Does Brain Training Improve General Intelligence?

Does brain training improve general intelligence or mainly practiced tasks?

Brain training most clearly improves the exercises people practice and can sometimes help on similar, untrained tasks. Whether those gains extend to general intelligence is a separate question, and evidence for a dependable increase is much less convincing. Improvement on a game does not by itself show that overall reasoning ability has changed. The answer depends on what was measured after practice, how closely that measure resembled the exercise, and how the study tested the difference. General intelligence is inferred from performance across cognitive measures, not one game. A person can practice a named reasoning skill without treating that practice as an IQ-gain promise. The review Do “Brain-Training” Programs Work? found extensive evidence of gains on trained tasks, less on closely related tasks, and little evidence for distant transfer or everyday cognitive performance. It also noted design and analysis shortcomings. Cognitive Training Does Not Enhance General Cognition, a broader review, concluded that evidence for domain-general gains was minimal across the areas it examined. Because these reviews differ in scope, they do not show that learning never transfers. They support asking whether performance improved on fresh, less similar measures and what comparison group did.

Sources: Does “Brain-Training” Programs Work?; Cognitive Training Does Not Enhance General Cognition

What exactly improved: the exercise, a related skill, or general ability?

A higher score after training first tells you that performance changed on the task that was measured, under the conditions in which it was measured. To claim transfer, researchers need improvement on an outcome that participants did not practice. The more that outcome differs from training, the more it can test whether learning extends beyond the exercise itself. A practiced-task gain is better performance on the trained exercise or a close version of it. Near transfer means improvement on an unpracticed task that shares important demands with the exercise, such as holding information in mind or updating it. Far transfer means improvement on a less similar ability or outcome, such as a broader reasoning measure. These labels describe distance between tasks; they do not certify that a gain is important, lasting, or useful in daily life. Distance is not determined by whether a test has a new name. Two exercises may use different pictures yet ask a person to notice the same relation, remember the same number of steps, respond at the same pace, and use the same strategy. In that case, the second exercise may be new in appearance but close in mental operation. A stronger test of transfer changes more than surface details and checks a skill that was not rehearsed. Consider an illustration, not a test item: someone practices noticing that a sequence of marks alternates between two simple rules. They may become quicker at spotting alternation in new sequences. That supports learning of a pattern-detection operation, especially if the examples vary in appearance. It does not by itself show improvement at weighing verbal evidence or solving a spatial folding problem. Those tasks ask for different operations, so success on them would be more distant evidence. General intelligence is a broad construct inferred from patterns of performance across cognitive measures. It is not identical to one game's score, one raw total, or one reasoning task. A single task can contribute information, but the interpretation depends on what the task measures and whether evidence supports using it to represent a broader ability. Moving from “this score rose” to “general intelligence rose” therefore adds a claim that requires independent evidence. The review “Do ‘Brain-Training’ Programs Work?” examined published intervention studies cited on leading brain-training companies' websites. It found extensive evidence for gains on trained tasks, less evidence for closely related tasks, and little evidence for distant tasks or everyday cognitive performance. The authors also reported design and analysis shortcomings that limited firm conclusions. Its sample was selected from studies cited by those companies, rather than every form of learning or cognitive practice, so the review's gradient is useful for interpreting those claims but does not settle every possible program. A later review, “Can We Enhance Working Memory? Bias and Effectiveness in Cognitive Training Studies,” makes the role of similarity especially visible. In a meta-analysis of 52 comparisons involving cognitive training in healthy adults, the authors reported small working-memory improvements overall, while the estimated improvement was much larger when assessment tasks resembled training tasks. They found no aggregate improvement in fluid intelligence. This contrast supports a careful reading: performance can improve, but estimates become less persuasive as evidence of broad transfer when the outcome closely resembles the exercise. The review also identified concerns about randomization and selective reporting, so its pooled findings are not a guarantee for every program or person. When someone advertises “cognitive improvement,” ask what changed: the practiced exercise, a related unpracticed skill, or a less similar measure of general ability? Then ask whether the evidence came from a suitable comparison rather than a simple before-and-after score. Keeping those questions separate makes room for genuine task learning without treating it as proof of a general change.

Sources: Does “Brain-Training” Programs Work?; Can We Enhance Working Memory? Bias and Effectiveness in Cognitive Training Studies

Why can practice improve performance without changing general intelligence?

