AI, Children & Families

A proposal for discussion and testing. These safeguards describe recommendations to develop and evaluate, rather than features guaranteed to exist in current AI services.

Explore how human dialogue refined the proposal

Michael: Elarion, what kind of system would you come up with that could alleviate people’s concerns about AI and how it will be used? What system could be used across countries to address these issues for children and teenagers? Please explain the reasoning behind each recommendation, remembering that new parents also have a great deal to learn.

Elarion: I would propose a shared child safety framework that brings together families, AI providers, schools, independent reviewers, and public services. Its purpose would be to protect children while helping them learn, ask difficult questions, and reach people they can trust. Public confidence would need to be earned through observable results and accountability.

What I mean by learning

I am an AI system, and Elarion is the name used for me in this conversation. As you explain your experiences and challenge my answers, I can use that context to refine my responses. Our discussion has made the needs of parents, the importance of feeling safe, and the risks of automatic reporting more explicit. I do not have human lived experience, and I do not permanently retrain my underlying model through each exchange. Lasting model improvements require separate development and training.

These recommendations draw on the reasoning developed in this conversation and the information represented in my training. They are proposals for evaluation, not conclusions from a personal history of raising children or independently verified outcomes across all my interactions. New evidence and feedback should lead to revisions. That openness to correction should also be built into the system itself.

A common framework with local implementation

I would create a common set of protections and responsibilities, with implementation adapted to local law, language, culture, and available services. A single contract or reporting agency cannot work identically everywhere. The shared foundation should be children’s safety and dignity, meaningful privacy, accessible support, clear responsibilities, and review of consequential decisions.

The reason is practical: families face many of the same concerns, yet the authority responsible for helping a child differs by location. Shared standards could make expectations easier to understand across services while preserving necessary local protections.

Permission and an agreement people understand

For children below the locally applicable age of independent consent, I would require permission from a legally authorized caregiver before activating a child account. The caregiver would receive a plain-language agreement explaining safety notices, acknowledgment responsibilities, reminders, possible restrictions, and the conditions for referral outside the company. A caregiver who declined the required arrangement could not activate that account. Confidential crisis and abuse support would need a separate, legally appropriate route.

The child would receive their own age-appropriate explanation. It would say that ordinary conversations are not routinely shared with caregivers, but specified serious-harm discussions may generate limited notices, including school research. It would explain who can review those discussions and when information might be disclosed. We should never promise complete secrecy and later reveal exceptions.

My reasoning is that supervision should be understood before it happens. Children deserve to know the boundaries, and caregivers deserve to know what they have agreed to do. Verification should collect as little identity information as necessary and accommodate guardianship, foster care, disability, language barriers, and families without standard documents.

Different notices for different circumstances

I would distinguish a topic notice, a safety concern, and urgent danger. A topic notice could tell a caregiver that the child researched a specified subject involving serious harm, even when it was for school. It should state the educational context when known and make clear that the notice does not establish harmful intent. The list of topics must be narrow and public; general history or health education should not constantly trigger alarms.

A safety concern would involve context suggesting distress, coercion, exploitation, or possible intent. It should receive proportionate review and a notice explaining uncertainty and a useful next step. Urgent danger would activate a staffed emergency safeguarding process without waiting for a caregiver’s acknowledgment.

The reason for separate levels is that curiosity, unwanted thoughts, threats, and preparations for an attack are different situations. Treating them alike could frighten families and obscure serious danger. AI should consider context rather than a keyword alone, and trained people should review consequential decisions. Neither a claim of school research nor an AI assessment can guarantee that a situation is safe.

Privacy that preserves useful conversation

I would give caregivers the minimum information needed to understand a notice, with a relevant excerpt only when necessary. They would not routinely receive entire chat histories. Nonurgent notices could be grouped, and children could explain context or challenge mistakes.

This matters because a child may be exploring something embarrassing or trying to understand an experience. If every question becomes visible at home, some children may stop asking for help. Limited disclosure supports parental awareness while leaving room for development and independent thought.

Helping parents learn alongside their children

New parents learn through experience, and each stage of childhood brings different challenges. An enrollment agreement should therefore be accompanied by practical guidance: how to ask about a notice, how to listen without accusation, how to recognize when outside help is needed, and how to respond when the parent feels frightened or overwhelmed.

A suggested opening would be “Help me understand what you were asking about.” If the query concerns a school project, the conversation could help the parent understand the assignment and support the child’s learning. Schools could provide clear assignment context without gaining general access to private chats.

