Two trials called evoke and evoke+ tested whether a daily semaglutide tablet could slow early Alzheimer’s disease. They were the largest test the hypothesis has ever received, they finished, and their results are published. Both missed. The record states that both trials have been discontinued due to negative clinical outcome, and the authors’ own interpretation is that oral semaglutide was not efficacious in slowing clinical progression in early Alzheimer’s disease.[1]
That is worth reading closely rather than filing, because the evidence that motivated these trials was not weak, and the way it failed says something about how to read every other secondary claim made for this class — including the ones in the mood and behavior article.
The mechanism argument was better than most
The rationale had three legs. The GLP-1 receptor is expressed in the brain; animal models showed effects on neuroinflammation and neurodegeneration; and people with type 2 diabetes or obesity treated with these drugs appeared in observational data to develop dementia less often. The trials were designed around a disease-modifying, multimodal effect running through neuroinflammatory, vascular and other Alzheimer-related processes rather than through amyloid clearance.[2]
The strongest version of the prior was randomized rather than observational. A 2025 systematic review and meta-analysis pooled 26 randomized trials of cardioprotective glucose-lowering drugs that reported dementia or cognitive outcomes. Overall, those drugs were not associated with less cognitive impairment or dementia (OR 0.83; 95% CI, 0.60 to 1.14). But split by class, the 10 GLP-1 receptor agonist trials gave OR 0.55 (95% CI, 0.35 to 0.86) for all-cause dementia, while the 12 SGLT2 inhibitor trials gave OR 1.20 (95% CI, 0.67 to 2.17), with a P value for heterogeneity of 0.04.[3]
A 45% reduction, drawn from randomized data, with a comparator class pointing the other way. That is a serious signal, and it is the kind of finding that justifies spending several years and several thousand participants to test directly. It is also, in retrospect, a demonstration that dementia counted as an adverse event in trials designed to measure heart attacks is not the same measurement as a cognitive scale in trials designed to measure cognition.
What the trials were built to do
evoke and evoke+ were multicenter, randomized, double-blind, placebo-controlled phase 3 trials run across 566 sites in 40 countries. Eligible participants were aged 55 to 85 with mild cognitive impairment or mild dementia due to Alzheimer’s disease and amyloid abnormality confirmed by positron emission tomography or cerebrospinal fluid. Assignment was 1:1 to once-daily oral semaglutide titrated to 14 mg or to placebo — 3 mg for four weeks, 7 mg for four weeks, then 14 mg — for up to 156 weeks, structured as a 104-week main phase plus a 52-week blinded extension.[2]
The primary endpoint was change from baseline to week 104 in the Clinical Dementia Rating Sum of Boxes, a scale on which a higher number means more impairment, so a negative treatment difference favors the drug. Plasma biomarkers were collected from everyone and a cerebrospinal fluid substudy was planned in about 210 participants. Enrollment ran from May 18, 2021 to September 8, 2023.[2] The two trials shared their inclusion criteria with one exception: evoke+ was designed to also admit participants with significant small vessel pathology, the group the vascular half of the mechanism argument pointed at most directly.[4]
The dose is the one already studied in cardiovascular outcomes. Oral semaglutide at 14 mg is the same strength discussed in the SOUL trial article, and the delivery differences between a daily tablet and a weekly injection are set out in the oral versus injectable article.
What they found
Of 9,981 people screened, 3,808 were randomized: 1,855 in evoke (928 semaglutide, 927 placebo) and 1,953 in evoke+ (976 semaglutide, 977 placebo). Mean age was 72.2 years (SD 7.1) and mean Clinical Dementia Rating Sum of Boxes score was 3.7 (SD 1.6) at baseline. Over 104 weeks, that score rose 2.3 (SE 0.1) on semaglutide against 2.3 (0.1) on placebo in evoke, and 2.2 (0.1) against 2.1 (0.1) in evoke+.[1]
The estimated differences were −0.08 (95% CI, −0.35 to 0.20; P = 0.57) in evoke and +0.10 (95% CI, −0.17 to 0.38; P = 0.46) in evoke+.[1] Two trials with the same protocol, the same dose and the same endpoint produced point estimates on opposite sides of zero. Neither interval excludes no effect, both intervals are narrow, and they overlap almost completely. This is not a near miss or an underpowered result; it is the signature of a treatment effect that is not there.