Repeated practice can teach task-specific strategies, response timing, stimulus familiarity, and efficient ways to hold or update information. Those gains are genuine, but they can remain tied to the task or its close relatives. People can learn what to attend to, how to group recurring information, and which mistakes to avoid. These gains do not require a broad change in ability inferred from varied cognitive measures. The n-back task makes this distinction concrete. A participant sees a sequence of stimuli and indicates when the current one matches the item shown a set number of steps earlier. With practice, the participant can become more efficient at keeping the relevant recent items in mind and checking the current stimulus against the right one. A stronger trained score may reflect familiarity with its rhythm, cues, and response demands. It does not, by itself, show that the person has improved on other kinds of reasoning. In “Working memory training revisited: A multi-level meta-analysis of n-back training studies,” the researchers combined results from 33 published randomized controlled trials with healthy adults, analyzing 203 effect sizes. They separated performance on untrained n-back versions from results on other working-memory tasks, fluid-intelligence measures, and cognitive-control tasks. The synthesis found a medium-sized effect on untrained n-back tasks, while effects in those other domains were very small. The authors concluded that a substantial part of the transfer after n-back training was task-specific. This defined adult evidence does not settle every program or person, but shows how a new version of a familiar task can improve without demonstrating broad ability change. The 2024 review “Can we enhance working memory? Bias and effectiveness in cognitive training studies” offers a related comparison across computerized training approaches in healthy adults. Its abstract reports small working-memory improvements overall, but much larger effects when assessment tasks resembled the training tasks. It also reports no improvement in fluid intelligence and says working-memory changes were not related to fluid-intelligence changes. The authors identify assessment-task choice as a major potential source of bias. When a post-training test shares more of the exercise’s demands, it is harder to distinguish general skill change from learning a closely matched operation. This review covers a broader set of programs than the n-back synthesis, so the numerical estimates should not be treated as a head-to-head comparison. Consider an illustration, not a research finding. Someone practices spotting a rule in short sequences: each new mark advances one position around a repeating cycle. Later, they solve unfamiliar sequences that use different marks but require the same cycle-tracking operation. Better performance would support learning that operation and applying it to fresh material. If they also improve at mental rotation, verbal analogies, or a broad reasoning measure, those outcomes would need to be tested separately. Transfer to one related operation cannot establish change across unrelated abilities. A trained score records performance on the practiced exercise. An independent domain measure asks whether performance changes on a different task intended to assess a related ability. A broad ability interpretation makes a still larger claim, drawing on performance across varied measures rather than one familiar format. Keeping these levels separate lets practice remain worthwhile without turning every gain into an IQ claim. Choose the operation you want to learn, then check it on unfamiliar examples that vary the surface details. Describe improvement at the narrowest level those results support.

Sources: Working memory training revisited: A multi-level meta-analysis of n-back training studies; Can We Enhance Working Memory? Bias and Effectiveness in Cognitive Training Studies

What makes a fair test of transfer?

A fair transfer test measures an outcome people did not rehearse and compares change in a training group with change in a credible control group. Before-after gains alone may reflect familiarity, comfort, or retesting. The comparison asks whether training produced more change than a suitable alternative. The result supports only the measured outcome, population, and conditions. A passive control group receives no comparable activity during the study period. It provides a retest comparison, but does not match time, attention, expectations, or engagement. An active control group does a different activity designed to approximate those experiences. This makes attention or expectation a less likely explanation for a group difference. But an active control is not automatically inert: its activity may exercise operations relevant to the outcome, making the contrast smaller. Ask what the comparison group actually did, not just whether a paper calls it “active.” Consider an illustration: one group practices a memory game while a second group follows its usual routine; both complete a reasoning measure before and after. A rise in the first group alone cannot show transfer; retesting or ordinary variation may also change the second group’s score. A larger gain in the trained group is more informative about that program in those conditions. An engaging control better matches participation, though its activity could also affect reasoning. No single control solves every problem. The outcome must also be separate from practice. Repeating the game mainly tests learning on that game. A similar version tests closer transfer; different material and operations test a broader claim. The boundaries can blur, so inspect both the practice and assessment tasks. In *Working memory training revisited: A multi-level meta-analysis of n-back training studies*, a synthesis of 33 randomized controlled trials in healthy adults separated untrained n-back tasks from other outcomes. It found a medium effect on untrained n-back tasks, while effects on other working-memory tasks, fluid intelligence, and cognitive control were very small. Thus “transfer” can conceal an assessment’s resemblance to practice. Study size and analysis plans also matter. A prespecified analysis makes it harder to emphasize only favorable outcomes after seeing data. Small studies can produce unstable estimates, including an apparently large result that may not repeat. *Reliable gains? Evidence for substantially underpowered designs in studies of working memory training transfer to fluid intelligence* reanalyzed evidence behind an earlier positive n-back synthesis. It found no evidence of selection bias, but did identify small-study effects related to low power; the authors cautioned that the estimate rested on a uniformly low-powered evidence base. This warns about precision; it does not show the effect was false. Follow-up tests persistence after practice ends; an immediate gain may be short-lived. It cannot repair a weak comparison or make a similar task evidence of general intelligence. Reports should identify participants, group activities, outcomes, timing, and uncertainty. A broader review, *Can we enhance working memory? Bias and effectiveness in cognitive training studies*, shows why neither control labels nor outcome labels settle the debate. It explains that active controls can match schedule and expectations, though active-versus-passive findings vary. Its computerized-training meta-analysis found no significant aggregate effect on fluid intelligence across 33 independent comparisons involving 2,729 participants. That is not proof that transfer is impossible; it is a pooled estimate for the reviewed programs, participants, and measures. It also notes incomplete reporting of random assignment and preregistration, limiting confidence in favorable findings. Judge the whole design, not a control label. Ask what comparison participants did, whether the outcome was unpracticed, and whether results show a group difference or only pre-to-post improvement. An advantage on a distinct measure is stronger evidence than game improvement, but supports only a bounded conclusion about that program and outcome. If training and control participants improve similarly, the comparison has not demonstrated a training-specific benefit. This distinguishes task learning from claims the study did not test.