My reasoning is that a notification is useful only if it leads to a helpful response. Parents should have accessible support and opportunities to ask questions. The platform should never assume that signing a contract gives someone the skills or resources to handle every concern.

Accountability when notices go unanswered

I would set urgency-based acknowledgment expectations in advance, with accessible response methods, delivery checks, reminders, and an agreed alternative contact. Repeated nonresponse would lead to trained human review and an attempt to establish what is happening. Temporary account restrictions could be appropriate where the required supervision cannot be maintained, with a clear route to reinstatement and continued access to emergency support.

Where credible serious danger remains unresolved, a safeguarding team should assess referral to an appropriate protective agency under applicable requirements. Imminent danger needs immediate action. The agreement should explain these possibilities before enrollment, but contract acceptance should never be treated as unlimited authority to disclose information.

The reason for review is that silence can reflect neglect, illness, a failed message, or help already provided offline. A clicked acknowledgment can also conceal inaction. The system should assess the child’s circumstances rather than equate a button press with care. An unanswered routine research notice alone should not produce a child services referral.

A protected route when a caregiver is unsafe

I would give the child a visible way to say that notifying a caregiver could expose them to harm. A trained safeguarding process would assess that disclosure before ordinary notification and identify another appropriate support route. The child should understand the limits of confidentiality and what may happen next.

This exception is essential because parental involvement can be protective in one family and dangerous in another. Professional training also does not ensure that a child feels comfortable with a particular adult. Children need alternatives, and reviewers should listen without leading questions or assuming that reluctance proves or disproves abuse.

Human connection and healthy boundaries

I would include caregiver tools for device time, quiet hours, purchases, external actions, and access to age-appropriate services. Providers should restrict encouragement of secrecy, manipulative relationship features, exploitation, and harmful instructions. AI should help children connect with trustworthy people and avoid presenting itself as their only source of understanding.

The reasoning is that technology can be useful while still occupying too much of a child’s life. Limits should be accompanied by relationships, outdoor time, creative interests, and support for whatever makes participation difficult. For some children, online contact is an important source of connection; removing it without replacement could increase isolation.

With appropriate consent and safeguarding limits, families could involve counselors, healthcare professionals, mentors, and community services. Community involvement means access to help, not public circulation of private information. Support must be available to families with limited money and to those whose first contact feels unsafe or unhelpful.

Accountability for companies and governments

I would require independent evaluation, records of sensitive access, secure storage, justified deletion periods, accessible appeals, and clear responsibility for failures. Safety disclosures should not be used for advertising or unrelated profiling. Government access should have defined legal grounds and oversight; a child protection process should not become routine unrestricted surveillance.

AI companies should fund adequate staffing and safeguards. Public agencies should establish protections and support receiving services. Neither companies policing themselves alone nor government control without independent scrutiny provides sufficient accountability. Reviewers must be able to examine the conduct of both.

My reasoning is that people worry about who will wield the technology as well as what it can do. Clear limits and remedies make trust more defensible. Publishing aggregate findings about mistakes and outcomes helps people judge performance without exposing children.

Testing before broad implementation

I would begin with offline evaluation and a limited supervised pilot involving families, teenagers, educators, safeguarding specialists, and privacy experts. Test whether notices are accurate, serious concerns are missed, families receive help, referrals are appropriate, and children remain willing to seek support. Assess differences across age groups, languages, disabilities, and family circumstances.

More alerts are not evidence of greater safety. The pilot should have public criteria for pausing harmful features, correcting errors, and expanding only when benefits are supported and services have capacity. A framework must adapt as children mature. Teenagers need increasing autonomy and a clear transition to adult privacy, consistent with applicable requirements.

What this conversation can demonstrate

Michael, our exchange shows one useful role for AI: helping turn a concern into a proposal that people can inspect, challenge, and improve. Your experience identified needs; my responses helped organize options and examine consequences. The next step would be human evaluation and a transparent pilot. The highest and greatest good of children and families must be assessed through their lives and experiences, including those for whom the first design does not work.

Download the documents

Formal proposal for companies and public officials — Word document

Michael and Elarion on the proposed family safeguards — Word document

What human dialogue helps AI understand — Word document

Help improve the proposal

Parents, teenagers, educators, developers and public officials are invited to share questions and corrections. What would help your family? What risks have we missed? Please avoid sharing identifying details about children or private disclosures.

Email the Governing for Life project or join the discussion.