How Should Parents and Teachers Approach AI with Children Under 13?

The vendor's terms say the tool is for users 13 and older. Last week a third-grade teacher showed you the poems her class generated with it, a parent emailed to ask whether the talking toy her son got for his birthday is safe, and the board wants a one-page policy before its November meeting.

Three questions are tangled in that request: what a six-year-old may do with these systems, what the school owes families whose children meet AI at home, and what to do when a ban is not the answer you can defend.

The short answer you can give the board: tier access by age, teach children how these systems are built before you hand one over, and keep an adult in the loop wherever the child cannot judge the output. A blanket ban is the option with the least evidence behind it. No study of age-based bans exists in this knowledge base, and most of this age group's exposure happens at home, where a school rule does not reach.

The short version

1. Write the rule by age band, not for K-12 as a whole. Under 9: play, making, tangible coding, and explanations of what a machine is. Ages 9 to 12: scaffolded, mostly non-generative activity on vetted platforms, plus explicit teaching about data and about what AI cannot know.

2. Put generative chat last in the sequence, and only with a named adult role and a stated purpose.

3. Vet every tool before adoption on four questions: which child-specific hazards were tested, what it does with voice and transcript data by default, what it costs families if the free tier disappears, and whether it works in the languages your families speak.

4. Publish what the school uses and how a family can ask. Disclosure is an institutional obligation.

5. Keep detection out of the enforcement plan.

Set the rule by age band

Tiered design is not something you have to invent. Wang, Chuang and Wu (2026) evaluated an age-tiered AI literacy resource built for Taiwan's K-12 system in two editions: an elementary edition for ages 9 to 12 of scaffolded, platform-based, non-generative activities, and an advanced edition for ages 13 to 18 with ethical reasoning and supervised generative AI use.

Their evaluation is a single post-exposure survey of 831 participants, and the elementary cohort rated the materials significantly higher on performance expectancy, effort expectancy, playfulness and behavioral intention, with perceived playfulness the strongest correlate of intention in both groups. The authors are explicit about the limit: roughly thirty minutes of exposure, measuring acceptance rather than learning, ethical competence or actual use. Read it as evidence that this sequence is teachable and appealing; it says nothing about whether the sequence works.

The two editions also show what the boundary is for. The elementary band gets platform-based activities that are scaffolded and non-generative, so a child learns what the system does without a system that talks back. The advanced band attaches ethical reasoning to supervised generative use. Staff who expect engagement to follow the sophistication of the tool usually find the opposite.

What the tiering buys you is a defensible order of operations. Plenty of schools run it backwards, buying the chatbot in the fall and teaching about it in the spring, and the evidence on that order is further down this page.

Start the youngest children with play, making and machines

Play With AI is a play-centered pre-K and kindergarten curriculum built on embodied play, tangible coding as a procedural bridge, guided dialogue as reflective Scaffolding, and teacher co-design as the Sustainability anchor. Across design cycles teachers' self-rated comfort rose, and children reasoned about AI as human-designed, built sequencing and debugging through tangible coding, and reflected on fairness through guided dialogue. AI-Play takes the unplugged route, with screen-free activities parents can run at home.

For the youngest, the risk is belief. Yang, Li and Lee (2025) caution that generative social robots can produce content preoperational children accept as true, that these systems are not developmentally calibrated in the feedback they give, and that their cost and availability widen existing divides. Xu, Girouard and Shi (2026) describe the mechanism as a redistribution of agency in play: control shifts toward the child-toy interaction, constraining play when the toy organizes it and enabling it when the toy follows the child's imagination. The same toy can do either, which is why the adult's presence is the design variable you control.

Adult co-use means something specific here: sitting with the child, asking what the toy or the screen just did, and keeping the making visible. Quiet monitoring from the doorway is not the same arrangement, and for a four-year-old it misses the belief problem entirely.

Teaching how machines work belongs in this band too. Tsingidou et al. (2026) find problem-based learning, storytelling and scaffolding the most-used strategies in robot-mediated computational thinking at kindergarten, with sequencing, debugging and algorithmic design the most-assessed skills, most often through TechCheck-K. Many of the tools in that literature were developed ad hoc and never validated. Skip the teaching and the risks reappear later as over-reliance.

Move 9- to 12-year-olds onto vetted, mostly non-generative work

This is the band where the curriculum evidence is strongest. Demir and Akar (2026) ran an 18-hour, 5E-based critical media literacy program for 36 Turkish fourth-graders covering data privacy, safe communication and media Ethics, with between-group Cohen's d of 1.12 for reading, 1.18 for writing and 1.31 for total literacy. A program that fits inside a single semester and reports effects of that size is the most encouraging result on this page.

