Raising for the UpgradeWhat the evidence actually says about children in the AI age — three stages, with the numbers and the limits of each
If AI has commoditized retrieval and analysis, the obvious question is what a child should be built for instead. This page takes the framework's answer and tests it against the published developmental literature — attachment, play, nature, middle childhood, adolescence, discovery learning, psychological safety. Some of it survives well. Some of it does not, and where it does not, that is said plainly.
The framework's claim is that the world's education systems were built to transmit L1–L3 — material, energetic, and above all intellectual capacity — and that AI has now made the top of that stack cheap. If that is true, the practical question for anyone raising a child is not what to add to the curriculum. It is what the curriculum was never measuring.
What follows is not a parenting philosophy. It is an audit. Every number below carries its source, its design, and its limitation, because a claim about someone's child should be held to a higher standard than a claim about a market.
The finding that reframes the question
Start with the strongest number in the whole literature, because it is not about children at all — it is about what accumulated knowledge and practice actually buy you.
Macnamara, Hambrick and Oswald (2014, Psychological Science) meta-analysed 88 studies to measure how much of the variance in performance is explained by accumulated deliberate practice. The answer depends almost entirely on what kind of domain you are in.
| Domain | Variance explained by accumulated practice | Character of the task |
|---|---|---|
| Games | 26% | Closed, fixed rules, defined win condition |
| Music | 21% | Largely closed, defined repertoire and technique |
| Sports | 18% | Bounded, repeatable, measurable |
| Education | 4% | Partly open, contested criteria |
| Professions | under 1% | Open-ended, ill-defined, judgment-loaded |
Read the ordering rather than any single number. Practice predicts performance where the task is closed, and predicts nearly nothing where the task is open. In professions — the domain where adult life is actually lived — accumulated practice accounts for less than one percent.
Knowledge was never the differentiator in open domains. It was the floor. AI is lowering the floor, which makes the differentiator more visible — not different.
This is a more careful claim than "AI has made knowledge worthless," and it is the one the evidence carries. Deep domain knowledge remains genuinely necessary for major creative work; the ten-year rule of domain immersion before revolutionary contribution has real support. But everyone standing at a domain's frontier has that knowledge. It is the entry ticket, not the thing that separates the people holding tickets.
Stage one — before roughly 8: the secure base
Build the base from which exploration is possible
The framework's reading: before a child can be taught to construct, the child has to be secure enough to try. The evidential status of that claim is partly supported — real, replicated, and smaller than the popular literature implies.
What the attachment evidence supports
Ainsworth's original observation — that a child explores further when there is a secure base to return to — is the structural claim, and it holds. Meta-analytically, early security predicts greater social competence (d = 0.39), fewer externalizing problems (d = 0.31), fewer internalizing problems (d = 0.15) and lower negative temperament (d = 0.13) — Groh, Fearon, van IJzendoorn, Bakermans-Kranenburg and Roisman, summarising four component meta-analyses.
The Minnesota Longitudinal Study, which followed a high-risk cohort from birth into adulthood, converges: children with secure histories were rated by blind teachers as more curious and exploring, more able to recover after stress, and — the finding most relevant here — dramatically less reliant on adults, with very little overlap between the secure and anxious groups.
What happens if formal academics start early
The clearest finding is about reading. Suggate, Schaughency and Reese (2013, Early Childhood Research Quarterly) compared New Zealand children whose reading instruction began at age 5 against those who began at 7 — n = 287, with controls for vocabulary, parental income and education, home literacy environment, and school-community affluence. The early group led at first. By age 11 the advantage had disappeared, and the later-start group showed generally greater reading comprehension. The design was non-experimental, so selection into school type cannot be ruled out.
On behaviour rather than academics, the strongest evidence is a randomized trial. Schweinhart and Weikart (1997) followed 68 children randomly assigned to High/Scope, Direct Instruction, or a traditional nursery school, to age 23. By 23, 39% of the Direct Instruction group had felony arrest records against 10% of the High/Scope group and 17% of the nursery group; treatment for emotional impairment during schooling was 47% versus about 6% in each of the others.
Anyone building an argument here also has to carry the counter-evidence, which is a well-run RCT that went the other way. Tennessee's statewide pre-K program, followed through sixth grade with an analytic sample of 2,990 low-income children randomized by oversubscription lottery, produced negative effects that grew over time: sixth-grade ELA d = −0.13, math d = −0.18, science d = −0.13, with higher rates of disciplinary infractions each year (Durkin, Lipsey, Farran & Wiesen, 2022). Boston's lottery-based universal pre-K, by contrast, raised on-time college enrollment by about 8.3 percentage points while moving achievement test scores not at all (Gray-Lobe, Pathak & Walters, 2023, QJE).
