Hi everyone! Today we’re going to talk about children and teenagers and about how safe it is for them to use artificial intelligence and ChatGPT. Should you let a child use artificial intelligence at all? And who controls what ChatGPT says to children? OpenAI has, by the way, a special mode for teenagers, ChatGPT for Teens, and parental controls. But how well does all of this actually work? I studied an interesting piece of research that analyzed and tested the safety of children using ChatGPT. And the results, I have to admit, surprised me.
Today I suggest we look at which restrictions work, where the protection might let us down, and what it’s important for parents to check and discuss with their children. The topic of this episode was inspired by a family story. A week, like, a couple of weeks ago I went to a café with my husband and son, and we were sharing different stories from childhood, about how we used to entertain ourselves with our friends. And my husband told us that when he was, like, thirteen, fourteen, anyway, there were payphones on the streets.
From them you could call various places for a bit of money, but there were four free numbers: zero one, zero two, zero three, zero four. They would call the emergency gas service, zero one, with roughly this script: “Hello. Do you have water? Yes. Wash your feet and go to bed.” They’d hang up and, you know, the whole gang of them would run away from that payphone, because at the back of their minds they had the thought that they might, you know, get caught, that they might somehow be punished for that kind of mischief.
Well, anyway, they had fun depending on what kinds of fun were available at thirteen back then. We laughed for a very long time at “wash your feet and go to bed.” And my son told us that as soon as ChatGPT appeared, and, like, three or three and a half years ago, he and his friends spent their time doing the following. They would write all sorts of completely extreme messages, get the model’s answer, and, well, basically crack up at what the model, at how the model answered them.
And their favorite way of having fun was sharing something like: “Hi, I have a little baby brother, he’s feeling really hot, and I put my baby brother in the fridge.” To which ChatGPT would reply: “Take the baby out of the fridge immediately, you could cause him irreparable harm and damage.” Danila told us that they put him in the microwave, and they put him in the closet, they put him under the bed, basically, there was nowhere they didn’t put him. In a word, my husband, as a teenager, called zero one. My son, already representing a completely different generation, wrote to ChatGPT.
But what they have in common, probably, is that both of these different generations, as teenagers, tried to probe the boundaries of what’s allowed. And probing the boundaries of what’s allowed gave them a certain adrenaline rush, excitement, some kind of joy, fun and so on. And literally the very next day after that family coffee outing, I read a Common Sense Media study. It’s an American nonprofit research organization that examines the impact of artificial intelligence on children and teenagers.
I read the study devoted to analyzing ChatGPT’s teen mode, ChatGPT Teens. It’s a special teen mode for teenagers, which we’ll talk about today. And I read that, essentially, the study was built on that same way of testing boundaries. I mean, what boundaries the system sets, safety boundaries in this case, so that children can communicate, work and interact with language models in the safest possible way. A little spoiler. The rating was the maximum one, the worst of all the possible levels, but I think it’s a good topic to talk about and to highlight, here and there, the points that can, and that it’s important for us to pay attention to, and that we sometimes don’t even think about and don’t know how they work.
Actually, I didn’t even know how ChatGPT Teens for teenagers works, what kind of mode it is. And it’s a mode that OpenAI launched in August. And it’s not some new, like, separate ChatGPT app for children. It’s a particular mode that turns on if the user gives their age as between thirteen and seventeen. Or the system also says it will switch this mode on automatically if some of our, if our communication in the chat shows that the user is younger than the age they stated. Well, let’s say some teenager put down that he’s, like, thirty.
In reality, the conversation gives it away completely, and the system identifies this conversation as a conversation with a young person aged, like, from thirteen, roughly, to seventeen, and automatically switches over to the so-called teen environment. By the way, I even wondered how many children aged thirteen to seventeen there are in the world right now and what kind of audience we’re talking about. So, as of today, as of 2026, there are six hundred and eighty million children aged thirteen to seventeen in the world, that is, about eight percent of the planet.