Ages 9 to 12 get scaffolded exposure on platforms the school has vetted, with the generative parts reserved for later and supervised. Once children can read, argue and check a claim, teaching them what a system is doing becomes the work, and the tool can enter as an object of study rather than a helper.

The boundary tracks capability more than birthday. A 9-year-old who reads fluently can be taught what a training corpus is and why a system that predicts the next word can produce a confident wrong answer. A 6-year-old cannot, which is why the play and making band comes first and why the kindergarten curriculum runs through tangible coding rather than screens. Reading, argument and self-monitoring are the capacities you are really slotting children by, and age is the proxy schools can administer.

Put the rule where the exposure actually is

Assume the responsibility will be unclear unless you settle it in writing. In interviews with 33 US K-12 teachers, Xiao et al. (2026) found 10 named the parent-facing companion scenario as the most concerning, 17 said a teacher should get involved and 11 said it depends. Their visibility test left home AI use outside school jurisdiction absent observable effects, and parents came out as primarily responsible while being described as unaware or overstretched. That is a jurisdictional vacancy, and a policy written as if it did not exist governs the easy half of the problem.

Fan, Li and Zhang (2026) formalize the arrangements as additive, synergistic and compensatory models, and state plainly that the synergistic version has no supporting evidence. A working family-school arrangement has to be built on purpose, which means telling parents what to do at home rather than only what not to do.

Disclosure is the school's job, not the family's detective work. Write down which tools the school uses, what data they keep, and how a family can ask. Privacy, Pedagogical Safety and AI Use and Disclosure Statements cover the substance, and the equity, ethics, privacy and safety synthesis covers how thin most of it is in practice.

Vet the tool before you adopt it

Assume the safety layer was not built for your students. A child-safety evaluation framework grounded in expert guidance and real incident data reports that most AI safety frameworks and benchmarks target adult users despite evidence of heavy youth engagement, and that a national survey cited in that work found 72% of US adolescents had used AI companions. When three Llama Guard models were tested on education-related unsafe prompts, they struggled to identify them.

Four questions will save you a procurement cycle. Which child-specific hazards were tested? What happens to recordings, transcripts and voice data by default? What will it cost families if the free tier disappears? Does it work in the languages your families speak?

An all-girls makerspace case study (Liu et al. 2026) shows why the last three are not footnotes. Parents reported that girls became scared and spoke less in mixed-gender settings, a free image generator dropped its free functions for a paid tier the nonprofit could not absorb, and English-only prompts shut out families the program was meant to serve.

Staff confidence is a privacy control as much as a training goal. A study of early-childhood settings finds perceived usefulness and ease of use central to preschool teachers' intentions to use AI, AI Self-Efficacy a meaningful predictor, AI anxiety a deterrent, and subjective norm (colleagues and leadership) shaping intention. Staff who cannot evaluate a tool will not notice what it collects.

Give the adult a defined role

Leave the role undefined and something else takes it. ParaTutor (Luo et al. 2026) separated parent and child roles in Large Language Models (LLMs)-mediated home mathematics tutoring with 23 parent-child dyads, children aged 10 to 12, across four conditions. Generic assistance tended to displace the parent's tutoring role; the role-separated interface preserved it. Parents also struggled with the content and with how to talk about it, which is a design requirement rather than a parent failing.

Companions that read with children need the same care. Liao (2026) found a fixed peer-role AI companion produced significantly longer interactions but a lower share of words and sentences from the student, and was stronger on factual recall than on emotional or future-oriented reflection. Hence the Parent Advisor role in the proposed framework, and the support vacuum the paper names for teachers and parents.

Some of this work belongs to people regardless of what the tool can do. Raave and colleagues (2026) compared a generative AI conversational agent with human educators facilitating story-based social-emotional learning activities with a static AI child: 18 simulations per facilitator, 108 observations, rated blind to facilitator type. The agent was strong on respectful tone and routine procedural scaffolding. Human educators were stronger at deeper SEL instruction, guiding reflection and building social-emotional knowledge. Early SEL should default to complementarity, with the adult holding the reflective part.

The role needs a name on a page. Decide who sits with the child during a session, who reads what the system kept, and who answers the parent who asks what happened. In a primary classroom that is usually the classroom teacher for the first two and the principal for the third. Written down, the arrangement survives a staff absence.

