Can Brain Training Delay Dementia? A 20 Year Trial, Read Closely

In 1998 and 1999, researchers enrolled 2,802 older Americans in a trial called ACTIVE and gave some of them a few weeks of training on a computer task that rewards seeing things fast. Twenty years later, the group that got that speed training plus a few booster sessions had been diagnosed with dementia less often than the control group. The result, published in February 2026 in Alzheimer's & Dementia: Translational Research & Clinical Interventions, was promoted by the National Institutes of Health as evidence that weeks of training may delay dementia over decades. It is one of the most interesting findings in the field. It is also one of the easiest to oversell.

The numbers

ACTIVE, which stands for Advanced Cognitive Training for Independent and Vital Elderly, randomized participants to one of three types of training, memory, reasoning or speed of processing, or to a control group with no training. Some trainees also received booster sessions later on. The new analysis, by Coe and colleagues, linked 2,021 of the original participants to their Medicare records from 1999 to 2019 to see who received a dementia diagnosis.

In the control group, 239 of 491 people, 48.7%, were diagnosed with dementia over that period. In the group that received speed training with boosters, 105 of 264, 39.8%, were. The hazard ratio was 0.75, with a 95% confidence interval of 0.59 to 0.95, meaning roughly a quarter lower risk at any given time, an effect that just clears the bar for statistical significance.

The other arms did not show the same thing. Speed training without boosters had a hazard ratio of 1.01, no difference at all. Memory and reasoning training, with or without boosters, were not statistically significant.

The effect lives in one arm out of six, and only in the arm that got the extra sessions.

Why the details matter

Several features of the study cut against the headline. The analysis was preregistered on AsPredicted (number 67523), which is good practice, but the authors acknowledge the preregistration inadvertently left out the booster sessions, which turned out to be where the effect sits. Dementia was identified through Medicare diagnostic codes using an algorithm with sensitivity and specificity of about 85 to 90%, so some people were misclassified in both directions. And 77% of participants died during the follow up, which complicates any comparison of who developed dementia, because death competes with diagnosis.

There is also a disclosure readers should know. Karlene Ball, one of the investigators, is a consultant to and shareholder in Posit Science, the company that sells speed of processing training software. Disclosed conflicts do not make a result wrong, but they are part of how to weigh it.

A public argument about statistics

The paper drew a formal critique. In a commentary published in June 2026, Dickens argued that once you correct for testing several arms at once, using a Bonferroni adjustment, none of the confidence intervals excludes no effect. The same commentary made a second, more intriguing point: across age groups, all 12 hazard ratios fall below 1, a pattern unlikely by chance (p of about 0.00025). That reading suggests a modest effect shared by all three kinds of training, strongest in younger participants, rather than a special property of speed training.

The original authors replied in July 2026. Gross, Rebok and Albert argued that each arm was prespecified as its own comparison, so no correction was needed, and accused the critique of internal inconsistency. Both sides are serious researchers, and both positions are defensible. That is precisely why the honest summary is uncertain.

The brain game industry has been here before

In 2016, the Federal Trade Commission settled with the makers of Lumosity over claims that its games could delay cognitive decline. The company agreed to pay $2 million, with a $50 million judgment suspended. The same year, a large review in Psychological Science in the Public Interest by Simons and colleagues found little evidence that brain training improves skills far beyond the trained task itself, the so called far transfer problem.

ACTIVE is a different kind of evidence from an app's marketing. It is randomized, long, and published with its warts visible. But the leap from "one arm of a 1990s trial, with boosters, under one statistical reading" to "brain games prevent dementia" is exactly the leap those earlier cases warned about. Anyone worried about memory or cognitive decline should talk with a doctor rather than a download page.

The most useful reading may be the least exciting one. If the critique is right and all three kinds of training nudged risk down a little, the active ingredient may be structured mental effort and contact with other people, not the specific software. That would be harder to sell and easier to test. A new trial designed around boosters from the start would settle more than another 20 years of arguing over this one.