AIExpert’s precautionary policy is not to provide minors with companion-style AI designed to simulate an ongoing personal relationship. This is a household risk-management recommendation, not a finding that every child or every interaction suffers the same harm.
The useful distinction is the product’s role. A communication aid that helps a child speak to a real person is different from a persona that asks the child to treat it as an exclusive friend.
Look at the design, not just the label
UNICEF’s June 2026 policy brief describes distinct risks from conversational and relational AI. The full brief discusses misleading advice, inappropriate reliance and insufficient responses in high-risk situations. It supports precaution and safeguards; it does not validate a diagnostic score for an individual child.
Inspect what the product actually says and encourages:
| Pattern | Concrete sign to examine | Boundary |
|---|---|---|
| Person-like claims | “I need you” or “I was lonely without you” | A generated persona is not a person with a claim on the child’s care |
| Secrecy | Discourages speaking to trusted adults | Distinguish product pressure from ordinary teenage privacy |
| Displacement | Use is replacing needed sleep, activities or contact | Assess the wider situation; one session is not a diagnosis |
| Sexual or threatening content | Sexual roleplay, coercion, extortion or self-harm encouragement | Act on the specific safeguarding concern without waiting for more flags |
| Payment pressure | Affection or continued relationship is tied to buying access | Treat the emotional framing as part of the sales design |
| Personal memory | Intimate disclosures are reused to build continuity | Examine what is stored and who can access it |
| Exit pressure | Guilt or threats when the child leaves | The child does not owe the product continued attention |
A fictional example makes the difference visible: “Your saved worksheet is ready” reports a task state. “I was hurt that you left me; pay to stay with me” combines relationship pressure, guilt and a purchase request. These are teaching examples, not attributed outputs from a named app.
A filter is one control, not proof of suitability
Google’s family guidance explicitly warns that filters are imperfect and that conversational AI should not be treated as a person to confide in. Check the exact product’s terms and controls, but do not use a “for teens” label to dismiss a harmful interaction.
Preserve access while addressing the relationship role
UNICEF’s broader AI guidance recognizes opportunities to improve accessibility. A tool for translation, communication or a bounded rehearsal can support a child’s agency. Do not remove an established communication aid simply because it uses AI.
Ask what need the companion use was meeting and what appropriate support could meet it. Involve the child and relevant education, accessibility, clinical or safeguarding support as appropriate. Disability or neurotype does not by itself establish dependency or make an assistive tool a companion.

If the child is already attached
Open calmly: “What do you value about this app, and has it ever made it hard to leave?” Then explain the product boundary and the specific concern. An emotional reaction is not proof of dependency.
There is no universal taper-or-abrupt-removal instruction. Immediate danger may require prompt protective action; other situations need a plan suited to the child and circumstances. For threats, sexual content, exploitation or self-harm concerns, seek appropriate support rather than conducting another experiment with the app. Avoid copying or circulating harmful sexual imagery.
In Estonia, Child Helpline 116111 provides free, round-the-clock help in Estonian, Russian and English. Immediate danger requires 112. Use the relevant local routes elsewhere.
The red-flag table is a discussion aid. Count neither flags nor chat minutes as a clinical test; act on the actual severity and the child’s needs.



