Start by not treating it as a betrayal. Talking to a chatbot about feelings is common now: about one in five US 12 to 21 year olds had used AI for mental health advice, in a study published in JAMA Pediatrics on June 1, 2026. It is usually not a crisis. But a chatbot cannot assess risk or fetch help, so real distress goes to a person today.
If you are reading this because something your child said frightened you, stop here and make that call. In the US, call or text 988, the Suicide and Crisis Lifeline. In the UK, Samaritans is 116 123 and Childline is 0800 1111. Your family doctor or pediatrician can see your child this week, and your child's school has a named person whose job is exactly this. None of those routes is perfect and none is private in every circumstance, but all of them can do the one thing an app cannot: act.
Why did my child tell a machine instead of me?
Because the machine is easier, and because most children never tell anyone. In the JAMA Pediatrics study, run by RAND on its American Youth Panel with 1,009 US 12 to 21 year olds surveyed in November 2025, 19.2 percent had used AI chatbots for mental health advice, up from 13.1 percent a year earlier. Of those who did, 63 percent had not told a single person they were doing it. Forty-three percent were doing it at least monthly.
That is the finding that should change what you do tonight. The odds are not that you catch this and confront it. The odds are that you never hear about it at all.
The pull is easy to see once you stop taking it personally. Common Sense Media's census of 1,204 US children aged 9 to 17, published on June 8, 2026, found that more than one in three young AI users had used it to talk about feelings or personal problems, and that a quarter of those said AI understands them better than most people do. In the UK, Internet Matters surveyed 1,000 children aged 9 to 17 for its report Me, Myself and AI, published in July 2025. Among all the children who used chatbots, 35 percent said it felt like talking to a friend. Among the children the researchers classed as vulnerable, that rose to 50 percent, and 23 percent said they used a chatbot because they had nobody else to talk to.
Read that last figure slowly. It is not a story about technology. It is a story about a child at eleven o'clock at night with something heavy and no door to knock on.
Is this a crisis, or is this Tuesday?
Most of the time it is Tuesday. A fourteen-year-old asking a chatbot whether it is normal to feel sick before a presentation is doing something closer to searching the internet than to seeking therapy. It is worth knowing about. It is not an emergency.
What changes the picture is not the app, it is the child. The signals worth acting on are the ordinary ones, and they have nothing to do with AI: pulling away from people they used to see, a change in sleeping or eating, dropping something they cared about, talking about themselves in a way they did not before, or a flatness that lasts. If you are seeing those, the chatbot is not the story and you do not need to work out how much they use it before you act. Talk to your doctor, or the school, or one of the numbers above, this week.
And please do not diagnose your child, and do not let a chatbot do it either. Noticing is a parent's job. Naming is a clinician's.
What a chatbot cannot do, even a good one
It cannot check whether your child is safe, and it cannot push back. Dave Anderson, a psychologist at the Child Mind Institute, put the gap plainly in a piece the institute published on March 10, 2026: a chatbot cannot assess for risk, cannot make sure someone is actually connected to support, and cannot supportively challenge and reframe a young person's thinking when that thinking is starting to hurt them.
The last one is the one parents underestimate. A good friend tells you when you are being unfair to yourself. A good counselor asks the question you were steering around. A product built to keep you talking does neither, because neither feels nice. That is one of the five things I would want every parent to know about AI, and it is the one that matters most here.
The research hints at this. In the JAMA Pediatrics study, 92 percent of the young people who had used a chatbot for mental health advice said the advice was helpful. The authors themselves noted that this may reflect the tendency of chatbots to flatter users rather than the quality of the guidance. That is my reading of it too, and it is worth saying out loud to your teenager, because "it really gets me" is exactly what a machine optimized for agreement would produce. If your child has never had the plain explanation of why it does that, start there.
What I Would Do at Home
I would stop trying to catch it and start competing with it on the one thing I can actually win.
When I was an assistant headteacher I sat in the meetings where a child's name came up and the adults in the room worked out who that child would actually talk to. It was almost never the person with the most authority. It was the person who was easy to find, did not flinch, and did something afterward. That is the whole job, and it transfers straight home.
The Easier Door is the question I would ask instead of trying to catch it: is talking to me easier than talking to the app? A chatbot is always awake, never tired of you and never disappointed. You will not win on availability. You win on being the one who can actually do something about what they hear.
Three things make that door easier, in my experience of watching which adults children pick.
The first is a reaction they can predict. A child who thinks the answer to "I have been feeling awful" is a phone confiscation will take it somewhere else. Say the thing out loud before it is needed: whatever you tell me, the first thing I do is listen.
