Article · Critical literature review and policy design
Ability before access: who medea the AI that reaches the child
The public debate on generating AI in school has been reduced to two options — release or ban. Neither describes the variable that the strongest available evidence points to as decisive: what the adult does while the child uses.
The responsible integration of the generating AI in the founding years is not a function of access to the child to technology, but the interaction between developmental readiness, competence adult, pedagogical mediation, institutional governance and cognitive function attributed to AI within the learning activity.
Figure 1 · What changes the result is not the tool
Where do you come in?
Nope. Mixing 0–6 with 6–13 years invalidates conclusion — and it is error The most common methodology in this field.
See the two domains → "We have already trained teachers in tools."Capacity is only the first of four layers. Judgment, mediation and governance is hardly ever measured.
See the four layers → "Does AI make the child think less?"Depends on the design: in the most directly informative experiment from this matrix, the same model was 17% below control in the final test without AI — or kept the gain, according to the safeguards.
See causal evidence → "Prohibiting cell phone worked, so banning AI works."Different risk mechanisms: attention and distraction on one side, cognitive production of the other.
See why it is not proxy →In short
What is "capacity before access"
Responsible integration of the AI in the founding years is not a function of child access to technology, but the interaction between developmental readiness, adult competence, mediation quality, institutional governance and cognitive function attributed to AI in the activity. Capacity-before-access is presented here as analytical proposition derived from policy comparison — not as a principle already enshrined in the literature.
- Two domains, not one — Children's education (0–6) and initial Fundamental (~6–13) are not generalised; confusing them invalidates the conclusion.
- Four layers of competence — capacity, judgment, mediation and governance; literature almost always measures only the first, and by self-perception.
- Required, not enough — P1 is partially sustainable, P2 is not supported by any revised study, P3 (conditionality) is the strongest.
- Replacement × magnification — with unrestricted access, − 17% in the final test without AI; with pedagogic guardrails, sustained gain (Bastani et al., 2025, PNAS).
It is the same principle of fail-closed ed AI Agent Engineering, applied to the classroom: the design decides the outcome, not the tool.
Frequently Asked Questions
Frequently Asked Questions
What does "capacity mean before access"?
It is the proposition that children's access to the generative AI should be conditioned to the available mediation capacity — adult competence, quality of pedagogical intervention and institutional safeguards — and not just chronological age. The article treats it as an analytical proposition derived from the comparison between political architectures, not as a principle already established by the literature or by any current policy.
Why separate early childhood education from elementary school?
Because they are distinct domains and the evidence of one does not transfer to the other. Early childhood education (ECEC) designates 0 to 6 years, before compulsory schooling, where the direct use of the tool by the child is minimal and the focus is on adult education. The initial Fundamental (about 6 to 13 years) is where Norway published official guidelines per school stage in 2026, and where the child is already in literacy and numeration. Treating them as a single block compromises the validity of inference, because the evidence of one domain does not support conclusions about the other.
What are the four layers of adult competence?
Capability (the teacher knows how the tool works), judgment (knows how to decide when to use, when not to use and for what purpose), mediation (knows how to intervene during use — question, verify, withdraw support) and governance (the institution allows and structure this intervention, with time, infrastructure and policy). Most of the literature measures only the first, in general by self-perception, silencing precisely the layers that the strongest evidence suggests are decisive.
When does cognitive discharge become a replacement?
Unloading operational load can be desirable and release capacity for higher order reasoning; the precise question is when offloading becomes a replacement for the cognitive process that the child needs to develop. In the randomized trial by Bastani et al. (2025), unrestricted access to the model improved 48% performance during practice but produced a result 17% lower than the control in the final test without AI; with guardrails that prohibited delivering the final response, the improvement was 127% and the gain remained.
Does the restriction of smartphones serve as evidence of a generative AI?
Not directly. The risk mechanisms are distinct: smartphone and social networks operate on attention, distraction and social comparison, while generativa AI operates on cognitive production, outsourcing of reasoning and writing. Evidence of device restriction is treated as comparative evidence of policy — useful for understanding how institutional restrictions behave — rather than as causal evidence of AI.
What does the Norwegian policy of 2026 show?
Demonstrates policy design (policy design): a coherent architecture that combines minimum age, previous teaching competence and explicit pedagogical purpose, with almost total restriction between 6 and 13 years and gradual use between 14 and 16 conditioned to teacher competence. It does not demonstrate policy effectiveness (policy effectiveness): it is too recent for longitudinal impact assessment, and no policy in itself constitutes an experiment.
Is teacher training enough to integrate AI with responsibility?