The subgroup the second trial existed for barely enrolled
evoke+ had one job the first trial did not: to include people with significant small vessel pathology on magnetic resonance imaging. In the event, 54 of 1,953 participants — 2.8% — met that criterion.[1] The design feature that distinguished the two trials applied to fewer than three participants in a hundred, which is the most economical explanation for why the two trials returned near-identical answers.
That is not a criticism of the sponsors so much as a lesson about enrichment. The vascular argument was the most testable part of the rationale, and the trial designed to test it recruited a population in which it was almost absent. Nothing in this published record establishes what would have happened in a trial that actually enrolled the vascular subgroup, and nothing in it licenses the claim that such a trial would succeed.
A smaller trial had already shown the same shape
The phase 2b ELAD trial randomized 204 participants with mild to moderate Alzheimer’s disease syndrome and no diabetes to daily liraglutide injections or placebo for 52 weeks. Its primary outcome was change in cerebral glucose metabolic rate, and it found no significant difference between groups. On secondaries, the Alzheimer’s Disease Assessment Scale executive domain favored liraglutide at 0.15 (95% CI, 0.03 to 0.28; unadjusted P = 0.01), while the Clinical Dementia Rating Sum of Boxes — the endpoint evoke would later use as its primary — showed −0.06 (95% CI, −0.57 to 0.44; unadjusted P = 0.81).[5]
One unadjusted secondary reached significance and the endpoint that mattered did not. Read on its own, that trial reads as encouraging. Read next to evoke, it reads as the ordinary behavior of secondary endpoints in a trial that missed its primary, which is the same reading problem that recurs in the smoking cessation evidence.
What the trials did establish, which is about safety
Treatment-emergent adverse events were reported in 1,729 (91.2%) of 1,896 participants receiving semaglutide against 1,613 (84.8%) of 1,902 receiving placebo, and the authors conclude that safety and tolerability in early Alzheimer’s disease were consistent with studies in other indications. Five fatalities were considered treatment-related by investigators: one in the semaglutide group and four in the placebo group.[1]
A pooled analysis assembled while the trials were still running gives the wider context for this age group. Across three phase 3a semaglutide programs, 3,529 participants aged 65 or older had adverse events at 73.6% to 92.4% against 73.2% to 90.8% in the overall populations, but permanent discontinuation for an adverse event was more frequent in the older group, at 9.3% to 12.4% against 5.7% to 8.7%. Estimated weight loss in those aged 65 and over was 3.8% at week 52 against 0.1% on placebo.[6] Dosing considerations for this population are covered in the older adults article.
What this settles, and what it does not
It settles the treatment question as asked. In adults aged 55 to 85 with amyloid-confirmed mild cognitive impairment or mild dementia, oral semaglutide 14 mg daily for two years does not slow progression on the Clinical Dementia Rating Sum of Boxes. Anyone offering a GLP-1 receptor agonist as a treatment for diagnosed Alzheimer’s disease is offering something that has been tested at scale and did not work.
It does not settle prevention. Every participant already had amyloid-confirmed disease and a mean impairment score of 3.7, and a drug that fails to slow an established neurodegenerative process may still, in principle, act on the metabolic and vascular risk that precedes it by decades. That is a different trial in a different population, and no such trial has reported. Nor does the result transfer automatically to other molecules or routes: evoke tested one drug, one dose, one delivery form.
For anyone buying a GLP-1 through a cash subscription, the practical content is short. No product in this class is FDA-approved for Alzheimer’s disease, cognition or dementia prevention, so any prescription written for those purposes is off-label. Most product sold through cash telehealth is compounded, which is not FDA-approved and is not reviewed by the FDA for safety, efficacy or quality before it is dispensed. A cognitive claim attached to a subscription now sits against two completed phase 3 trials, and how figures on this site are established before publication is set out in the methodology.