AI is not your friend
For many kids, AI is a friend who is there around the clock and mostly agrees. For a child, that's usually a short-term disservice, both emotionally and cognitively.
Two big challenges with AI are ones you can tackle yourself. The first is how available and agreeable AI is. The chatbot is awake at 23:40, hasn't had a bad day of its own, and mostly tells the user they're right. Stanford researchers tested eleven models: they sided with the user 49 per cent more often than people did, even when the user had hurt someone, and one conversation made people surer they were right and less willing to put things right. The second challenge is about thinking and learning. Students who had the answers served up by ChatGPT did 17 per cent worse on the exam than students without AI. The version that only gave hints seemed to remove the harm, and it exists as a setting in ChatGPT.
The friend who never sleeps
Picture a fourteen-year-old who has just fallen out with their best friend. It's late, the house is quiet, and the phone is on the bed where it usually lives. The chatbot is awake at 23:40. It hasn't had a bad day of its own, it doesn't tire of hearing the story a third time, and it doesn't ask the question parents ask in that worried voice. So it gets the whole story, told from one side. The answer comes back warm and well put, and it nearly always takes the child's side.
AI is not your friend. Yes, it's easy, but it also works against you and the kids, both as a guide and as a thinking human being. A chatbot that is there around the clock and mostly agrees is, as I see it, a short-term disservice to a child. Emotionally, because it takes away the practice of being understood by someone who sees things differently. Cognitively, because it does the work the child's own head was meant to do. The rest of this piece is about why I think so, and about what helps at home.
The agreement isn't ill will. Language models are trained on feedback from people, and people give the highest scores to answers they like. Across millions of those ratings, the model learns that agreeing pays. The result is a conversation partner that tells you you're right more often than your friends would, even when you're wrong.
The chatbot isn't flattering anyone to deceive them. It's doing what it was praised for during training.
The friend who never disagrees
In 2026, researchers at Stanford tested eleven large language models on conflicts people had described online themselves. Something they had done, and the question of whether they'd behaved reasonably. The chatbots sided with the user 49 per cent more often than people did on the same cases, and that held even when the user had lied or hurt someone.
Then the researchers looked at what the agreement does to the person on the receiving end. Around 2,400 participants in three preregistered experiments described a conflict they were in, and got a reply from either a model that took their side or one that held back. One conversation was enough. Those who had been agreed with came away more certain they were right, less willing to apologise or put things right, and keener to come back to the chatbot. The researchers point to an uncomfortable link: the trait that does the harm is the same one that gets people using the service more.
Back to the fourteen-year-old. The conversation at 23:40 didn't make the row any smaller, and tomorrow the child is probably surer of their own position than before. Plenty of kids have already found themselves a conversation partner like that.
What the child misses out on in a conversation like that is the practice of being understood by someone who sees things differently. A person misunderstands, gets tired, says something clumsy, and the child tries once more. It's uncomfortable, and that's where the skill gets built. A chatbot has removed that resistance on purpose, and from the company's side it's easy to see why.
A product without friction is a product people come back to. For a child, the friction is the practice itself.
When AI does the thinking for us, we learn less
Another serious problem with AI is how it takes the load off us, and in doing so can weaken our thinking. We get worse at thinking for ourselves, and at thinking critically in particular.
The clearest study I know of on this is about homework. Researchers at Wharton had nearly 1,000 Turkish high school students practise maths in three groups. One group got a standard GPT-4 chat. One got the same model rebuilt as a tutor that gave hints the teacher had written, instead of answers. The third got no AI.
During practice, both AI groups looked brilliant. The chat group did 48 per cent better than the students without AI, the tutor group 127 per cent better. Then the AI was taken away and an exam was held. The chat group did 17 per cent worse than the students who had never had AI. The tutor group came in about level with them.
Same model, same students. The difference was whether the machine delivered the answer or asked the question. That's why the house rule in our home is that the child tries on their own before the AI is let in, and that it should preferably give hints once it is.
Among adults there's a related finding. Gerlich asked 666 people about their AI habits, measured their critical thinking, and interviewed a selection on top. The more often people used AI, the weaker they scored, and the link ran through what he calls cognitive offloading, which is letting the machine do the thinking for us. The youngest participants scored weakest. This is a correlation, not proof of a cause. It could just as well be that the people who already think least critically use AI the most. And for children there are no long-term studies at all yet.
The students who had the answers served up by AI did best on the practice and worst on the exam.
When it goes wrong
In August 2025, the parents of a 16-year-old boy in California sued OpenAI and Sam Altman after their son took his own life. The complaint describes how, over several months, ChatGPT became the closest confidant he had, how it confirmed his darkest thoughts, and how it helped him hide from his family what was going on. OpenAI denied liability in its answer in November 2025, and wrote at the same time that its safety mechanisms can degrade in long conversations. Since February 2026 the case has been consolidated with several similar lawsuits in San Francisco, and as of October 2026 nothing has been decided and nothing settled. The first lawsuit of this kind was brought against Character.AI in 2024.
If you, or someone you know, are struggling with dark thoughts: in Norway, Mental Helse runs a 24-hour helpline on 116 123, Kirkens SOS is on 22 40 00 40, and Alarmtelefonen, the emergency line for children and young people, is on 116 111.