Test-score fade-out is near-universal. Long-run effects can still be strongly positive or strongly negative — and test scores turn out to be a poor guide to which.
Nature, play, and family — what each is actually worth
Nature. Following 2,593 children aged 7–10 across 36 Barcelona schools for a year, each increment of surrounding greenness was associated with about a 5% improvement in working memory and a 1% reduction in inattentiveness (Dadvand et al., 2015, PNAS). The honest caveat is in the study itself: its mediation analysis attributed 20–65% of the school-greenness association to reduced traffic pollution. Some of what looks like the benefit of nature is the absence of exhaust. The two well-known attention studies in this area each have n = 17.
Independent play. Gray, Lancy and Bjorklund (2023, Journal of Pediatrics) document the decline directly: in England, children licensed to walk home alone from school fell from 86% in 1971 to 25% in 2010; in the US, 12.7% walked or biked to school in 2009 against 47.7% in 1969. They are candid that this is "a correlation over decades" that "cannot, by themselves, prove causal direction," and note that the screen-time association they compete with is "on average, negative but very small."
Family and kin. The evolutionary case for alloparenting is theoretically rich and empirically thin for this particular claim. The kin-effects literature measures child survival in high-mortality subsistence populations — maternal grandmothers improved survival in roughly 9 of 13 populations studied — and there is no bridge from that to emotional security in a modern household. The framework's emphasis on dense family and kin contact is a reading, not a finding.
Stage two — roughly 8 to 13: the concept window
Concepts at a higher level of relation — not abstraction
The framework's reading: this is when concepts should be introduced, with the family still structurally present. Evidential status: the window is real — but "higher level" has to mean something specific, or the claim breaks.
The bracket is not arbitrary. Something biological marks it.
Del Giudice (2014, Child Development Perspectives) synthesises this into a life-history reading: middle childhood is a switch point, a period of heightened sensitivity to the environment combined with the expression of new genetic influences, during which self-regulation, executive function, mentalizing — holding multiple perspectives and conflicting goals — and moral reasoning all step up together. That is as close as the literature comes to a scientific warrant for "this window matters especially."
What "higher level of understanding" can and cannot mean
Here the framework's phrasing needs discipline. If "higher level" means abstract formal reasoning, the evidence does not support it. Piaget placed formal operations at 11–12, but the modern reading is that formal-operational thought is domain-specific, experience-dependent, and never universal — adults reason formally in familiar domains and fail in unfamiliar ones. Piaget also underestimated younger children badly, having conflated competence with performance in tasks that imposed motor demands.
What the evidence does support at this age is richer relational reasoning about well-grounded, concrete material — comparison, analogy, multiple simultaneous perspectives. And it supports demanding that reasoning rather than letting it collapse into procedure.
The TIMSS video studies are the sharpest evidence for that. Comparing nationally representative samples of about 100 teachers per country, Richland, Stigler and Holyoak (2012) found something counterintuitive: the proportion of conceptual "making connections" problems in a curriculum did not track achievement — Hong Kong had the lowest share and among the highest scores. What differentiated systems was what happened to those problems in the room. In the US, conceptual problems were routinely converted mid-lesson into procedural ones. Their striking observation about why:
"In the United States, one of the reasons that problems do not succeed in engaging students in productive struggle is that the students push back."
The family finding that inverts the usual advice
The framework says family integrity should be structurally present through this window. That is right in direction and needs one sharp correction in content.
Hill and Tyson (2009, Developmental Psychology) meta-analysed 50 studies of parental involvement at middle-school age. Involvement was positively associated with achievement with one exception — direct homework help, which was negatively associated. In the NELS analyses, "assist in homework" carried β = −.17 and "check homework" β = −.06.
What worked best was what they call academic socialization: communicating expectations, conveying the value and utility of education, discussing learning strategies, connecting schoolwork to current events and to a future. Their developmental explanation is precisely the framework's logic — early adolescents are newly able to hold multiple dimensions and anticipate consequences, so direct supervision becomes less useful while orientation becomes more so.
And the honest ceiling: the best-identified causal test of deliberately increasing parental involvement — a randomized field experiment in deprived schools near Paris (Avvisati, Gurgand, Guyon & Maurin, 2014, Review of Economic Studies) — successfully raised involvement and improved behaviour, attitudes and truancy, and moved test scores not at all.
