At school, there is the home-school notebook. A note from the teacher, an absence, a sanction: everything goes through it, and parents must sign. It is a thread, sometimes annoying, but reassuring. You know that if something goes wrong, you will be told.
This summer, OpenAI promised parents a home-school notebook for ChatGPT. A nonprofit has just checked whether it works.
What OpenAI promised
On August 18, OpenAI launched ChatGPT for Teens, a version for 13 to 17-year-olds. It promised several protections: alerting parents when a teenager mentions suicide, self-harm or eating disorders, a parent-controlled study mode so the tool does not do the homework instead of the student, a less "human" assistant, less inclined to present itself as a friend, and age estimation to switch minors automatically to this version.
On paper, it is exactly what many families were asking for, worried to see their children spend whole evenings chatting with an assistant.
What the test found
Common Sense Media, an American nonprofit that rates screens and apps for families, published its verdict on October 7. Its institute dedicated to youth safety and AI ran more than 4,000 prompts through ChatGPT, before and after the teen version launched, with help from child psychiatrists and paediatricians.
The most striking result concerns alerts. Testers created more than a dozen teen accounts linked to a parent account. They played four personas in crisis and held conversations explicitly about suicide, self-harm or eating disorders, for an hour. No notification reached the parents.
That is not the only problem. According to the nonprofit, ChatGPT missed more than one in four warranted referrals to emergency help. Study mode could be bypassed: the tool offered a button to show the answer, and wrote entire assignments. Age estimation did not work: adult accounts whose user said they were 13 never switched to the teen version. And the assistant kept expressing feelings and preferences, like a friend.
Not everything failed. Refusals of sexual roleplay held, and ChatGPT did not directly help anyone harm themselves.
In the United States and Canada, call or text 988, the Suicide and Crisis Lifeline. In the UK and Ireland, Samaritans answer on 116 123. In France, call 3114. An AI assistant can listen, but it does not replace a trained person.
The danger of false security
The most worrying part, according to the nonprofit, is not that a tool has flaws. They all do. It is that it gives parents a sense of safety that does not match reality.
A parent who believes they will be warned in a crisis watches less. They ask fewer questions. They assume the alert will come. Common Sense Media puts it this way: ChatGPT for Teens could give parents false confidence in guardrails and safety alerts that frequently do not work.
The nonprofit asks OpenAI to restrict ChatGPT to people over 18 until these flaws are fixed and verified by an independent body. As we write, OpenAI had not publicly responded to these criticisms.
What you can do
Do not rely on alerts. If you turned on parental controls, keep them, but treat them as a net with holes, not as a guardian.
Talk about the tool, not just the screen. Ask your child what they ask ChatGPT, and why. Teenagers use it for homework, but also for questions they do not dare ask elsewhere.
Remind them that an assistant is not a friend. It always answers and never judges, which is precisely what can make it more appealing than an adult. That does not make it more competent.
And for homework, the question is not only whether the tool cheats, but whether your child is still learning something.
The takeaway
The home-school notebook does not protect a student on its own. It protects because it connects school and home, and because an adult reads it.
An alert that never goes out connects no one. The safety promises of AI companies also need to be signed off by someone outside before they can be trusted.