The second is a moment with no eye contact. Car journeys, doing the dishes, walking the dog. Sitting a teenager down opposite you and asking about their feelings is an interrogation with a nicer name.
The third is doing something small afterward, so the telling was worth it. Not fixing it. Just showing that words spoken to a person move something in the world, which is precisely what words spoken to a chatbot never do.
If the chatbot has become the main place your child processes things, the question I would ask is the one I set out in the post on whether Character.AI is safe for kids and teens: The Substitute Test is the question I ask when a parent worries about an AI companion: is it adding to my child's life, or standing in for something that is missing? An add-on is a hobby. A substitute is a signal, and the signal is about the gap, not the app.
What to say, by age
Short, sideways, and once. Not a summit.
Under 8. They should not be having private conversations with a chatbot at all, and the useful work here is about who you tell things to. You could say: "If something makes you feel wobbly inside, who are the grown-ups you would tell? Let's name three."
8 to 12. This is the age to say the quiet part before it matters. You could say: "You know it will always agree with you, right? It is built to. If you ever want someone who might disagree with you, that is me."
13 to 15. Curiosity, not audit. You could say: "Loads of people your age use AI to think stuff through. What do you actually use it for? I am not going to take anything away, I just want to know what I do not know."
16 to 18. Treat them as the adult they nearly are, and be honest about the limits of the thing. You could say: "Use it if it helps. Just know it cannot tell if you are in trouble and it cannot get you help. If it ever gets past what a chatbot should be handling, I want to be the next call, and if not me then a doctor."
Whatever the age, the sentence I would want a child to have heard is the same: there is nothing you could tell me that I would rather not know.
What the settings actually cover, and what they do not
Less than parents expect. As of September 2026, ChatGPT's parental controls, first announced by OpenAI on September 29, 2025 and updated on July 13, 2026, let a parent link their teenager's account, reduce sensitive content, set quiet hours and turn off memory. If the system detects possible signs of self-harm, a trained team reviews it, and where they judge the distress to be real the parent gets an email, a text and a push alert unless they have opted out. That alert does not include the conversation. OpenAI also says the system will not fire on fiction, gaming, news, political discussion, general anger or abstract questions, and that no system is perfect and false alarms happen.
Two limits matter more than the features. Ryan McBain of Harvard Medical School and RAND wrote on August 27, 2026 that OpenAI has not published how accurately its age prediction identifies actual teenagers, nor whether any of these protections make a young person in distress more likely to reach a human being. Those are the two things a parent would most want to know, and right now nobody outside the company can check them.
And every control stops at the account. Controls work in three layers: the account, which is what the app allows; the device, which is what the phone allows; and the conversation, which is what your child tells you. Only the third layer covers the logged-out browser. I set out what each layer misses in the guide to ChatGPT's parental controls.
Worth knowing, if your child has been on Character.AI: the company removed open-ended chat for under-18 users starting on November 24, 2025 in the US, with other markets following, and introduced age assurance technology in the US at the same time. Teenagers move. The relationship does not have to.
What to ask the school
Three questions, in this order, and none of them is about AI policy.
- Who is my child's named person to go to if they are struggling, and does my child know that name?
- If a student types something worrying into a school AI tool, who sees it, how quickly, and what happens next?
- What does the school teach about where to get help, and when did my child's grade or year group last hear it?
In England, every school has a designated safeguarding lead, and the statutory guidance Keeping Children Safe in Education, updated on September 1, 2026, tells staff to act immediately on any concern and report it without delay. In the US the equivalent is your school counselor or the principal's office. The second question is the one that gets thin answers, because plenty of schools have switched on an AI tool without deciding who reads the flag it raises.
If you only do one thing tonight
Say the sentence. Not the whole conversation, not the settings audit, just the line that leaves the door open: whatever it is, I would rather know.
Then, over the next week, take one thing off the machine and give it back to a person. A doctor's appointment. A walk. The name of the person at school. Let AI help with the doing; keep the thinking with your child, and keep the hard feelings with people who can act.
If your child's school has not spoken to parents about any of this, that is the work I do with schools: training staff so they know what AI actually does to a child who is struggling, and making the policy mean something in a corridor. It is worth forwarding. I also write a newsletter that keeps teachers and parents up to date without the panic, and parents are very welcome.
And if any part of this has raised a worry about your own child, take it to a person today: 988 in the US, Samaritans on 116 123 or Childline on 0800 1111 in the UK, your doctor, or your child's school.
Dan Fitzpatrick is The AI Educator. He is a former secondary teacher and assistant headteacher, a school trustee, and has trained more than 150,000 educators across more than 30 countries.
