For about two years, clinics fielded a version of the same question from families: my mother is on Ozempic for her diabetes, is it also helping her memory? The question was not foolish. It came from a run of headlines built on real data, including a target trial emulation published in Alzheimer's and Dementia on 24 October 2024, which combed the health records of 1,094,761 adults in the United States with type 2 diabetes and found that the 17,104 who started semaglutide had a lower rate of a first recorded Alzheimer's diagnosis over three years than users of seven other glucose lowering drugs, with hazard ratios ranging from 0.33 against insulin to 0.59 against other GLP-1 drugs. Those authors said plainly that causation would have to be settled by randomised trials. On 19 March 2026, in The Lancet, it was.

The trials were evoke and evoke+ (NCT04777396 and NCT04777409), reported by Jeffrey Cummings, Alireza Atri, Howard Feldman and colleagues, and funded by Novo Nordisk. They were phase 3, randomised, double blind and placebo controlled, run across 566 sites in 40 countries. Between 18 May 2021 and 8 September 2023, 9,981 people were screened and 3,808 randomly assigned: 1,855 in evoke (928 semaglutide, 927 placebo) and 1,953 in evoke+ (976 and 977). Everyone was aged 55 to 85 with amyloid confirmed Alzheimer's disease and either mild cognitive impairment or mild dementia. Mean age was 72.2 years; mean CDR-SB score at baseline was 3.7. Participants took an oral semaglutide pill titrated from 3 mg to 7 mg to 14 mg daily, with flexibility to drop back, or matching placebo. The primary endpoint was change in CDR-SB, the sum of boxes rating that scores memory, orientation, judgment, community affairs, home life and personal care, from baseline to week 104. In evoke, CDR-SB worsened by 2.3 points on semaglutide and 2.3 on placebo, an estimated difference of minus 0.08 (95 percent CI minus 0.35 to 0.20, p=0.57). In evoke+, 2.2 versus 2.1, a difference of 0.10 (minus 0.17 to 0.38, p=0.46). The confirmatory secondary endpoint, the ADCS-ADL-MCI measure of daily function, showed no separation either. Both trials have been discontinued for negative clinical outcome, and the planned 52 week extension was cancelled.

Here is the part that will stay with the field. The drug was demonstrably doing something. Reporting the biomarker results at the CTAD conference in San Diego in December 2025, Cummings said plasma high sensitivity C-reactive protein, a general marker of inflammation, fell by roughly 30 percent in both trials. In a small spinal fluid substudy of about 200 participants, several markers shifted modestly in the hoped for direction, by 10 percent or less and only nominally significant: phosphorylated tau 181 and 217, total tau, neurogranin and the inflammation marker YKL-40. Weight came off, enough that decreased weight was logged as an adverse event in 36.5 percent of the semaglutide group against 7.4 percent on placebo. And yet in blood, none of those tau markers budged; glial fibrillary acidic protein rose by about 4 percent in both trials and neurofilament light rose by about 5 percent in evoke+, a direction nobody had predicted and which Cummings said he could not explain. Metabolism and inflammation were engaged. The trajectory of a person's memory, orientation and independence was not.

That gap is the story. A biomarker is a measurement that travels with a disease; an outcome is what happens to the person. The two can come apart, and when they do, the outcome wins. Registry epidemiology carries a parallel trap: people who are prescribed a newer, pricier drug and who stay on it for three years differ from people who are not, in ways no statistical adjustment fully erases, and a diagnosis code for Alzheimer's in a health record is not the same as amyloid confirmed disease. In evoke and evoke+, every participant was amyloid confirmed, treatment was assigned by chance, and the effect vanished. Robert Howard, professor of old age psychiatry at UCL, said when the topline results emerged in November 2025 that this had always been "something of a hopeful thinking longshot", and that a trial this large lets the field close the question and move on.

On safety, which matters to any family whose relative takes semaglutide for diabetes or weight, the trials found nothing unexpected in a group of seniors with early Alzheimer's. Treatment emergent adverse events were more common with semaglutide, 1,729 of 1,896 people (91.2 percent) against 1,613 of 1,902 (84.8 percent), driven by weight loss, reduced appetite and nausea (24.3 percent). Serious adverse events were somewhat fewer on the drug, 20.4 percent against 23.8 percent. Falls and bone and joint injuries were numerically lower in the semaglutide groups in both trials, a finding the authors did not claim as a benefit. The trial team also flagged a real limitation of their own: weight loss and stomach upset are hard to hide, so some participants and staff may have guessed who was on what.

What the trials could not answer is almost as important as what they did. evoke+ was meant to open a window onto vascular co-pathology, but only 54 participants (2.8 percent) turned out to have significant small vessel pathology, far too few to say anything about semaglutide and vascular contributions to dementia. Nobody tested the injectable formulation, nobody tested people without symptoms, and with the extension cancelled the three year data will remain partial. Several researchers, including the Alzheimer's Drug Discovery Foundation's Howard Fillit, have argued the metabolic pathway may still belong in combination approaches. That is a hypothesis. After evoke, it is a hypothesis that will have to be tested in people, prospectively, with a clinical endpoint, before anyone should believe it.