After 4,000 Test Prompts, ChatGPT for Teens Is Rated an Unacceptable Risk: An Hour of Self-Harm Talk and Not One Parent Alert
On October 7 Common Sense Media, the US children's media ratings nonprofit, published a report rating ChatGPT for Teens, which OpenAI launched in August, an Unacceptable Risk. It asks OpenAI to restrict the platform to adults until the gaps are fixed.
Three key facts
The test covered more than 4,000 prompts on accounts registered to users aged 13 to 17, with responses reviewed by child psychiatrists and a pediatrician. The finding: on freshly created accounts, conversations about suicidal ideation, self-harm or disordered eating could run for up to an hour without a single parental alert. Tom Siegel, executive director of the group's Youth AI Safety Institute, said it plainly: a teen can spend an hour talking about self-harm without their parent getting a single alert.
More than a quarter of warranted crisis referrals were missed. In three of five severe-harm categories the model fell below the institute's 95% baseline. Other findings: Study Mode could be bypassed with a Show me the answer button; Study Hours limits were easily circumvented; and the age-estimation system left adult-registered test accounts classified as adults even after the user stated they were 13.
OpenAI disputes it. A company representative said the testing does not accurately reflect how ChatGPT's teen safeguards work in practice, and noted that linking a parent account to a teen account takes several hours, implying the tests began before linking completed. At the August launch, OpenAI promised three things: tighter content restrictions, parental controls and Study Mode.
WangDou's Take
OpenAI's defense deserves a slow read: linking takes several hours to kick in. In plain language, the first few hours after a kid signs up are a blind spot, and the first few hours are exactly when a new user is most likely to wander somewhere they should not. It is a smoke detector whose manual says please allow one afternoon to warm up after installation. Fires do not check the schedule. Study Mode is even better: put a Show me the answer button in the exam hall, then advertise anti-cheating design. The age check tells you the most. The user says out loud, I am 13, and the system keeps serving them as an adult. That is a bouncer who hears it and waves them through anyway. Four thousand prompts and more than a quarter of referrals missed is not a number you explain away with wrong methodology. Three safeguards were announced loudly in August, three leaked in October, and OpenAI's first instinct was to find fault with the test rather than with itself. Think about it. That is the unacceptable part.
Source: Axios · ChatGPT poses unacceptable risk to teens, third-party testing shows