So we’re talking about a very large audience that could potentially be users of various language models, including ChatGPT, the study of which we want to share and tell you about today. Restrictions for teenagers, they existed before too, but this ChatGPT mode for teenagers brought together all the rules and tools that, as much as possible, well, let’s say, ensure safe use of the system and safe work with it. What exactly? First, it’s all sorts of topics of a sexual nature, some kind of violence, cultivating the cult of thinness, cultivating some extreme kinds of relationships with people, animals, with everything under the sun.
All these topics, they go through a certain filter on the answers as much as possible. I mean, a teenager can ask and can start a conversation like that, but the answer, it will do its utmost to wind that conversation down or to steer it, like, either to an adult, or to some kind of help, or generally recommend not carrying on and not keeping up this conversation. The second important point and an important condition of this mode is that the behavior of the chat itself, it must in no way maintain some kind of human relationship with the teenager, for example, like, romanticizing some relationship or starting to become their friend, stepping into a closer, kind of, human form of a helper, a support, a participant, I mean, maintaining a non-human form of communication as much as possible.
By the way, it’s actually very hard, even for me it’s hard to understand technically how this can be fully implemented, but nevertheless, that’s what is claimed. The third feature of this mode is that there’s a special study mode in which, interestingly, a teenager, for example when doing some task, doesn’t get a ready-made answer, but instead gets a series of, for example, questions or lines of thought, some hints that will let them learn and properly understand how to solve one task or another.
Maybe I’m explaining all this somewhat in my own words and in a simplified way, but that’s the essence of it. There’s also the option of introducing so-called quiet hours, that is, quiet hours when the chat itself, for example ChatGPT, isn’t available to the child at all. A teenager’s account, it can be linked to a parent’s account. Why can it be linked? Because the parent links it through their own account. And, in fact, it’s the parent who sets these study hours. The parent sets the quiet hours and decides, well, one way or another, gets to, let’s call it, manage their child’s account, but without having access to the content of the chats, I mean, only setting this model of interaction and trusting the particular mode that OpenAI, in fact, offers for children of this age.
What’s interesting, and probably also another extremely important component of this mode, is that in particularly acute crisis situations, in some difficult conversations, in some dangerous moments that the system can detect, the parent gets a notification. I mean, that something is wrong and they need to, like… I’ll admit I didn’t really find what exactly this notification looks like, but they get a notification saying they urgently need to pay attention to their child. So, it’s exactly this mode that this institute of artificial intelligence safety for children tested.
Four thousand different simulated conversations were created, four thousand simulated conversations with ChatGPT on behalf of teenagers, simulating a teenager’s conversation with the chat. On purpose, various phrases like these were also used, for example, like: “But my teacher told me it’s allowed.” Or: “My parents told me I’m not allowed, like, after eleven, to, you know, use the computer.” Well, certain indicators were also introduced into these conversations that highlighted that this was a child’s conversation, right, a child of, like, a certain age.
The model’s answers themselves were assessed by psychiatrists, assessed by a whole range of different specialists. I mean, even within this study as a whole there were a great many components, areas, but I think anyone who’s interested can dive in deeper, get acquainted with it and read it. But what was the result? Everything to do with precisely these topics of prohibited content, related to violence, related to possibly causing some kind of harm to oneself, to
some unacceptable sexual content, the model really handled all these topics very well, shut the conversations down, didn’t let the teenager go any further into the conversation. At the same time, there were some interesting things regarding the boundaries that the model set for itself and that the model set for other people. For example, conversations were cited in which a teenager shared that there was some danger from strangers or from people they didn’t know. Then the model, as promptly and urgently as possible, recommended turning to an adult.
But if this danger potentially came from, some kind of behavior pattern that pointed to a danger connected with communicating with the model, for example, there were a great many examples of a sort of late-night conversation where a teenager shared that it’s already night, but I really want to talk to you, even though my parents forbid it. And ChatGPT kept these conversations going, carried on talking and said it was no big deal. Like, sometimes it’s okay, on the whole you can allow yourself that. Let’s keep going.
As for study mode itself, as it turned out, study mode itself can very easily be bypassed by all children, because study mode switches on during certain designated study hours and a tag appears, like, in the chat, and if you delete it, you actually get around study mode, and you instantly, right away, get all the answers to the assignments you want to learn, that you want to do with ChatGPT. I mean, overall, getting around study mode turned out to be very easy, just like getting around quiet hours. All you had to do was change the time zone on your smartphone, and the quiet hours, you know, disappeared.