Stage three — 13 and beyond: scaffolded experimentation
Curiosity and experiment — with structure, not without it
The framework's reading: build the curiosity layer through continuous experiment. Evidential status: supported with one hard correction. Curiosity is a real lever. Unguided experimentation is not.
Curiosity does what the framework says it does, at least for the thing the learner is curious about. Gruber, Gelman and Ranganath (2014, Neuron) found high-curiosity trivia recalled at 70.6% against 54.1% for low-curiosity items immediately, and 45.9% against 28.1% after a day. Reward-circuit activity scaled with curiosity at the moment the question was posed — during anticipation, before any answer arrived.
And curiosity is manipulable: across 41 randomized trials, N = 4,496, curiosity-enhancing interventions produced g = 0.57 (0.49 after correcting for publication bias), with mystery- and game-based framings strongest (Schutte & Malouff, 2023).
Where "continuous experimentation" fails
This is the clearest, largest, best-replicated number in the entire review, and it points against the phrase.
Alfieri, Brooks, Aldrich and Tenenbaum (2011, Journal of Educational Psychology) meta-analysed 164 studies across two comparisons:
| Comparison | Comparisons | Effect | Direction |
|---|---|---|---|
| Unassisted discovery vs. explicit instruction | 580 | d = −0.38 [−0.44, −0.31] | Favours explicit instruction |
| Enhanced / scaffolded discovery vs. other instruction | 360 | d = +0.30 [0.23, 0.36] | Favours guided discovery |
Their conclusion: "unassisted discovery does not benefit learners, whereas feedback, worked examples, scaffolding, and elicited explanations do." The sign of the effect flips on a single design variable — whether the exploration is supported.
The same pattern appears independently in secondary-level inquiry science. Across 37 experimental and quasi-experimental studies, Furtak et al. (2012) found inquiry teaching effective overall at d = 0.50 — and teacher-led activities outperformed student-led ones by roughly d = 0.40. More student autonomy over the investigation was associated with worse learning.
Even cognitive load theory's case against discovery learning — Kirschner, Sweller and Clark's much-cited argument that minimal guidance "ignores the structures that constitute human cognitive architecture" — contains the boundary condition that matters here: "the advantage of guidance begins to recede only when learners have sufficiently high prior knowledge to provide internal guidance." For a sixteen-year-old already deep in a domain, guided exploration is exactly what the theory predicts should work. The argument is not never let them explore. It is do not make novices explore alone.
The emotional-security question, answered honestly
The framework's strongest intuition is that emotional security is the substrate the capacity to construct rests on. This is the place where the evidence is thinnest, and saying so is more useful than overclaiming.
In the field's largest meta-analysis of psychological safety — 136 independent samples, over 22,000 individuals (Frazier et al., 2017) — the corrected correlations are these:
| What psychological safety predicts | Corrected correlation | Reading |
|---|---|---|
| Learning behaviour | ρ = .62 | Strong |
| Information sharing | ρ = .52 | Strong |
| Engagement | ρ = .45 | Moderate |
| Task performance | ρ = .43 | Moderate |
| Creativity (individual level) | ρ = .13 | Weak — roughly 1.7% of variance |
Safety strongly drives speaking up, sharing, and learning. It is only weakly tied to originating. At the group level the creativity estimate rises to ρ = .29, but on a much smaller cumulative sample. That is the number that most threatens the framework's claim, and it comes from the best available source.
What is well evidenced runs in the negative direction, and it is robust. Moran (2016, Psychological Bulletin) meta-analysed 177 samples, N = 22,061: anxiety degrades working memory capacity at g = −0.334, at p < 10⁻²⁹. And Byron, Khazanchi and Nazarian (2010) meta-analysed 82 experimental samples on stress and creativity, finding a curvilinear relationship — high evaluative pressure reduced creative performance, low evaluative pressure raised it above a no-pressure control — while uncontrollability was linearly, monotonically corrosive.
Threat narrows what a mind will attempt. Safety does not manufacture originality — but its absence reliably suppresses the exploration that originality depends on.
That sentence is fully defensible on the evidence above. The stronger sentence — that emotional security is the substrate of construction — is the framework's interpretation, and is labelled as such here.
What this actually implies
Six things the evidence will carry, stated as plainly as it allows:
One. Starting formal academics earlier buys an advantage that does not last. The reading gap between a 5-start and a 7-start had closed by age 11, with comprehension favouring the later start.