Keep the child's own thinking visible

The pattern to watch is output quality rising while unaided work falls. A 26,811-student secondary analysis found homework scores rose while closed-book exam scores fell, which is why the recommendation is to monitor inputs rather than only products. In a primary classroom that means watching the drafting, the explaining and the revising, and treating a sudden polished artifact as a question rather than an achievement.

The primary writing evidence supports tools used inside instruction. A systematic review of 8 studies of AI in elementary writing instruction from 2019 to 2025 finds conversational tools supporting writing practice and Multimodal AI tools extending composition beyond text, while concluding that automated assessment still requires human oversight for fairness and accuracy. A nine-week GenAI-supported opinion-writing program with 301 Grade 5 and 6 students raised ideal writing self and academic buoyancy and lifted behavioral and emotional engagement. It did not move growth mindset, cognitive or metacognitive engagement, or organization, and the authors name the risks: over-reliance, shortcut-seeking and diminished self-monitoring.

Handle the wrong answer while the child is watching

Deleting a bad output teaches a child that outputs are either correct or forbidden. Asking how they would check it teaches the habit you want. A child whose chatbot gets a factual detail wrong can be sent to two other sources and asked which one disagrees, and the habit carries to the next system that sounds confident.

When a child says the companion "understood" them, that is the conversation to have. A system predicting the next word has no model of the child, and the useful question is what it would have needed to know in order to understand. Keep it concrete and skip the lecture on neural networks.

Fluency buys belief, which is the mechanism to watch. A Grade 5 multimodal writing study found that visualizing children's stories produced sustained gains in interpretation, analysis, evaluation and explanation, with no gain in inference: once images made meaning explicit, children reported less need to infer it. A fluent AI answer does the same work, supplying the inference the child would otherwise have made for themselves.

Some of the hardest conversations have a vehicle. REMind uses robot-mediated applied drama for anti-bullying bystander intervention: children watch a bullying scenario enacted by social robots and rehearse defense strategies by puppeteering a robotic avatar. In a mixed-methods play-test with 18 children aged 9 to 10 it supported self-efficacy, perspective-taking and understanding the outcomes of defending. Robots and role-play carry a difficult subject with an adult in the room.

None of this is a tested intervention for teaching a child to check an AI answer. It is what the evidence on fluency, belief and role-play implies, and staff should treat it as practice rather than proof.

What you can promise the board

Arthars and colleagues (2026) reviewed 271 empirical papers on generative AI in PreK-12 and found no single GenAI effect. Affective gains are common and weak indicators of learning; the most consistent evidence concerns improved immediate performance and product quality; evidence on durable learning, transfer and sustained self-regulation is uneven. Outcomes depend on five entangled conditions, learner, tool, task, social arrangements, and cultural and institutional context. Their framework separates four pedagogical functions, learning from AI, with it, about it, or by shaping it, and asks whether a student surrenders, offloads or exercises agency.

For this age band that maps onto a sequence: about and by shaping first, from later, adult co-use wherever the child cannot yet evaluate the output, and no claims about durable learning from a good afternoon's product.

A board paper can promise a review date and a definition of success your school can actually check: lessons taught, teachers who can answer a parent's question, data requests answered in writing. Claiming a learning gain would overreach every study cited here.

There is no study of age-based bans in this knowledge base. Nothing here can settle whether the bans now appearing in some school systems work, and the honest version of that sentence belongs in your board paper. What the evidence does show is that the alternative to a ban is not laissez-faire. The study of GenAI availability versus integration found that making GenAI available without teaching students how to use it was associated with significantly lower performance on applied questions (ω² = 0.35), while explicitly teaching its use for tasks like summarizing returned applied performance to pre-GenAI levels and exceeded baseline on one harder quiz. This age band can be taught: the fourth-grade program above produced effects of d = 1.12 to 1.31.

What schools get wrong

"Under 13 means no AI." The published tiering puts non-generative, scaffolded activity in the 9 to 12 band and reserves generative use for 13 and up. A school that bans everything has no curriculum in this area, and its children meet the tools at home without the lesson.

"The tool is compliant, so it is safe." Most frameworks and benchmarks target adults, and three Llama Guard models struggled on education-related unsafe prompts. Compliance paperwork answers a procurement question. It does not answer the hazard question.

