A working-memory game can make its progress feel unmistakable. The sequence grows longer, the level rises and a task that once felt exhausting becomes manageable.

It is tempting to call that increased intelligence. Yet the scientific question is more demanding: does improvement travel beyond the game to unfamiliar tests of reasoning, or has the player mainly become skilled at the thing being practiced?

After years of studies and competing meta-analyses, the most defensible answer is that working-memory training produces real learning, but the learning is usually narrow. People often improve on the trained task and sometimes on closely related tasks. Evidence that these gains reliably raise general intelligence is much weaker.

Why the intelligence claim seemed plausible

Working memory is the limited mental workspace used to hold and manipulate information for a short time. It helps when following a complicated instruction, calculating without paper or keeping several parts of a problem in mind.

Fluid intelligence refers more broadly to reasoning through unfamiliar problems without relying mainly on previously learned facts. The two abilities are correlated, so researchers had a sensible hypothesis: strengthening working memory might also improve fluid intelligence.

One popular training method is the n-back task. A participant sees or hears a stream of items and must identify when the current item matches one presented a certain number of steps earlier. Adaptive versions become harder as performance improves.

The difficulty is that a correlation between two abilities does not guarantee that training one will change the other. Both may depend partly on shared processes, while practice can also teach strategies that are highly specific to one task.

An early meta-analysis found a small benefit

A 2015 meta-analysis of 20 n-back training studies reported a small but statistically significant improvement on measures of fluid intelligence among healthy adults aged 18 to 50. The authors concluded that the evidence supported a modest transfer effect, while also noting that the size of the gain varied with study design.

This result mattered because it pooled evidence and gave researchers a testable estimate rather than a vague claim that games were “good for the brain.”

But a meta-analysis inherits the strengths and weaknesses of the studies it includes. Small samples, differences in intelligence tests and weak comparison groups can all affect the combined result. A person who knows they are receiving the interesting training may try harder at the final assessment. Repeated testing alone can also improve a score.

Those concerns made the choice of control group central to the debate.

Active control groups changed the picture

In 2016, researchers conducted a much broader meta-analytic review covering 87 publications and 145 experimental comparisons. They separated near transfer, meaning improvement on tasks similar to the training, from far transfer to abilities such as nonverbal reasoning, reading and arithmetic.

The review found reliable immediate gains on measures of verbal and visuospatial working memory. That is evidence of learning, not nothing. However, when working-memory trainees were compared with treated control groups who also completed an activity, the researchers found no convincing improvement in nonverbal ability, verbal ability, reading comprehension, word decoding or arithmetic.

An active control is important because it helps match expectations, effort, contact with researchers and familiarity with computerized tasks. A passive control group that simply waits cannot account for those influences.

The review also found no evidence that gains in working memory mediated gains on far-transfer measures. Its conclusion was not that participants failed to improve. It was that improvement did not travel as far as commercial descriptions of brain training often implied.

Two years of practice did not unlock broader transfer

Perhaps short studies simply ended too soon. A 2022 longitudinal controlled study tested that possibility with ninth-grade students. A training group of 112 students practiced a varied set of validated working-memory tasks every two weeks for two years, while 113 students in a control group completed the pretest and post-test assessments.

The trained students made substantial and reliable progress on the practiced tasks. The researchers also found improvement in a latent working-memory factor, which attempts to capture the ability shared across several measures. Yet the study found no transfer to either fluid or crystallized intelligence.

Because the study was not described as a randomized trial, it cannot eliminate every pre-existing difference between groups. Still, its length and training dose weaken the simple explanation that far transfer appears only after enough practice.

A preregistered, double-blind randomized controlled study published in 2025 reached a similar result. Longer training helped participants reach higher levels on the trained n-back task, but the researchers found no near or far transfer at the post-test or one-month follow-up. Neither initial cognitive ability nor belief that intelligence is malleable meaningfully changed the outcome.

Practice effects can resemble general improvement

Transfer exists on a gradient. Performing the identical task again is closest to training. A new task with similar rules is farther away. An unfamiliar reasoning test that shares few surface features is farther still.

A 2017 multilevel meta-analysis of 33 randomized n-back trials illustrated that gradient. It found a medium transfer effect to untrained versions of n-back, but only very small effects on other working-memory tasks, cognitive control and fluid-intelligence measures. Much of what transferred was specific to the task family.

That pattern is exactly what ordinary skill learning would predict. Practice can improve attention to relevant cues, reduce confusion about instructions and produce efficient strategies. None of those changes is fake. They simply do not establish that a general mental capacity has increased.

Intelligence tests also differ from one another. A small improvement on one reasoning measure should not automatically be described as a change in intelligence as a whole. This is one reason stronger studies use several outcomes and analyse whether the shared ability across tests has changed.

A narrow training result does not mean intelligence is fixed

The weak evidence for far transfer from working-memory games is sometimes interpreted too broadly. It does not show that intelligence-test performance can never change, or that environments and education are irrelevant.

A 2018 meta-analysis of 42 datasets involving more than 600,000 participants used several quasi-experimental approaches to estimate the effect of education. Across the designs, an additional year of education was associated with roughly one to five extra IQ points.

Education is not a longer brain game. It involves varied knowledge, reasoning, language, social demands and repeated encounters with new problems. The comparison suggests that broad, sustained experiences may influence cognitive test performance in ways a repetitive drill does not.

There are limits here too. Many studies measured performance on particular tests rather than a single underlying quantity, and estimated gains do not tell us that every student responds in the same way. As ScienceBlog has previously explored, intelligence is not well reduced to mental speed or one score.

The practical lesson is modest. Training can be worthwhile when someone wants to improve at the activity being trained. What the evidence does not justify is treating a higher game level as proof of a general upgrade in reasoning.

People learn the tasks they practice. The farther a claimed benefit moves from that task, the stronger the evidence needs to be.