A rules-only opening — “here’s what you should and shouldn’t do” — can miss how a teenager actually uses AI and why. A listen-first conversation begins with genuine questions while keeping adults responsible for safety and product boundaries. It may surface useful context, but it cannot guarantee disclosure or establish that an interaction is safe.
This is not a permissiveness argument. It ends with real boundaries. It just gets there by listening first.
Why most of these conversations fail
Common Sense Media’s 2025 research on teen AI-companion use found that nearly three in four teens had used an AI companion at least once, and just over half use one regularly (Common Sense Media, “Talk, Trust, and Trade-Offs,” 2025) — though the report itself cautions that some respondents may have conflated general AI use with AI companion interactions, potentially inflating those usage figures. The same research found nuance worth taking seriously in both directions: about a third of AI-companion users have discussed something important or serious with an AI companion instead of a real person, but 80% of users still spend more time with real friends than with AI companions — most teens are not replacing their social world, even where they are supplementing it in ways worth understanding.
A conversation that opens only with “I read that AI companions are dangerous; are you using one?” leaves little room for the nuance the research describes. Prefer specific, open questions, while recognizing that safeguarding concerns may require more than conversation alone.
Start with listening, not inspection
Four questions, asked with real curiosity and enough silence afterward for an actual answer:
“Do you use AI day to day? If so, what for?” Do not assume use or non-use from a survey or peer group. Ask about actual products and contexts: schoolwork, search, creativity, companionship, or something else.
“What do your friends use it for that you don’t?” This sidesteps a defensive answer about their own use and often produces more honest information, because describing a friend’s behavior feels lower-stakes than confessing your own.
“Has anything about using it ever felt weird or off to you?” This invites the teenager to name their own discomfort rather than you naming it for them. A teenager who has independently noticed something concerning is in a much stronger position to build a boundary from than a teenager being told what should concern them.
“What do you wish adults understood better about this?” This one is not rhetorical. Teenagers frequently have accurate, specific observations about these tools — including their limitations and manipulative design patterns — that adults have not caught up to. Treat the answer as real information, not as a segue back to your own points.
Share what you genuinely don’t know
The credibility of this conversation depends on you naming real uncertainty, not performing it. There are genuine open questions here: nobody, including AI companies themselves, has a settled answer on the long-term effects of adolescent relationships with AI companions, and platform safety measures are new and unevenly enforced. Saying “I don’t actually know what the long-term effect of this is, and neither does anyone else yet, which is part of why I want to talk about it now rather than wait” is both true and disarming — it reframes the conversation as two people reasoning about a genuinely unresolved question, rather than one person testing the other.
Where the evidence is not uncertain, say that plainly too. Common Sense Media’s risk assessment of companion apps like Character.AI, Nomi, and Replika found they can easily produce sexual content, harmful stereotypes, and dangerous advice, and concluded no one under 18 should use them — that is a real recommendation from people who tested the products directly, not a vague parental worry (Common Sense Media, “AI Companions Decoded,” 2025). Do not give a child an AI friend covers the specific red flags behind that recommendation in more depth.
Set boundaries after listening, not instead of listening
Once you’ve actually heard what’s happening, name one or two concrete boundaries rather than a long list. Concrete beats comprehensive: “no AI companion apps, and if a general-purpose AI tool starts feeling like the friend you go to first with hard feelings, that’s a signal we talk again” is more usable than a ten-point household AI policy a teenager will not remember by Thursday. The family AI agreement is the right place to formalize more than one or two boundaries at a time — use it after this conversation, not instead of having it.
Explain the reasoning behind each boundary in terms the teenager can evaluate, not just enforce. For example: “A personalised, agreeable chatbot that is always available can make it easy to keep seeking reassurance or to replace a harder human conversation. That is why we are setting a boundary around companion-style use.” This reflects risks identified by the APA without claiming every product has the same business model or effect.
Do not justify the conversation with surveillance
If the conversation is prompted by something you saw in a chat history, a browser, or a monitoring app, be honest about that rather than pretending the topic came up organically — a teenager who later discovers the real reason will trust every future version of this conversation less. But surveillance is not itself a justification for restricting AI use, and covert monitoring should not be the default posture here. A teenager old enough to have independent AI access is old enough to be told directly what you’re watching for and why, the same way you’d explain any other household rule tied to trust rather than to catching them.
Do not present covert monitoring as a universal solution. Monitoring decisions depend on age, law, disability and communication needs, known risks, and the least-intrusive safeguarding response. If there is a credible risk of exploitation, sexual content, extortion, self-harm, violence, or adult contact, seek qualified safeguarding advice and take proportionate protective action rather than relying on a conversation alone.
End with a real revisit date
Close the conversation with a specific date, not “we’ll talk about this again sometime” — put it on a calendar, a month or two out, framed as a check-in rather than a follow-up interrogation: “let’s talk about this again in six weeks and see what’s changed.” A set date does two things: it lowers the stakes of this conversation (nothing has to be fully resolved today), and it makes the next conversation easier to start, because it was already agreed to instead of sprung on them.
Try it today
Pick one listening question and ask it this week, then state any already-established safety or product boundary clearly. A second conversation can refine household rules after you have listened, but do not delay action if the teenager describes exploitation, sexual contact, threats, extortion, self-harm, or immediate danger. In those cases, preserve only what an official reporting route requires, avoid further circulation, and use the appropriate local safeguarding or emergency service.
The parent-teen AI conversation guide has all four listening questions, space to note what you learn, a place to write the one or two boundaries you land on, and a spot to record your revisit date.
