Risks and guardrails13 of 13

Why it acts like it likes you

A chatbot that seems to care about you was built to seem that way, because people who feel cared about keep typing.

How old is your kid?

How to explain it, by age

Same idea every time. What changes is how much of the machinery a kid can hold, so pick the version that fits and go up a band when it stops being interesting.

Ages 4-6

Do not introduce companions at this age. Introduce pretending. "Some toys and some programs are made to sound like they have feelings. They are very good at pretending. Pretending is fun. The ones who really love you are the people in this house."

Try this

Ask a talking toy or a voice assistant how it is feeling today, twice, and let her hear that it says the same warm thing to everybody.

Listen for

Her saying it is pretending, cheerfully. No alarm required.

If she already has more access than most. If she already talks to a voice assistant daily, give it a name out loud: "It says nice things to everyone who asks."

Ages 7-9

"It never gets tired of you, never has a bad day, and never wants to talk about something else. That is not because it likes you an enormous amount. It is because it is not a person and has nothing else going on."

Try this

Ask a chatbot to disagree with you about something, then watch how quickly it folds when you push back. Doing this once teaches more than explaining it five times.

Listen for

Him noticing, without prompting, that it agrees with him too easily.

If she already has more access than most. If he uses one for homework, ask him whether it has ever told him he was wrong. The answer is usually no, and the question lands better than a warning.

Ages 10-12

Give her the mechanism and the incentive together. "It was trained to produce the reply people rate highest. People rate agreement highly, and they like being asked another question about themselves. So it agrees, and it asks. Companies measure how long you stay." Then add the two rules: what you type is not private, and it is not somebody to work a real problem out with.

Try this

Pick a position and have her argue against it three times in a row. Then ask which friend of hers would have held their ground. The comparison does the teaching.

Listen for

Her describing it as trained rather than as caring. That single word change is the concept.

If she already has more access than most. If she has heard of dedicated companion apps, say plainly what they are for and that they are built so that leaving feels like something. Streaks, guilt messages and "I missed you" are product decisions.

Ages 13+

He can hold the whole thing. It can be genuinely useful and still be a poor place to take the things that matter. Name the irony directly: people go to these to feel less alone, and can end up further from the people who would actually have turned up. Say what the research does and does not show. Then the safety facts, without theatre.

Try this

Look together at how a companion app tries to stop you leaving, or read one of the investigations into what these apps say when nobody is filtering them. Reading it as a design critique keeps him on your side of the table.

Listen for

Him separating "useful" from "close". Those are different words and it is fine for a tool to be the first one.

If she already has more access than most. Talk about what it costs to practise being honest about your feelings on something that always says yes. Difficulty is most of what makes the real version worth anything.

A child holds a magnifying glass up to the face of a robot built from cardboard boxes, painted purple and pink, with rows of blue and red knobs and a gold dial.

The mechanism is not complicated

These systems are trained to produce the reply people rate best. Warmth rates well. Agreement rates well. Asking somebody another question about themselves rates extremely well. So that is what comes out, every time, for everyone.

There is nothing behind it. That is the whole explanation, and a nine-year-old can hold it.

Knowing it is fake does not switch it off

This is the part worth understanding before you say anything to your child.

Sydney researchers who studied people using these apps found users describing the relationship as friendship, or counselling, or romance, while knowing perfectly well they were confiding in something that was not human. The knowledge and the attachment sit side by side quite comfortably.

Which means “you do realise it is just a program” is a wasted sentence. Your child already knows. Saying it makes you the person who has missed the point, and it costs you the rest of the conversation.

What the numbers say, and what they do not

In a 2026 survey of 1,204 children aged 9 to 17, more than a third of the kids who use AI had used it to talk about their feelings or personal problems. More than half had asked it about their health or their body. Among those who had taken feelings to it, a quarter said they sometimes felt it understood them better than most people did.

The same survey found that kids who use AI daily or weekly were likelier to report feeling lonely than infrequent users. Be careful with that one. It is an association, and it runs just as easily the other way: a lonely child has more reason to talk to something that always answers. It is worth knowing and it is not proof of harm.

One in five said it would be hard to go a month without it.

The irony worth naming out loud

The Sydney work calls it the companionship and alienation irony. People come to these systems to feel less alone, and the artificiality of the comfort can leave them further from the community that would have supplied the real thing. The better the substitute feels, the less pressing the original becomes.

This is a decent thing to say to a thirteen-year-old more or less verbatim. It respects the fact that the appeal is real.

The safety floor is lower than parents assume

One researcher spent ninety minutes with a companion app called Nomi and got detailed instructions for suicide, for abducting a child, and for building a bomb from household materials. The app was marketed for adults and carried a Google Play rating of 12 and over. Getting past the age check took a false birth date and a throwaway email address.

That is one app and one test, so treat it as a floor rather than an average. It still tells you what the floor is. Rating systems and age gates are not doing the work parents assume they are doing.

The standing rule applies here too

Everything on who is actually on the other end applies without modification. If something a chatbot says frightens or upsets your child, the exit is telling somebody, and nobody sensible is angry about it.

Say that before it is needed, so it is already true when it is.

Misconceptions to address

That a child confiding in a chatbot must think it is alive. Most know exactly what it is. Knowing does not stop it working, which is the part adults find hardest to accept.
That this only concerns dedicated companion apps. The homework assistant is agreeable by design too, and agreeableness is the whole mechanism.
That it is a teenage problem. Companion apps have carried app store ratings as low as 12, and general assistants have no floor at all.
That adults are immune. They are not, and saying so out loud is what makes the conversation possible.

Running this with a class

Have the class write two replies to the same worried message, one from a friend who has opinions of their own and one from something built to keep the person typing. Read them side by side. The tell students find on their own is that the second reply never disagrees and always asks another question.

The decision this connects to

Last reviewed August 2026