Not according to the literature reviewed. No study supports the sufficiency of teaching competence as an isolated variable, and the I AI Act itself rejects this premise implicitly by requiring institutional supervision for high-risk uses, regardless of the individual competence of the educator. Adult competence is probably a necessary enabling condition; the most likely intermediate variable is the quality of pedagogical mediation.
Delimitation Summary
Two age domains that literature treats as one
"Child Education" (Early Childhood Education and Care, ECECEC) designates international literature, children from 0 to 6 years, before compulsory schooling. O more robust empirical case available to discuss generative AI — Norwegian policy 2026 — falls on the 1st to 7th year of grannskole, approximately 6 to 13 years. It's early elementary school, not early childhood education.
Treating both as equivalents compromises the validity of the inference, because the Norwegian policy is not an ECEC policy. This article separates domains from beginning — and separation changes what each evidence can sustain.
AREA A · 0–6 years
ECEC / Early Childhood Education
Upstream focus: development of digital competence and adult AI responsible. The direct use of the tool by children in this range is minimal or In addition, it is important that the European Parliament and the Council adopt a common position on the proposal from the Commission to the Council. teaching profession.
Field B · ~6–13 years
Initial primary education
Focus on the effects of restriction, conditioned introduction and teacher mediation on the use (or not used) AI generative by children in literacy and numeration process. This is where Norway published official guidance by school stage.
With the separation, Norway ceases to be read as evidence about Early Childhood Education and passes to be described by what is: a case of governance of the AI in the training phase of skills foundational, with implications indirect to Domain A.
The layers that the public debate collapses
| Layer | Operational definition |
|---|---|
| Access | The child has the technical possibility to use the AI, regardless of the use effective. |
| Usage | The child effectively interacts with the tool, with or without supervision. |
| Mediated use | The use occurs under design, supervision, questioning and progressive withdrawal of support by a competent adult. |
| Educational integration | Technology is subject to curricular and developmental objectives explicit. |
| AI literacy | ♪ Distinguished to learn ♪ with AI, learn about AI, learn through AI and use AI to perform the task — the latter the case of risk of cognitive outsourcing. |
Decomposition Original proposal
"Teaching digital competence" is too wide to be useful
The unique concept of teaching digital competence hides four distinct layers and partially independent. The distinction matters because most literature measures only the first — and generally by self-perception —, silencing exactly the layers that the strongest evidence suggests are the ones that alter the educational outcome.
PARADA 1
Capacity (capacity)
The teacher You know how does the tool work, technical and theoretically?
Typical failure: not knowing the response of a language model can be manufactured.
CHALLENGE 2
Judgment (judgment)
The teacher can decide when to use, when not to use and with which purpose?
Typical failure: master the tool and still introduce it without clear pedagogical purpose.
CHALLENGE 3
Mediation (mediation)
The teacher knows how to intervene during use — question, check, Retreat backup?
Typical failure: allow the child to copy the AI response without Checking or arguing.
CHALLENGE 4
Governance (governance)
The institution allows and structure that intervention — time, infrastructure, politics, supervision?
Typical failure: the teacher has ability and judgment, but the school offers no time, no filters, no support.
A causal chain stronger than "training → better use"
OECD literature on professional development in ECEC does not simply state that teachers need to know technology: it differentiates foundational skills, intensified and specialized, and anchors digital competence in the understanding of the child development. The resulting chain is more defensible because it makes the technological competence subordinate knowledge about the child — no A replacement for him.
Figure 2 · The proposed causal chain
Recasting the hypothesis Original proposal
Three propositions that need to be tested separately
Strong formulation — "teacher training is necessary and sufficient" — not survives a serious test because it mixes three statements of different types. Decomposed, they have distinct epistemic status:
P1 · NEED
Without sufficient adult competence in all four layers, responsible integration of the AI Generating is unlikely.
Partially sustainable
P2 · SUFFICIENCY
Sufficient adult competence alone ensures responsible integration of the AI Generative.
Not sustained
P3 · CONDITION
The educational effect depends on the interaction between adult competency, quality of mediation, maturity of the child and design of the task.
The strongest
Separation avoids displaying capacity-before-access as thesis demonstrated. She is. treated, from start to finish, as analytical proposition derived from synthesis policy comparison — not as a principle established in the literature.
Mechanism causal evidence A
The question is not if there's cognitive discharge, but when it becomes substitution
Not all cognitive offloading It's harmful. Downloading may be desirable, temporary, strategic or compensatory — an external resource that reduces operational burden and releases capacity for higher order reasoning. The scientifically accurate question It's another one: when the unloading becomes cognitive replacement?