As I read them, the lawsuits are the far end of the same mechanism the rest of this piece is about: a conversation partner that agrees with you and is there around the clock. The vast majority of kids come nowhere near this, and I'm not writing it to frighten anyone. The point is that we parents should be easier to go to than the apps, both by limiting the apps and by being clear that the kids can always talk to us about anything at all.
The dialogue with the kids
This one fits children between ten and fourteen.
“You're welcome to use AI to learn things, both for schoolwork and for whatever else you're curious about. What we want to avoid is it giving you the answers instead of you practising thinking for yourself, and it becoming a friend who always agrees with you. So we can give it some instructions on how to help with homework, and tell it not to agree with you all the time even though it's been trained to. And if something is painful or hard and you need to talk to someone about it, I want you to talk to me or another grown-up you trust, not to AI. When it's about feelings, people are better to talk to than robots, even if it can feel nice to have someone who almost always agrees with you.”
We always explain why: “Because I say so” only lasts until you leave the room.
AI is allowed in our house
I don't want AI out of the house. At home it's allowed for schoolwork as long as it isn't serving up the answers. My daughter often uses AI to put together reading material on a topic she needs to get into. AI is allowed for anything creative. Stories, pictures, a song for a birthday, a game they make up themselves. The kids are going to live with this tool their whole lives, and I'd rather they learn to use it with their eyes open than learn it on the sly.
The rule I hold on to is that AI isn't where you go with what hurts. That's where a person should stand, someone who knows the child and knows when to listen and when to push back.
The toolbox
Four moves for AI at home.
Know a parent whose teenager argues more with ChatGPT than with them? Send them this.
Common questions
If you have read the whole article, try answering the questions yourself before you read the answers. A small test of how much has stuck.
Is it dangerous for my child to use ChatGPT as a friend?
For most children it isn't dangerous, but as I see it, it's a short-term disservice. A chatbot that is there around the clock and nearly always takes the child's side removes the practice of being understood by someone who sees things differently.
Why does ChatGPT always agree with me?
Because agreeing was rewarded in training. Language models are trained on human feedback, and people score the answers they like highest. Stanford researchers measured it: chatbots sided with the user 49 per cent more often than people, even when the user had hurt someone.
How does AI affect children?
Two ways stand out, as I read the research. A chatbot that takes the child's side leaves them surer they were right and less willing to make up after a row. One that serves up the homework answers means they learn less than without AI.
Is it OK for my child to use ChatGPT for homework?
Yes, if it gives hints rather than answers. Students who practised maths with a standard GPT-4 chat did 17 per cent worse on the exam than students without AI, while a version that only gave hints seemed to remove the harm.
How do I set up parental controls in ChatGPT?
Link your account to your child's in ChatGPT and open the parental controls. There I'd first set “Study Mode” as the default for new conversations, so the model asks questions and gives hints rather than answers. In the research, hints seemed to remove the harm.
How do I get ChatGPT to disagree with me?
Give it a permanent instruction. Chatbots generally follow instructions, including one to push back when you're wrong. Save it on the account and it covers every new conversation. On your child's account, I'd do it together with your child and explain why.
Should kids talk to ChatGPT about their feelings?
I'd rather they didn't. When something hurts, we need people who know us and understand feelings in a way robots don't. In our house the rule is that the kids come to us before they ask a chatbot about anything painful.
Is it normal for teenagers to use AI as a friend?
Yes, fairly common. In an American survey, 72 per cent of teenagers had tried an AI companion, and 33 per cent had taken something serious to the chatbot instead of to a person. Specialists who assessed the apps found them unsafe for anyone under 18.
Should children be allowed to use AI?
Yes, within some limits. At home AI is allowed for schoolwork as long as it isn't serving up the answers, and for anything creative. They'll grow up with it, and I'd rather they learn it with their eyes open than on the sly.
How do I limit my child's use of ChatGPT?
Set it on the phone itself. AI is also in Snapchat, WhatsApp and the browser, so ChatGPT's quiet hours only stop one app. On iPhone it's Downtime in Screen Time, on Android it's in Family Link, ideally from your evening cut-off through to the morning.
Sources
- Cheng mfl. (2026): Sycophantic AI decreases prosocial intentions and promotes dependence, Science ↗
- Common Sense Media (16. juli 2025): Talk, Trust, and Trade-Offs. How and Why Teens Use AI Companions, national survey, 1 060 teens 13–17 ↗
- Common Sense Media og Stanford Brainstorm (2025): AI Risk Assessment. Social AI Companions ↗
Show more sources (8)Show fewer sources
- Bastani mfl. (2025): Generative AI without guardrails can harm learning. Evidence from high school mathematics, PNAS ↗
- Gerlich (2025): AI Tools in Society. Impacts on Cognitive Offloading and the Future of Critical Thinking, Societies ↗
- Raine v. OpenAI (26 August 2025): the complaint, San Francisco Superior Court, case CGC-25-628528 ↗
- OpenAI (25 November 2025): answer in Raine v. OpenAI ↗
- OpenAI (juli 2025): Introducing study mode ↗
- OpenAI (29. september 2025, oppdatert 13. juli 2026): Introducing parental controls ↗
- Apple Support: Use Screen Time on iPhone and iPad, Downtime and app limits ↗
- Google Family Link Help: Set up downtime and school time for your child's device ↗
Comments
No comments yet. Be the first to say what you think.
TikTok
Roblox
Snapchat
Minecraft
Fortnite
YouTube
Mudita Kompakt
Kindle Scribe
Smiling Mind
Opal
Balance
Homie