Does working-memory or brain training transfer to intelligence and achievement?

Verdict

No-effect on far transfer and on g — not debunked, and the choice of word is doing real work.

Brain training is not a fake phenomenon. Training reliably improves the trained task, and it improves tasks that resemble the trained task, in proportion to the resemblance. That is a genuine, replicated finding. What does not exist, at any dose, in any population, from any programme, is transfer to general cognitive ability or to academic achievement. Owen et al. found improvement on every one of the trained tasks in 11,430 people and transfer to none of the untrained ones. That is the shape of the entire literature.

So the honest verdict is a bounded one: near transfer is real, far transfer is zero. The commercial claim — that training working memory raises intelligence and school achievement — is what fails, and it fails against well-powered randomised evidence rather than by argument.

What the evidence shows

Source	Design	Grade	Key effect
Sala & Gobet 2020	Multilevel meta, typically developing children, 41 studies / 393 ES / 2,375 children	B	Near transfer small–medium, scaling with training–outcome similarity. Far transfer 0.001 (SE 0.055), p = .982, τ² = 0.000 vs active controls
Roberts 2016	Population-based RCT, 452 children aged 6–7, 44 schools, 2-year follow-up	A	Visuospatial STM 0.43 → 0.49 → null at 24 months; maths −3.0 (95% CI −5.4 to −0.7) at 2 years; A$1,035/child
Sala 2019	Second-order meta, k = 233 across all training types	B	Far transfer zero, with zero true variance, once placebo and publication bias controlled
Melby-Lervåg 2016	Meta, 87 publications / 145 comparisons, treated controls	B	No far transfer; WM gain does not predict far transfer — the mechanism does not operate
Kassai 2019	Meta of children's EF training	B	Near g+ = 0.44 (k = 43, p < .001); far g+ = 0.11 (k = 17, p = .11)
Owen 2010	Online RCT, n = 11,430 (adults)	A	Improvement on every trained task; no transfer even to closely related tasks
Redick 2013	Placebo-controlled replication of Jaeggi (adults)	B	High power, no positive transfer to any ability test
Aksayli 2019	Meta of Cogmed specifically, 50 studies / 637 ES	B	Near-zero on ability and on maths/language; near transfer decays with distance
Cortese 2015	Meta of 16 RCTs, 759 children with ADHD	B	ADHD symptoms SMD 0.37 unblinded → 0.20 blinded; inattention → non-significant; academics ns
Sala & Gobet 2019	Synthesis of the authors' meta-analytic series	B	Minimal effect on domain-general skills; heterogeneity explained by design quality + artefacts
Simons 2016	PSPI review of the industry's own cited evidence	B	Zero of the industry-cited studies met all best practices
Birtwistle 2025	Meta of children's EF training, N = 57	C	g = 0.23, claims near and far transfer — the steelman
Jaeggi 2008	Original n-back → Gf claim (adults)	C	Reported dose-dependent Gf gain; failed direct replication

The child-specific evidence is the strongest and the most deflationary. Sala and Gobet's multilevel meta is restricted to typically developing children and reports the single most useful number in this file: with active controls, far transfer to cognitive ability and to mathematics and language is 0.001, with τ² = 0.000. Zero effect and zero real variation — which forecloses the field's standard escape, that it works for some programme or some subgroup.

Roberts et al. is the trial that should have settled it favourably. 44 schools, 452 randomised children, concealed allocation, intention-to-treat, 88% retention at two years, 90.3% completing at least 20 sessions, and a target population selected for low working memory — regression to the mean should have flattered the intervention. One outcome of four moved (visuospatial STM, 0.43 at 6 months, 0.49 at 12), and it was gone by 24 months. Nothing else moved. Mathematics was 3.0 points worse in the trained arm at two years, which the authors attribute to lost classroom time. At A$1,035 per child, that is the actual purchase decision.

The mechanism has been tested directly and it is absent. Melby-Lervåg et al. asked whether children who gained more working memory showed more far transfer. They did not. If working memory were the bottleneck the theory says it is, that mediation would have to appear. It does not.

Kassai 2019 closes the last escape route. Training one executive function does not move the other executive functions — a far shorter jump than "working memory training raises maths." Near g+ = 0.44, far g+ = 0.11 (ns). If transfer fails at that distance, the longer jumps were never plausible.

