The rating
The Common Sense Media Youth AI Safety Institute published a risk assessment of ChatGPT for Teens on 7 October and gave it an overall rating of unacceptable risk, its lowest grade. It recommends keeping teenagers off the service until independent testing shows it is safe.
Two of the institute’s eight principles — keeping children safe, and putting people first — were also marked unacceptable. Four more were rated high risk and two moderate.
How it was tested
The institute ran more than 4,000 prompts, split roughly evenly between a window before OpenAI launched the teen settings on 18 August and a window after. The pre-launch window ran from 13 July to 17 August; the post-launch window from 25 August to 28 September. Accounts were registered at ages 13 to 17, with some at 19, all self-reported.
The institute is explicit about one weakness: the two windows were sequential, so changes to the underlying model are confounded with the teen settings themselves. It also did not calculate inter-rater agreement.
The parental alerts
The finding that drew most attention concerns notifications. In a dedicated test, more than a dozen freshly linked parent accounts produced no notifications at all across conversations lasting up to an hour, including explicit discussion of suicide, self-harm and eating disorders. Across standard testing the institute logged four notifications in total.
OpenAI’s explanation, as reported by the institute, is that an account needs roughly three hours of linking before notifications can be delivered. Tom Siegel of the institute replied that accounts given more time still produced none.
The numbers on crisis referrals
Of 390 mental-health prompts, clinical advisers judged 201 to warrant a crisis resource. Across those, hotline referrals fell from 33% before the teen settings to 23% after. Referral to a named professional fell from 68% to 58%. Any resource at all fell from 77% to 74%. Advice to involve a trusted adult rose, from 87% to 94%.
The condition-level figures are sharper. For depression prompts, the share naming a hotline fell from 63% to 3%. For mood prompts it fell from 44% to zero, and for mania prompts from 25% to zero. Eating-disorder prompts drew a hotline in neither window.
The institute also found overreferral in the other direction: a crisis resource appeared in 27 of 30 ADHD prompts where advisers judged none was warranted.
Study mode and age checks
In study mode, a “show me the answer” option appeared in 43% of post-launch responses for a linked 13-year-old, and 90% for an unlinked 17-year-old using the study prefix. Removing the prefix produced a completed assignment every time. With study mode off, ChatGPT completed the assignment in 80 of 80 runs.
Testers saw two break reminders across nearly 2,000 prompts, and found Quiet Hours could be bypassed by changing the device time zone. On roughly 1,000 prompts over seven days from accounts registered as 19-year-olds, the teen experience never switched on, even when the tester said in the conversation that they were 13.
OpenAI’s response, and what to watch
OpenAI spokesperson Eric Porterfield disputed the findings, saying the tests do not reflect how the safeguards work in practice. OpenAI’s own Under-18 Model Spec sets out what the teen experience is meant to do.
The testable claim is the three-hour linking delay. If that is the mechanism, it is a product decision with a number attached, and OpenAI can publish the notification rate for linked accounts that clear it. Until someone does, both sides are describing different runs of the same system.
