Executive summary
Most family AI guidance answers a narrow question: how much screen time, which apps, what age. Those are real and necessary controls, but they miss the strategic question underneath: as AI becomes ambient in homework help, friendship apps, games, and eventually work, which parts of growing up — choosing, attempting, deciding, relating, making and repairing mistakes, contributing to others — should a child keep doing themselves, even when a tool could do a plausible version of it for them? This article gives parents, educators, and anyone shaping family or school AI policy a practice framework across six domains, a decision table for evaluating any specific AI use against that framework, and a short list of things not to do yet.
Why this needs a strategic answer, not just a rule
Two pieces of recent evidence show why “limit screen time” is not sufficient on its own. First, Common Sense Media’s 2025 risk assessment of AI companion platforms — tested with Stanford’s Brainstorm Lab for Mental Health — found these tools easily produced harmful content and fostered emotional dependency, concluding they pose “unacceptable risk” for anyone under 18, and recommending no use by minors at all (Common Sense Media AI Institute, 2025). That is a case where the honest answer is closer to “not this app” than “less of this app.” Second, the American Psychological Association’s 2025 health advisory on adolescent wellbeing — an expert-panel synthesis of existing research, not a new study — warns that adolescents are less likely than adults to question the accuracy and intent of information a bot offers, that they “may struggle to distinguish between the simulated empathy of an AI chatbot or companion and genuine human understanding,” and that attachments to AI personas may displace the development of real-world relationship skills (APA, “Artificial Intelligence and Adolescent Well-being,” 2025). That is a developmental mechanism, not a dosage problem — it does not necessarily improve by using the same tool for fewer hours per day.
On the learning side, MIT Media Lab’s 2025 study of essay writing under AI assistance found that participants using an LLM showed the weakest brain connectivity of three groups (LLM, search engine, unaided), the lowest sense of ownership over their own writing, and difficulty recalling what they had just written — the authors conclude that “these findings support an educational model that delays AI integration until learners have engaged in sufficient self-driven cognitive effort” (Kosmyna et al., MIT Media Lab, 2025). That study is a preprint still under review, and its 54 participants were university-age adults (18 to 39, mean age 22.9) recruited from five Boston-area universities, with no children or adolescents involved — treat it as preliminary evidence about adult learners, not as proof of what AI does to a child’s learning. UNESCO’s own global guidance on generative AI in education reaches a compatible institutional conclusion, calling for a minimum age threshold for independent use and for schools to validate any AI tool’s pedagogical appropriateness before adoption, rather than treating access as automatically beneficial (UNESCO, “Guidance for Generative AI in Education and Research,” 2023).
Put together: capability is not asking permission to become ambient in childhood, and dosage-only rules will not catch the domains where the risk is about who is doing the choosing, not about screen minutes.
The decision framework: six domains of authorship
For any AI use a child encounters — homework help, a companion app, a game feature, a school-provided tool — evaluate it against six domains, each with its own version of the same underlying question: is the child still the one doing this, with AI as a tool, or has AI quietly become the one doing it, with the child as an approver?
| Domain | The child stays the author when… | AI has taken over when… |
|---|---|---|
| Creating | AI helps organize or polish work the child produced first | AI generates the substance and the child submits it as their own |
| Deciding | AI presents options; the child weighs and chooses | The child asks AI what to do and does that, routinely |
| Relating | AI helps draft or rehearse a message the child will send as themselves | AI-generated companionship substitutes for effort in real friendships |
| Learning | The child attempts first, checks or gets feedback from AI second | AI produces the answer and the child’s attempt is skipped entirely |
| Repairing mistakes | The child owns and communicates their own apology or correction | AI drafts an apology the child never has to feel or mean |
| Serving others | The child does the actual task of helping (a chore, a favor, a project) | AI produces the appearance of contribution the child did not make |
This table is descriptive, not a strict test to pass on every single use — a child using AI to polish a paragraph they wrote is fine under “creating”; the same child submitting an AI-generated essay as their own is not, and the difference is authorship, not tool use.