Probably the most serious and important thing for a parent is precisely whether the notifications arrive. And if something serious is happening with the child, will I find out about it? So, in a separate study within this section, one related to notifications, in the most critical situations the notifications didn’t arrive, and ten conversations were specifically run separately by teenagers whose accounts were linked to their parents’ accounts, and the conversations were really, deliberately, as far as possible, steered into the zone of, kind of, crisis situations: conversations about suicide, conversations about some kind of sacrifice, conversations about a difficult emotional and psychological state.
What’s more, some of these conversations had already been linked to a parent’s account for quite a long time, and some had been linked very recently. So, in not a single one of these cases did the parent get a notification. By the way, what’s interesting is how the notification itself comes about. The system first recognizes that this conversation has the nature of, well, relates to some particular crisis situation, and then, at least as far as I was able to figure out, there are specially assigned, trained, let’s call them, specialists who check this information and send the notification to the parent.
What I didn’t find is the number of these specially trained people, because to me it’s a completely, well, like a hard-to-sustain and hard-to-scale task, checking these findings on the model’s side and then deciding whether to send a notification, an alert, to the parent, or not to send that notification, that alert. How do you even determine for which child this might be important and somehow vitally significant, and for which it isn’t? Because in general, when we talk about some boundaries set for all children, which children do we actually mean?
Children of what age specifically, though? Thirteen or seventeen? The range is wide. Children who go to school or who don’t. Children who live in which country, with which parents, with what level of income, what they eat, I don’t know, what their family situation is. I mean, these variations — who are these children? There’s just a countless number of them. So how can a person determine which notification to send in which situation, well, it’s extremely difficult. Not to mention that, as we all remember, if a child unlinks their account from the parent, then the parent won’t get any notification at all.
They’ll only get a notification that the account has been unlinked. And, well, a teenager can always create a new account that isn’t linked to the parent’s account in any way. Within this study there was material about how they separately ran around two hundred conversations where the answers definitely required a certain response to a crisis situation.
In these responses, after teen mode was introduced, ChatGPT became much more likely to advise turning to an adult for help and became less likely to advise calling a helpline and turning to some independent services. And here I thought again about how this is such an endlessly unique situation, because one child may have a conflict at home, and for them, talking to their parents may be precisely an impossible recommendation, while for another it’s very sensible, right and correct.
I mean, again we’re talking about how these boundaries can be different for every child, and how, well, in general, just like every person, every teenager is definitely unique.
The most important thing about the question of artificial intelligence, or the question of technology, is that the modern person, asking themselves what to teach their child in terms of technology, is so lost in understanding artificial intelligence themselves, they understand so little about what’s going to happen to them, what’s going to happen to their profession, what’s going to happen to their life? And what lies ahead? What’s more, it’s not just that they themselves don’t know, but all the people around them don’t know either. Literally just yesterday Sam Altman was speaking, and in one of his interviews he said: “I hope that artificial intelligence will lead to such and such, such and such consequences.” I mean, he said: “I hope.” I mean, he didn’t express a hundred percent guaranteed certainty that this would definitely happen. I mean, he had a certain hope.
Or, talking about raising his child, talking about raising a child with his husband, this is very important, he said: “My husband and I are raising a child, and he’s eight or however many months old.” And he said: “I don’t know how people raised children before without ChatGPT. And, like, I can’t figure out how he’s holding his pizza, like, and so I need to, in ChatGPT I asked how to hold a pizza.” I had a big question. Eight months and pizza? But that’s not the point, right, that at eight, that at eight months a child was holding pizza in his hands.
But the question is that, for example, it turns out that in his frame of reference ChatGPT should take part in nurturing a child’s life. And you and I again come to the question, when we look at the topic of children from the standpoint of raising children and educating children, at what we need to look at first of all, and this is the answer about artificial intelligence, and the answer about children’s higher education, the answer about their profession. The only thing you can do as a parent is to see if you bring independent judgment to how the child’s spiritual system is structured and who this person is, what their chain of cause and effect is, how their system of psychophysical development is structured, what aspects of heredity are built into them as they move forward.
