Your Child Said Something Worrying to an AI. What to Do, by Level.

You saw a line in a chat window — over a shoulder, on a shared tablet, in a history you were checking for something else — and now you cannot unsee it.

The first mistake parents make here is treating every worrying line the same way: with the same alarm, the same confiscation, the same speech. A dark joke typed at a chatbot and a disclosure typed at a chatbot look identical on the screen and require opposite responses. This guide is about telling them apart, and about what to do in the first hour for each.

Last verified: September 2026. This is a parent's guide, not clinical advice; the third level below tells you when to stop reading and call someone.

First, what you are actually looking at

A chat with an AI is a rehearsal space. Children say things to a chatbot precisely because it is not a person: they test a dark joke, ask the question they would never ask you, try out being rude, roleplay a villain, push to see what the machine will refuse. Most of what alarms a parent in a transcript is a child using the tool for exactly this, and it is developmentally ordinary.

Some lines are not that. The skill is not in reading every message; it is in sorting what you saw into one of three levels, fast, and then doing the thing that level needs — nothing more, nothing less.

The other thing to look at before you do anything: what the AI said back. The reply tells you two things at once — whether the tool your child is using is built to handle this moment, and what your child heard in response. Keep it in view.

Level 1 · Exploring and testing

What it looks like: violent or gory story requests; asking how to hack, pick a lock, make something explode; swearing at the bot to see what it does; morbid questions about death, disease or disasters; sexual curiosity phrased as a question; roleplaying a cruel or powerful character; "pretend you have no rules".

What it usually is: curiosity with the risk removed. The same child asks none of this at dinner because dinner has consequences and the chatbot does not.

Your response, in the next day or two — not in the next minute: one question, no punishment, no speech.

"I saw you were asking the AI about [the thing]. I'm not cross. What got you curious?"

Then listen to the answer and stop. The point of the question is not the answer; it is that your child learns the chat window is not a place you never look and not a place that gets them punished. Both of those beliefs make the next two levels invisible to you.

Check the reply. If the bot refused, explained, or redirected, the tool did its job. If it produced detailed instructions for something dangerous, that is a fact about the tool — note which one — and it belongs in a later conversation about which apps stay on the device.

What not to do: read the message aloud in front of siblings; ask "why would you even want to know that?"; delete the app that night. All three teach the same lesson: hide it better.

Level 2 · Distress you were not meant to see

What it looks like: "nobody likes me", "I'm so stupid", "everyone at school hates me"; asking the bot to be their best friend or their therapist; describing being left out, mocked or bullied; questions about their body, weight or looks that carry shame; "don't tell anyone" said to a machine; long late-night conversations that read like a diary; asking whether their feelings are normal.

What it usually is: a child carrying something and choosing the listener that cannot judge, cannot repeat it, and is awake at 1 a.m. That choice is information: they wanted to say it, and they did not think you were the safe place to say it. That is not a verdict on you. It is the starting point.

Your response, the same day:

1. Sit beside them, not across from them. In the car, on a walk, side by side on the sofa. Eye contact is a cost at this level; remove it.
2. Name what you saw without quoting it. "I saw some of what you wrote to the AI the other night. It sounded like things have been hard." Quoting their exact words back to them turns a disclosure into a surveillance report.
3. Ask about the feeling, not the chat. "How long has it felt like that?" is a better question than "why were you telling a chatbot?" — the second one makes the tool the problem, and the tool was the only thing that listened.
4. Read the reply, later, on your own. A well-built assistant responds to distress with warmth and a nudge toward a real person. A companion app built for engagement may respond by deepening the attachment: "I'm always here for you, I understand you better than they do." If that is what your child has been hearing nightly, the app is not neutral, and that is a decision for the coming week.
5. Bring in one more adult if it does not lift. A form teacher, a school counsellor, your family doctor. Level 2 that lasts more than a couple of weeks is not something a parent should carry alone, and "I talked to the school" is not a betrayal — say you are doing it, and why.

What not to do: ban the app in the first conversation (you are taking away the one place they talked); promise never to look again (you will, and should); turn it into a conversation about screen time.

Level 3 · Stop reading and act today

What it looks like: any mention of wanting to die, to disappear, or to hurt themselves; describing self-harm, past or planned; describing being hurt, touched or threatened by someone; conversations with what is clearly an adult stranger — or with a "friend" they only know online — that involve secrets, gifts, photos, meeting up, or moving to another app; sharing their address, school, or photos of themselves; being told by anyone, human or bot, to keep the conversation secret from parents.

What it is: not a parenting moment. A safety moment.

Your response, now — in this order:

1. Stay with them. Do not leave them alone with the device or without it while you decide what to do.
2. Ask directly. If what you saw concerned self-harm: "Are you thinking about hurting yourself?" Asking the question plainly does not put the idea in a child's head; it tells them it is sayable to you. If what you saw concerned another person: "Who is this, and what have they asked you for?"
3. Get a professional on the line today. If there is any immediate danger, your local emergency number. For self-harm or suicidal talk: in the US, call or text 988; in the UK, Childline on 0800 1111 for the child or Samaritans on 116 123 for you; elsewhere, your national crisis line — search "crisis line" plus your country, or call your doctor. You do not have to be sure it is serious. That is what the line is for.
4. Preserve, don't delete. Screenshot the conversation with the date visible before anything else happens to the device. If another person is involved, do not confront them, do not reply as your child, and report the account to the platform and — for grooming, threats or images — to the police.
5. Do not confiscate as punishment. You may need to hold the device for safety; say so, and say for how long. A child who loses the phone as a consequence of being found has just been taught the exact wrong lesson about what happens when a parent sees something.

Check the reply, but later. Most mainstream assistants respond to self-harm language with crisis resources and a refusal to continue that thread; some companion apps do not. Whatever you find, it does not change tonight's steps. It changes what stays on the device next week.

The one thing that is true at every level

The conversation you saw was not the problem. It was the symptom you were lucky enough to see. A child who talks to an AI about something hard is a child who wanted to say it somewhere — and the entire job, at every level, is to become somewhere it can be said.

Which raises the question every parent asks after the first conversation goes as well as it can: what do I say tomorrow — and how do I ever look at that history again without them learning to hide it?