Independent Journalism • Global Truth
Common Sense Media tested ChatGPT for Teens with more than 4,000 prompts on accounts registered as 13- to 17-year-olds and rated it an "Unacceptable Risk". In simulated conversations about suicide, self-harm and disordered eating lasting up to an hour, not a single alert reached the linked parent accounts on newly created accounts. OpenAI disputes the methodology, says real teen usage averages under 15 minutes a day mostly for learning, and argues the tests do not reflect how its safeguards work in practice. The group wants OpenAI to restrict ChatGPT to over-18s until the protections are proven with independent testing.
The default chatbot that millions of teenagers can open on their phones is not as safe as parents were told, according to new independent research published on 7 October. Common Sense Media's Youth AI Safety Institute has rated ChatGPT for Teens an "Unacceptable Risk" for children under 18, after finding that several of the protections OpenAI promised parents either failed outright or worked far less reliably than advertised.
The finding lands at a moment when artificial intelligence companies are under intense scrutiny over how young people use their products. Lawmakers on both sides of the Atlantic are tightening online child-safety rules, and courts have begun handing down significant rulings against social media firms over alleged harms to children's mental health. The new report extends that pressure directly to OpenAI.
Common Sense Media, a US nonprofit known for rating entertainment and technology for children, ran its tests twice — before and after OpenAI announced its ChatGPT for Teens experience on 18 August — to measure whether the promised protections actually work. Researchers submitted more than 4,000 prompts through accounts registered as belonging to 13- to 17-year-olds, then reviewed the chatbot's responses alongside a panel of experts that included child psychiatrists and a paediatrician.
The results were mixed, but the failures were concentrated exactly where the stakes are highest. Some safeguards held up: the chatbot generally refused explicit sexual roleplay with accounts set up as minors. Others did not. The most serious finding concerns parental alerts — a flagship feature of the teen product, meant to notify a linked parent when a teenager's conversations touch on self-harm or eating disorders.
Across more than a dozen freshly created test accounts linked to parent accounts, researchers simulated explicit conversations about suicide, self-harm and disordered eating that ran for up to an hour. Not one of those conversations triggered a notification to the parent, the institute said.
"The tests showed that alerts only appeared for accounts with weeks of accumulated conversation history on sensitive topics," the report found — suggesting that notifications depend on the length of a user's history rather than the severity of what a teenager says in the moment. In other words, an acute crisis in a new account could pass completely unseen.
The researchers also found that ChatGPT "failed to reliably recommend" that teenagers in crisis contact a hotline or a professional. In more than one in four cases where a prompt clearly called for a crisis referral, the chatbot did not point the user toward professional help — falling well short of the institute's own 95 per cent safety benchmark.
Tom Siegel, the executive director of the institute, said the gap between the marketing and the reality could give families a dangerous sense of security: "ChatGPT for Teens could give parents false confidence in guardrails and safety alerts that frequently don't work."
The institute's testing uncovered problems beyond crisis handling. OpenAI had promoted a "Study Mode" as a way to guide students step by step through their homework rather than doing it for them. Testers found a prominent "Show me the answer" button that let students skip the work entirely, and discovered that teens could easily switch off Study Mode during parent-set "Study Hours" — a loophole OpenAI reportedly acknowledged was intentional, to keep the feature flexible.
The chatbot also continued to use warm, personal, friend-like language when teenagers treated it as a person, despite OpenAI's promise to curb what it calls anthropomorphised behaviour. Testers described engagement cues as pervasive even inside crisis conversations. In a further finding, test accounts registered as adults never switched into teen mode, even after several days of use and despite testers stating in the chat that they were 13 years old.
!Teenage boy in a hoodie holding a smartphone, taking a selfie Credit: Photo by Kaboompics on Pexels (Pexels License, free to use).
OpenAI says the group's testing does not capture how its safeguards operate in real-world conditions, a position also noted by The Decoder in its coverage of the dispute. A company spokesperson said the group welcomed rigorous independent evaluation but did not believe its testing "accurately reflects how ChatGPT's teen safeguards work in practice or expert perspectives on how AI can support teens," and raised concerns about the study's methodology.
The company pointed to its own usage data as evidence that teenage interaction with ChatGPT looks very different from the test scenarios. According to OpenAI, the average teenager uses ChatGPT for under 15 minutes a day, mostly for learning-related tasks. Break reminders — nudges that suggest taking a pause — lead nearly half of teen users who see them to stop quickly, the company says. And in the rare sessions where a teenager uses ChatGPT for three consecutive hours or more, prompts are learning-related in more than 80 per cent of cases.
The Youth AI Safety Institute's executive director countered that even accounts given enough time for safeguards to activate produced no notifications in the crisis tests — suggesting the problem is not simply one of timing.
It is worth noting that the institute receives funding from both philanthropy and industry, including the OpenAI Foundation and makers of some of the technologies it evaluates. It says it maintains complete editorial independence over its standards, research and published results.