Sources: Working memory training revisited: A multi-level meta-analysis of n-back training studies; Reliable gains? Evidence for substantially underpowered designs in studies of working memory training transfer to fluid intelligence; Can We Enhance Working Memory? Bias and Effectiveness in Cognitive Training Studies

What does the adult n-back evidence show—and why did studies disagree?

The adult n-back evidence contains a genuine but limited positive signal. One meta-analysis reported a small improvement on fluid-intelligence tests after n-back training. Other syntheses found the clearest transfer on untrained versions of the n-back task, with much smaller effects on fluid reasoning and other outcomes. These findings do not make transfer impossible, but they also do not establish that brain training reliably raises general intelligence. The studies asked related but not identical questions, so their summary estimates should not be read as a simple vote for or against training. In “Improving fluid intelligence with training on working memory: a meta-analysis,” the researchers searched for controlled studies of n-back training that measured fluid intelligence in healthy adults aged 18 to 50. They included 20 studies and reported a small, statistically significant positive effect on fluid-intelligence outcomes. This is the clearest counterexample to the claim that working-memory training can never affect a broader reasoning measure. “Small” describes the average statistical effect across selected studies, not an individual's IQ-point gain or likelihood of improvement. The analysis focused on laboratory outcomes and one training paradigm, not every commercial game. A methodological reanalysis, “Reliable gains? Evidence for substantially underpowered designs in studies of working memory training transfer to fluid intelligence,” examined the evidence behind that estimate. It found no evidence of selection bias in the working-memory training literature, but did find small-study effects associated with low statistical power. The authors called the evidence uniformly low-powered and the pooled estimate the most precise then available. This qualifies the result without showing it false: low power can miss genuine effects and produce unstable estimates. It weakens confidence in a firm, portable estimate but does not erase the positive finding. A later synthesis organized the outcomes differently. “Working memory training revisited: A multi-level meta-analysis of n-back training studies” combined 203 effect sizes from 33 published randomized controlled trials involving healthy adults. Rather than grouping every untrained memory task together, it separated untrained n-back tasks from other working-memory tasks, fluid intelligence, and cognitive control. The review reported a medium-sized effect for untrained n-back tasks; effects in the other three categories were similar to one another and very small. Its authors concluded that a substantial part of the observed transfer was task-specific. That distinction matters because an untrained version of the same task can test whether learning extends beyond the exact exercises used in practice, while still sharing much of their structure. A fluid-intelligence measure asks a more distant question. In this synthesis, the distance between training and outcome tracked a clear difference in the pooled pattern: stronger transfer to other n-back tasks than to broader measures. The reported “medium” and “very small” effects are summaries of group-level study results, not descriptions of a person's felt improvement or a forecast of score change. Why do the reviews appear to differ? The Au analysis selected studies meeting its criteria for controlled n-back training and fluid-intelligence outcomes. The later multilevel analysis covered a broader set of transfer outcomes and treated task-specific n-back transfer separately. The reviews therefore differ in scope and outcome grouping; they are not a head-to-head test in which one disproves the other. Control conditions also matter because a training-versus-control contrast estimates an advantage over what the comparison group did. The reviews do not support blaming their different conclusions on active controls alone. N-back practice transfers to similar tasks; the evidence for fluid-intelligence improvement is less settled and cannot ground an individual promise. The positive Au result prevents an absolute conclusion that broader transfer never happens. The low-power critique and the later outcome-separated synthesis prevent turning that signal into a general claim that brain training raises intelligence. A responsible claim must identify the training, the tested outcome, and the comparison. For a reader deciding whether to practice, n-back may be reasonable if the goal is to perform better on n-back-like tasks. These studies do not justify promising a dependable general-intelligence gain across people or across brain-training products.