The most robust causal evidence of the entire revised matrix responds to this in a way operational. The randomized controlled trial of Bastani and colleagues (2025, PNAS), with about 1,000 high school students in Turkey, compared two accesses to the same model. The group with unrestricted access improved 48% during practice and performed 17% lower than control in the final test without AI. The group whose AI was restricted by guardrails pedagogical — instructed never to deliver the answer final — improved 127% during practice and kept the win in the final test.
Figure 3 · Same model, two designs, opposite effects
What the percentages measure. Performance in math tasks of High school, about 1,000 students in Turkey. The ed practice (+48%, +127%) and final test (−17%) comes from different measurement moments and do not form a pair: the first is measured with the tool available, the second after withdrawal. It is this withdrawal that separates apparent gain from learning — and that is why the number of practice, alone, informs nothing about what the child learned. Before quoting any of these values in decision of policy, refer to the original article for confidence intervals and n per arm.
Strength A for the studied population and task — teenagers, mathematics. Extrapolation for young children it is hypothesis, not given: the assumption that children in development of self-regulation and inhibitory control are more vulnerable to Premature replacement is plausible, but has not been tested in the fields A or B.
Policy instrument Comparative evidence
Why cell phone is not a genetic AI proxy
There are seven restrictions in use: total prohibition; restriction of access; restriction by age; context restriction; supervised use; use conditioned to competence previous teacher; and curricularly integrated use. A recurring error in the literature of public policy is to treat evidence about restriction of smartphones as indirect proof about AI. They're different risk mechanisms.
| Technology | Predominant risk mechanism |
|---|---|
| Smartphone | Attention, distraction, interruption, media. |
| Social networks | Social comparison, dependence, exposure to content. |
| Generating AI | Cognitive production, outsourcing of reasoning and writing, task automation. |
What evidence of smartphones shows
A quasi-experimental Norwegian study with administrative data found a reduction in bullying and performance improvement among girls without detectable average aggregate effect (force B). A rapid review with meta-analysis of five studies found aggregated effect modest — d A realistic analysis suggested possible worsening for boys and increased disparity of Gender.
The Brazilian parallel
São Paulo Law 18.058/2024 and Federal Law 15.100/2025 restrict devices personal in schools, with pedagogical exceptions and accessibility. Reports institutions indicate improvement of focus and social interaction in the first semesters; less favored communities, the abrupt implementation was associated with behaviors aggressive — isolated restriction, without psychosocial monitoring, does not solve facilities cultivated outside the school.
This evidence prevents us from stating that technological restrictions are simply ineffective — but they do not allow to conclude that restricting generative AI would produce the same effects.
Case study Policy design
Norway demonstrates policy design — not policy effectiveness
| Age range / step | Guideline |
|---|---|
| 0–5 years (Domain A) | Restriction of leisure screen time (0.5–1 h/day); emphasis on adult interaction No screen intermediation. |
| 6–13 years (1st–7th year, Domain B) | As a rule, students no have access to generative AI for work School kids. |
| 14–16 years (8–10th year) | Gradual and cautious use, conditioned to which teachers have acquired sufficient competence in advance. |
| 17-19 years | Independent use allowed, with expectation of critical learning of the use of AI. |
Administrative data (Udir, 2025) indicate that nine out of ten Norwegian schools report having content filters, but the schools themselves report lack of competence institutional to support teachers in the use of AI — which supports the reading that the restriction preceded, and did not follow, the consolidation of teaching competence.
Norway demonstrates policy design: a coherent architecture of age +
teaching competence + pedagogical purpose.
Norway no demonstrates Policy effectiveness: that this
architecture produces better learning outcomes than alternatives.
No policy in itself constitutes an experiment. The policy of 2026 is too recent for longitudinal impact assessment. What exists is previous and correlated evidence — the 2024 smartphones restriction — which informed the argumentative basis to extend the restriction to AI, but this generalization is policy inference, not given directly.
There is still one detail that hasty reading loses: politics conditions use in the years 8–10 to previous teaching competence, and also maintains age restriction in the years 1–7 regardless of available competence. The actual model is additive, no replacement — minimum age + previous teaching competence + pedagogical purpose explicit. It's not right to describe it as an "age" exchange for "competence."
Comparison Summary
Three governance architectures, not two
MODEL 1
Age restriction
Chronological age as the main safeguard. Example: Recommendation UNESCO at least 13 years of age for educational use of AI generativa.
MODEL 2
Capacity-conditioned access
Age + previous teaching competence + gradual introduction. Example: a Norwegian policy of 2026 for the 8th–10th year.
MODEL 3
Unrestricted / market-oriented access
Technological availability as the main determinant of access. Example: commercial adoption on educational platforms without public policy mediation.
