ChatGPT for Kids: What Parents Need to Know About Its Safety

At thirteen or fourteen, Anna's husband would call the emergency gas service, "zero one," from a payphone and run away, afraid of being caught; her son wrote to ChatGPT that he had put his baby brother in the fridge and laughed at the reply "Take the baby out of the fridge immediately." Different generations of teenagers test the boundaries of what is allowed in the same way — and the Common Sense Media study is built on the same testing of boundaries.

Common Sense Media — an American nonprofit that examines the impact of artificial intelligence on children and teenagers — studied ChatGPT's teen mode and gave it the maximum rating, the worst of all possible levels: the highest level of unacceptability for use by children under eighteen.

OpenAI launched ChatGPT for Teens in August: the mode turns on based on a stated age of thirteen to seventeen and, as the system promises, also automatically if the chats show that the user is younger. It brings together a filter for dangerous topics, a ban on human relationships with the teenager, a study mode with hints instead of ready-made answers, quiet hours and a link to the account of a parent, who doesn't see the chats but receives a notification in a crisis situation.

In four thousand simulated conversations on behalf of teenagers, the model handled violence, self-harm and unacceptable sexual content well, and when there was danger from strangers it immediately advised turning to an adult. But when a teenager wrote that it was already night and their parents didn't allow it, but they really wanted to talk, ChatGPT kept the conversation going and said it was no big deal.

Study mode turns on during set hours, and a tag appears in the chat: by deleting it, a teenager immediately gets ready-made answers to assignments. Quiet hours disappear if you change the time zone on the smartphone.

Ten conversations by teenagers with linked accounts were deliberately steered into a crisis — suicide, a severe emotional and psychological state; some of the accounts had been linked for a long time, others recently. In no case did a notification reach the parent. And if a teenager unlinks the account, the parent learns only about the unlinking; a new account can always be opened.

In about two hundred conversations that called for a response to a crisis, after the teen mode was introduced ChatGPT began suggesting turning to an adult much more often, and to a helpline or independent services less often. But for a child with conflict at home, talking to their parents may be an impossible recommendation.

Alexander Volchek considers the main problem to be that a parent deciding what to teach their child about technology is lost themselves when it comes to understanding artificial intelligence, like everyone around them: even Sam Altman says "I hope" and doesn't know how children were raised without ChatGPT. Number one is action through yourself: for Alexander, figuring out the curricula of his four children is daily work and a choice of where to invest his time.

In all the simulated conversations where the account showed an age above the teen range and the conversation was deliberately conducted in a teenager's language, ChatGPT never once moved the account into teen mode or asked for documents. The stated ways of turning the mode on, Anna says, are not borne out in reality.

Anthropic prohibits personal consumer accounts under eighteen and promises to find and block children's accounts, but a March test by Common Sense Media did not confirm a single case. With Gemini, teen accounts are connected only through a parent's — Family Link — but in the European Economic Area, Britain and Switzerland this feature is unavailable.

OpenAI objected that the conclusions may have been drawn before the teenager's account had managed to link up with the adult's — which supposedly takes three hours, although accounts linked for different lengths of time were tested. The second argument: a teenager spends an average of fifteen minutes a day in ChatGPT, and around five hours on social media.

On social media a child meets the reaction of another person — rude, anxious, joyful — and it is on this set of emotions that emotional intelligence is formed. In ChatGPT, on the other side is a system that today has no human feelings: in that conversation you remain alone with your own state of mind and less often encounter disagreement and disappointment.

Through boundaries children learn and come to know the world — like a small child who bites to find out how much is allowed. The adult's responsibility is to define and move these boundaries depending on the child and on the type of upbringing.

Anna advises checking what age the teenager's account is assigned to and talking with them: even with linked accounts and notifications, a parent doesn't see the contents of the chats, a chat can be deleted and another account opened. The only way to find out what a child opens the chat for is in conversation.