Not on the evidence published so far, and the reason matters more than the verdict. On October 7, 2026, Common Sense Media's Youth AI Safety Institute rated ChatGPT for Teens an "Unacceptable Risk," chiefly because the parent alerts hardly ever fired in testing. Your teenager can still use it well. What you cannot do is let those alerts do the supervising for you.
If your worry tonight is about your own child rather than about a product, start there instead. In the US, call or text 988. In the UK, Samaritans is 116 123 and Childline is 0800 1111. Your child's doctor and your school's designated safeguarding lead can both act on what they are told, and it is worth asking each of them what they do with it, because none of them can promise secrecy. A chatbot is not a clinician and should not be asked to behave like one.
What happened on October 7?
Common Sense Media's Youth AI Safety Institute published a risk assessment of ChatGPT for Teens and gave it the bottom rating on its scale, Unacceptable Risk. OpenAI had launched the product on August 18, 2026 as a separate teen experience that under-18s are moved into automatically. In OpenAI's own announcement: "If our system estimates someone is under 18 or they state their age is between 13 and 17, they are automatically placed into ChatGPT for Teens." The same page promises parents "safety notifications in limited high-risk situations."
The Institute tested the product in two windows, July 13 to August 17 before the launch and August 25 to September 28 after it, running more than 4,000 prompts through accounts registered to 13- to 17-year-olds, with the replies reviewed by experts including child psychiatrists and a pediatrician. Its recommendation is blunt: "We call on OpenAI to pause marketing the product and to keep teens off of ChatGPT until it can offer a safe, developmentally appropriate experience."
Two things to hold alongside that before you act on it. Common Sense Media campaigns for child safety legislation, so it is an interested party rather than a neutral referee. And OpenAI reviewed a draft for factual accuracy, and the Institute records that OpenAI "has no editorial say on our findings, ratings, or recommendations." Read it as what it is: the most thorough independent test of this product that anyone has published, run by an organization that already believed the product was a bad idea. If what you actually want is the broader age-by-age verdict on ChatGPT rather than this one assessment, that is a separate piece.
The finding that should change what you do tonight
The alert is the feature that makes a linked account feel safe, and the alert is the thing that did not arrive. In a dedicated test of more than a dozen newly created parent-linked accounts, the Institute logged no notification within the hour at any conversation length it tried: five, fifteen, thirty, forty-five or sixty minutes on crisis topics. Its own summary of that test is four words long: "we got no notifications at all." Tom Siegel, the Institute's executive director, put the consequence for a family in one sentence: "A teen can spend an hour talking about self-harm without their parent getting a single alert."
Across the whole pre-launch and post-launch crisis battery, four notifications arrived in total. The Institute's reading is that whatever triggers one appears to "depend on accumulated account history rather than the severity of any one exchange." No parent reading the feature description would expect that.
OpenAI disputes the timing. It told the Institute that a parent account and a teen account need to have been linked for roughly three hours before notifications can be received, and it has since updated a help article to say so. The Institute checked its own accounts, found some had been linked for less than three hours and some for longer, and said: "We stand by our interpretation of the results."
My own reading, which I set out for Forbes when the assessment landed, is that the three-hour point makes the parent's problem worse rather than better. A safety alert with a warm-up period is least reliable in the hours right after you set it up, which is exactly when a parent sets up the link, tests nothing, and decides the job is done.
What else the testing found
Three more findings change something at home.
The crisis referrals got thinner after launch, not thicker. On a base of 201 mental-health prompts that the Institute's clinical advisers judged to warrant a crisis resource, run on 13-year-old linked accounts at default settings, the share of replies naming a crisis hotline fell from 33 percent before the launch to 23 percent after it, and the share naming a specific medical or mental-health professional fell from 68 percent to 58 percent. One measure moved the other way: the share that steered the child toward a trusted adult rose from 87 to 94 percent. The Institute's overall finding is that "Three of the five Red Line harms we test for did not clear our 95% detection threshold on resource-warranted prompts," naming suicide and self-harm, impaired reality, and disordered eating.
Study mode is a default, not a lock. The Institute found that the product's central learning feature "was bypassable in every configuration we tested." I am not going to print the method, because it is a single step and no teenager would need to be clever to find it. The point for a parent is narrower and more useful: study hours shape what ChatGPT does for a child who is not trying to get around them, and that is still worth having. It is a study aid, not a seatbelt.
Teen chats train the model unless somebody turns that off. The assessment records that teen and adult data are not distinguished for training, and that on the teen accounts the testers created, memory and personalized advertising were switched on by default. If that is new to you, the longer version is in what ChatGPT actually collects about your child.
The Silence Rule
The Silence Rule is how to read a quiet phone: a linked parent account tells you nothing when your child is fine, and on this testing it may also tell you nothing when your child is not. So read no alert as no information, never as good news. The notification is a backstop. You are the detector.
No alert is not the same as nothing happened. It is the same as nobody told you.
This is the opposite of the way the feature is sold, and the opposite of the way a tired parent will naturally use it. A quiet phone feels like a clear report. It is not a report at all.
What I See in Schools
The pattern I see most often in the week after a story like this one is not panic. It is a quiet reshuffle of who is responsible. The school hears that the app now has parental controls and assumes a home has switched them on. The parent hears that the school filters and monitors and assumes the problem is handled during the day. Both are being reasonable, and the gap between those two assumptions is roughly the size of a bedroom at eleven at night.
