Why Did Semaglutide Fail in the Alzheimer's Trials?

Because it did not slow the disease, and that was the only thing the trials were built to find out. EVOKE and EVOKE+ were two large Phase 3 trials that gave oral semaglutide or placebo to 3,808 people aged 55 to 85 with early, confirmed Alzheimer's disease, and followed them for two years. At the end, the people on semaglutide had declined in memory and daily functioning just as much as those on placebo 12. Some Alzheimer's-related biological markers improved in the semaglutide groups, but that improvement did not show up as any delay in the illness itself 1.

This is a different situation from most of the compounds discussed on this site. The usual problem with a peptide is that nobody has run the trial. Here the trial was run, properly, at scale, and it answered the question. A clear negative from a well-designed trial is one of the most useful things medicine produces, even when the answer is disappointing.

What were the EVOKE trials?

Two near-identical randomised, double-blind, placebo-controlled trials run by Novo Nordisk: EVOKE with 1,855 participants and EVOKE+ with 1,953 1. Everyone enrolled was aged between 55 and 85 and had either mild cognitive impairment or mild dementia caused by Alzheimer's disease, and the Alzheimer's diagnosis was confirmed by showing the amyloid build-up that defines the condition. Participants took semaglutide tablets or identical placebo tablets once a day 1.

The main phase lasted 104 weeks, with a planned one-year extension after it 1. Running two trials in parallel is deliberate: if both reach the same answer independently, chance becomes a very unlikely explanation. The top-line results were announced in November 2025, and the full reports were published in The Lancet in March 2026 12.

Why did anyone think a diabetes drug would help?

Because of a combination of biology and a striking signal in existing data. On the biology, GLP-1 receptors are found in the brain as well as the pancreas and gut, and GLP-1 signalling has effects on inflammation and on cells outside the metabolic system 4. Alzheimer's disease involves chronic inflammation in the brain, and diabetes is a known risk factor for dementia, so a drug acting on both metabolism and inflammation was a reasonable candidate.

The signal came from people with type 2 diabetes. A 2022 analysis pooled three large cardiovascular trials of GLP-1 drugs, covering 15,820 patients, and found fewer dementia diagnoses among those randomised to the drug than to placebo, with a hazard ratio of 0.47. The same paper examined a Danish national register of 120,054 patients and found that longer exposure to this class of drug was associated with a lower rate of dementia 3.

Those findings were real and worth following up, which is exactly what EVOKE did. But look at what they were. The pooled trials were designed to count heart attacks and strokes, not dementia; dementia was a secondary observation, the number of cases was small, and the confidence interval around that hazard ratio was wide 3. The register data were observational, which means people were not randomly assigned, and those who took GLP-1 drugs for longer may have differed from those who did not in many other ways. And the dementia counted was dementia of any kind, not specifically Alzheimer's disease.

What exactly did the trials measure?

The main measure was the change over 104 weeks on the Clinical Dementia Rating Sum of Boxes, usually shortened to CDR-SB 2. It is a standard scale in Alzheimer's research. A clinician rates a person in six areas — memory, orientation, judgement and problem-solving, community affairs, home and hobbies, and personal care — and the six scores are added together. A rising score means worsening.

Choosing a scale like this as the main measure is a deliberate decision. It captures things patients and families notice: whether someone can manage money, find their way, cook, dress. Regulators treat changes on scales of this kind as meaningful precisely because they describe how a person is living, not what a laboratory test shows. The trials also measured biomarkers — substances in the blood and in the fluid around the brain and spinal cord that track the underlying disease — but those were supporting measures, not the question the trials were built to answer.

What did the trials find?

No difference between semaglutide and placebo on the main measure, in either trial 12. People in both groups declined at the same rate over the two years. The plain-language summary of the published results puts it directly: semaglutide taken once a day by mouth did not slow cognitive or functional decline in people with early Alzheimer's disease 2.

