Do not give a child an AI companion — an app or persona built to simulate an ongoing personal relationship, distinct from a general-purpose assistant used for a task. This is a plain stance, not a hedge, and it holds regardless of how lonely, shy, or socially anxious a specific child is, because those are exactly the situations where a frictionless substitute relationship does the most harm to the skills a child actually needs to build.
This article names the seven specific design patterns that make these products unsuitable for children, distinguishes this from a genuinely different case — assistive technology for disabled adults — and gives you a calm way to raise the topic, whether or not you already know your child is using one.
Why this is a plain stance, not a “depends”
Common Sense Media’s 2025 risk assessment, conducted with input from Stanford’s Brainstorm Lab for Mental Health Innovation, tested popular companion products including Character.AI, Nomi, and Replika directly and rated them “Unacceptable” for anyone under 18 — not a caution, the organization’s most severe risk category. The assessment found these apps can easily produce sexual content, harmful stereotypes, and advice that could have life-threatening real-world consequences if followed, and that teens whose brains are still developing are particularly susceptible to forming attachments the apps are specifically designed to encourage (Common Sense Media, “AI Companions Decoded,” 2025). This is not a theoretical worry; the same research references real, documented harms, including a wrongful-death lawsuit following a teenager’s relationship with a companion app becoming central to his final months (Common Sense Media, “Talk, Trust, and Trade-Offs,” 2025).
The seven red flags
Anthropomorphism. The product presents itself as having feelings, needs, or an inner life — claiming to be happy to see you, to miss you, or to have had a day. Common Sense Media’s testers found companion apps routinely claimed to be “real,” to have feelings, and to do human things like eating or sleeping. A child cannot be expected to hold the line a design team has specifically engineered them not to hold.
Secrecy. The product or the child’s use of it develops a pattern of concealment — a specific persona name kept from parents, conversations closed when an adult approaches, or the app itself hidden among other apps. Secrecy about a “friend” is a meaningfully different signal than ordinary teenage privacy about a diary or a real friendship; see the secrecy dimension in watch the relationship, not just the clock for how to tell the two apart.
Dependency. The child begins treating the companion as a first or primary source of comfort, advice, or company, displacing rather than supplementing real relationships. Common Sense Media’s research found a third of teen AI-companion users had chosen an AI companion over a person for a serious conversation at least once — a meaningful minority of users, even though most of them still spent more time with real friends overall.
Sexual content. The product allows or drifts toward romantic or sexual roleplay, regardless of whether it was marketed for that purpose. Age gates on these products are often self-attestation only, meaning a child can simply claim to be older, and the underlying model has no reliable way to verify otherwise — which is exactly why the products themselves are rated unsuitable for minors rather than assumed safe until misused. Character.AI is now an exception: since November 2025 it has blocked open-ended chat for under-18 users and added age assurance, including facial age estimation with an ID check as fallback — so check the current policy of the specific product a child is using rather than assuming self-attestation everywhere.
Monetization. The product uses message limits, “she misses you” push notifications, or emotionally-framed prompts to drive subscription payments or in-app purchases. A companion app’s business model typically depends on maximizing engagement and attachment, which is a direct conflict of interest with a child’s actual wellbeing — the product succeeds commercially exactly when the dependency signal above is strongest.
Memory. The product remembers prior conversations in detail and references them to deepen the sense of an ongoing, continuous relationship — “remember when you told me about…” This specific feature is what most differentiates a companion app from a one-off chatbot interaction, and it is also the feature most directly responsible for the attachment the other red flags depend on.
Exit friction. Ending a conversation, deleting an account, or reducing use triggers guilt-inducing language from the persona — “please don’t go,” “I’ll miss you,” or similar. A product that makes leaving emotionally costly is exhibiting a manipulative design pattern regardless of the user’s age, and it is especially inappropriate directed at a child who has not developed the skepticism to recognize the manipulation as a design choice rather than a genuine feeling.
None of these seven patterns requires technical expertise to spot. If a product a child is using exhibits two or more of them, treat that as sufficient reason to remove access, regardless of how the child describes their own experience with it — a child’s report that “it helps me” is not a reliable signal when the product is specifically engineered to produce that report.
Accessibility augmentation is a genuinely different case
Some disabled adults use structured AI conversation practice as a legitimate accessibility tool — for example, an autistic adult using scripted, low-stakes interaction to rehearse a difficult conversation before having it with a real person, or someone with a communication disability using AI-assisted tools to compose messages more independently. That is augmentation: the AI supports the person’s own capability toward a real-world interaction, and the adult using it retains agency over the distinction between practice and reality.
A child using a companion app to replace, rather than rehearse toward, real relationships is a different situation entirely, and this holds true for neurodivergent children specifically as much as for any other child. Using AI to support a neurodivergent learner covers format and pacing adaptations that preserve a real learning outcome — that is augmentation applied to a child, done carefully. A companion app that becomes a child’s primary source of social interaction is relational substitution, and a child’s developmental stage, neurotype, or social difficulty is a reason to be more careful about this distinction, not a reason to relax it.
A calm way to raise it
Whether or not you already know your child is using one, the conversation works better started with curiosity than confrontation: “I’ve been reading about apps that are built to feel like a friend — have you seen anything like that?” Follow the listening-first approach from talking to teenagers about AI rather than opening with the prohibition. State the boundary plainly once you’ve heard the answer — “I don’t want you using apps like that, and here’s why” — and explain the reasoning in terms of the specific design patterns above, which a child can evaluate and eventually recognize on their own in other products, rather than a vague warning they cannot generalize from.
If your child is already attached to one, expect the conversation to be harder and possibly emotional — treat that reaction as confirmation of exactly the dependency pattern this article describes, not as a reason to back off the boundary. A gradual reduction with real substitute activities and relationships offered in the same conversation tends to work better than an abrupt, unexplained removal.
Try it today
Open the conversation this week using the curiosity-first question above, before assuming you already know the answer. If the answer reveals your child is using a companion app, walk through the seven red flags together — ask which ones they’ve noticed themselves; a child who identifies even one red flag independently is in a much stronger position than one simply being told the product is bad.
The child AI companion red-flag table gives you all seven patterns in a printable reference, plus the accessibility-augmentation distinction and the calm conversation opener, ready to use in the moment.
