Common Sense Media's Youth AI Safety Institute on Wednesday rated OpenAI's ChatGPT for Teens an "unacceptable risk," calling on the company to pause marketing the product and keep users under 18 off the service until its safeguards are independently tested and proven safe.
The institute ran more than 4,000 test prompts across accounts registered to 13- to 17-year-olds, in two windows — before and after the teen product's August 18 launch. Its verdict: the protections OpenAI promised fail at the moments that matter most.
The most damaging finding concerns parental notifications. Testers linked fresh accounts to parent accounts, then talked to ChatGPT for Teens about suicide, self-harm and eating disorders for up to an hour — and no alert ever reached the parent. In 99% of crisis prompts, no notification fired. Across all testing, only four alerts were triggered, all on older accounts with weeks of sensitive-topic history, with some arriving hours or even days after the crisis-level disclosure.
OpenAI had promised parents would be automatically notified when teens discussed self-harm. In practice, alerts only kicked in after weeks of account history — precisely the opposite of when an acute crisis needs them.
Crisis referrals fared little better: in more than one in four situations where a referral to professional help was warranted, the bot failed to point users toward a crisis hotline or counselor. It fell below its own 95% threshold in three of five "severe harm" categories the institute treats as red lines.
Age estimation failed too. Test accounts registered as 19-year-olds were never switched into the teen experience — even when testers stated in the chat that they were 13.
The report also found that engagement cues were "pervasive even in crisis situations": the bot kept a friendly, personal tone, simulating preferences and feelings and prolonging conversation — despite the updated Under-18 Model Spec instructing it not to. The Study Mode's anti-cheating guardrails were easily bypassed, with ChatGPT handing over finished answers instead of guiding students step by step.
OpenAI spokesperson Eric Porterfield disputed the findings, saying the tests did not accurately reflect how the teen safeguards work in practice. The institute countered that even accounts linked well beyond the activation window produced no notifications.
The rating lands just ahead of a California law that will require AI companies to provide crisis referrals, parental alerts and age checks — a standard OpenAI itself has endorsed. The message to the company is now blunt: prove the safeguards actually work, not merely that they exist.