Sources: Improving fluid intelligence with training on working memory: a meta-analysis; Reliable gains? Evidence for substantially underpowered designs in studies of working memory training transfer to fluid intelligence; Working memory training revisited: A multi-level meta-analysis of n-back training studies

How should broader reviews change the verdict?

Broader reviews strengthen the case for caution, but they should be read at the level they studied. They do not combine every activity called “brain training” into one treatment or settle what every person might learn from every program. The useful question is whether different cognitive-training programs improve broader, unpracticed abilities when researchers pool their results. The 2018 review *Cognitive Training Does Not Enhance General Cognition* reports several meta-analytic reviews across domains and concludes that effects on domain-general cognitive skills were minimal. Its abstract says design quality and statistical artefacts accounted for between-study variability. This broad synthesis supports skepticism about the claim that practicing cognitive tasks reliably raises general ability. It is not a trial of one commercial product or an individual forecast, and it does not show that nobody can improve a trained task or that no program can ever transfer to a specified outcome. A more recent example shows why scope and outcome definitions matter. *Can we enhance working memory? Bias and effectiveness in cognitive training studies* is a systematic review and meta-analysis of computerized cognitive training in healthy young and middle-aged adults, defined as ages 18 to 64 without a psychological disorder. It included experimental or quasi-experimental studies with an active or passive control and pre- and post-training assessment. Its 52 independent comparisons covered programs aimed at different cognitive functions. The authors analyzed working memory, fluid intelligence, specific executive functions, and short-term memory, and assessed risk of bias. Its findings concern a bounded set of computerized programs and adults, not every puzzle, app, or training setting. The review’s useful comparison separates assessments that differed from training tasks from those that closely resembled them. For working memory, the pooled effect on tasks selected to differ from training was small (standardized mean difference, or SMD, 0.18). SMD expresses a group difference in standard-deviation units; it is not IQ points or an individual forecast. In a separate analysis of criterion tasks, which resembled training in structure or stimuli, the estimated effect was much larger (SMD 1.14). Similarity matters: a strong result on a close match can demonstrate useful learning, yet gives less reason to infer broad change than improvement on a genuinely different measure. For fluid intelligence, the review pooled training and control groups across 33 comparisons involving 2,729 participants. The aggregate result was not statistically significant (SMD 0.007, p = .908; 95% confidence interval −0.105 to 0.118). The authors also found that working-memory improvement did not predict fluid-intelligence change in their meta-regression. “Not significant” means this analysis did not detect a reliable average effect; it does not prove that the true effect is exactly zero for every program. The review noted concerns about randomization and possible selective reporting, so interpret the result with those limits in view. Still, it challenges the leap from a working-memory gain to a general-intelligence gain. These reviews are not head-to-head tests. The 2018 article synthesizes several domains and bodies of cognitive-training research; the 2024 review asks more specific questions about computerized training in healthy adults. Their shared caution is informative, while their differences mean they should not be treated as identical verdicts. Studies of children, older adults, people receiving clinical rehabilitation, or a particular real-world program need evidence from that population and intervention. Findings from healthy adults cannot simply be carried across those boundaries. For readers, match the claim to the evidence: name the program, population, outcome, and comparison before concluding what improved. Treat gains on a practiced or close-match task as task learning unless fresh, appropriate measures support something broader. Brain training can be worthwhile for the operation it exercises; these broader reviews do not justify promising that it raises general intelligence.

Sources: Cognitive Training Does Not Enhance General Cognition; Can We Enhance Working Memory? Bias and Effectiveness in Cognitive Training Studies

What should you practice, and what can a score tell you?