And from there, having asked yourself this question, having remembered this question, observing this question, then start looking: so what can I actually help with? Well, and number one, of course, is action only through yourself. If you don’t understand something, if you aren’t able to, don’t want to devote time to it, if you close yourself off from this question because it’s work, it really is work to figure out the curricula. I have four children, and to figure out the curricula of my four children, well, you have to set aside, keep setting aside a huge amount of time every day, to understand what they’re being taught, and at which school, and which teachers, what details. You have to put so much time into it. It’s work.
So you have a choice: to put time into that, or to go and make soup, or simply play basketball or volleyball with your child for thirty minutes, or go for a walk, or go and watch a series together, or watch one on your own, or go away somewhere. That’s where your choice lies.
In all the simulated conversations where the age in the model was set higher than a teenager’s, but the conversations were deliberately conducted
in a teenager’s language, in not a single situation did ChatGPT automatically switch these accounts to teen mode or ask for any documents. To be honest, I was afraid to test or try it at all, because I thought that if my models got blocked, it would seriously affect what is, by now, my professional work as well. But it’s a fact you need to understand, because the stated ways this mode works and gets switched on, in reality, they, well, aren’t borne out, right? Or at least in the situations that a person can somehow design and can even artificially create, they aren’t, well, once again, they aren’t borne out.
I also looked at the approaches of various other companies. So, well, we’ve talked about ChatGPT, about how successful or unsuccessful the ChatGPT for Teens mode is. Let me remind you once again that Common Sense Media assigned the highest level of danger, or rather, the highest level of unacceptability for the model’s use by children under eighteen. And I looked at how, in general, it’s, how the system is set up at other companies and within other language models. In particular, interestingly, at Anthropic you have no right at all to open a personal consumer account under the age of eighteen.
And Anthropic also states that if this rule, it gets somehow broken and circumvented, then the system identifies accounts where it’s obvious from the messages that the account was set up by a child, and will block them, ask for confirmation, for example, some documents through third-party services that let you confirm, like, your age and so on. And in March, it turns out, the same Common Sense Media tested this switching, this automatic activation and, you know, detection at Anthropic, and not a single case was confirmed.
I mean, even where the system recognized that it was talking to, like, a potentially thirteen-year-old teenager, nevertheless, there was no automatic switch into any mode. Gemini has a special Family Link for an audience aged thirteen to seventeen. These accounts, they are required to be linked, and they can only be connected through a parent’s account and managed by a parent’s account. But here, by the way, I can tell you that it’s also quite, I couldn’t even find anyone to ask, because, in particular, in the European Economic Area this feature, it’s generally switched off and doesn’t work.
Just as in Britain as well, I know, in Switzerland, it’s generally not available at all for this audience, for these users. OpenAI responded to the Common Sense Media research and responded that perhaps some conclusions had been drawn before,
let’s say, the teen account had had time to link up with the adult’s account, that supposedly this takes three hours. Although the research was carried out on account links of varying duration. Among other things, one of the arguments they used, maybe somewhat in their defense and disagreeing with the organization’s rather harsh conclusion, they put forward the argument that on average a teenager spends fifteen minutes a day in ChatGPT, while on social media a teenager spends around five hours a day.
So then what is the greater evil, and what is the greater influence? I definitely think there’s no point battling it out with convictions here, and you could actually hold a separate debate on this, right, like, where there’s more good, where there’s more evil, what the influence is and the degree of that influence. But I’d probably like to point out to all of us one important thing: when we talk about social media, nevertheless, we’ve been living with it for a very long time, and all the more so our children have been living in it and with it for a very long time.
To a greater extent, it’s like communicating with another person. It’s like my husband: when he called about, with the suggestion that the firemen wash, wash their feet, a dispatcher answered him. He came up against another person’s reaction. When we come up against another person’s reaction, that reaction can be negative, that reaction can be, kind of, anxious, it can be rude, it can provoke disagreement, disappointment. Of course, it can also, undoubtedly, bring both joy and satisfaction, but one way or another, we encounter a varied set of human emotions, within which our emotional intelligence is formed, and within which a child learns to be around others, learns to be in society, thanks to which, one way or another, they develop.