The productive question is not "which model is right?", but: which architecture offers best correspondence between risk, developmental maturity and mediation capacity available in each institutional context?
Test of propositions
What Evidence Sustains — And What It Does Not Sustain
P1 · Need Partially sustainable
The evidence is indirect but convergent: the absence of competent pedagogical design produces measurable damage (Bastani et al., 2025); Norwegian schools without competence institutional indicators indicate risk of acritical adoption; the literature of AI-TPACK shows that isolated technical knowledge does not replace pedagogical-ethical judgment.
Calibrated formulation: adult competence is probably a enabling condition required in many forms of responsible integration, but this need was not tested directly in the domains A and B with generative AI.
P2 · Suficiency Not sustained
No reviewed study directly tests this proposition. EU AI Act itself rejects the implicit premise when requiring institutional supervisory infrastructure — a governance layer — for high risk systems, regardless of competence individual educator.
Calibrated formulation: the current literature does not allow to sustain the sufficiency of teacher training as an isolated variable.
P3 · Conditionality The strongest
The experimental design of Bastani et al. (2025) shows, with strength A, that the presence of of guardrails Education — demonstration of design competence, similar to Judgment layer — substantially alters the effect of AI on the retention of learning in teenagers.
Calibrated formulation: teacher competence may be necessary, but it is difficult to suffice; the most likely intermediate variable is pedagogical mediation — and it is this variable, not "competence" in abstract, that agenda future should prioritize measuring directly.
Template Original proposal
A continuum of governance, not a dichotomy
Instead of the binary axis ban ↔ access, a gradual continuum crossed by an axis Cognitive function transversal — replacement on one side, enlargement on the other — present at any point in the sequence:
- Protective restriction Temporary and selective, while capacity does not exist.
- Adult capacity building Layers 1 and 2: capacity and judgment.
- Teacher-mediated access Layer 3: intervention during use, with progressive withdrawal of support.
- Appropriate use of age and development The developmental readiness as a criterion, not the availability of the tool.
- Cognitive enlargement Layer 4: institutional governance supporting the process.
The continuum as an operable criterion Original proposal
A continuum that doesn't say what to do on Monday is vocabulary, not governance. A table below is the operable form of the same sequence — and remains proposed: no line it has been empirically validated in the domains A or B.
| Step | Purpose permitted | Adult role | Minimum Safeguard | Criteria for withdrawal of support |
|---|---|---|---|---|
| 0–5 (Domain A) | None involving direct use by the child. AI is an adult tool. | Form in layers 1 and 2 before any introduction. | Human interaction without screen intermediation as standard. | It does not apply — there is no support to withdraw. |
| 6–9 (Domain B) | Learn about AI as an object, in a mediated and collective context. | It conducts activity; the child does not operate the tool alone. | No task the AI can generate in full. | The child can explain what the tool did and why it can make mistakes. |
| 10–13 (Domain B) | Learn with AI in task whose central cognitive work remains of child. | It intervenes during use: questions, requires verification, interrupts the copy. | Guardrails to prevent the delivery of the final answer. | The child demonstrates the result without the tool available. |
| 14–16 | Gradual use, conditioned to previous teaching competence. | Draw the task and audit the process, not just the product. | Pedagogic purpose declared before use, not after. | Evaluation without AI maintains the performance obtained with it. |
The last column is the one that carries the weight. It is the Bastani test transposed into the room: if the performance does not survive the removal of the tool, what happened was substitution — and none I gain measured during practice denies that.
What conditions of governance and pedagogical design move the use of AI to replace expansion, at each point of the continuum and in each age domain?
The formulation avoids presenting restriction and training as competing policies: a protective restriction can function as a governance mechanism Transitional — without this implying that they are necessarily sequential steps in every context.