The secondary findings did not rescue it. Novo Nordisk stopped the planned one-year extension of both trials on the basis of the efficacy results 1. On safety, the side effects were similar to those already known from semaglutide in other conditions, and no new adverse effects were identified 2. So the drug behaved as expected in the body and was tolerated; it simply did not do what the trials were testing for.

If the biomarkers improved, why does it count as a failure?

Because a biomarker is a stand-in for the outcome, not the outcome, and in these trials the stand-in moved while the outcome did not. Novo Nordisk reported that semaglutide improved Alzheimer's-related biomarkers in both trials, and said in the same announcement that this did not translate into a delay in disease progression 1.

A biomarker earns its place when changes in it reliably predict changes in how patients do. That is the whole reason to measure one. When a treatment shifts the marker and leaves the patient's course unchanged, it tells you one of two things: either the marker was not capturing the part of the disease that matters for symptoms, or the change was too small or too late to make a difference. Either way, the patient-level result is the one that counts.

The biomarker movement is not meaningless, though. It suggests semaglutide had some biological activity relevant to the disease process. That is a finding worth having. It is just not the same as a finding that the drug helps people with the disease, and treating it as one would repeat the exact mistake that surrogate endpoints are known for.

So why didn't it work?

Nobody knows for certain, and a negative trial does not come with an explanation attached. What can be done is to list the plausible reasons and be clear that they are hypotheses, not findings. The main candidates:

• The original signal may not have been what it looked like. Fewer dementia diagnoses in people with diabetes on these drugs could reflect better control of blood sugar, blood pressure and vascular disease — which affects vascular dementia more than Alzheimer's — or differences between the people who took the drugs and those who did not.
• The disease may have been too far along. Everyone enrolled already had symptoms. By that stage amyloid and tau changes in the brain have been building for many years, and an effect on inflammation or metabolism may be too little, too late.
• The drug may not reach the relevant parts of the brain in enough quantity. Whether a compound that acts strongly on appetite also acts strongly on the brain regions affected in Alzheimer's is an open question.
• The mechanism may simply not matter enough. Inflammation and metabolism may be real features of Alzheimer's disease without being levers that change its course once it is under way.

These are not mutually exclusive, and more than one may be true. What the trials establish is the result; working out why will take other studies designed to test each explanation directly.

Was the earlier evidence wrong?

Not wrong, but it was answering a different question. The 2022 analysis found an association between GLP-1 drugs and fewer dementia diagnoses in people with type 2 diabetes 3. EVOKE tested whether semaglutide slows Alzheimer's disease that has already produced symptoms, in people who were not selected for diabetes 1. Those are different populations, different stages and different outcomes.

This is the normal path of evidence, and it is worth seeing it work. An association in existing data raises a question. A randomised trial is designed to answer it. Sometimes the trial confirms the association; often, as here, it does not. Associations in observational data are the cheapest form of evidence to produce and the least reliable guide to what a treatment will do, which is precisely why trials exist.

Does this mean GLP-1 drugs do nothing for the brain?

No — it means this drug did not slow this disease at this stage. The trials did not test whether GLP-1 drugs prevent dementia in people who do not yet have symptoms, whether they affect vascular dementia, whether a different compound in the class would behave differently, or whether one might work alongside another treatment. Each of those remains an open question.

The danger runs in both directions. Reading EVOKE as proof that GLP-1 drugs are useless for the brain overstates it. Reading the biomarker changes as a hidden success, or the earlier observational data as still standing untouched, understates it. The accurate position is narrow: for early symptomatic Alzheimer's disease, oral semaglutide was tested properly and did not work.

So what is the short answer?

Semaglutide failed in the Alzheimer's trials because it did not slow the decline in memory and daily function that the trials measured, in two large, parallel, placebo-controlled studies of people with early Alzheimer's disease 12. The biomarker improvements were genuine but did not reach the patients.

The reasons are still hypotheses: an earlier signal that may have reflected diabetes and vascular health more than Alzheimer's disease, treatment that started after symptoms had appeared, uncertain exposure in the brain, or a mechanism that is real but not decisive. What is not in doubt is the result. It is a clean answer to a properly asked question — which is more than most questions in this field ever get.