An exercise programme can show a better average score on a thinking task without proving that it will improve your everyday planning. That gap matters when a class, app or headline promises a sharper brain. The useful decision is whether the evidence supports the particular benefit being offered, rather than whether exercise deserves a single positive or negative verdict.
The short answer
Research supports the possibility of modest average cognitive benefits from repeated exercise, but estimates depend on the population, comparison and analytical choices. A broad 2025 umbrella review reported a positive executive-function estimate. A narrower 2023 review of healthy populations found that its overall cognitive estimate became much smaller and inconclusive after publication-bias adjustment. Those results answer different questions. [S2, S3]
Neither result gives a personal percentage improvement or validates a paid brain-training addition. If you already follow an appropriate activity plan, these findings are not a reason to replace it with a programme selling an untested cognitive promise. This explainer helps you evaluate that promise; it does not prescribe exercise or assess memory symptoms.
Definitions and scope
Executive functions help control thinking and action: keeping information available, resisting an unhelpful response and changing approach when a task demands it. In the 2025 review, working-memory tasks were classified within executive function. Memory and general cognition were analysed separately. A result for one domain should not be relabelled as an improvement in all thinking skills. [S2]
A laboratory score is also different from everyday independence. Remembering a sequence of digits, completing a switching task and managing a complicated appointment are not interchangeable outcomes. Improvement on a measured task does not establish the last outcome unless the study measured it.
Here, repeated training means a programme continued over weeks. Singh and colleagues’ 2025 review required at least four weeks and excluded single exercise bouts. Ciria and colleagues’ 2023 review allowed programmes lasting at least two weeks. Neither synthesis supplies a direct answer to whether one walk before an examination will improve a student’s performance. [S2, S3]
The participants matter just as much. The 2025 review included healthy and clinical populations across a wide age range. The 2023 review restricted its question to healthy populations and excluded mind-body activities and programmes combined with cognitive training. Its conclusion should not be exported to a dementia-treatment population; the broader review’s result should not be assumed to describe every healthy adult equally. [S2, S3]
How it works
There are several possible routes from an exercise programme to a different test score. The reviews discuss physiological explanations, including changes in blood supply and neuroplasticity, alongside cognitive engagement and social contact. These are proposed contributors. Pooling a programme’s test results does not identify which mechanism produced them. [S2, S3]
The comparison group helps separate those explanations. A no-programme control may differ from an exercise class in attention, expectations and social contact as well as movement. An active comparison supplies some alternative activity. It asks a more specific question: did this exercise package outperform that alternative? Neither design makes every other feature identical. [S2, S3]
Random assignment strengthens an inference about the programmes actually compared. It does not automatically randomize the features used in a later review’s subgroup analysis. For example, a comparison between studies using different intensities may also compare different populations and control activities. A subgroup ranking is not sufficient reason to choose a supposedly optimal intensity for yourself.
One original trial illustrates this distinction. Stern and colleagues randomized 132 cognitively normal adults aged 20–67 with below-median aerobic capacity to six months of aerobic exercise or stretching and toning. The author abstract reports an executive-function benefit that increased with age, rather than benefits across every cognitive measure. This was a sustained programme against an active alternative. Our use of this trial remains abstract-scoped: its full adherence, attrition and safety findings were not evaluated here. [S4]
A second trial, by Kleinloog and colleagues, used an eight-week supervised cycling period and a no-exercise period in a randomized crossover design. Seventeen sedentary men aged 60–70 completed it; two others withdrew. Response latency improved on one multitasking measure, while other executive-function measures, memory and psychomotor speed did not change. The small, screened, men-only sample limits transfer to other readers. Blood flow was the primary outcome; the study was not sized to settle the relationships between brain changes and cognitive performance. It therefore illustrates a particular task result, rather than demonstrating the mechanism or a broad improvement in daily life. Its supervised protocol is not a routine prescribed here. [S6]
What the evidence supports
A positive average with important quality limits
Singh and colleagues combined 133 systematic reviews. The authors reported 2,724 unique component randomized trials across that review collection. Those are the authors’ counts, not a new experiment or a participant pool we have independently deduplicated. Their executive-function synthesis pooled 117 meta-analyses and estimated a standardized mean difference of 0.24, with a 95% confidence interval of 0.21 to 0.27. [S2]
A standardized mean difference expresses a group contrast in standard-deviation units so that studies using different tests can be synthesized. 0.24 does not mean 24% better thinking, 24 IQ points, or a 24% chance of benefiting. It also does not tell a reader how much easier an ordinary task will become. The confidence interval concerns the pooled estimate under the analysis, not a range of likely personal gains.
