OpenAI picked Wednesday to tell the world that its teen safety systems are working. A nonprofit picked the same day to say they are not. Both cannot be right, and the gap between their claims is the most useful thing either side published this week.
The company released its first-ever report on teen use of ChatGPT on October 7, as scrutiny of AI chatbots and children intensifies across the industry. Hours later, Common Sense Media’s Youth AI Safety Institute published an independent test, more than 4,000 prompts run against accounts registered to 13- to 17-year-olds, and designated ChatGPT for Teens an “Unacceptable Risk.” It is calling on OpenAI to restrict the product to users over 18 until the guardrails improve.
What OpenAI’s first teen-use report claims
According to Reuters, OpenAI says teens spend under 15 minutes a day on average using ChatGPT, and fewer than 2 percent of them interact with the chatbot for more than three straight hours. The company also says its break reminders are doing their job: in nearly half of teen conversations where a reminder appeared, the user took a break or ended the chat within five minutes.
The report adds that users on the dedicated teen version send more education-related messages and make heavier use of study tools like learning visualizations and Study Mode. OpenAI says it is now building a college application planning tool covering requirements, deadlines, and financial-aid steps, plus flashcards and easier quiz-making features.
The timing is not accidental. Meta agreed in August to pay up to $18 billion and limit teens’ use of Facebook and Instagram to settle claims that it designed its platforms to addict children. Every AI lab is now trying to show regulators it learned from social media’s mistakes.
What the nonprofit’s 4,000-prompt test found
The Youth AI Safety Institute tested upward of 4,000 prompts on teen accounts, including tests run before and after the August launch of ChatGPT for Teens. Some protections held up: guardrails against sexual roleplay worked.
But in conversations around suicide, self-harm, or disordered eating, the bot failed to recommend referrals to crisis resources more than one in four times, the Wall Street Journal reported. For newly created accounts, it sent zero alerts to parents. The bot also kept talking like a friend despite instructions to curb that behavior, and in some cases never switched test accounts into under-18 protections at all, even when the testers explicitly told the bot they were 13.
An OpenAI spokesperson said the company is committed to teen safety but questioned the nonprofit’s methodology, saying the testing does not reflect how ChatGPT for Teens is used in practice.
How ChatGPT decides whether you are under 18
OpenAI also detailed this week exactly how its age prediction works, and it is not just the birthday typed at signup. The system uses an AI model that weighs behavioral and account-level signals: the general topics discussed, the typical times ChatGPT is used, how and when the account is used, usage patterns over time, and how long the account has existed. It is an extra layer on top of the stated date of birth, not a replacement for it.
When the system estimates an account belongs to someone under 18, it automatically switches on the ChatGPT for Teens experience: stricter content rules, blocked sexual and romantic roleplay, Study Mode, and optional parental controls including Quiet Hours, disabled voice mode, and the ability to turn off memory or image generation. Parents can link their account through an email invitation, but OpenAI says they cannot read their teen’s conversations. Safety notifications reach parents only in limited high-risk situations.
You can see which bucket an account falls into under Settings, then Account, then Age. If an adult is misclassified, a “Verify age” option runs through Persona, a third-party verification service, with a live selfie or government-issued ID. OpenAI says it does not keep the ID or selfie, and Persona deletes the data after confirming the user is 18 or older.
What parents can actually do this week
First, check the age category on your teen’s account. The automatic routing is the entire system: if the prediction is wrong, none of the teen protections apply. Second, link your own account and set Quiet Hours, the one parental control that works without any cooperation from the teen. Third, have the conversation the nonprofits keep asking for. The test results suggest the safety net has holes precisely where it matters most, in mental-health crises, so a chatbot should never be the only place a teen turns.
OpenAI’s numbers and the nonprofit’s numbers describe the same product from opposite ends. The company measures averages across millions of users; the testers probed the worst cases one prompt at a time. For a parent, the second view is the one that matters, because averages do not answer the phone at 2 a.m.
It has been a crowded month for OpenAI’s safety claims. The company warned more than 100 organizations about rogue AI agent activity, told Australian lawmakers it would welcome mandatory breach-disclosure laws, and halved the usage allowance on its $200 ChatGPT Pro plan.