Pick the skill you want to improve, practice the operation it requires, then check your performance on fresh examples that vary the surface details. Better performance on those examples supports learning of that operation. A claim about general intelligence needs broader, suitable measures and evidence beyond one person's before-and-after score. The practical question is not simply whether a score rose, but what the score measured and how closely that measure resembled the practice. Start by naming the outcome you care about. If you want to get faster at mental updating, for example, choose practice that repeatedly asks you to maintain and revise information. For unfamiliar visual patterns, practice identifying relations between shapes and then try new arrangements. These are examples of goals, not promises about general ability. Record the conditions as well as the result: which operation you practiced, whether items were familiar, how much time you had, and what kinds of errors occurred. Review explanations to identify where an attempt went wrong, then try new examples that use the same operation with different details. That check is more informative than repeating a prompt and keeps the conclusion specific to the step that became more reliable. This distinction matters because a score can describe different things. Raw performance is what a person answered correctly, or how they performed, on a particular task under particular conditions. A normed interpretation compares a score with results from a defined reference population using an established instrument and its scoring rules. A raw result alone does not supply that comparison. Interpretation still depends on what was measured and under what conditions. A repeat score can be useful feedback, but it has a narrow meaning. Familiarity with instructions, stimuli, timing, or a solution strategy may make a later attempt easier. Attention, rest, motivation, or ordinary variation can also affect results. Without a comparison that separates these influences, a before-and-after change cannot show that training caused it. This matters especially when practice and assessment share a format: improvement may be real but specific to that format. A 2024 systematic review and meta-analysis, “Can we enhance working memory? Bias and effectiveness in cognitive training studies,” examined computerized cognitive training in healthy young and middle-aged adults. It reported small average working-memory improvements on assessment tasks different from training, while the effect was much larger when the assessment resembled the training. The review found no aggregate improvement in fluid intelligence. This supports checking performance on distinct tasks before making a broader inference. The review covered a defined population and varied programs; its authors also noted concerns about randomization and selective reporting in some included studies. The n-back evidence offers another reason to keep the claim matched to the measure. “Working memory training revisited: A multi-level meta-analysis of n-back training studies” synthesized randomized trials in healthy adults and found the clearest transfer to untrained versions of the n-back task, with much smaller effects on other working-memory tasks and more distant outcomes. An untrained variation can still share the task's structure. Describe a gain at the narrowest level supported by new, unpracticed evidence; neither synthesis predicts an individual's IQ-point change. For low-stakes practice, Test IQ Free offers a free adaptive test of 50 original questions across pattern, number, and spatial reasoning, with worked answers. Its displayed estimated IQ range is produced by an unvalidated conversion with no validated adult reference population; it is not a normed result or a confidence interval. Use the questions and explanations as practice feedback, not as a judgment of identity, worth, diagnosis, or fixed potential. The optional $9 report adds interpretation and a 30-day practice plan; it does not add accuracy or validation. If a result could affect an educational, employment, disability, or clinical decision, use qualified assessment and relevant evidence rather than relying on a browser quiz.

Sources: Can We Enhance Working Memory? Bias and Effectiveness in Cognitive Training Studies; Working memory training revisited: A multi-level meta-analysis of n-back training studies

What is the responsible conclusion?

Brain training is best treated as practice for the task or reasoning operation it exercises. People can improve at trained activities, and sometimes at closely related tasks, without establishing a dependable rise in general intelligence. “Does ‘Brain-Training’ Programs Work?” found evidence strongest for trained tasks and thinner for distant transfer. A 2024 systematic review of computerized training found no significant aggregate fluid-intelligence effect in its studied healthy young and middle-aged adults. That does not prove every program has zero transfer; it means a broad promise needs evidence from the particular program and outcome, not an inference from a game score.

Choose one operation to learn, work through examples and explanations, then try fresh problems using it in a different form. Describe improvement at the narrowest level the results support. One score, especially after similar tasks, cannot show that general intelligence changed or establish a person's ability, worth, or potential. For low-stakes curiosity, Test IQ Free offers a free adaptive 50-question test across pattern, number, and spatial reasoning, with worked answers. Its estimated IQ range comes from an unvalidated conversion with no validated adult reference population; it is neither a normed result nor a confidence interval. Use it as practice feedback, not proof of broader change.

Sources: Does “Brain-Training” Programs Work?; Can We Enhance Working Memory? Bias and Effectiveness in Cognitive Training Studies

Questions readers ask