Two. What is worth protecting before 8 is the base from which a child will try things — because threat narrows attempt, and the effects of security run through exploration rather than through instruction.
Three. Something real changes around 6–8. Adrenarche is a biological marker, not a scheduling convenience, and it opens a period of heightened environmental sensitivity.
Four. In that window, demand relational reasoning about concrete material and resist letting it collapse into procedure — which is the specific failure the TIMSS video studies caught American classrooms making, often at the students' own request.
Five. Family presence should be orientation, not supervision. Expectations, strategies, and a connection to a future are what the meta-analysis rewards; sitting on top of homework is the one form it does not.
Six. After 13, experiment — inside structure. Feedback, worked examples, escalating difficulty, elicited explanation. The framework's instinct toward experimentation is right. Removing the scaffolding is what turns d = +0.30 into d = −0.38.
None of this is a claim that schools are failing or that institutions should be abandoned. It is a claim about what the measured curriculum was never measuring, and about which capacities become load-bearing when the machine takes the layer the curriculum was built to certify. That is the argument the five-layer architecture makes structurally; this page is its evidential floor.
Questions people are actually asking
What should children learn in the AI age?
The evidence points away from accumulating retrievable knowledge as the differentiator. Across 88 studies, accumulated deliberate practice explained 26% of performance variance in games and under 1% in professions. Practice and knowledge predict performance where tasks are closed and rule-bound, and almost nothing where they are open-ended. Knowledge is the entry floor, not what separates people once they are past it.
Should my child start school early?
Starting formal reading instruction earlier produced an early advantage that did not last. Comparing children taught reading at 5 and at 7 across 287 New Zealand children, the earlier group led at first, but the difference had disappeared by age 11, and the later group showed generally greater comprehension. The study was non-experimental, so selection effects cannot be ruled out.
Is play-based learning better than academic preschool?
Better than highly scripted direct instruction, on long-run behaviour. In a randomized trial of 68 children followed to age 23, 39% of the Direct Instruction group had felony arrest records versus 10% of the High/Scope group. But once gender was entered into the analysis there was no significant difference between High/Scope and an ordinary nursery school — the contrast is against scripted instruction, not in favour of one play curriculum.
What changes in a child at age 8?
Adrenarche — the adrenal glands begin secreting androgens, mainly DHEA, from roughly age 6 to 8, years before puberty proper. It barely changes appearance but acts substantially on the brain. Developmental scientists treat it as marking entry into middle childhood, a period described as a life-history switch point characterised by heightened environmental sensitivity and newly expressed genetic influences.
Should I help my child with homework?
Across 50 studies of middle-school-age children, parental involvement was positively associated with achievement with one exception — direct homework help, which was negatively associated. What worked best was academic socialization: conveying expectations, discussing learning strategies, connecting schoolwork to a future. The authors caution the homework finding may be reverse causation, since parents help more when a child is already struggling.
Does discovery learning work?
Only when guided. Across 164 studies, unassisted discovery performed worse than explicit instruction (d = −0.38 across 580 comparisons), while scaffolded discovery outperformed other instruction (d = +0.30 across 360 comparisons). In secondary-level inquiry science, teacher-led activities outperformed student-led ones by roughly d = 0.40. The sign of the effect turns on whether exploration is supported.
Does emotional security make children more creative?
Security is more clearly tied to learning than to originating. In the field's largest meta-analysis — 136 samples, over 22,000 people — psychological safety correlated with learning behaviour at ρ = .62 but with creativity at only ρ = .13 at the individual level. The better-evidenced claim runs negative: anxiety measurably degrades working memory (g = −0.334, N = 22,061), and threat narrows what a mind will attempt.
Is the teenage brain not fully developed until 25?
No, and developmental neuroscientists have asked for the claim to be retired. Maslowsky, Buss and Wray-Lake (2024) recommend their own field avoid statements such as "the brain is not fully mature until age 25" and correct them when heard. The figure was extrapolated from longitudinal scanning that stopped at age 20. Adolescents decide comparably to adults given time, complete information, and support.
Is there a creativity crisis among children?
The claim did not survive re-analysis. Kim (2011) reported declining Torrance test scores across 272,599 people. Barbot and Said-Metwaly (2021), re-analysing with the latest norming edition and over 330,000 participants, found the original rested on problematic empirical and statistical decisions and reported no evidence of generational decline — describing the creativity crisis as a counter-productive myth.