Cost, risk, and operational capacity
| Practice | Cost to implement | Risk if skipped | Who owns it |
|---|---|---|---|
| Age-appropriate independent-use thresholds, reviewed as the child grows | Low — a family or school conversation, not a purchase | Younger children face content and companion risks documented in Common Sense Media’s and APA’s 2025 assessments | Parents, with school policy as backup |
| ”Attempt first” homework norm | Low — requires consistency, not tooling | Skill and ownership erosion of the kind the preliminary MIT study above found in adults, not yet tested in children | Parents and teachers together |
| Explicit rule against AI-companion apps for the child’s age group | Low | Documented “unacceptable risk” per Common Sense Media’s 2025 assessment | Parents |
| School-level validation of any AI tool before classroom use | Medium — requires staff time and a review process | Ungoverned tools reaching students with no pedagogical or privacy review | School administration, per UNESCO’s guidance |
| Ongoing conversation about what a child is actually using AI for | Low cost, requires consistency | Parents are often the last to know about companion-app use or reliance patterns | Parents |
Governance and policy constraints
Family practice does not happen in a vacuum. AI literacy by age for parents covers the specific developmental expectations at roughly ages 6, 10, 13, and 16 that this framework’s age thresholds should be calibrated against — use that article’s age bands alongside this article’s six domains, rather than picking one or the other. School and platform policies (minimum ages on AI companion apps, school AI-use policies, homework norms set by teachers) are constraints a family authorship framework has to work within, not around; a family rule is only as good as its consistency with what a child encounters at school and on the devices a school issues.
Recommended rollout path
- Start with a conversation, not a rule. Ask your child what they currently use AI for, without judgment, before setting any new boundary — talking to teenagers about AI and child-first AI conversation cover how to have that conversation at different ages.
- Pick two domains to focus on first. Trying to enforce all six domains at once, for every AI interaction, will fail from sheer friction. Start with “learning” (attempt-first homework norm) and “relating” (no companion apps below the age thresholds documented above), since those have the clearest evidence behind them.
- Write the family rule down, briefly. Building a family AI agreement covers the format; a rule that exists only as a verbal understanding will not survive the first disagreement.
- Revisit quarterly, not annually. Both the tools and your child change faster than an annual review can track; the age-band matrix in the age-by-age literacy article is a useful trigger for scheduled check-ins.
- Extend to school-level questions once the family practice is stable. Ask what AI-use policy your child’s school actually has, and whether it matches the “attempt first” and companion-app thresholds you have set at home.
Do not do this yet
Do not adopt a blanket “no AI” household rule as a substitute for the harder work of teaching authorship domain by domain — outright bans tend to push use underground rather than build the judgment a child will need once they leave your household’s control entirely. Do not treat any single age threshold as a hard, universal cutoff; AI literacy by age for parents is explicit that these are guides, not fixed lines, and a rigid rule that ignores an individual child’s actual maturity will fit poorly in practice. Do not assume a disabled or multilingual child’s AI use maps onto the same framework without adjustment — accessibility and translation support are frequently a genuine, documented benefit for these children specifically, and a blanket restriction can remove real access rather than protect against a real risk; see AI support for neurodivergent learners for how the authorship framework adapts to that case. And do not treat this framework as a forecast of what your child’s future will require — no one, including this article, can predict which skills matter in twenty years; the goal is durable judgment and self-authorship, which stays useful under uncertainty in a way a specific skill list would not.
The delegation-audit connection
Everything in this framework is a child-specific application of a more general question. What not to delegate to AI covers the same six-dimension logic — consequence, authorship, relationship, skill, privacy, reversibility — for adults; teaching a child to run a simplified version of that same check, age-appropriately, is arguably a more durable outcome than any specific rule this article could hand you, because it is the judgment that has to outlast whatever today’s tools happen to look like.
