Model before monitoring
children observe how adults check their cell phones, check an answer, and tolerate not knowing. The everyday example is also digital education.
Case studies · Families
Accompanying is not watching.
It is teaching to decide.
AI can spark curiosity, adapt an explanation and open up ways of access. The family task is to ensure that this help does not replace conversation, practice, privacy or the confidence to ask for human support. This case study looks at how mediation changes with age, task, and demonstrated autonomy.
Autonomy does not appear suddenly. It builds from shared use to gradual responsibility, with boundaries that change based on age, task, and observable maturity.
family compass
Preadolescence
The family teaches to delay the solution, contrast and respect the center's rules.
Adult mediation78
Autonomy45
Data protection94
Unaided test92
Reading
Guidance priorities for family design; They are not a psychological scale or a child's score.
Role of the family
Less digital police.
More everyday architecture.
Good support does not consist of reading all conversations or prohibiting by default. It consists of making purposes, limits, checks and ways to help visible before a problem appears.
Talk about the process
Asking “what were you thinking before?” or "how would you check?" It reveals more than inspecting only the final product and allows intervention without turning each use into an interrogation.
Design the environment
Decisions about predetermined schedules, accounts, privacy, spaces, and tools reduce conflict because they do not depend solely on willpower.
Withdraw help
An educational aid is safer when it can be reduced. Re-explaining, deciding or creating without AI proves that the ability still belongs to the person.
Meta-analysis · 100 studies · early childhood
The screen context modifies the association.
The synthesis by Mallawaarachchi and collaborators invites us to overcome a single question about minutes. Who accompanies, what content appears and what activity they displace also matter.
The coefficients r of the graph describe grouped associations. They do not prove that a behavior alone causes a result nor do they translate directly to any use of AI.
JAMA Pediatrics, 2024 · full study ↗NegativePositive
Viewed programs
−0,16
Television in the background
−0,10
Joint use
"Joint use" means interaction with the caregiver around the content, not mere physical presence.+0,14
Development and experience
Age guides the type of presence.
The observable capacity adjusts the pace.
These bands are not clinical borders. A child may show a lot of autonomy in a familiar topic and need intense support in another new, sensitive one or with important consequences.
family role
The family teaches to delay the solution, contrast and respect the center's rules.
Digital autonomy is growing, but the fluidity of AI can be confused with authority. It is advisable to practice the habit of formulating a hypothesis before consulting and opening the sources afterwards.
- Use for
- Graded clues, review questions, comparison of explanations and preparation of practice cases.
- Protect
- Authorship, intermediate reasoning, sustained reading, privacy and the right to ask for human help.
- Find out
- Redo the exercise with new data and indicate which statements were verified.
!
The minimum age of a platform is not a universal pedagogical recommendation.
UNESCO guidance ↗Before creating an account or allowing independent conversations, please check the current terms of service, applicable regulations and center rules. Also adjust privacy, maturity and ability to ask for help.
family agreement
A useful agreement fits in real life.
Select the clauses that you can support. The goal is not to achieve a perfect score, but to convert diffuse expectations into observable and reviewable behaviors.
Click to include or remove
67%of the agreement
seleccionado
seleccionado
Draft for 10–12 years
Our AI Use Agreement
- Before opening the tool we will say what we use it for and what part we will not delegate.
- We will not share other people's full names, center, location, images, conversations, private documents or data.
- Important statements will be verified in an appropriate source; the fluency of the answer will not count as evidence.
- If an interaction makes you uncomfortable, scary, pressured, or asks for secrecy, it can be closed and told without fear of automatic punishment.
Revision
Conversation tool, not a contract, parental control or validated evaluation.Check back in two weeks: which rule helped, which created friction, and which exception needs to be written down?
Study and homework
Helping with homework does not require becoming an examiner.
The familiar question may be simple: “what part do you need to learn and what part do you just need to produce?” If learning is the goal, the solution must come later than the attempt.
Track Tutor
Ask one question at a time, wait for the attempt, and offer the minimum help necessary. The student retains the decisions.
Practice Builder
Create examples and variations after learning the procedure. The answers are reviewed with reliable material.
Draft Critic
It is used on one's own text to locate ambiguities, assumptions or counterarguments, without replacing the author's voice.
Immediate solver
Deliver the product before the effort that was to build the capacity. It can improve the task and worsen the reading of what has been learned.
Domestic protocol
Before → Hint → I explain → I check
Four movements that fit into a short conversation.
The child says what he understands, what he has tried and where the blockage appears.
A minimal question or help is requested; the final work has not yet been delivered.
The tool is closed and the step is reconstructed with words, drawing or your own example.
A source is revised or a variant is resolved without AI.
↔
Family and center need a common language.
Open the case study Students ↗Ask what uses are permitted, how aid should be declared, and what evidence of the process is expected. Avoid punishing a rule that was never explained or doing at home what the activity was intended to teach.
Observable signs
A sign is not a diagnosis.
It is an invitation to look at the pattern.
Observe frequency, intensity, duration, context and consequences. An isolated episode calls for conversation; A pattern that grows or causes damage requires reducing risk and seeking appropriate support.
Understanding and authorship
- You cannot explain words or decisions in a work that you present as your own.
- Consult before attempting even tasks that you previously solved.
- Confuse a well-written answer with a reliable source.
Regulation and coexistence
- Use repeatedly displaces sleep, meals, movement, play, or conversations.
- They increase secrecy, conflict, or difficulty stopping an interaction.