There was substantial variation between results: the reported executive-function I² was 70.34%. That statistic concerns heterogeneity; it is not the percentage of participants who failed to improve. Differences in tests, participants and interventions make the average less informative about any particular programme. [S2]
The review assessed the included reviews using AMSTAR-2: 94 were rated critically low, 19 low and 20 high confidence. These ratings concern review methods and reporting. They do not mean that every underlying trial received the same rating. The high-confidence-review subgroup’s executive-function estimate was smaller: 0.16, with a 95% confidence interval of 0.10 to 0.23. That remained positive, while offering a more restrained picture than the full collection. [S2]
The authors also conducted a post hoc trim-and-fill analysis addressing possible publication bias. Its executive-function estimate was 0.20. This is a model-based sensitivity result, not a measurement of the unknowable true effect. The review’s conclusions favour cognitive benefits, while its limitations include weak review methodology, varied tests and limited evidence specifically about healthy middle-aged adults. [S2]
Why a healthy-population review reached a more cautious answer
Ciria and colleagues re-examined 24 meta-analyses and their eligible primary studies. Their overall cognitive estimate was 0.22, with a 95% confidence interval of 0.16 to 0.28. A model allowing for control activity and baseline differences gave 0.13, with a confidence interval of 0.07 to 0.20. These are overall cognition estimates, not the same executive-function-only analysis reported above. [S3]
The researchers then used a Bayesian approach that averaged across models addressing publication bias. It produced an estimate of 0.05, with a 95% credible interval of −0.09 to 0.14. Their evidence for a positive effect was inconclusive. That is not proof that the effect is exactly zero. Different correction models also made different assumptions and produced different estimates; the smallest number should not be selected simply because it makes a striking counter-headline. [S3]
An author correction published on September 2, 2026 clarifies the study accounting: 109 studies met eligibility criteria, while 100 non-redundant studies entered the quantitative analyses after insufficient statistical information and duplicate reports were handled. The authors state that the clarification does not change the reported estimates or conclusions. We have read the correction’s primary text; we have not rerun the analyses or audited the updated repository. [S5]
The two reviews therefore do not form a clean before-and-after contest. Eligibility differed, including clinical populations and mind-body activities. One pooled review-level estimates; the other reanalysed eligible primary-study effects and examined analytical choices. Their searches also ended at different times: November 2023 for the broader review and July 2022 for the narrower one. Their underlying trials can overlap. Adding their counts together would inflate the apparent amount of independent evidence. [S2, S3]
Misconceptions and uncertainty
A statistically significant result need not be a noticeable everyday change. A test can detect an average difference without establishing a useful effect on a reader’s work, study or independent living. The outcome and its practical meaning have to be checked separately.
More publications do not automatically settle a disagreement. A large synthesis may cover more kinds of participant and intervention while answering a broader question. Quality, control conditions and analysis still matter. The 2025 authors report only slight overlap within their own review collection, but that does not establish independence between every publication discussed here. [S2, S3]
A subgroup result is not a personal prescription. The 2025 review examined age, clinical status, mode, intensity and other characteristics. It did not find statistically significant executive-function differences for several scheduling variables. Failure to detect a difference is not proof that every dose is equally effective, or that the least demanding programme is optimal. [S2]
An uncertain cognitive benefit is not a verdict on every reason to move. These analyses address measured cognitive outcomes. They do not decide which activities a person enjoys, what is appropriate for their health, or whether an existing clinician-supported programme should continue. Equally, a plausible brain mechanism does not turn a sales promise into a demonstrated benefit.
A separate September 2026 review considers exercise-only and combined programmes in older adults without dementia. Its author abstract is available, but the full methods and numerical comparisons remain unreviewed here. It is limited context, not the main evidence for this explainer or a source of missing effect estimates. The earlier blocked news draft is not being presented as a completed evaluation of that paper. [S1]
Practical meaning and sources
When a programme advertises a cognitive benefit, ask four questions:
- Who took part? Match age, health status and the relevant starting characteristics before applying the result to yourself.
- What was the alternative? Distinguish no programme, usual care, another activity and exercise without an added component.
- What actually changed? Look for the named task or domain and the size and uncertainty of the difference. Everyday planning needs evidence about everyday planning.
- Does the claim identify the same programme? A broad exercise review does not demonstrate the value of an unrelated app, paid add-on or specific class.
Consider Nadia, a fictional reader who already attends an activity suited to her circumstances. A provider offers a paid brain-game addition and cites the 0.24 estimate as evidence of better everyday planning. Nadia can now identify two gaps: that number is a standardized executive-function average, and it does not isolate the provider’s addition or establish its planning claim. She asks for the exact programme study, comparator and measured outcome before deciding whether to pay. Her completed action is a clearer purchase question, not a demonstrated cognitive gain. This invented example illustrates the reading aid; it does not validate an intervention or recommend a product.
Keep established professional advice and activity restrictions in view. This article gives no exercise dose, intensity or progression schedule. It is not a diagnosis, dementia-prevention claim or substitute for assessment or treatment. If symptoms, a health condition or uncertainty make changing activity a safety question, seek appropriate qualified guidance. Persistent or worsening cognitive difficulties also deserve professional assessment rather than a self-test inferred from these findings. Do not push through concerning symptoms or override advice because a programme promises cognitive improvement.
The useful outcome is a more precise claim: a defined programme, tested against a defined alternative, in a relevant population, with a measured benefit and an uncertainty interval. Where those pieces are missing, leave the promised benefit unresolved.
Sources
- S2 — Singh and colleagues (2025). Systematic umbrella review. The archived open full-text XML’s methods, results, sensitivity analyses and discussion/limitations were read for this revision. Supplementary analyses were not independently replicated. This is the main broad-population source.
- S3 — Ciria and colleagues (2023). Healthy-population umbrella review; archived author manuscript. The relevant results, methods and discussion were read, with selected page images inspected. The preserved manuscript predates the correction below.
- S5 — Ciria and colleagues (2026). Author correction. Nature’s indexed primary text was read after its metadata screen; the direct publisher open failed. It clarifies accounting and reports that estimates and conclusions are unchanged. No repository reanalysis is claimed.
- S4 — Stern and colleagues (2019). Randomized trial; indexed author abstract. Abstract-only illustration, with its access and safety limits retained.
- S6 — Kleinloog and colleagues (2019). Randomized crossover trial. Full primary methods, cognitive results, attrition and limitations read; the restricted sample and outcome remain explicit.
- S1 — Hernández-Gamboa and colleagues (2026). Older-adult review; indexed author abstract. Limited context only. Full methods and numerical uncertainty remain unavailable to this article; no new access recovery is claimed.
Topic hubs
Tags