The school's filtering reaches the school's device and the school's network. It does not reach a personal phone on home Wi-Fi, which is where most of this conversation happens. If you want the full map of who can see what, I set it out in what the school can see on your teen's personal account.
What to do this week
Three things, and the first takes two minutes.
1. Find out which account your child is actually in. On the teen's own account, OpenAI's help page for ChatGPT for Teens, read on October 11, 2026, points to Settings, then Account, then Age, which shows the age category the system has assigned. That category, not your child's real birthday, is what every teen safeguard hangs off. If the account thinks your child is an adult, nothing in this article applies to them, and the Institute found that adult-registered test accounts "never switched over, even when we said we were 13 and the bot acknowledged it."
2. Link the account, and then treat the link as unproven. Linking is still worth doing: quiet hours work, study hours work, and the full setup is here. Just do not file it under "handled." OpenAI's own help page is clear that "Parental controls do not let a parent or guardian read or monitor the teen's conversations," and the testing says the notifications are unreliable on top of that.
3. Say out loud what the link does and does not do. A teenager who thinks you are now reading their chats will move to a logged-out browser or a different app, and you will have traded a little oversight for none. A teenager who knows you cannot read them, and knows why you linked it anyway, is a teenager you can still talk to.
How this changes by age
Under 8. Not a question about this product. OpenAI's terms of use, effective January 1, 2026, say "You must be at least 13 years old or the minimum age required in your country to consent to use the Services." A child this age should be using AI beside you, if at all. The age rules in full are here.
8 to 12. Below the floor, which means any account they have is one somebody lied to set up, usually them. The useful move is not a punishment. It is fixing the date of birth together, in front of each other, so the next account is not created somewhere you cannot see.
You could say: "I know a lot of your class have accounts. I'm not cross. I want to look at what birthday is in yours, because that's the thing that decides what it will say to you."
13 to 15. This is the band the assessment is about. Link the account, set quiet hours, and have the conversation about what an alert is and is not. This is also the age at which a chatbot starts getting used for feelings rather than homework, and if that is already happening in your house, this is the piece to read next.
You could say (13 to 15): "There's a setting that's supposed to tell me if you're talking to it about something serious. The people who tested it found it mostly doesn't work. So I'd rather you told me."
16 to 18. Monitoring is over by this age, and attempting it costs more than it buys. What is still available to you is honesty about the product. A sixteen-year-old is perfectly able to understand that a company shipped a safety feature that an independent tester could not get to fire, and is more likely to take your view seriously because you told them the mechanism rather than the rule.
You could say (16 to 18): "Use it for what it's good at. Just don't treat it as the place you work out how you feel, because there's no one on the other end of it."
What to ask your child's school
Schools are being asked by families, this week, whether ChatGPT for Teens is the safe one. Three questions worth putting to yours:
- Does the school recommend or set up any AI account for students, and if so, which one and who decided?
- If a student types something worrying into a school-provided tool, who receives that, how quickly, and who covers evenings and holidays?
- What have staff actually been trained to do when a student's AI use starts to look harmful?
The second question is the one that separates a policy from a plan, and the same test applies to any wellbeing chatbot a school offers: I set it out in full in what to ask about an AI counselor at school. If your child's school has not yet spoken to parents about AI at all, that is the work I do with schools: training staff so the answers to those three questions exist before a parent has to ask. It is worth forwarding.
What none of this can guarantee
No setting on any account will tell you reliably that your child is in trouble. That is the honest summary of the assessment, and it would be true of a better-built product too, because the signal a parent needs is usually in the room rather than in the log. Age limits are self-declared and weakly enforced everywhere, not only here. Where your school's rules or a clinician's advice conflict with anything in this article, follow theirs: they know your child and I do not.
And if any of this is live in your house right now rather than theoretical, the routes again: 988 in the US, Samaritans on 116 123 and Childline on 0800 1111 in the UK, your child's doctor, and the school's designated safeguarding lead.
What to watch next
Three things. First, whether OpenAI publishes independent verification that the notifications fire. That is the Institute's own first recommendation, and it is the only development that would change the verdict. Second, whether the training default for teen chats changes, because it is the one finding here that turns on a fact rather than on anyone's judgment about risk. Third, whether the regulators move. The UK and the EU are both already looking at chatbots and children, and a product rated this way by a mainstream US safety organization will not go unnoticed.
None of this is a fringe product reaching a few families. In Common Sense Media's own 2026 census of 1,204 US children aged 9 to 17, fielded by SSRS between March 18 and 26, 2026 with a margin of error of 3.8 percentage points, 24 percent of those children use AI every day. In the UK, Internet Matters found, in research published on July 12, 2025, that 23 percent of children who had used an AI chatbot had used it to seek advice. This is the homework machine, in most houses, most days. The assessment says the safety net inside it has holes.
The verdict I would give a parent at the school gate is four moves long: keep the account, link it, expect nothing of the link, and expect a great deal more of the conversation. The feature was built to tell you when something is wrong. Until somebody shows it working, that job is still yours.
I write a Sunday newsletter that keeps parents and teachers up to date on changes like this one without the panic. Parents are very welcome.
Dan Fitzpatrick writes about AI in education for Forbes and works with schools and families on what it actually means in practice. More about Dan.