Cortese 2015 is where the reported benefit comes from. The same 16 RCTs give ADHD symptom SMD = 0.37 when scored by the rater closest to treatment and 0.20 when scored by a probably blinded rater; inattention drops from 0.47 to non-significant. Academic performance was never significant at all. Half the visible effect is measurement.

Hereditarian-lens assessment

Risk: low. This verdict rests on randomised trials — Owen (n = 11,430), Roberts (n = 452, concealed allocation), Redick (active placebo), and meta-analyses whose decisive moderator is active versus passive control, an experimental feature rather than a family characteristic. Genes cannot differ between randomised arms, so nothing here is explicable by selection.

The hereditarian premise is nonetheless load-bearing in a different way: this is the archive's cleanest confirmation of premise (2), that claims of durably raising g are near-certainly wrong. Brain training is the best-funded, most-tested version of that claim in modern psychology, and after two decades the corrected far-transfer estimate is zero with zero variance.

Keep the archive's standing distinction visible. "No durable gains in g" is not "nothing raises IQ scores" — schooling raises IQ test scores ~1–5 points per year of education and the effect persists, but via directly taught skills, not via g. Brain training is the mirror image: it is a pure attempt at the g route with no taught curricular content, and it produces nothing. Read together, the two findings say the same thing — what transfers is knowledge and skill, not capacity.

One further note against a tempting misreading: the design-quality gradient in Sala & Gobet 2017 reproduces identically across chess, music and working-memory training. Raw effects are small-to-moderate; they shrink as design quality rises. That is the signature of selection and expectancy, not of a trainable capacity.

Boundaries & what critics say

• The steelman is Birtwistle 2025, a 2025 meta of 57 studies reporting g = 0.23 with claimed near and far transfer including numeracy and literacy. It is graded C because it pools published articles without the control-type moderation and publication-bias correction that Sala et al. show are decisive. Its own oddest result argues against it: non-adaptive training produced larger effects than adaptive training, which no capacity-training mechanism predicts and which artefact explains easily.
• The sympathetic reading is Smid 2020 — the field's failure is theoretical, not substantive; it needs process-based accounts and attention to individual differences in trainability. Note what that defence does not claim: no durable far-transfer effect is offered. And "individual differences in responsiveness" is precisely the untested trainability assumption this archive flags elsewhere; no genetically informative design exists in this literature either.
• Two flagship trials are in adults. Owen 2010 and Redick 2013 are out of the 4–18 band, and are retained because they are the largest and the most decisive experimental tests. The verdict does not depend on them: Roberts 2016, Sala & Gobet 2020, Kassai 2019 and Cortese 2015 are all in children and all point the same way.
• Near transfer is real and should not be denied. If a child needs to be better at digit span, digit-span training works. The claim under test is that this buys something else.
• Clinical populations are a live question. Cortese's blinded estimates are small but not uniformly zero, and ADHD symptom management is a different decision from academic achievement. This topic's verdict is about learning outcomes in schools.
• Absence of far transfer is not absence of value in the activities. Chess and music appear in the same design-gradient analysis; that they do not raise IQ is not an argument against playing them.

Practical guidance

• Do not purchase brain-training software, including Cogmed, and do not accept it as an intervention for a struggling student. Simons et al. evaluated the exact studies the industry itself cites as its best evidence: none met all the best-practice criteria.
• Protect the classroom time. The most concrete finding in this file is not a null — it is −3.0 points in maths at two years. Twenty to twenty-five 45-minute sessions have to come from somewhere.
• If the target is maths, teach maths; if it is reading, teach reading. The archive's fadeout and persistence findings apply: the interventions that hold are the ones that build content, not capacity.
• Interrogate any positive claim on two axes: was the control group active, and was the outcome rated blind? Cortese shows blinding halves the effect; Sala shows active controls erase it.
• Distinguish "trained task improved" from "the child got smarter" in every report you are shown. A vendor graph of n-back scores rising over 20 sessions is not evidence of anything but n-back practice.

Open questions

• Whether cognitive training has value for specific clinical populations on non-academic outcomes (ADHD symptom burden, post-injury rehabilitation) is separable from this verdict and less settled.
• Whether individual differences in trainability exist at all in this domain is untested with adequate designs — the same gap this archive flags in talent and trainability.
• Nobody has established why near transfer decays with task distance so lawfully; the overlap gradient is well documented and unexplained, and it is the only real phenomenon the field has.