- The tool becomes an automatic response to boredom or discomfort.
Privacy and security
- Share images, documents, location, educational center or your own or someone else's privacy.
- An account, bot or character asks for secrecy, external contact, money or sexual content.
- Use AI to harass, impersonate, manipulate images, or harm another person's reputation.
Authority and link
- Accept health, risk or relationship indications without going to a responsible person.
- Attributes consciousness, loyalty or emotional obligation to the system and avoids human support.
- An AI response decides for the family instead of providing information for conversation.
FirstDescribe without accusing
‘I’ve noticed that lately you ask before starting and then find it hard to explain. I want to understand what is helping you and what is not.’
AfterChange a condition
Reduce access, accompany, remove data, change tools or introduce a test without AI for an agreed period.
If it persists or there is damageEscalate support
Coordinate with the center or relevant professionals. In the event of threat, exploitation, sexual content, self-harm or immediate danger, prioritize safety and urgent local help.
judgement games
judgement is also trained by playing.
These practices are not intended to demonstrate that family “beats” AI. They teach how to predict, contrast, explain and recognize uncertainty without turning error into shame.
From 6 years · 10 minutes
The prediction envelope
Separate what we thought from what the tool suggests.
- Write or draw an answer before consulting.
- Ask the AI for two possible explanations.
- Compare matches and differences without erasing the prediction.
- Decide what you would change and what evidence you would need.
So that it continues to be a game
Do not use errors to ridicule or score intelligence.
Alternate topics that people other than the family know.
Celebrate a good correction as much as an initial success.
Stop if the activity produces tension or invades privacy.
Conditions and inequality
Capacity is not measured by the device available.
Time, income, language, health, accessibility, digital literacy, and school coordination change what a family can do. Recommendations must be adapted without lowering rights or expectations.
Time and care
A recommendation that requires constant monitoring may not be feasible. Prioritize few observable rules, well-configured accounts and a weekly moment of conversation.
Access and cost
A better subscription, your own device or a stable connection should not be confused with greater capacity. Every educational activity needs an equivalent alternative.
Language and culture
Quality varies between languages, varieties and local contexts. A response that ignores the family reality requires more contrast, not that the family abandons its language.
Disability and accessibility
Reading aloud, simplification of the format, dictation or translation can increase autonomy. Support must remove a barrier without lowering the intellectual objective.
Educational capital
Knowing how to ask questions and recognize a source is learned knowledge. It is advisable to teach it explicitly and not attribute errors to a lack of interest or intelligence.
Coordination with the center
Home and school rules may differ. Before sanctioning, clarify what uses are permitted, how aid is declared, and how the process will be evaluated.
Minimum viable plan
If there is only time for three things
A rule: Do not share personal data or documents of other people.
A weekly conversation: teach a use that helped and another that raised doubts.
A test: explain or redo without AI something important that was done with it.
Evidence and limits
What does this case study support?
And where the evidence ends.
Specific research on generative AI in family life is still limited. That is why we combine direct studies of learning with evidence on mediation, screens, memory and children's rights, indicating when a conclusion is a prudent inference.
01Meta-analysis · early childhood
Context matters besides time
In 100 studies, background television and program volume were negatively associated with cognitive outcomes; joint use was positively associated. They are small or moderate associations, not automatic causality.
Mallawaarachchi et al. · JAMA Pediatrics, 2024 ↗02Policy and technical reviewThink about digital ecosystems
The AAP proposes examining the child, caregivers, product design and social context, distinguishing development-centred experiences from designs intended to maximise engagement.
American Academy of Pediatrics · Pediatrics, 2026 ↗03International guide · educationHuman and age-appropriate use
UNESCO recommends data protection, pedagogical validation and age limits for independent conversations, along with critical literacy about what these systems can and cannot do.
UNESCO · Guidance for Generative AI, 2023 ↗04Children's rights guideSecurity, privacy, explanation and inclusion
The third edition formulates ten requirements for child-centered AI and incorporates AI companions, accessibility, non-discrimination and well-being.
UNICEF Innocenti · Guidance on AI and Children 3.0, 2025 ↗05Cohort · 10,048 adolescentsFamily practices are related to the pattern of use
Adult screens, in meals and bedrooms were associated with more problematic use; monitoring and limits, with less. The main cross-sectional design does not allow causality to be attributed.
Nagata et al. · Pediatric Research, 2025 ↗06Field experiment · learningSafeguards change the effect of AI
In secondary mathematics, unguarded access improved practice but continued to perform worse when AI was removed; a tutor with clues mitigated the problem.
Bastani et al. · PNAS, 2025 ↗07Experiments · child memoryRetrieval strengthens learning
Children around ten years old remembered more after practicing retrieval than after studying again; the benefit appeared with different levels of reading comprehension and speed.
Karpicke et al. · Frontiers in Psychology, 2016 ↗08Health guide · under 5Development requires movement, sleep and play
The WHO integrates sedentary exposure to screens into a full day of activity, rest and play; the guide does not specifically study generative AI.
OMS / WHO · 2019 ↗!
We are not claiming that AI harms the brain or that one rule fits all families.
An association, a change in behavior, or a worse outcome without help does not equate to general impairment. The design of the study, the task, the age, the starting point and the alternative with which it is compared matter.
Common theoretical foundation
Do you need the theoretical explanation?
Cognitive development, intelligence, individual differences, screens, and cognitive offloading are developed in detail in the Theoretical Foundations.