide was associated with less dementia than SGLT2 inhibitors (hazard ratio 0.66) and than semaglutide (0.69, just reaching significance); the abstract reports no dementia result for semaglutide against SGLT2 inhibitors. The authors flag the subtlety that makes this useful. SGLT2 inhibitors have themselves looked protective in earlier work, so beating them is a harder test than beating insulin.
PMID 41544899 ↗DOI 10.1016/j.diabres.2026.113083 ↗
#8
da Silva AMP, Januário Campos Cardoso L, Batista S, et al. · J Diabetes Complications · 2026
What it measured: New diagnoses of mild cognitive impairment, dementia or Alzheimer's
A direct tirzepatide-against-semaglutide comparison across 44,470 matched pairs with type 2 diabetes, counting diagnoses made at least a year after starting. Tirzepatide was associated with much lower risk ratios for mild cognitive impairment (0.12) and dementia (0.15), and a non-significant one for Alzheimer's (0.48), with low absolute risks throughout. Two drugs, same database, same design, and therefore the same problem: whatever distinguishes people prescribed one from people prescribed the other is not in the record. The authors call the findings descriptive and hypothesis-generating.
PMID 41825212 ↗DOI 10.1016/j.jdiacomp.2026.109306 ↗
#9
Lee WT, Wang JT, Tsai MH, et al. · Nephrol Dial Transplant · 2026
What it measured: Dementia of any kind, and Alzheimer's specifically
A study limited to patients who had both type 2 diabetes and stage 3 or worse chronic kidney disease, a group at raised dementia risk. Compared with new users of DPP-4 inhibitors, GLP-1 users had a 20% lower hazard of dementia (0.80) and a 24% lower hazard of Alzheimer's (0.76). Vascular dementia, frontotemporal dementia, Lewy body dementia and Parkinson's showed no difference, which does not fit a simple circulation story.
PMID 41697144 ↗DOI 10.1093/ndt/gfag032 ↗
#10
O'Mara A, Mody BP, Mammi M, et al. · Neurol Sci · 2026
What it measured: Pooled cognitive score change versus placebo
The pooled randomized evidence in people with cognitive impairment but not diabetes, which amounts to four trials and 112 patients. Cognitive scores did not differ from placebo (standardized mean difference −0.10, 95% CI −0.53 to 0.34), and the Alzheimer's cognitive subscale showed no effect either. With so few patients the estimate is wide. It marks where the small-trial evidence stood before the phase 3 trials reported, and the phase 3 result pointed the same way.
PMID 41524953 ↗DOI 10.1007/s10072-025-08602-z ↗
About this list
Obesity medicine throws up the same handful of questions over and over, and we answer them with ranked lists anchored to the papers themselves. On 2026-10-30 every citation below was put back through PubMed. Where a trial has a registry entry, the card carries that link too, alongside the PubMed one.
Browse our full index of research lists or our long-form research articles.