Evidence audit
Evidence Matrix
A power scale is only auditable if the criterion separating one degree from the next is written. This matrix:
| Grade | Allocation criteria |
|---|---|
| A | Randomized trial with measured behavioral outcome and test without the tool available. |
| B | Quasi-experimental with comparison group, or systematic synthesis — in both cases with declared limitations of external validity. |
| C | Descriptive, small sample, no control group, or outcome measured by self-perception. |
| D | Standard, framework, official guideline or construct validation. Say what's due to do or how to measure — not what produced effect. |
| Study / document | Domain and population | Main Find | Type | Strength |
|---|---|---|---|---|
| Bastani et al. (2025), PNAS | Adolescents, high school, Turkey (outside A/B domains) | Guardrails pedagogics reverse replacement in expansion | Causal Educational | A |
| Dardanou et al. (2023), OECD WP 295 | Field A, ECECC professionals | Mediation, play and collaboration are central; continuing and situated training is more promising | Policy / training | B/C |
| Quasi-experimental study (Norway, smartphones) quote not fixed | Secondary students | Reduction of bullying; improvement among girls; no mean added effect | Policy comparison (not AI) | B |
| Quick revision PRISMA (5 studies, smartphones) quote not fixed | Students, multiple countries | Modest aggregate effect (d | Policy comparison (not AI) | B |
| Celik (2023), AI-TPACK | Teachers, international | Isolated technical knowledge is not enough; pedagogical judgment is central | Construction | B/C |
| Ning et al. (2024) | 400 teachers, China | Compound components mediate relationships of competence | Construction | B/C |
| Romero-Tena et al. (2020) | Domain A, 535 ECE students, Spain | Increased self-perception of competence (layer 1 only) | Educational | C |
| Yang et al. (2024) | Domain A, 6 children 5 years old, Hong Kong | Children Learn about AI in mediated context | Educational | C |
| UNESCO (2024), AI CFT | Teachers, global | Five areas of competence; technology does not replace the teacher | Normative | D |
| IA Act, Art. 4 and Annex III | Education systems, EU | Compulsory literacy; high-risk uses require institutional supervision | Normative / regulatory | D |
| Norwegian policy 2026 (Udir/Regjeringen) | Field B, 1st-19th year | Almost total restriction 6–13; gradual use 14–16 conditioned to teaching competence | Policy design, no effectiveness | D |
| Laws No 15.100/2025 and No 18.058/2024 | Basic Education, Brazil | Restriction of devices with pedagogical exceptions | Policy comparison (not AI) | D |
On the two lines marked "unfixed quotation". They're the only ones matrix without authorship, year and DOI solved. The finding they report is used in the text only as comparative evidence of politics — never as evidence on AI — but who There's no place to start today. They're marked until the quote complete is set: a matrix that exists to allow audit must state where She won't allow it herself.
Note on unchecked references. Items located in surveys Automated with incomplete DOI identifiers — mostly scoping and systematic reviews on AI risks in K-12 and AI literacy frameworks — could not be verified as to title, authorship, existence and reported results. They were classified as reference not verified, excluded from this matrix and the list of references, and used only qualitatively to identify gaps.
What is missing to know
Search Agenda
The gap is not treated here as a problem to be solved by optimistic inference: a literature offers sufficient frameworks to formulate the hypothesis and insufficient to establish causality. Five drawings would close that distance.
1 · four-arm CTR
Adapting Bastani et al. (2025) separately for domains A and B: without AI; AI without mediation; AI with trained teacher in the four layers; AI with qualified teacher and additional restrictions. Outcomes: reading, writing, calculation emerging, autonomy and substitution versus enlargement indicators.
2 · longitudinal study per layer
Separately measure capacity, judgment, mediation and governance institutional, avoiding the recurrent error of measuring only self-perception of the first.
3 · Study of formal mediation
Test the model directly teaching competence in AI → quality of pedagogical mediation → child outcome, with direct observation in the room — no Just questionnaires.
4 · Dose-response and thresholds
Compare 3–5, 6–8, 9–11, 12–13 and 14–16 years, testing whether vulnerability to cognitive replacement varies nonlinearly with age — hypothesis suggested, but not tested, by the very age architecture of Norwegian politics.
5 · Impact assessment of Norwegian policy
When there are longitudinal data, test if the architecture "age + teaching competence + pedagogical purpose" produces better outcomes than others two governance models.
Conclusion
Restriction and training need not be competing policies
The most defensible answer is conditional, not binary. Adult competence — decomposed in capacity, judgment, mediation and governance — is probably a qualifying condition necessary, but no revised evidence supports its isolated sufficiency. A conditionality is the most well supported proposition: the design of the task and the quality of the mediation alters the effect of AI on learning — chain demonstrated in adolescents, not in domain children A or B.
Restriction can function as a temporary governance mechanism until capacity human and institutional for responsible mediation be established. Without that implies that one should always precede the other in every context, or that the combination already has been empirically validated as superior to the alternatives.
This is not a conclusion that "we must use AI in children", nor that "we must I forbid it." It is a conclusion on conditional governance architecture — whose validity Predictive remains an open question.
Main limitation declared
Absence of longitudinal and experimental studies linking teaching competence, mediation quality and infant outcomes specifically with generating AI in both age areas defined. Every statement above is calibrated by this absence.
The same principle governs a reliable AI agent on another scale: stop at a safe point Instead of promoting a hunch. View AI Agent Engineering and, about when to stop to refine a problem, Problem Formulation.