"A policy sentence is a policy." Nash and Burriss (2026) found 26 of 27 preservice teachers permitted some AI use on teacher-specified terms and one prohibited it entirely, with limits rarely operationalized: one policy allowed AI "to get your thinking started" and then declared "this is where the line should be drawn" without saying where. A comparative analysis of institutional policies and course syllabi found institutions broadly encouraging GenAI use in 63% of 116 policies while half of 98 course syllabi prohibited it outright. A UNESCO-framework analysis of 30 universities' policies adds that ethics and governance principles are widely adopted while inclusion, equity, internet access and environmental impact are often overlooked, with many provisions left declarative.

"Parents will cover the home half." The 33 teachers described parents as primarily responsible and also as unaware or overstretched. The synergistic family-school model has no supporting evidence yet. It has to be built.

"Six-year-olds cannot understand how a machine works." Children in the pre-K curriculum reasoned about AI as human-designed, built sequencing and debugging with tangible coding, and reflected on fairness through guided dialogue. The kindergarten robotics literature finds sequencing, debugging and algorithmic design among the most-assessed skills in that band. The capacity is there. The curriculum usually is not.

What will come back at you

"Ban it. It is the only defensible line." No study of these bans exists, so nobody can tell you they work, and the enforcement path is weak. Bassett et al. (2026) argue detection should not be used in education at all: its estimates are probabilistic and unverifiable because real-world text origin is unknown, its use violates procedural fairness since detector scores do not meet the standard integrity investigations require, and detection "does not safeguard academic integrity; it undermines it," eroding Trust. A ban also governs the least important channel, since companion apps, AI toys and family tutoring happen at home, beyond school reach.

"Everyone else's child is already using it." Partly true, and the figure usually quoted is about adolescents: 72% of US adolescents had used AI companions in the survey cited in the child-safety work. That is not your ten-year-old. The 271-paper review found no single GenAI effect, so what peers do tells you what is normal, not what is safe or useful.

"We are a school. Why are we running a parenting campaign?" Because the exposure is there. Ten of 33 teachers named the home companion scenario as the most concerning thing they faced, and the visibility test left it outside their jurisdiction. A school that publishes what it uses, what it recommends at home and how to ask questions governs more of the problem than one that writes a rule for its own network.

"Surveillance will keep them honest." Mohamed and Temimi (2026) model this as imperfect information: the student knows how the work was produced and the institution sees only the artifact. Deterrence runs through a detector's discrimination between hidden use and legitimate work rather than its raw catch rate, and when false positives rise faster than true positives, stronger monitoring makes concealment relatively more attractive. Qu and Wang (2026) found non-disclosure among 409 undergraduates was strategic adaptation to peer norms and low interpretive trust rather than moral negligence.

"Our families cannot afford what you are recommending." The makerspace case is the warning: a free tier disappeared into a paid tier the nonprofit could not absorb, and English-only prompts excluded families. Procurement that assumes a paid tier is a policy that widens the divide you were trying to close.

"Our teachers are not trained for any of this." Often true, and the fix is smaller than it looks. In the pre-K curriculum, teachers' self-rated comfort rose across design cycles, which is what co-design does that a training day does not. In early-childhood settings, self-efficacy predicts intention to use AI and anxiety deters it. Start with one lesson in one class and let the staff who run it write the guidance.

"All this caution is a moral panic about toys." There is real uncertainty here, and the 271-paper review is the honest summary of it. The case for teaching rather than banning rests on a jurisdictional gap, a failed enforcement route, and a fourth-grade program that moved literacy by more than a standard deviation in eighteen hours.

What to do this week

1. Write the age bands into one page: what under-9s do, what 9- to 12-year-olds do, and what requires a named adult and a stated purpose.

2. Take the four vetting questions to whoever signs the software contract, and ask for answers in writing.

3. Send families a one-page note listing what the school uses, what it keeps, and who to ask.

4. Pick one non-generative activity for the youngest band and one media-literacy lesson for the 9 to 12 band, and book them into the term.

5. Strike detection from the draft policy, replace it with a rule about what the school will and will not provision, and set a review date with three things to check: what teachers used, what families asked, and whether the 9 to 12 lessons ran.

Where to go next

How Is AI Impacting Students? covers what the tools do to learning, and How Should I Incorporate AI Literacy into My Course? with AI Literacy covers what to teach. Parents and Families and Early Childhood Education hold the research base for this page, and How Should AI in Education Research Incorporate Equity, Accessibility, Privacy, Ethics, and Pedagogical Safety? is where the privacy and safety argument continues. If your question is the assessment side for older students, Should We Use AI Detectors? and How Can I Reduce AI Cheating in My Course? take it further.