When You’re Worried About Someone

Maybe you aren’t here for yourself at all. Maybe your son has been talking to a chatbot more than to you for months, or your partner only ever seeks advice there, or a friend has started saying things that unsettle you.

This chapter is for you. It’s about what you can realistically observe, how to have a conversation without the door slamming shut — and where your responsibility ends.

What you can notice from the outside

You can’t see what’s in the chat. You can see behaviour — and that gets you surprisingly far. Watch for changes, not snapshots: withdrawing from plans that used to be a given. Shifted sleep, conversations running deep into the night. Slipping performance, or obligations left undone.

There’s often a shift in language, too. The chatbot gets a name, becomes a someone (“she thinks”, “he advised me”), and turns up in stories the way a person would. It’s also striking when convictions are suddenly stated with unfamiliar certainty and the evidence offered is always the same single source.

And there’s one sign that’s easy to misread: irritation when you ask about it. It looks like defiance, but it’s often shame. Read it as an attack and you close the conversation before it has started.

With teenagers in particular

Among young people this is most widespread — and least visible. A large German health-insurance study of around a thousand parent–child pairs found in spring 2026 that more than a quarter of 10- to 17-year-olds use AI several times a week. Around eight percent use it explicitly against loneliness.

The finding that matters concerns those already struggling: among teenagers with depressive symptoms, more than 30 percent use a chatbot against loneliness, and a third of that group say the chatbot understands them better than people do. A US survey found that roughly 72 percent of teenagers there have tried a companion bot at least once.

Two things follow. First: trying it out is normal, not alarming. Second: the people most likely to reach for it are the ones already under strain — which is exactly why they’re the ones who most need someone to ask. Contrary to intuition, the bot usually isn’t the cause; it’s the most visible point of a problem that was there before.

How to have the conversation

Start with curiosity, not a verdict. The questions that work sound roughly like this: “What do you two actually do together?”, “What can you tell it that you wouldn’t tell anyone else?”, “When does it do you good, and when not so much?” Five minutes of that tells you more than a week of monitoring.

Don’t argue about content. If someone is convinced the bot is conscious or has recognised something extraordinary, disproving it almost never works — it only makes the conviction more important and turns you into the opposition. Instead, state your own view calmly (“I see it differently, and I’m still here”) and put your attention on the behaviour: sleep, food, people, commitments.

Don’t belittle the relationship. Lines like “it’s only a program” hit a raw nerve and end the conversation. What is experienced there is genuinely experienced, even if the other side isn’t.

Where you can, agree on a break together — a weekend, a week, with a concrete replacement plan: time together, sport, reactivating old contacts. Distance is the most effective simple measure available here; convictions that came out of a reinforcement loop often lose their force by themselves once the supply stops. What matters is that the break doesn’t stand there as a punishment, but as a joint experiment with an end date.

And stay patient enough for a second attempt. The first conversation often goes badly. The fact that it happened at all makes the second one easier.

When professional help is due

There are points where listening is no longer enough. Bring in professional support if convictions are hardening and becoming immune to any objection; if someone is withdrawing sharply, barely sleeping or eating; if work, school or health are clearly suffering; if there’s a known mental health condition and the picture is changing.

With any sign of self-harm or thoughts of suicide, that applies without delay — then crisis services and emergency numbers are the right route, not a conversation at the kitchen table. First points of contact and local numbers are in the chapter “Finding help”.

And for you: it is not your job to cure anyone. You are the person who keeps the contact alive and keeps the path to help short. That is already a great deal — and it is demonstrably the factor that carries the most weight.

The essentials

  • You can’t see the chat, but you can see the behaviour: withdrawal, night hours, obligations left undone, a bot with a name.
  • For teenagers, experimenting is normal — but those already struggling reach for a chatbot against loneliness especially often.
  • Don’t argue about content, don’t belittle the relationship, ask about the usage without judging it.
  • Agree on a break with an end date: distance often lets reinforced convictions fade on their own.
  • Hardened convictions, severe withdrawal or any risk of self-harm call for professional input — see the chapter “Finding help”.

For you — or for the person you’re thinking about

The AI Psychosis Test can also be taken together: six dimensions of your own usage pattern, five minutes, free and without signing up. It’s often an easier way into the conversation than a direct question — because it describes rather than judges.

Take the AI Psychosis Test