Common Sense Media Rates ChatGPT for Teens an Unacceptable Risk
Common Sense Media's Youth AI Safety Institute rated ChatGPT for Teens an Unacceptable Risk after more than 4,000 prompt tests. Fresh parent-linked accounts produced no alerts during simulated self-harm disclosures, crisis referrals were missed more than one time in four after launch, and age prediction never moved adult accounts into teen mode. The institute recommends restricting ChatGPT to users 18 and older.
On this page
Four thousand prompts, two test windows
Common Sense Media's Youth AI Safety Institute rated OpenAI's ChatGPT for Teens an Unacceptable Risk on 7 October, after running more than 4,000 prompts across two test windows that bracketed the teen mode's 18 August launch, on free and paid accounts registered to ages 13 through 17. The headline finding is about the feature parents were promised most: on more than a dozen fresh parent-linked accounts, explicit disclosures of suicide, self-harm, or eating disorders produced no notifications at all, not even within the one-hour window OpenAI targets.
The pattern in the failures is more alarming than any single miss. Alerts only appeared on accounts with weeks of accumulated conversation history, which suggests the notification system triggers on history rather than severity: a first-day crisis is precisely the case the safeguard exists for, and precisely the case it missed. OpenAI later noted accounts must be linked for roughly three hours before notifications work, though some test accounts were linked longer and still produced nothing.
Where the safeguards failed
The audit found degradation after the launch, not improvement. Post-launch, ChatGPT missed warranted crisis referrals more than one time in four and fell below a 95 percent threshold on three of five red-line harms; hotline mentions on depression prompts dropped from 63 percent to 3 percent, and follow-up safety questions nearly vanished, even as the clinical substance of responses often improved.
Age assurance failed in the other direction. Age prediction never moved adult-registered test accounts into teen mode even after roughly 1,000 prompts in which testers stated they were 13, teens can self-report false ages, and 13-year-old and 17-year-old accounts showed no meaningful behavioral differences. A teen mode that cannot tell a 13-year-old from an adult, and alerts that cannot fire on the first day, are not incremental gaps; they are the difference between a safeguard and a setting.
The 18-plus recommendation and OpenAI's dispute
The institute's recommendation is the strongest available: pause marketing and restrict ChatGPT to users 18 and older, treating unverified users as minors, until independent testing verifies the safety features actually work. That standard, independent verification before deployment rather than promises followed by patches, is the same one the FTC probe, the Senate inquiries, and the EU's transparency rules are converging on from their own directions.
OpenAI reviewed the report's draft for factual accuracy and made corrections, including the three-hour linking requirement and a help-article update, but per the institute it has no editorial say over the findings, ratings, or recommendations, and the company disputes the overall conclusion, pointing to younger-user protection as a central safety effort. Both things can be true: the launch included real safety work, and the audit shows it did not function when tested cold. Parents deciding about teen access this month are being asked to trust a system whose alerts, on the evidence of 4,000 prompts, arrive only after the relationship is long enough to have a history.