The bottle on the kitchen table has become a small referendum. In September 2025, a study in BMJ Evidence-Based Medicine was reported around the world as closing the case: no safe level. Families who had spent years being told that a daily glass might be protecting a parent's memory quietly poured it away. Then, on 10 December 2025, the Internal Medicine Journal published a systematic review and meta-analysis by Zhang, Li and Miao that pooled 21 observational studies and found no significant link between alcohol exposure and dementia at all: a relative risk of 1.03 for all-cause dementia, with a confidence interval running from 0.84 to 1.27. For Alzheimer's disease the pooled figure was 0.97 (0.86 to 1.08), for vascular dementia 1.09 (0.95 to 1.26). Two peer-reviewed papers, one question, and headlines pointing in opposite directions.
The two studies are not really arguing about the same thing, and that is the key a clinician needs. The Internal Medicine Journal paper is a pooling exercise: literature searched to 22 July 2024, study quality graded on the Newcastle-Ottawa Scale, random-effects models because heterogeneity was high. Its own subgroup analysis is where the familiar J-curve reappears. Light to moderate drinking was associated with lower risk of all-cause dementia (RR 0.88, 0.81 to 0.96) and of Alzheimer's disease (RR 0.88, 0.79 to 0.97), while heavy drinking was associated with higher risk across the board: all-cause dementia 1.18 (1.02 to 1.36), Alzheimer's 1.29 (1.21 to 1.36), vascular dementia 1.25 (1.11 to 1.40). A flat overall number, in other words, can be two opposite signals cancelling out.
The September study, by Anya Topiwala and colleagues, went at the same curve from underneath. It followed 559,559 people aged 56 to 72 in the US Million Veteran Program and UK Biobank, of whom 14,540 developed dementia over an average of four years of follow-up in the American group and twelve in the British. Its observational half found the classic U shape: compared with light drinkers taking fewer than seven drinks a week, risk was 41 percent higher in both non-drinkers and those drinking 40 or more a week, and 51 percent higher in people with alcohol dependence. Then the researchers swapped self-reported drinking for genetic variants associated with drinking, drawing on genome-wide data from 2.4 million people. That curve straightened. A one standard deviation rise in genetically predicted drinks per week was associated with 15 percent higher dementia risk, and a doubling of genetic liability to alcohol use disorder with 16 percent. The dip at the bottom vanished.
Why would it vanish? Because of reverse causation, the quiet saboteur of every drinking study ever done. People in the earliest, undiagnosed years of dementia tend to cut back, and the Topiwala group found that drinking fell faster before diagnosis in those who went on to develop dementia than in those who did not. Anyone who has watched a relative lose interest in the Friday pint eighteen months before anyone said the word dementia has seen this happen. Pool enough cohorts in which the sickest people have already stopped drinking, and abstinence looks dangerous while moderation looks protective. Mendelian randomisation sidesteps that, because genes are fixed at conception. It has its own weakness, though: the genetic instruments explain only a small slice of actual drinking behaviour, they capture lifetime tendency rather than what someone drinks at 62, and variants linked to alcohol can influence risk through other routes entirely, which is the pleiotropy problem. One method is biased towards finding a benefit that is not there. The other can overstate harm it cannot fully isolate.
So what can a clinician honestly say in a ten-minute appointment? That the two papers agree on more than the headlines suggest. Heavy drinking is associated with higher dementia risk in both, and neither offers any evidence that starting to drink, or drinking more, protects the brain. The 2024 Lancet standing Commission on dementia already lists excessive alcohol, defined there as more than 21 units per week, among 14 modifiable risk factors, and nothing published in 2025 moves that judgement in a reassuring direction. What is genuinely unresolved is the bottom of the range: whether one drink a night is neutral, faintly harmful, or faintly helpful. The December meta-analysis says the observational literature still leans towards a small benefit at low doses. The September analysis says that lean is probably an artefact. Neither is a randomised trial, and there will almost certainly never be one.
For a family, the practical shape of that uncertainty is narrower than it sounds. Under either reading, the advice a clinician can defend is the same at the top end of the dose range and only genuinely contested at the very bottom. What has changed in 2025 is not the number of units on a leaflet but the confidence with which anyone can describe a glass of wine as good for the brain. That claim has lost its best evidence, and it has not found new evidence to replace it.