AI in education, AI tutors, AI feedback, AI study companions, AI grading, is being adopted across schools, edtech platforms, and home learning environments faster than the field is producing the independent evaluation work needed to understand what those systems are actually doing for children.
This article walks through how the Foundation evaluates educational AI for age-appropriate content and design. The framing is: what would good practice look like for educators, parents, and developers considering or building AI for educational use with children?
What age-appropriate actually means
Age-appropriate is not the same as age-restricted. Age-restricted means content marked for users above a certain age. Age-appropriate means content, interactions, and design choices that fit the developmental capacity, emotional range, and cognitive needs of children at a specific age.
A 7-year-old is not a small adult. A 13-year-old is not a young adult. Each age range has specific developmental considerations that affect how AI should communicate, what content is appropriate, what feedback is constructive, what kinds of failure are tolerable, and what kinds of success are meaningful.
The five dimensions of age-appropriateness for educational AI
1. Content level
Information presented at a level the child can engage with, vocabulary, sentence complexity, conceptual abstraction, prior knowledge assumed. AI tutors that present information too far above or below the child's level fail in different ways, both common in current products.
Educational AI should adapt to the child's actual demonstrated level, not to the level the curriculum says they should be at.
2. Emotional intensity
The emotional intensity of the AI's interaction, warmth, encouragement, correction, disappointment-handling, calibrated to the child's age. Younger children typically need more warmth and less explicit correction. Older children can engage with more direct feedback.
Educational AI that treats all ages the same misses the developmental specificity that genuinely supportive teaching requires.
3. Feedback style
How the AI handles wrong answers, mistakes, or struggling moments. Some children respond well to immediate correction; others shut down. Some need encouragement before correction; others find encouragement without correction patronizing. Good educational AI adapts feedback style. Poor educational AI applies a default style that works for some children and undermines others.
4. Pacing and frustration tolerance
How the AI calibrates difficulty, when to make tasks harder, when to step back, when to celebrate, when to take a break. Young children especially benefit from AI that recognizes when frustration is building and responds appropriately. AI optimized for engagement metrics tends to push too hard; AI optimized for child wellbeing knows when to step back.
5. Content boundaries
What topics and material the AI engages with. Educational AI for a 7-year-old should not engage with content about violence, sexual material, or adult themes, even when generative AI's broad capability could produce such content. Educational AI for a 16-year-old can engage with more material but should still calibrate to developmental appropriateness.
Failure modes in current educational AI
● Adult-pitched content presented to children with vocabulary and complexity beyond their level
● Feedback styles that work for the average user but undermine specific children, particularly children who already struggle academically
● Engagement design that pushes children past their frustration threshold to maximize session length
● Bias in feedback, Foundation audits have documented AI giving systematically different feedback to children based on apparent socioeconomic, racial, or linguistic markers in their work
● Content boundary failures, generative educational AI surfacing inappropriate content in response to ordinary educational queries
● Over-reliance encouragement, AI tutors that solve problems for children rather than supporting children to solve problems themselves
● Privacy practices that retain detailed records of every child's learning struggles indefinitely
● Generative AI producing factually incorrect content presented to children as authoritative
● Speech and writing recognition that works less well for children with accents, dialectal English, or non-English-speaking families
What good educational AI looks like
● Content adapted to the child's demonstrated level, not to a generic curriculum tier
● Feedback style adapted to the individual child, with appropriate fallback when the right style isn't yet known
● Engagement design optimized for learning outcomes, not session length, including encouraging breaks and recognizing diminishing returns
● Bias testing across the child user base, Foundation findings suggest most products fail this category until they specifically work on it
● Content boundaries appropriate to the child's age, with conservative defaults and parental visibility into what the AI is engaging with
● Scaffolding rather than solving, AI that supports the child to do the work, not AI that does the work for the child
● Privacy practices appropriate to long-term educational records, retention with purpose, deletion that works, profiling limited to in-product use
● Fact-checking and humility, generative educational AI that acknowledges uncertainty rather than confidently producing incorrect content
● Equity in recognition and response, speech and writing recognition that works across the actual diversity of children using the product
How educators and parents can evaluate educational AI
When considering an educational AI tool for a child, at home or in a classroom, useful questions include:
● What is this product designed to do, specifically? Vague claims like 'helps children learn' don't pass scrutiny
● What evidence supports the educational claim? Internal product metrics are insufficient; look for independent evaluation
● How does the product handle different children? Testing should cover diverse age, ability, language, and background contexts
● What does the product do with data the child generates? Acceptable answers are specific and minimal; unacceptable answers are vague or comprehensive
● What is the product's track record on safety and content appropriateness? Look for actual incident history, not absence of advertised incidents
● How does the product handle struggling children? AI that punishes struggle is the wrong choice; AI that supports through struggle is the right one
● Who can see what the AI sees? Family visibility into the AI's interactions with the child should be available, not hidden behind paywalls
● What happens when the child outgrows the product? Data retention, account closure, and graceful handoff all matter
What developers can do
● Build age-appropriate design into the product from the design phase, not as a retrofit
● Test with diverse children at the age the product is designed for, not with adults imagining how children might respond
● Engage child development expertise, not as advisory window-dressing, but as a substantive part of design and evaluation
● Submit to independent third-party evaluation and publish the findings
● Be honest about what the product is and is not. Marketing claims that exceed the product's actual age-appropriateness create field-wide trust failures
● Design for the long term, children using the product today will become adolescents and adults; the relationship the product builds with the child should withstand that transition
The shift to make
Stop treating educational AI as adult AI repurposed for children.
Start treating it as a distinct product category with specific developmental requirements, with safety evaluation that goes beyond adult moderation policies, and with the discipline that children's learning experiences deserve.
Done well, educational AI can expand access, support learning, and serve children whom traditional education serves poorly. Done badly, it can entrench bias, push children past developmental thresholds, exploit attention, and create educational experiences that no thoughtful parent or teacher would design. The difference is not the AI capability; it is the design discipline applied to using that capability for children.