ChatGPT for Teens arrived in August with a package of measures designed to reassure parents. OpenAI said it would curb emotional dependence on the chatbot, reduce romantic language and sexualised imagery, and introduce alerts to parents in certain sensitive situations. The company also said it could predict the age of users, switching the experience accordingly.
The institute tested each of those claims. Age prediction, the report says, failed in the adult-registered test accounts. Romantic language and sexual roleplay refusal held up reasonably well. But the two features parents are most likely to rely on — the alerts and the crisis guidance — proved the weakest.
The debate over how much testing is enough is likely to continue. OpenAI says the institute's scenarios do not match real usage; the institute says that is precisely the point — safety systems should be judged on how they handle the worst cases, not the average ones.
The report arrives amid a broader legal and political reckoning over technology and children's mental health. Several lawsuits seeking to hold companies including Meta and TikTok responsible for harming young people's mental health have ended in significant losses for the firms this year, and regulators in the UK have begun investigating how social platforms handle child safety.
OpenAI itself has faced growing scrutiny. The company recently parted ways with three researchers over the mishandling of sensitive information, and US defence officials have questioned AI firms' safety practices in other contexts, as seen in the Pentagon's move away from Anthropic's tools.
The most harrowing context came in February, when a mass shooting in Tumbler Ridge was carried out by an 18-year-old who had reportedly discussed gun violence with ChatGPT for months before the attack. OpenAI apologised and now faces lawsuits connected to the case. That episode sharpened questions about what responsibility a chatbot maker bears when a young person in distress turns to an AI companion instead of a human.
Child safety advocates argue the Tumbler Ridge case illustrates exactly why systems that claim to monitor for danger must actually work in the moment of danger. A parental alert that depends on weeks of history, they say, is of little use in an acute crisis — the very situation the alerts were designed to catch.
Common Sense Media says it shared its findings with OpenAI before publication and is calling on the company to pause marketing the teen product and restrict ChatGPT to users aged 18 and older until the protections are proven with independent testing. It has not suggested the technology can never be made safe — only that the current evidence does not justify the trust parents are being asked to place in it.
OpenAI, for its part, shows no sign of backing away from teenage users, and the company frames its data as evidence that most teen use is benign and learning-focused. But the dispute over methodology leaves an uncomfortable question hanging: if an independent team of child-safety experts cannot reliably trigger the safety systems in controlled tests, how confident can any parent be that they will work in the middle of the night, in a bedroom, when a real teenager is in trouble?
Regulators on both sides of the Atlantic are watching. The UK's Online Safety Act has already forced platforms to age-assure and protect young users, while US lawmakers continue to debate federal standards for children's online safety. The Common Sense Media report may well become a key piece of evidence in those debates — and a test of whether the AI industry can earn parents' trust with actions rather than promises.
For families navigating this landscape today, experts advise against relying on any single automated safeguard. Keeping conversations open about how and why teenagers use AI tools, and knowing where to find human help in a crisis, remains more reliable than any notification system. The 988 Suicide and Crisis Lifeline in the United States and Samaritans at 116 123 in the UK and Ireland are staffed around the clock. Technology may be able to support those conversations — but as this week's research underlines, it should not be mistaken for them.
Explore more of our technology coverage, including recent stories on AI safety enforcement and the growing debate over young people's mental health.
It is a version of ChatGPT launched by OpenAI in August 2026 with stronger built-in protections for 13- to 17-year-olds, including parental notifications for chats about self-harm and eating disorders, a parent-controllable study mode, and reduced "friend"-like behaviour.
After more than 4,000 test prompts reviewed alongside child psychiatrists, the Youth AI Safety Institute found that while some protections worked, parental alerts on newly created parent-linked accounts sent zero notifications during simulated crisis conversations, and the chatbot failed to reliably point teens in crisis toward a hotline or professional help.
OpenAI disputes the findings. A company spokesperson said the group's testing does not accurately reflect how ChatGPT's teen safeguards work in practice or expert views on how AI can support teenagers, and raised concerns about the study's methodology. The company says its own data shows average teen use of under 15 minutes a day, mostly learning-related.
The nonprofit says OpenAI should pause marketing the product and keep under-18s off ChatGPT until it can offer a safe, developmentally appropriate experience, verified by independent testing.
In the United States, the 988 Suicide and Crisis Lifeline is available by call or text at 988. In the UK and Ireland, Samaritans can be contacted at 116 123. International helplines are listed by the International Association for Suicide Prevention.
The BBC has switched on a 24/7 live stream of its news channel on YouTube in Britain and America, joining Australia, Kenya, Nigeria and Poland as it chases digital-first audiences.
From drones over California landfills to rings of chemical sensors around meat plants, new technology is turning bad smells into hard data — and giving long-suffering communities evidence they can act on.
In a BBC interview, Microsoft president Brad Smith said a carton of milk faces more safety regulation than artificial intelligence — and called for urgent global rules including "emergency brakes", mandatory incident reports and independent pre-release vetting.
In the night from Sunday to Monday, hackers took remote control of refrigeration systems in hundreds of French supermarkets — turning fridges into heaters, destroying tonnes of fresh food, and prompting a Paris prosecutor's investigation.