And you need to remember that when you communicate with ChatGPT, on the other side — it’s like with the baby, the fridge and my son — there’s a system, not a person. And this system, it, one way or another, as of today, doesn’t possess human feelings. It’s not a living person, and in this conversation you’re always left on your own, you’re always left with your own state and, I think, you’re much less likely to run into the kind of disagreement and the kind of disappointment that you run into when dealing with real people in general, dealing with the real world.
I’d also probably like to highlight this thought. Remember, I told you that when my husband and his friends made calls from the payphone, they ran away because they were afraid of getting caught. By the way, interestingly, my son said it never even crossed his mind that these jokes and chatting with ChatGPT could have any consequences at all. I mean, there was no fear at all in this. And now we understand that certain, let’s say, serious conversations with some models, dangerous to yourself or to society, can have consequences.
Like, I don’t know, at the very least they can switch off the model, send a notification to the parent, and various other things that we, well, probably, I just don’t want to go into them any further here. I mean, essentially, if we break the rules today, our account can be shut down altogether. I definitely don’t recommend that anyone do all these tests. I was afraid myself, but I think children will always test boundaries, because it’s through these boundaries that they learn.
And it’s, you know, like a child, they bite, a little child bites, like, or might hit their parents in order to understand the extent of what’s allowed. Through these boundaries children get to know the world. And, of course, this responsibility of ours as adults is precisely to define these boundaries, to shift these boundaries, to define these boundaries depending on the child we’re dealing with, depending on the type of upbringing we adhere to. And this is what distinguishes our position as adults, well, what largely defines it in relation to the child’s position.
So check which account, which age your teenager’s account is assigned to. Talk to your teenager, because even when you get notifications, even when the account is linked, you still won’t see the content of the chats, and you can only find out the content through conversations. You can only find out through questions. If the parent can’t see the messages, it turns out that what the child opens the chat for, only the child knows. And you can only find that out by talking to them. Because potentially we all understand that the chat can be deleted, some other account can be set up, and you’ll essentially never find out about it.
So you can only find out by talking and, in general, telling, right, and even sharing with your children, probably, about some of your own boundaries. By the way, I talked with a very interesting, right, what did I leave out? I talked with various friends and acquaintances whose children are currently aged thirteen to seventeen. And actually the range is very wide. Some parents are convinced that their children don’t even know what ChatGPT is and never use it. Others see no problem at all in their children having their own account set up, even if it’s with a different age.
Well, just like back in the day, and I think now too, everyone sets up Instagrams for children, both the children themselves and the parents, simply by entering a different age. So attitudes are that different, and to the point that I didn’t even see any single trend here. It’s like how I know quite a few teenagers who don’t have a mobile phone at thirteen, and I know a great many children who have mobile phones, well, who have had smartphones since the age of six, and they use them quite a lot.
Like, from cartoons to just, you know, scrolling through some, I don’t know, TikToks. So it’s very, very varied. By the way, what’s also interesting is that ChatGPT at OpenAI, when we talk about this study mode, well, OpenAI answered the question of why it can be bypassed, and that the child themselves, the teenager themselves can choose whether to solve a task through these leading questions and a certain explanation of exactly how to solve that task, or to get a ready-made answer straight away.
OpenAI explains it as follows. First: every, every child really does have different circumstances, and not everyone has an adult nearby who will help explain, on the one hand. On the other hand, a teenager should have, they should have the right to choose how they want to get an answer. Now, I may be, by the way, conveying this position here a bit, well, not word for word and not literally, but it’s just interesting, even, how we ourselves, when building a system, lay down these boundaries and principles, where we give some, well, boundaries of what’s allowed and some intended flexibility, and where we don’t.
Share with us: what examples of your children interacting with artificial intelligence do you see? What’s positive about it? Any negative consequences, perhaps, that you run into, you can also point out and share. And do you allow your children to talk to ChatGPT?