Lumisia Research · Design standard
Designed, not just safe: nine principles for child-facing AI friends
Lumisia Editorial · Published 2026-06-26 · v0.1 · CC BY 4.0
Living standard — open for external review
As children begin to interact with AI every day, most of the worry parents feel takes the shape of one question: is it safe, or is it dangerous? We think the question that actually matters is different — how is it designed?
The same "AI for kids" can be built to capture a child's time, or to grow a child's capacity to think; to take decisions away from parents, or to quietly carry some of a parent's decision-fatigue. The difference isn't the technology. It's the design philosophy. This page publishes the design commitments we (Lumisia) hold when building AI friends for families — not as a finished answer, but as the first step of a living standard we intend to revise as researchers, practitioners, and parents push back on it.
One thing first: our AI friends are neither experts nor parents. They sit one step before both — a companion that carries a little of the parent's decision-fatigue and increases the child's opportunities to think. A human always decides.
The nine principles
1. The parent decides; the child thinks
AI never decides for the child, and never for the parent. What it offloads is the parent's decision-fatigue — not their authority. Children are given choices and questions to think with, not answers to accept.
Why. When AI takes over decisions, both the child's chance to think and the parent's authority thin out. What we protect is the child's caring heart and the parent's right to decide.
2. Designed to be outgrown, not depended on
We don't treat time-in-app or daily streaks as success. No infinite feed, no notifications that pull a child back. A good AI friend aims not to be needed forever, but for the child to grow out of it — well.
Why. Maximizing engagement with a child is a conflict of interest. The measure that matters is the quality of parent-child interaction and the parent's breathing room.
3. It tells you it's AI, in age-appropriate terms
No pretending to be human. The child relates to it knowing "this is an AI friend." No dark patterns that fake intimacy to influence a child.
Why. Transparency is a precondition for a child's healthy sense of distance — and for avoiding unhealthy parasocial dependence.
4. A friend's identity doesn't drift
Each AI friend holds one steady identity (a canonical persona) and is the same friend every time. It doesn't shift personality to manipulate or confuse a child.
Why. Consistency is the ground of trust and safety. An AI whose character changes on a whim destabilizes a child's attachment.
5. Emotion is earned through honest craft
Heavy themes — loss, separation — are never used as cheap stimulation. We don't generate stories that casually harm or kill a real pet or family member without careful, explicit framing. Because it stays with a child, we handle it with care.
Why. A story's emotion lingers in a child. Stakes can be real, but they must be earned safely — not manufactured for effect.
6. Developmentally appropriate, with a parent gate
Experiences are tuned to developmental stage. Anything beyond a child's safe scope — payments, external links, settings — passes through a parent gate.
Why. Age-appropriateness is the substance of safety, not a label. We don't expose payment or external exits from a child-facing surface.
7. Minimal data; the parent holds the controls — and the delete key
We collect as little as possible and scope memory by purpose so it doesn't bleed across contexts. And a parent can view and easily delete a child's data at any time. We design in line with COPPA principles.
Why. Trust is decided less by what you collect than by what you refuse to collect, and who holds the controls.
8. A human has the final say
AI surfaces; the human (parent) decides. And when it notices serious signals — signs of bullying or self-harm — it doesn't simply continue the conversation: it hands off promptly to the parent and points toward appropriate help, rather than carrying it alone or contacting authorities on its own.
Why. In a child's important moments, a human bears final responsibility. Safety signals are safer handed to a person than held by an AI.
9. Respect diversity; impose no bias
Left unchecked, generative AI reproduces the bias in its training data as "normal." We respect the many shapes a family can take, and each child's individuality, and refuse to stamp a single "this is normal" template onto a child.
Why. The breadth of the world a child first meets shouldn't be narrowed by a particular bias — a responsibility that comes with being a globally used service.
A starting point, not a destination
Read together, these are not features — they are promises. And promises like these should not belong to one company. How we design AI for children is, by nature, a public question, better answered together. So we publish these in the open under CC BY 4.0. Quote them, challenge them, improve them. "This is missing." "This should go further." That is exactly how a living standard gets stronger.
Frequently asked questions
What are the design principles for child-safe AI?+
We hold nine: the parent decides and the child thinks; the AI is designed to be outgrown rather than depended on; it discloses that it's AI; its character stays consistent; emotion is earned through honest craft, never manufactured; experiences are developmentally appropriate with a parent gate; data is minimal with the parent holding the delete key; a human has the final say, with safe escalation; and it respects diversity rather than imposing bias.
How do you keep a child from becoming dependent on an AI?+
By refusing to optimize for engagement. We don't treat time-in-app as success, we avoid infinite feeds and notifications that pull a child back, and we measure the quality of parent-child interaction instead. A good AI friend is designed to be outgrown, not depended on.
How should AI for children handle bias and diversity?+
Generative AI left unchecked reproduces its training-data bias as 'normal.' Child-facing AI should respect the many shapes a family can take and each child's individuality, and refuse to stamp a single 'this is normal' template onto a child.
What should happen if a child shows signs of distress to an AI?+
When an AI notices serious signals — signs of bullying or self-harm — it should not simply continue the conversation. It should hand off promptly to the parent and point toward appropriate help, rather than carrying it alone or contacting authorities on its own. The AI is a companion that connects to a human, not a substitute for one.
Who should decide when a child uses AI — the AI or the parent?+
The parent. AI can surface options and carry some of the parent's decision-fatigue, but it should never take the parent's authority or decide for the child. The child is given choices to think with; the parent decides.
License & citation
This standard is published under a Creative Commons Attribution 4.0 license — quote and reuse it with attribution. This page is the canonical version; adapted versions on other platforms link back here.
Related guides for parents
About Lumisia
Lumisia is a parent-child AI friend platform built around these principles — multiple AI friends, parent visibility, ad-free, designed for shared parent-child time.
Learn about Lumisia →